Add GLM-5.2 BF16 KLD reference logits 20260708

README.md

@@ -0,0 +1,102 @@
+---
+pretty_name: GLM-5.2 BF16 KLD Reference Logits 20260708
+license: other
+tags:
+- glm-5.2
+- kld
+- vllm
+- wikitext
+- reference-logits
+- bf16
+---
+
+# GLM-5.2 BF16 KLD Reference Logits 20260708
+
+This dataset contains the current GLM-5.2 BF16 reference prompt logits used for
+the July 2026 vLLM/Blackwell KLD checks.
+
+The reference cache is intended for candidate-side KLD comparisons without
+rerunning the expensive BF16 reference pass.
+
+## Files
+
+```text
+reference-logits/logits_0.safetensors
+reference-logits/manifest.json
+generation-log/config.env
+generation-log/scoremode_kld.log
+```
+
+## Reference Generation
+
+| Field | Value |
+|---|---|
+| Source model | `zai-org/GLM-5.2` |
+| Snapshot | `4d67f66cc64d3219133b767c253b2ad1425c6c88` |
+| Local source | `/root/.cache/huggingface/hub/models--zai-org--GLM-5.2/snapshots/4d67f66cc64d3219133b767c253b2ad1425c6c88` |
+| Generated at | 2026-07-08 00:04 UTC |
+| Docker image | `voipmonitor/vllm:eldritch-enlightenment-v7-vllme2e2eaf-b12x26144c0-cu132-20260707` |
+| vLLM version string | `v0.11.2.dev279+eldritch.enlightenment.v7.vllme2e2eaf.b12x26144c0.cu132.20260707` |
+| Dataset | `Salesforce/wikitext`, `wikitext-2-raw-v1`, split `test` |
+| Context length | 2048 |
+| Stride | 512 |
+| Windows | 1 |
+| Tensor key | `logits` |
+| Tensor shape | `[2047, 154880]` |
+| Tensor dtype | `float32` |
+| Attention backend | `B12X_MLA_SPARSE` |
+| KV cache dtype during capture | `fp8` |
+| Tensor parallel | 16 |
+| DCP | 1 |
+| Load format | `safetensors` |
+
+Tokenization first 16 ids:
+
+```text
+[284, 8396, 425, 10960, 465, 284, 14721, 8396, 425, 10960, 465, 374, 458, 6364, 4531, 1154]
+```
+
+GLM-5.2 IndexCache override:
+
+```json
+{"use_index_cache":true,"index_topk_pattern":"FFFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSS"}
+```
+
+## Download
+
+```bash
+mkdir -p /root/kld/glm52_current_bf16_returnlogits_ref_20260708T000424Z
+
+huggingface-cli download festr2/GLM-5.2-BF16-KLD-Reference-Logits-20260708 \
+  --repo-type dataset \
+  --local-dir /root/kld/glm52_current_bf16_returnlogits_ref_20260708T000424Z/hf
+
+ln -sfn \
+  /root/kld/glm52_current_bf16_returnlogits_ref_20260708T000424Z/hf/reference-logits \
+  /root/kld/glm52_current_bf16_returnlogits_ref_20260708T000424Z/ref
+```
+
+## Inspect
+
+```bash
+python3 - <<'PY'
+from safetensors import safe_open
+path = "/root/kld/glm52_current_bf16_returnlogits_ref_20260708T000424Z/ref/logits_0.safetensors"
+with safe_open(path, framework="pt", device="cpu") as f:
+    for key in f.keys():
+        tensor = f.get_tensor(key)
+        print(key, tuple(tensor.shape), tensor.dtype)
+PY
+```
+
+Expected:
+
+```text
+logits (2047, 154880) torch.float32
+```
+
+The canonical reproduction page is:
+
+```text
+https://github.com/local-inference-lab/rtx6kpro/blob/master/models/glm5.2/glm52-gguf-bf16-dequant-kld-2026-07-08.md
+```

generation-log/config.env

@@ -0,0 +1,25 @@
+IMAGE=voipmonitor/vllm:eldritch-enlightenment-v7-vllme2e2eaf-b12x26144c0-cu132-20260707
+MODEL=/root/.cache/huggingface/hub/models--zai-org--GLM-5.2/snapshots/4d67f66cc64d3219133b767c253b2ad1425c6c88
+REF=/root/kld/glm52_refs/bf16-b12xmlasparse-w1-ctx2048-s512-20260618
+REF_OUTPUT=/root/kld/glm52_current_bf16_returnlogits_ref_20260708T000424Z/ref
+OUT_ROOT=/root/kld/glm52_current_bf16_returnlogits_ref_20260708T000424Z
+LOGITS_OVERLAY=/root/vllm/worktrees/vllm-release-kld-logits-export-20260622
+OVERLAY_MODEL_PY=
+OVERLAY_DEEPGEMM_MOE_PY=
+GPU_DEVICES=0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15
+TP=16
+DCP=1
+GPU_MEMORY_UTILIZATION=0.98
+CPU_OFFLOAD_GB=80
+LOAD_FORMAT=safetensors
+MAX_MODEL_LEN=4096
+MAX_NUM_BATCHED_TOKENS=512
+MAX_NUM_SEQS=1
+QUANTIZATION=
+MOE_BACKEND=auto
+ATTENTION_BACKEND=B12X_MLA_SPARSE
+HF_OVERRIDES={"use_index_cache":true,"index_topk_pattern":"FFFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSS"}
+LLM_EXTRA_JSON={"decode_context_parallel_size":1}
+PROBE_RUNNER=collect
+FALLBACK_QUANTIZATION=auto
+FALLBACK_LLM_EXTRA_JSON={"decode_context_parallel_size":1,"moe_backend":"auto","enforce_eager":true,"disable_custom_all_reduce":true,"cpu_offload_gb":80}

generation-log/scoremode_kld.log

@@ -0,0 +1,700 @@
+applied_logits_overlay=/root/vllm/worktrees/vllm-release-kld-logits-export-20260622
+/opt/venv/lib/python3.12/site-packages/torch/cuda/__init__.py:63: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+return_prompt_logits_enabled True
+/opt/venv/lib/python3.12/site-packages/torch/cuda/__init__.py:63: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+collect_prefill_ref_start {"context_length": 2048, "extra": {"cpu_offload_gb": 80, "decode_context_parallel_size": 1, "disable_custom_all_reduce": true, "enforce_eager": true, "moe_backend": "auto"}, "hf_overrides": {"index_topk_pattern": "FFFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSS", "use_index_cache": true}, "max_windows": 1, "model": "/root/.cache/huggingface/hub/models--zai-org--GLM-5.2/snapshots/4d67f66cc64d3219133b767c253b2ad1425c6c88", "output_dir": "/root/kld/glm52_current_bf16_returnlogits_ref_20260708T000424Z/ref", "quantization": "auto", "stride": 512}
+tokenized {"num_tokens": 2048, "first16": [284, 8396, 425, 10960, 465, 284, 14721, 8396, 425, 10960, 465, 374, 458, 6364, 4531, 1154]}
+INFO 07-08 00:04:47 [api_utils.py:273] non-default args: {'trust_remote_code': True, 'load_format': 'safetensors', 'dtype': 'bfloat16', 'kv_cache_dtype': 'fp8', 'max_model_len': 4096, 'tensor_parallel_size': 16, 'enable_prefix_caching': False, 'cpu_offload_gb': 80, 'gpu_memory_utilization': 0.98, 'max_num_batched_tokens': 512, 'max_num_seqs': 1, 'max_logprobs': -1, 'disable_log_stats': True, 'hf_overrides': {'use_index_cache': True, 'index_topk_pattern': 'FFFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSS'}, 'enforce_eager': True, 'disable_custom_all_reduce': True, 'attention_backend': 'B12X_MLA_SPARSE', 'model': '/root/.cache/huggingface/hub/models--zai-org--GLM-5.2/snapshots/4d67f66cc64d3219133b767c253b2ad1425c6c88'}
+WARNING 07-08 00:04:47 [envs.py:2097] Unknown vLLM environment variable detected: VLLM_PCIE_DMA_FP8
+WARNING 07-08 00:04:47 [envs.py:2097] Unknown vLLM environment variable detected: VLLM_RTX6K_FUSED_ALLREDUCE_ADD
+WARNING 07-08 00:04:47 [envs.py:2097] Unknown vLLM environment variable detected: VLLM_CACHE_DIR
+WARNING 07-08 00:04:47 [envs.py:2097] Unknown vLLM environment variable detected: VLLM_CPP_AR_IGNORE_CUTOFF_MAX_ROWS
+WARNING 07-08 00:04:47 [envs.py:2097] Unknown vLLM environment variable detected: VLLM_RTX6K_FUSED_ALLREDUCE_ADD_END_BARRIER
+WARNING 07-08 00:04:47 [envs.py:2097] Unknown vLLM environment variable detected: VLLM_MEMORY_PROFILE_INCLUDE_ATTN
+WARNING 07-08 00:04:47 [envs.py:2097] Unknown vLLM environment variable detected: VLLM_CPP_AR_1STAGE_NCCL_CUTOFF
+INFO 07-08 00:04:47 [model.py:599] Resolved architecture: GlmMoeDsaForCausalLM
+INFO 07-08 00:04:47 [model.py:1726] Using max model len 4096
+INFO 07-08 00:04:47 [cache.py:279] Using fp8 data type to store kv cache. It reduces the GPU memory footprint and boosts the performance. Meanwhile, it may cause accuracy drop without a proper scaling factor
+INFO 07-08 00:04:47 [scheduler.py:252] Chunked prefill is enabled with max_num_batched_tokens=512.
+INFO 07-08 00:04:47 [vllm.py:1057] Asynchronous scheduling is enabled.
+WARNING 07-08 00:04:47 [vllm.py:1113] Enforce eager set, disabling torch.compile and CUDAGraphs. This is equivalent to setting -cc.mode=none -cc.cudagraph_mode=none
+WARNING 07-08 00:04:47 [vllm.py:1142] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored.
+INFO 07-08 00:04:47 [kernel.py:280] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native'])
+INFO 07-08 00:04:47 [vllm.py:1320] Cudagraph is disabled under eager mode
+INFO 07-08 00:04:47 [compilation.py:310] Enabled custom fusions: norm_quant, act_quant, allreduce_rms
+[transformers] The following generation flags are not valid and may be ignored: ['top_p']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
+/opt/venv/lib/python3.12/site-packages/torch/cuda/__init__.py:63: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+(EngineCore pid=235) INFO 07-08 00:04:58 [core.py:114] Initializing a V1 LLM engine (v0.11.2.dev279+eldritch.enlightenment.v7.vllme2e2eaf.b12x26144c0.cu132.20260707) with config: model='/root/.cache/huggingface/hub/models--zai-org--GLM-5.2/snapshots/4d67f66cc64d3219133b767c253b2ad1425c6c88', speculative_config=None, tokenizer='/root/.cache/huggingface/hub/models--zai-org--GLM-5.2/snapshots/4d67f66cc64d3219133b767c253b2ad1425c6c88', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=True, dtype=torch.bfloat16, max_seq_len=4096, download_dir=None, load_format=safetensors, tensor_parallel_size=16, pipeline_parallel_size=1, data_parallel_size=1, decode_context_parallel_size=1, dcp_comm_backend=ag_rs, disable_custom_all_reduce=True, quantization=None, quantization_config=None, enforce_eager=True, enable_return_routed_experts=False, kv_cache_dtype=fp8, device_config=cuda, structured_outputs_config=StructuredOutputsConfig(backend='auto', disable_any_whitespace=False, disable_additional_properties=False, reasoning_parser='', reasoning_parser_plugin='', enable_in_reasoning=False), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None, kv_cache_metrics=False, kv_cache_metrics_sample=0.01, cudagraph_metrics=False, enable_layerwise_nvtx_tracing=False, enable_mfu_metrics=False, enable_mm_processor_stats=False, enable_logging_iteration_details=False, jit_monitor_verbose=False), seed=0, served_model_name=/root/.cache/huggingface/hub/models--zai-org--GLM-5.2/snapshots/4d67f66cc64d3219133b767c253b2ad1425c6c88, enable_prefix_caching=False, enable_chunked_prefill=True, pooler_config=None, compilation_config={'mode': <CompilationMode.NONE: 0>, 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['all'], 'ir_enable_torch_wrap': False, 'splitting_ops': [], 'compile_mm_encoder': False, 'cudagraph_mm_encoder': False, 'encoder_cudagraph_token_budgets': [], 'encoder_cudagraph_max_vision_items_per_batch': 0, 'encoder_cudagraph_max_frames_per_batch': None, 'compile_sizes': [], 'compile_ranges_endpoints': [512], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': <CUDAGraphMode.NONE: 0>, 'cudagraph_num_of_warmups': 0, 'cudagraph_capture_sizes': [], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': True, 'fuse_act_quant': True, 'fuse_attn_quant': False, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': True, 'fuse_rope_kvcache_cat_mla': False, 'fuse_act_padding': False}, 'max_cudagraph_capture_size': 0, 'dynamic_shapes_config': {'type': <DynamicShapesType.BACKED: 'backed'>, 'evaluate_guards': False, 'assume_32_bit_indexing': False}, 'local_cache_dir': None, 'fast_moe_cold_start': False, 'static_all_moe_layers': []}, kernel_config=KernelConfig(ir_op_priority=IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']), enable_flashinfer_autotune=True, moe_backend='auto', linear_backend='auto')
+(EngineCore pid=235) INFO 07-08 00:04:58 [multiproc_executor.py:140] DP group leader: node_rank=0, node_rank_within_dp=0, master_addr=127.0.0.1, mq_connect_ip=192.168.0.175 (local), world_size=16, local_world_size=16
+/opt/venv/lib/python3.12/site-packages/torch/cuda/__init__.py:63: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+/opt/venv/lib/python3.12/site-packages/torch/cuda/__init__.py:63: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+/opt/venv/lib/python3.12/site-packages/torch/cuda/__init__.py:63: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+/opt/venv/lib/python3.12/site-packages/torch/cuda/__init__.py:63: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+/opt/venv/lib/python3.12/site-packages/torch/cuda/__init__.py:63: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+/opt/venv/lib/python3.12/site-packages/torch/cuda/__init__.py:63: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+/opt/venv/lib/python3.12/site-packages/torch/cuda/__init__.py:63: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+/opt/venv/lib/python3.12/site-packages/torch/cuda/__init__.py:63: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+/opt/venv/lib/python3.12/site-packages/torch/cuda/__init__.py:63: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+/opt/venv/lib/python3.12/site-packages/torch/cuda/__init__.py:63: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+/opt/venv/lib/python3.12/site-packages/torch/cuda/__init__.py:63: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+/opt/venv/lib/python3.12/site-packages/torch/cuda/__init__.py:63: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+/opt/venv/lib/python3.12/site-packages/torch/cuda/__init__.py:63: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+/opt/venv/lib/python3.12/site-packages/torch/cuda/__init__.py:63: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+/opt/venv/lib/python3.12/site-packages/torch/cuda/__init__.py:63: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+/opt/venv/lib/python3.12/site-packages/torch/cuda/__init__.py:63: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+(Worker pid=277) INFO 07-08 00:05:24 [parallel_state.py:1597] world_size=16 rank=0 local_rank=0 distributed_init_method=tcp://127.0.0.1:48353 backend=nccl
+(Worker pid=286) INFO 07-08 00:05:28 [parallel_state.py:1597] world_size=16 rank=9 local_rank=9 distributed_init_method=tcp://127.0.0.1:48353 backend=nccl
+(Worker pid=284) INFO 07-08 00:05:28 [parallel_state.py:1597] world_size=16 rank=7 local_rank=7 distributed_init_method=tcp://127.0.0.1:48353 backend=nccl
+(Worker pid=285) INFO 07-08 00:05:28 [parallel_state.py:1597] world_size=16 rank=8 local_rank=8 distributed_init_method=tcp://127.0.0.1:48353 backend=nccl
+(Worker pid=278) INFO 07-08 00:05:28 [parallel_state.py:1597] world_size=16 rank=1 local_rank=1 distributed_init_method=tcp://127.0.0.1:48353 backend=nccl
+(Worker pid=282) INFO 07-08 00:05:28 [parallel_state.py:1597] world_size=16 rank=5 local_rank=5 distributed_init_method=tcp://127.0.0.1:48353 backend=nccl
+(Worker pid=280) INFO 07-08 00:05:28 [parallel_state.py:1597] world_size=16 rank=3 local_rank=3 distributed_init_method=tcp://127.0.0.1:48353 backend=nccl
+(Worker pid=283) INFO 07-08 00:05:28 [parallel_state.py:1597] world_size=16 rank=6 local_rank=6 distributed_init_method=tcp://127.0.0.1:48353 backend=nccl
+(Worker pid=292) INFO 07-08 00:05:28 [parallel_state.py:1597] world_size=16 rank=15 local_rank=15 distributed_init_method=tcp://127.0.0.1:48353 backend=nccl
+(Worker pid=289) INFO 07-08 00:05:28 [parallel_state.py:1597] world_size=16 rank=12 local_rank=12 distributed_init_method=tcp://127.0.0.1:48353 backend=nccl
+(Worker pid=290) INFO 07-08 00:05:28 [parallel_state.py:1597] world_size=16 rank=13 local_rank=13 distributed_init_method=tcp://127.0.0.1:48353 backend=nccl
+(Worker pid=288) INFO 07-08 00:05:28 [parallel_state.py:1597] world_size=16 rank=11 local_rank=11 distributed_init_method=tcp://127.0.0.1:48353 backend=nccl
+(Worker pid=279) INFO 07-08 00:05:28 [parallel_state.py:1597] world_size=16 rank=2 local_rank=2 distributed_init_method=tcp://127.0.0.1:48353 backend=nccl
+(Worker pid=281) INFO 07-08 00:05:28 [parallel_state.py:1597] world_size=16 rank=4 local_rank=4 distributed_init_method=tcp://127.0.0.1:48353 backend=nccl
+(Worker pid=287) INFO 07-08 00:05:28 [parallel_state.py:1597] world_size=16 rank=10 local_rank=10 distributed_init_method=tcp://127.0.0.1:48353 backend=nccl
+(Worker pid=291) INFO 07-08 00:05:28 [parallel_state.py:1597] world_size=16 rank=14 local_rank=14 distributed_init_method=tcp://127.0.0.1:48353 backend=nccl
+(Worker pid=280) (Worker pid=278) INFO 07-08 00:05:28 [nccl.py:24] Found nccl from environment variable VLLM_NCCL_SO_PATH=/opt/libnccl-local-inference.so.2.30.4
+INFO 07-08 00:05:28 [nccl.py:24] Found nccl from environment variable VLLM_NCCL_SO_PATH=/opt/libnccl-local-inference.so.2.30.4
+(Worker pid=282) INFO 07-08 00:05:28 [nccl.py:24] Found nccl from environment variable VLLM_NCCL_SO_PATH=/opt/libnccl-local-inference.so.2.30.4
+(Worker pid=291) (Worker pid=286) (Worker pid=292) (Worker pid=279) INFO 07-08 00:05:28 [nccl.py:24] Found nccl from environment variable VLLM_NCCL_SO_PATH=/opt/libnccl-local-inference.so.2.30.4
+INFO 07-08 00:05:28 [nccl.py:24] Found nccl from environment variable VLLM_NCCL_SO_PATH=/opt/libnccl-local-inference.so.2.30.4
+INFO 07-08 00:05:28 [nccl.py:24] Found nccl from environment variable VLLM_NCCL_SO_PATH=/opt/libnccl-local-inference.so.2.30.4
+INFO 07-08 00:05:28 [nccl.py:24] Found nccl from environment variable VLLM_NCCL_SO_PATH=/opt/libnccl-local-inference.so.2.30.4
+(Worker pid=287) INFO 07-08 00:05:28 [nccl.py:24] Found nccl from environment variable VLLM_NCCL_SO_PATH=/opt/libnccl-local-inference.so.2.30.4
+(Worker pid=290) (Worker pid=277) INFO 07-08 00:05:28 [nccl.py:24] Found nccl from environment variable VLLM_NCCL_SO_PATH=/opt/libnccl-local-inference.so.2.30.4
+INFO 07-08 00:05:28 [nccl.py:24] Found nccl from environment variable VLLM_NCCL_SO_PATH=/opt/libnccl-local-inference.so.2.30.4
+(Worker pid=285) (Worker pid=283) INFO 07-08 00:05:28 [nccl.py:24] Found nccl from environment variable VLLM_NCCL_SO_PATH=/opt/libnccl-local-inference.so.2.30.4
+INFO 07-08 00:05:28 [nccl.py:24] Found nccl from environment variable VLLM_NCCL_SO_PATH=/opt/libnccl-local-inference.so.2.30.4
+(Worker pid=284) INFO 07-08 00:05:28 [nccl.py:24] Found nccl from environment variable VLLM_NCCL_SO_PATH=/opt/libnccl-local-inference.so.2.30.4
+(Worker pid=281) (Worker pid=288) INFO 07-08 00:05:28 [nccl.py:24] Found nccl from environment variable VLLM_NCCL_SO_PATH=/opt/libnccl-local-inference.so.2.30.4
+INFO 07-08 00:05:28 [nccl.py:24] Found nccl from environment variable VLLM_NCCL_SO_PATH=/opt/libnccl-local-inference.so.2.30.4
+(Worker pid=289) INFO 07-08 00:05:28 [nccl.py:24] Found nccl from environment variable VLLM_NCCL_SO_PATH=/opt/libnccl-local-inference.so.2.30.4
+(Worker pid=277) INFO 07-08 00:05:28 [pynccl.py:113] vLLM is using nccl==2.30.4
+(Worker pid=278) (Worker pid=286) WARNING 07-08 00:05:29 [symm_mem.py:66] SymmMemCommunicator: Device capability 12.0 not supported, communicator is not available.
+WARNING 07-08 00:05:29 [symm_mem.py:66] SymmMemCommunicator: Device capability 12.0 not supported, communicator is not available.
+(Worker pid=279) (Worker pid=290) (Worker pid=284) (Worker pid=288) (Worker pid=292) (Worker pid=283) (Worker pid=277) (Worker pid=280) (Worker pid=291) (Worker pid=289) WARNING 07-08 00:05:29 [symm_mem.py:66] SymmMemCommunicator: Device capability 12.0 not supported, communicator is not available.
+(Worker pid=282) (Worker pid=287) (Worker pid=281) WARNING 07-08 00:05:29 [symm_mem.py:66] SymmMemCommunicator: Device capability 12.0 not supported, communicator is not available.
+WARNING 07-08 00:05:29 [symm_mem.py:66] SymmMemCommunicator: Device capability 12.0 not supported, communicator is not available.
+WARNING 07-08 00:05:29 [symm_mem.py:66] SymmMemCommunicator: Device capability 12.0 not supported, communicator is not available.
+WARNING 07-08 00:05:29 [symm_mem.py:66] SymmMemCommunicator: Device capability 12.0 not supported, communicator is not available.
+WARNING 07-08 00:05:29 [symm_mem.py:66] SymmMemCommunicator: Device capability 12.0 not supported, communicator is not available.
+WARNING 07-08 00:05:29 [symm_mem.py:66] SymmMemCommunicator: Device capability 12.0 not supported, communicator is not available.
+WARNING 07-08 00:05:29 [symm_mem.py:66] SymmMemCommunicator: Device capability 12.0 not supported, communicator is not available.
+WARNING 07-08 00:05:29 [symm_mem.py:66] SymmMemCommunicator: Device capability 12.0 not supported, communicator is not available.
+WARNING 07-08 00:05:29 [symm_mem.py:66] SymmMemCommunicator: Device capability 12.0 not supported, communicator is not available.
+WARNING 07-08 00:05:29 [symm_mem.py:66] SymmMemCommunicator: Device capability 12.0 not supported, communicator is not available.
+WARNING 07-08 00:05:29 [symm_mem.py:66] SymmMemCommunicator: Device capability 12.0 not supported, communicator is not available.
+WARNING 07-08 00:05:29 [symm_mem.py:66] SymmMemCommunicator: Device capability 12.0 not supported, communicator is not available.
+(Worker pid=285) WARNING 07-08 00:05:29 [symm_mem.py:66] SymmMemCommunicator: Device capability 12.0 not supported, communicator is not available.
+(Worker pid=277) INFO 07-08 00:05:29 [cuda_communicator.py:246] Using ['PYNCCL'] all-reduce backends (in dispatch order) for group 'tp:0' out of potential backends: ['NCCL_SYMM_MEM', 'QUICK_REDUCE', 'FLASHINFER', 'CUSTOM', 'SYMM_MEM', 'PYNCCL'].
+(Worker pid=277) INFO 07-08 00:05:29 [nccl.py:24] Found nccl from environment variable VLLM_NCCL_SO_PATH=/opt/libnccl-local-inference.so.2.30.4
+(Worker pid=281) INFO 07-08 00:05:29 [nccl.py:24] Found nccl from environment variable VLLM_NCCL_SO_PATH=/opt/libnccl-local-inference.so.2.30.4
+(Worker pid=278) INFO 07-08 00:05:29 [nccl.py:24] Found nccl from environment variable VLLM_NCCL_SO_PATH=/opt/libnccl-local-inference.so.2.30.4
+(Worker pid=279) INFO 07-08 00:05:29 [nccl.py:24] Found nccl from environment variable VLLM_NCCL_SO_PATH=/opt/libnccl-local-inference.so.2.30.4
+(Worker pid=283) INFO 07-08 00:05:29 [nccl.py:24] Found nccl from environment variable VLLM_NCCL_SO_PATH=/opt/libnccl-local-inference.so.2.30.4
+(Worker pid=280) INFO 07-08 00:05:29 [nccl.py:24] Found nccl from environment variable VLLM_NCCL_SO_PATH=/opt/libnccl-local-inference.so.2.30.4
+(Worker pid=282) INFO 07-08 00:05:29 [nccl.py:24] Found nccl from environment variable VLLM_NCCL_SO_PATH=/opt/libnccl-local-inference.so.2.30.4
+(Worker pid=289) INFO 07-08 00:05:29 [nccl.py:24] Found nccl from environment variable VLLM_NCCL_SO_PATH=/opt/libnccl-local-inference.so.2.30.4
+(Worker pid=284) INFO 07-08 00:05:29 [nccl.py:24] Found nccl from environment variable VLLM_NCCL_SO_PATH=/opt/libnccl-local-inference.so.2.30.4
+(Worker pid=291) INFO 07-08 00:05:29 [nccl.py:24] Found nccl from environment variable VLLM_NCCL_SO_PATH=/opt/libnccl-local-inference.so.2.30.4
+(Worker pid=285) INFO 07-08 00:05:29 [nccl.py:24] Found nccl from environment variable VLLM_NCCL_SO_PATH=/opt/libnccl-local-inference.so.2.30.4
+(Worker pid=287) INFO 07-08 00:05:29 [nccl.py:24] Found nccl from environment variable VLLM_NCCL_SO_PATH=/opt/libnccl-local-inference.so.2.30.4
+(Worker pid=290) INFO 07-08 00:05:29 [nccl.py:24] Found nccl from environment variable VLLM_NCCL_SO_PATH=/opt/libnccl-local-inference.so.2.30.4
+(Worker pid=288) INFO 07-08 00:05:29 [nccl.py:24] Found nccl from environment variable VLLM_NCCL_SO_PATH=/opt/libnccl-local-inference.so.2.30.4
+(Worker pid=286) INFO 07-08 00:05:29 [nccl.py:24] Found nccl from environment variable VLLM_NCCL_SO_PATH=/opt/libnccl-local-inference.so.2.30.4
+(Worker pid=292) INFO 07-08 00:05:29 [nccl.py:24] Found nccl from environment variable VLLM_NCCL_SO_PATH=/opt/libnccl-local-inference.so.2.30.4
+(Worker pid=277) INFO 07-08 00:05:30 [cuda_communicator.py:246] Using ['PYNCCL'] all-reduce backends (in dispatch order) for group 'ep:0' out of potential backends: ['NCCL_SYMM_MEM', 'QUICK_REDUCE', 'FLASHINFER', 'CUSTOM', 'SYMM_MEM', 'PYNCCL'].
+(Worker pid=277) INFO 07-08 00:05:30 [parallel_state.py:1932] rank 0 in world size 16 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank 0, EPLB rank N/A
+(Worker pid=277) INFO 07-08 00:05:30 [gpu_worker.py:334] Using V2 Model Runner
+(Worker_TP5 pid=282) INFO 07-08 00:05:31 [model_runner.py:312] Loading model from scratch...
+(Worker_TP1 pid=278) INFO 07-08 00:05:31 [model_runner.py:312] Loading model from scratch...
+(Worker_TP13 pid=290) INFO 07-08 00:05:31 [model_runner.py:312] Loading model from scratch...
+(Worker_TP4 pid=281) INFO 07-08 00:05:31 [model_runner.py:312] Loading model from scratch...
+(Worker_TP9 pid=286) INFO 07-08 00:05:31 [model_runner.py:312] Loading model from scratch...
+(Worker_TP11 pid=288) INFO 07-08 00:05:31 [model_runner.py:312] Loading model from scratch...
+(Worker_TP12 pid=289) INFO 07-08 00:05:31 [model_runner.py:312] Loading model from scratch...
+(Worker_TP8 pid=285) INFO 07-08 00:05:31 [model_runner.py:312] Loading model from scratch...
+(Worker_TP5 pid=282) INFO 07-08 00:05:31 [cuda.py:420] Using AttentionBackendEnum.B12X_MLA_SPARSE backend.
+(Worker_TP0 pid=277) INFO 07-08 00:05:31 [model_runner.py:312] Loading model from scratch...
+(Worker_TP10 pid=287) INFO 07-08 00:05:31 [model_runner.py:312] Loading model from scratch...
+(Worker_TP6 pid=283) INFO 07-08 00:05:31 [model_runner.py:312] Loading model from scratch...
+(Worker_TP15 pid=292) INFO 07-08 00:05:31 [model_runner.py:312] Loading model from scratch...
+(Worker_TP3 pid=280) INFO 07-08 00:05:31 [model_runner.py:312] Loading model from scratch...
+(Worker_TP7 pid=284) INFO 07-08 00:05:31 [model_runner.py:312] Loading model from scratch...
+(Worker_TP2 pid=279) INFO 07-08 00:05:31 [model_runner.py:312] Loading model from scratch...
+(Worker_TP1 pid=278) INFO 07-08 00:05:31 [cuda.py:420] Using AttentionBackendEnum.B12X_MLA_SPARSE backend.
+(Worker_TP14 pid=291) INFO 07-08 00:05:31 [model_runner.py:312] Loading model from scratch...
+(Worker_TP13 pid=290) INFO 07-08 00:05:31 [cuda.py:420] Using AttentionBackendEnum.B12X_MLA_SPARSE backend.
+(Worker_TP4 pid=281) INFO 07-08 00:05:31 [cuda.py:420] Using AttentionBackendEnum.B12X_MLA_SPARSE backend.
+(Worker_TP9 pid=286) INFO 07-08 00:05:31 [cuda.py:420] Using AttentionBackendEnum.B12X_MLA_SPARSE backend.
+(Worker_TP11 pid=288) INFO 07-08 00:05:31 [cuda.py:420] Using AttentionBackendEnum.B12X_MLA_SPARSE backend.
+(Worker_TP0 pid=277) INFO 07-08 00:05:31 [cuda.py:420] Using AttentionBackendEnum.B12X_MLA_SPARSE backend.
+(Worker_TP0 pid=277) INFO 07-08 00:05:31 [mla_attention.py:452] Using fp8_ds_mla KV cache format for B12X_MLA_SPARSE backend.
+(Worker_TP12 pid=289) INFO 07-08 00:05:31 [cuda.py:420] Using AttentionBackendEnum.B12X_MLA_SPARSE backend.
+(Worker_TP8 pid=285) INFO 07-08 00:05:31 [cuda.py:420] Using AttentionBackendEnum.B12X_MLA_SPARSE backend.
+(Worker_TP10 pid=287) INFO 07-08 00:05:31 [cuda.py:420] Using AttentionBackendEnum.B12X_MLA_SPARSE backend.
+(Worker_TP6 pid=283) INFO 07-08 00:05:31 [cuda.py:420] Using AttentionBackendEnum.B12X_MLA_SPARSE backend.
+(Worker_TP15 pid=292) INFO 07-08 00:05:31 [cuda.py:420] Using AttentionBackendEnum.B12X_MLA_SPARSE backend.
+(Worker_TP7 pid=284) INFO 07-08 00:05:31 [cuda.py:420] Using AttentionBackendEnum.B12X_MLA_SPARSE backend.
+(Worker_TP3 pid=280) INFO 07-08 00:05:31 [cuda.py:420] Using AttentionBackendEnum.B12X_MLA_SPARSE backend.
+(Worker_TP2 pid=279) INFO 07-08 00:05:31 [cuda.py:420] Using AttentionBackendEnum.B12X_MLA_SPARSE backend.
+(Worker_TP14 pid=291) INFO 07-08 00:05:31 [cuda.py:420] Using AttentionBackendEnum.B12X_MLA_SPARSE backend.
+(Worker_TP5 pid=282) INFO 07-08 00:05:31 [cuda.py:420] Using AttentionBackendEnum.B12X_MLA_SPARSE backend.
+(Worker_TP13 pid=290) INFO 07-08 00:05:31 [cuda.py:420] Using AttentionBackendEnum.B12X_MLA_SPARSE backend.
+(Worker_TP1 pid=278) INFO 07-08 00:05:31 [cuda.py:420] Using AttentionBackendEnum.B12X_MLA_SPARSE backend.
+(Worker_TP0 pid=277) INFO 07-08 00:05:31 [selector.py:166] Using FLASH_ATTN MLA prefill backend.
+(Worker_TP4 pid=281) INFO 07-08 00:05:31 [cuda.py:420] Using AttentionBackendEnum.B12X_MLA_SPARSE backend.
+(Worker_TP0 pid=277) INFO 07-08 00:05:31 [deep_gemm.py:202] DeepGEMM PDL enabled on deep_gemm.
+(Worker_TP0 pid=277) INFO 07-08 00:05:31 [deep_gemm.py:120] DeepGEMM E8M0 enabled on current platform.
+(Worker_TP11 pid=288) INFO 07-08 00:05:31 [cuda.py:420] Using AttentionBackendEnum.B12X_MLA_SPARSE backend.
+(Worker_TP9 pid=286) INFO 07-08 00:05:31 [cuda.py:420] Using AttentionBackendEnum.B12X_MLA_SPARSE backend.
+(Worker_TP0 pid=277) INFO 07-08 00:05:31 [cuda.py:420] Using AttentionBackendEnum.B12X_MLA_SPARSE backend.
+(Worker_TP12 pid=289) INFO 07-08 00:05:31 [cuda.py:420] Using AttentionBackendEnum.B12X_MLA_SPARSE backend.
+(Worker_TP0 pid=277) INFO 07-08 00:05:31 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 3.
+(Worker_TP0 pid=277) INFO 07-08 00:05:31 [unquantized.py:247] Using FlashInfer CUTLASS Unquantized MoE backend out of potential backends: ['FlashInfer TRTLLM', 'FlashInfer CUTLASS', 'TRITON', 'BATCHED_TRITON'].
+(Worker_TP8 pid=285) INFO 07-08 00:05:31 [cuda.py:420] Using AttentionBackendEnum.B12X_MLA_SPARSE backend.
+(Worker_TP6 pid=283) INFO 07-08 00:05:31 [cuda.py:420] Using AttentionBackendEnum.B12X_MLA_SPARSE backend.
+(Worker_TP10 pid=287) INFO 07-08 00:05:31 [cuda.py:420] Using AttentionBackendEnum.B12X_MLA_SPARSE backend.
+(Worker_TP15 pid=292) INFO 07-08 00:05:31 [cuda.py:420] Using AttentionBackendEnum.B12X_MLA_SPARSE backend.
+(Worker_TP7 pid=284) INFO 07-08 00:05:31 [cuda.py:420] Using AttentionBackendEnum.B12X_MLA_SPARSE backend.
+(Worker_TP3 pid=280) INFO 07-08 00:05:31 [cuda.py:420] Using AttentionBackendEnum.B12X_MLA_SPARSE backend.
+(Worker_TP2 pid=279) INFO 07-08 00:05:31 [cuda.py:420] Using AttentionBackendEnum.B12X_MLA_SPARSE backend.
+(Worker_TP0 pid=277) INFO 07-08 00:05:31 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 4.
+(Worker_TP14 pid=291) INFO 07-08 00:05:31 [cuda.py:420] Using AttentionBackendEnum.B12X_MLA_SPARSE backend.
+(Worker_TP0 pid=277) INFO 07-08 00:05:31 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 5.
+(Worker_TP0 pid=277) INFO 07-08 00:05:31 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 7.
+(Worker_TP0 pid=277) INFO 07-08 00:05:32 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 8.
+(Worker_TP0 pid=277) INFO 07-08 00:05:32 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 9.
+(Worker_TP0 pid=277) INFO 07-08 00:05:32 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 11.
+(Worker_TP0 pid=277) INFO 07-08 00:05:32 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 12.
+(Worker_TP0 pid=277) INFO 07-08 00:05:32 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 13.
+(Worker_TP0 pid=277) INFO 07-08 00:05:32 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 15.
+(Worker_TP0 pid=277) INFO 07-08 00:05:32 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 16.
+(Worker_TP0 pid=277) INFO 07-08 00:05:32 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 17.
+(Worker_TP0 pid=277) INFO 07-08 00:05:33 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 19.
+(Worker_TP0 pid=277) INFO 07-08 00:05:33 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 20.
+(Worker_TP0 pid=277) INFO 07-08 00:05:33 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 21.
+(Worker_TP0 pid=277) INFO 07-08 00:05:33 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 23.
+(Worker_TP0 pid=277) INFO 07-08 00:05:33 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 24.
+(Worker_TP0 pid=277) INFO 07-08 00:05:33 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 25.
+(Worker_TP0 pid=277) INFO 07-08 00:05:33 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 27.
+(Worker_TP0 pid=277) INFO 07-08 00:05:33 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 28.
+(Worker_TP0 pid=277) INFO 07-08 00:05:34 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 29.
+(Worker_TP0 pid=277) INFO 07-08 00:05:34 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 31.
+(Worker_TP0 pid=277) INFO 07-08 00:05:34 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 32.
+(Worker_TP0 pid=277) INFO 07-08 00:05:34 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 33.
+(Worker_TP0 pid=277) INFO 07-08 00:05:35 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 35.
+(Worker_TP0 pid=277) INFO 07-08 00:05:35 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 36.
+(Worker_TP0 pid=277) INFO 07-08 00:05:35 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 37.
+(Worker_TP0 pid=277) INFO 07-08 00:05:36 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 39.
+(Worker_TP0 pid=277) INFO 07-08 00:05:36 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 40.
+(Worker_TP0 pid=277) INFO 07-08 00:05:36 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 41.
+(Worker_TP0 pid=277) INFO 07-08 00:05:36 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 43.
+(Worker_TP0 pid=277) INFO 07-08 00:05:36 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 44.
+(Worker_TP0 pid=277) INFO 07-08 00:05:37 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 45.
+(Worker_TP0 pid=277) INFO 07-08 00:05:37 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 47.
+(Worker_TP0 pid=277) INFO 07-08 00:05:37 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 48.
+(Worker_TP0 pid=277) INFO 07-08 00:05:37 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 49.
+(Worker_TP0 pid=277) INFO 07-08 00:05:37 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 51.
+(Worker_TP0 pid=277) INFO 07-08 00:05:37 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 52.
+(Worker_TP0 pid=277) INFO 07-08 00:05:37 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 53.
+(Worker_TP0 pid=277) INFO 07-08 00:05:37 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 55.
+(Worker_TP0 pid=277) INFO 07-08 00:05:37 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 56.
+(Worker_TP0 pid=277) INFO 07-08 00:05:38 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 57.
+(Worker_TP0 pid=277) INFO 07-08 00:05:38 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 59.
+(Worker_TP0 pid=277) INFO 07-08 00:05:38 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 60.
+(Worker_TP0 pid=277) INFO 07-08 00:05:38 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 61.
+(Worker_TP0 pid=277) INFO 07-08 00:05:38 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 63.
+(Worker_TP0 pid=277) INFO 07-08 00:05:38 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 64.
+(Worker_TP0 pid=277) INFO 07-08 00:05:38 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 65.
+(Worker_TP0 pid=277) INFO 07-08 00:05:38 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 67.
+(Worker_TP0 pid=277) INFO 07-08 00:05:39 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 68.
+(Worker_TP0 pid=277) INFO 07-08 00:05:39 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 69.
+(Worker_TP0 pid=277) INFO 07-08 00:05:39 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 71.
+(Worker_TP0 pid=277) INFO 07-08 00:05:39 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 72.
+(Worker_TP0 pid=277) INFO 07-08 00:05:39 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 73.
+(Worker_TP0 pid=277) INFO 07-08 00:05:39 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 75.
+(Worker_TP0 pid=277) INFO 07-08 00:05:39 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 76.
+(Worker_TP0 pid=277) INFO 07-08 00:05:39 [deepseek_v2.py:1179] Using index_topk_pattern/index_topk_freq to skip sparse MLA indexer computation on layer 77.
+(Worker_TP0 pid=277) INFO 07-08 00:05:40 [weight_utils.py:914] Filesystem type for checkpoints: EXT4. Checkpoint size: 1403.19 GiB. Available RAM: 911.37 GiB.
+(Worker_TP0 pid=277) INFO 07-08 00:05:40 [weight_utils.py:944] Auto-prefetch is disabled because the filesystem (EXT4) is not a recognized network FS (NFS/Lustre) and the checkpoint size (1403.19 GiB) exceeds 90% of available RAM (911.37 GiB).
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) INFO 07-08 00:10:51 [default_loader.py:451] Loading weights took 310.85 seconds
+(Worker_TP0 pid=277) INFO 07-08 00:10:51 [unquantized.py:312] Using MoEPrepareAndFinalizeNoDPEPModular
+(Worker_TP8 pid=285) INFO 07-08 00:10:53 [model_runner.py:333] Model loading took 89.59 GiB and 322.874489 seconds
+(Worker_TP8 pid=285) INFO 07-08 00:10:53 [interface.py:620] Setting kv cache block size to 64 for B12X_NON_COMPRESSED_INDEXER backend.
+(Worker_TP4 pid=281) INFO 07-08 00:10:53 [model_runner.py:333] Model loading took 89.59 GiB and 322.943765 seconds
+(Worker_TP10 pid=287) INFO 07-08 00:10:53 [model_runner.py:333] Model loading took 89.59 GiB and 322.934795 seconds
+(Worker_TP4 pid=281) INFO 07-08 00:10:53 [interface.py:620] Setting kv cache block size to 64 for B12X_NON_COMPRESSED_INDEXER backend.
+(Worker_TP10 pid=287) INFO 07-08 00:10:53 [interface.py:620] Setting kv cache block size to 64 for B12X_NON_COMPRESSED_INDEXER backend.
+(Worker_TP15 pid=292) INFO 07-08 00:10:53 [model_runner.py:333] Model loading took 89.59 GiB and 322.951303 seconds
+(Worker_TP3 pid=280) INFO 07-08 00:10:53 [model_runner.py:333] Model loading took 89.59 GiB and 322.959631 seconds
+(Worker_TP7 pid=284) INFO 07-08 00:10:53 [model_runner.py:333] Model loading took 89.59 GiB and 322.958586 seconds
+(Worker_TP3 pid=280) INFO 07-08 00:10:53 [interface.py:620] Setting kv cache block size to 64 for B12X_NON_COMPRESSED_INDEXER backend.
+(Worker_TP9 pid=286) INFO 07-08 00:10:53 [model_runner.py:333] Model loading took 89.59 GiB and 323.011432 seconds
+(Worker_TP15 pid=292) INFO 07-08 00:10:53 [interface.py:620] Setting kv cache block size to 64 for B12X_NON_COMPRESSED_INDEXER backend.
+(Worker_TP14 pid=291) INFO 07-08 00:10:53 [model_runner.py:333] Model loading took 89.59 GiB and 322.958301 seconds
+(Worker_TP1 pid=278) INFO 07-08 00:10:53 [model_runner.py:333] Model loading took 89.59 GiB and 323.002111 seconds
+(Worker_TP7 pid=284) INFO 07-08 00:10:53 [interface.py:620] Setting kv cache block size to 64 for B12X_NON_COMPRESSED_INDEXER backend.
+(Worker_TP5 pid=282) INFO 07-08 00:10:53 [model_runner.py:333] Model loading took 89.59 GiB and 323.039836 seconds
+(Worker_TP9 pid=286) INFO 07-08 00:10:53 [interface.py:620] Setting kv cache block size to 64 for B12X_NON_COMPRESSED_INDEXER backend.
+(Worker_TP0 pid=277) INFO 07-08 00:10:53 [model_runner.py:333] Model loading took 89.59 GiB and 322.983290 seconds
+(Worker_TP14 pid=291) INFO 07-08 00:10:53 [interface.py:620] Setting kv cache block size to 64 for B12X_NON_COMPRESSED_INDEXER backend.
+(Worker_TP1 pid=278) INFO 07-08 00:10:53 [interface.py:620] Setting kv cache block size to 64 for B12X_NON_COMPRESSED_INDEXER backend.
+(Worker_TP5 pid=282) INFO 07-08 00:10:53 [interface.py:620] Setting kv cache block size to 64 for B12X_NON_COMPRESSED_INDEXER backend.
+(Worker_TP12 pid=289) INFO 07-08 00:10:53 [model_runner.py:333] Model loading took 89.59 GiB and 323.012290 seconds
+(Worker_TP0 pid=277) INFO 07-08 00:10:53 [topk_topp_sampler.py:55] Using FlashInfer for top-p & top-k sampling.
+(Worker_TP0 pid=277) INFO 07-08 00:10:53 [interface.py:620] Setting kv cache block size to 64 for B12X_NON_COMPRESSED_INDEXER backend.
+(Worker_TP2 pid=279) INFO 07-08 00:10:53 [model_runner.py:333] Model loading took 89.59 GiB and 323.005942 seconds
+(Worker_TP13 pid=290) INFO 07-08 00:10:53 [model_runner.py:333] Model loading took 89.59 GiB and 323.033835 seconds
+(Worker_TP12 pid=289) INFO 07-08 00:10:53 [interface.py:620] Setting kv cache block size to 64 for B12X_NON_COMPRESSED_INDEXER backend.
+(Worker_TP11 pid=288) INFO 07-08 00:10:53 [model_runner.py:333] Model loading took 89.59 GiB and 323.024354 seconds
+(Worker_TP2 pid=279) INFO 07-08 00:10:53 [interface.py:620] Setting kv cache block size to 64 for B12X_NON_COMPRESSED_INDEXER backend.
+(Worker_TP6 pid=283) INFO 07-08 00:10:53 [model_runner.py:333] Model loading took 89.59 GiB and 323.029203 seconds
+(Worker_TP13 pid=290) INFO 07-08 00:10:53 [interface.py:620] Setting kv cache block size to 64 for B12X_NON_COMPRESSED_INDEXER backend.
+(Worker_TP11 pid=288) INFO 07-08 00:10:53 [interface.py:620] Setting kv cache block size to 64 for B12X_NON_COMPRESSED_INDEXER backend.
+(Worker_TP6 pid=283) INFO 07-08 00:10:53 [interface.py:620] Setting kv cache block size to 64 for B12X_NON_COMPRESSED_INDEXER backend.
+(Worker_TP0 pid=277) INFO 07-08 00:10:57 [gpu_worker.py:517] Available KV cache memory: 3.18 GiB
+(EngineCore pid=235) INFO 07-08 00:10:57 [kv_cache_utils.py:2202] GPU KV cache size: 63,232 tokens
+(EngineCore pid=235) INFO 07-08 00:10:57 [kv_cache_utils.py:2203] Maximum concurrency for 4,096 tokens per request: 15.44x
+(Worker_TP0 pid=277) INFO 07-08 00:10:57 [indexer.py:322] DSA indexer decode path: use_flattening=False (next_n=1, use_fp4_indexer_cache=False)
+(Worker_TP15 pid=292) INFO 07-08 00:10:59 [kernel_warmup.py:258] Warmed up 15 B12X sparse-indexer decode variants.
+(Worker_TP8 pid=285) INFO 07-08 00:10:59 [kernel_warmup.py:258] Warmed up 15 B12X sparse-indexer decode variants.
+(Worker_TP2 pid=279) INFO 07-08 00:10:59 [kernel_warmup.py:258] Warmed up 15 B12X sparse-indexer decode variants.
+(Worker_TP3 pid=280) INFO 07-08 00:10:59 [kernel_warmup.py:258] Warmed up 15 B12X sparse-indexer decode variants.
+(Worker_TP7 pid=284) INFO 07-08 00:10:59 [kernel_warmup.py:258] Warmed up 15 B12X sparse-indexer decode variants.
+(Worker_TP1 pid=278) INFO 07-08 00:10:59 [kernel_warmup.py:258] Warmed up 15 B12X sparse-indexer decode variants.
+(Worker_TP10 pid=287) INFO 07-08 00:10:59 [kernel_warmup.py:258] Warmed up 15 B12X sparse-indexer decode variants.
+(Worker_TP4 pid=281) INFO 07-08 00:10:59 [kernel_warmup.py:258] Warmed up 15 B12X sparse-indexer decode variants.
+(Worker_TP11 pid=288) INFO 07-08 00:10:59 [kernel_warmup.py:258] Warmed up 15 B12X sparse-indexer decode variants.
+(Worker_TP9 pid=286) INFO 07-08 00:10:59 [kernel_warmup.py:258] Warmed up 15 B12X sparse-indexer decode variants.
+(Worker_TP12 pid=289) INFO 07-08 00:10:59 [kernel_warmup.py:258] Warmed up 15 B12X sparse-indexer decode variants.
+(Worker_TP6 pid=283) INFO 07-08 00:10:59 [kernel_warmup.py:258] Warmed up 15 B12X sparse-indexer decode variants.
+(Worker_TP0 pid=277) INFO 07-08 00:10:59 [kernel_warmup.py:258] Warmed up 15 B12X sparse-indexer decode variants.
+(Worker_TP14 pid=291) INFO 07-08 00:10:59 [kernel_warmup.py:258] Warmed up 15 B12X sparse-indexer decode variants.
+(Worker_TP5 pid=282) INFO 07-08 00:10:59 [kernel_warmup.py:258] Warmed up 15 B12X sparse-indexer decode variants.
+(Worker_TP13 pid=290) INFO 07-08 00:10:59 [kernel_warmup.py:258] Warmed up 15 B12X sparse-indexer decode variants.
+(Worker_TP15 pid=292) 2026-07-08 00:10:59,130 - INFO - autotuner.py:651 - flashinfer.jit: [Autotuner]: Autotuning process starts ...
+(Worker_TP8 pid=285) 2026-07-08 00:10:59,130 - INFO - autotuner.py:651 - flashinfer.jit: [Autotuner]: Autotuning process starts ...
+(Worker_TP2 pid=279) 2026-07-08 00:10:59,130 - INFO - autotuner.py:651 - flashinfer.jit: [Autotuner]: Autotuning process starts ...
+(Worker_TP3 pid=280) 2026-07-08 00:10:59,131 - INFO - autotuner.py:651 - flashinfer.jit: [Autotuner]: Autotuning process starts ...
+(Worker_TP7 pid=284) 2026-07-08 00:10:59,131 - INFO - autotuner.py:651 - flashinfer.jit: [Autotuner]: Autotuning process starts ...
+(Worker_TP4 pid=281) 2026-07-08 00:10:59,131 - INFO - autotuner.py:651 - flashinfer.jit: [Autotuner]: Autotuning process starts ...
+(Worker_TP1 pid=278) 2026-07-08 00:10:59,131 - INFO - autotuner.py:651 - flashinfer.jit: [Autotuner]: Autotuning process starts ...
+(Worker_TP10 pid=287) 2026-07-08 00:10:59,131 - INFO - autotuner.py:651 - flashinfer.jit: [Autotuner]: Autotuning process starts ...
+(Worker_TP11 pid=288) 2026-07-08 00:10:59,131 - INFO - autotuner.py:651 - flashinfer.jit: [Autotuner]: Autotuning process starts ...
+(Worker_TP9 pid=286) (Worker_TP12 pid=289) 2026-07-08 00:10:59,131 - INFO - autotuner.py:651 - flashinfer.jit: [Autotuner]: Autotuning process starts ...
+2026-07-08 00:10:59,131 - INFO - autotuner.py:651 - flashinfer.jit: [Autotuner]: Autotuning process starts ...
+(Worker_TP6 pid=283) 2026-07-08 00:10:59,132 - INFO - autotuner.py:651 - flashinfer.jit: [Autotuner]: Autotuning process starts ...
+(Worker_TP0 pid=277) 2026-07-08 00:10:59,132 - INFO - autotuner.py:651 - flashinfer.jit: [Autotuner]: Autotuning process starts ...
+(Worker_TP14 pid=291) 2026-07-08 00:10:59,132 - INFO - autotuner.py:651 - flashinfer.jit: [Autotuner]: Autotuning process starts ...
+(Worker_TP13 pid=290) 2026-07-08 00:10:59,132 - INFO - autotuner.py:651 - flashinfer.jit: [Autotuner]: Autotuning process starts ...
+(Worker_TP5 pid=282) 2026-07-08 00:10:59,132 - INFO - autotuner.py:651 - flashinfer.jit: [Autotuner]: Autotuning process starts ...
+(Worker_TP2 pid=279) (Worker_TP15 pid=292) 
+(Worker_TP13 pid=290) 
+(Worker_TP13 pid=290) 
+2026-07-08 00:10:59,348 - INFO - autotuner.py:1351 - flashinfer.jit: [Autotuner]: Skipped 2 unsupported tactic(s) for trtllm::fused_moe::gemm1 (enable debug logs to see details)
+(Worker_TP7 pid=284) 
+(Worker_TP7 pid=284) 
+2026-07-08 00:10:59,398 - INFO - autotuner.py:1351 - flashinfer.jit: [Autotuner]: Skipped 2 unsupported tactic(s) for trtllm::fused_moe::gemm1 (enable debug logs to see details)
+(Worker_TP2 pid=279) 
+(Worker_TP2 pid=279) 
+2026-07-08 00:10:59,448 - INFO - autotuner.py:1351 - flashinfer.jit: [Autotuner]: Skipped 2 unsupported tactic(s) for trtllm::fused_moe::gemm1 (enable debug logs to see details)
+(Worker_TP15 pid=292) 
+(Worker_TP15 pid=292) 
+2026-07-08 00:10:59,498 - INFO - autotuner.py:1351 - flashinfer.jit: [Autotuner]: Skipped 2 unsupported tactic(s) for trtllm::fused_moe::gemm1 (enable debug logs to see details)
+(Worker_TP6 pid=283) 
+(Worker_TP6 pid=283) 
+2026-07-08 00:10:59,548 - INFO - autotuner.py:1351 - flashinfer.jit: [Autotuner]: Skipped 2 unsupported tactic(s) for trtllm::fused_moe::gemm1 (enable debug logs to see details)
+(Worker_TP10 pid=287) 
+(Worker_TP10 pid=287) 
+2026-07-08 00:10:59,597 - INFO - autotuner.py:1351 - flashinfer.jit: [Autotuner]: Skipped 2 unsupported tactic(s) for trtllm::fused_moe::gemm1 (enable debug logs to see details)
+(Worker_TP14 pid=291) 
+(Worker_TP14 pid=291) 
+2026-07-08 00:10:59,648 - INFO - autotuner.py:1351 - flashinfer.jit: [Autotuner]: Skipped 2 unsupported tactic(s) for trtllm::fused_moe::gemm1 (enable debug logs to see details)
+(Worker_TP3 pid=280) 
+(Worker_TP3 pid=280) 
+2026-07-08 00:10:59,698 - INFO - autotuner.py:1351 - flashinfer.jit: [Autotuner]: Skipped 2 unsupported tactic(s) for trtllm::fused_moe::gemm1 (enable debug logs to see details)
+(Worker_TP5 pid=282) 
+(Worker_TP5 pid=282) 
+2026-07-08 00:10:59,750 - INFO - autotuner.py:1351 - flashinfer.jit: [Autotuner]: Skipped 2 unsupported tactic(s) for trtllm::fused_moe::gemm1 (enable debug logs to see details)
+(Worker_TP12 pid=289) 
+(Worker_TP12 pid=289) 
+2026-07-08 00:10:59,798 - INFO - autotuner.py:1351 - flashinfer.jit: [Autotuner]: Skipped 2 unsupported tactic(s) for trtllm::fused_moe::gemm1 (enable debug logs to see details)
+(Worker_TP9 pid=286) 
+(Worker_TP9 pid=286) 
+2026-07-08 00:10:59,849 - INFO - autotuner.py:1351 - flashinfer.jit: [Autotuner]: Skipped 2 unsupported tactic(s) for trtllm::fused_moe::gemm1 (enable debug logs to see details)
+(Worker_TP4 pid=281) 
+(Worker_TP4 pid=281) 
+2026-07-08 00:10:59,899 - INFO - autotuner.py:1351 - flashinfer.jit: [Autotuner]: Skipped 2 unsupported tactic(s) for trtllm::fused_moe::gemm1 (enable debug logs to see details)
+(Worker_TP11 pid=288) 
+(Worker_TP11 pid=288) 
+2026-07-08 00:10:59,949 - INFO - autotuner.py:1351 - flashinfer.jit: [Autotuner]: Skipped 2 unsupported tactic(s) for trtllm::fused_moe::gemm1 (enable debug logs to see details)
+(Worker_TP8 pid=285) 
+(Worker_TP8 pid=285) 
+2026-07-08 00:11:00,000 - INFO - autotuner.py:1351 - flashinfer.jit: [Autotuner]: Skipped 2 unsupported tactic(s) for trtllm::fused_moe::gemm1 (enable debug logs to see details)
+(Worker_TP1 pid=278) 
+(Worker_TP1 pid=278) 
+2026-07-08 00:11:00,051 - INFO - autotuner.py:1351 - flashinfer.jit: [Autotuner]: Skipped 2 unsupported tactic(s) for trtllm::fused_moe::gemm1 (enable debug logs to see details)
+(Worker_TP0 pid=277) 
+(Worker_TP0 pid=277) 
+(Worker_TP14 pid=291) 2026-07-08 00:11:00,159 - INFO - autotuner.py:674 - flashinfer.jit: [Autotuner]: Autotuning process ends
+(Worker_TP13 pid=290) 2026-07-08 00:11:00,159 - INFO - autotuner.py:674 - flashinfer.jit: [Autotuner]: Autotuning process ends
+(Worker_TP2 pid=279) 2026-07-08 00:11:00,159 - INFO - autotuner.py:674 - flashinfer.jit: [Autotuner]: Autotuning process ends
+(Worker_TP4 pid=281) (Worker_TP6 pid=283) (Worker_TP11 pid=288) 2026-07-08 00:11:00,159 - INFO - autotuner.py:674 - flashinfer.jit: [Autotuner]: Autotuning process ends
+2026-07-08 00:11:00,159 - INFO - autotuner.py:674 - flashinfer.jit: [Autotuner]: Autotuning process ends
+(Worker_TP10 pid=287) 2026-07-08 00:11:00,159 - INFO - autotuner.py:674 - flashinfer.jit: [Autotuner]: Autotuning process ends
+2026-07-08 00:11:00,159 - INFO - autotuner.py:674 - flashinfer.jit: [Autotuner]: Autotuning process ends
+(Worker_TP7 pid=284) 2026-07-08 00:11:00,159 - INFO - autotuner.py:674 - flashinfer.jit: [Autotuner]: Autotuning process ends
+(Worker_TP12 pid=289) 2026-07-08 00:11:00,159 - INFO - autotuner.py:674 - flashinfer.jit: [Autotuner]: Autotuning process ends
+(Worker_TP9 pid=286) 2026-07-08 00:11:00,159 - INFO - autotuner.py:674 - flashinfer.jit: [Autotuner]: Autotuning process ends
+(Worker_TP3 pid=280) 2026-07-08 00:11:00,159 - INFO - autotuner.py:674 - flashinfer.jit: [Autotuner]: Autotuning process ends
+(Worker_TP15 pid=292) 2026-07-08 00:11:00,160 - INFO - autotuner.py:674 - flashinfer.jit: [Autotuner]: Autotuning process ends
+(Worker_TP5 pid=282) 2026-07-08 00:11:00,160 - INFO - autotuner.py:674 - flashinfer.jit: [Autotuner]: Autotuning process ends
+(Worker_TP8 pid=285) 2026-07-08 00:11:00,189 - INFO - autotuner.py:674 - flashinfer.jit: [Autotuner]: Autotuning process ends
+(Worker_TP1 pid=278) 2026-07-08 00:11:00,189 - INFO - autotuner.py:674 - flashinfer.jit: [Autotuner]: Autotuning process ends
+(Worker_TP0 pid=277) 2026-07-08 00:11:00,190 - INFO - autotuner.py:674 - flashinfer.jit: [Autotuner]: Autotuning process ends
+(Worker_TP12 pid=289) (Worker_TP0 pid=277) (Worker_TP3 pid=280) (Worker_TP2 pid=279) (Worker_TP4 pid=281) (Worker_TP13 pid=290) INFO 07-08 00:11:01 [jit_monitor.py:60] Kernel JIT monitor activated — Triton JIT compilations during inference will be logged as warnings.
+(Worker_TP7 pid=284) INFO 07-08 00:11:01 [jit_monitor.py:60] Kernel JIT monitor activated — Triton JIT compilations during inference will be logged as warnings.
+(Worker_TP11 pid=288) INFO 07-08 00:11:01 [jit_monitor.py:60] Kernel JIT monitor activated — Triton JIT compilations during inference will be logged as warnings.
+INFO 07-08 00:11:01 [jit_monitor.py:60] Kernel JIT monitor activated — Triton JIT compilations during inference will be logged as warnings.
+INFO 07-08 00:11:01 [jit_monitor.py:60] Kernel JIT monitor activated — Triton JIT compilations during inference will be logged as warnings.
+INFO 07-08 00:11:01 [jit_monitor.py:60] Kernel JIT monitor activated — Triton JIT compilations during inference will be logged as warnings.
+INFO 07-08 00:11:01 [jit_monitor.py:60] Kernel JIT monitor activated — Triton JIT compilations during inference will be logged as warnings.
+INFO 07-08 00:11:01 [jit_monitor.py:60] Kernel JIT monitor activated — Triton JIT compilations during inference will be logged as warnings.
+(Worker_TP6 pid=283) (Worker_TP8 pid=285) (Worker_TP1 pid=278) (Worker_TP5 pid=282) (Worker_TP10 pid=287) (Worker_TP14 pid=291) INFO 07-08 00:11:01 [jit_monitor.py:60] Kernel JIT monitor activated — Triton JIT compilations during inference will be logged as warnings.
+INFO 07-08 00:11:01 [jit_monitor.py:60] Kernel JIT monitor activated — Triton JIT compilations during inference will be logged as warnings.
+INFO 07-08 00:11:01 [jit_monitor.py:60] Kernel JIT monitor activated — Triton JIT compilations during inference will be logged as warnings.
+(Worker_TP9 pid=286) INFO 07-08 00:11:01 [jit_monitor.py:60] Kernel JIT monitor activated — Triton JIT compilations during inference will be logged as warnings.
+INFO 07-08 00:11:01 [jit_monitor.py:60] Kernel JIT monitor activated — Triton JIT compilations during inference will be logged as warnings.
+INFO 07-08 00:11:01 [jit_monitor.py:60] Kernel JIT monitor activated — Triton JIT compilations during inference will be logged as warnings.
+INFO 07-08 00:11:01 [jit_monitor.py:60] Kernel JIT monitor activated — Triton JIT compilations during inference will be logged as warnings.
+(Worker_TP15 pid=292) INFO 07-08 00:11:01 [jit_monitor.py:60] Kernel JIT monitor activated — Triton JIT compilations during inference will be logged as warnings.
+(EngineCore pid=235) INFO 07-08 00:11:01 [core.py:344] init engine (profile, create kv cache, warmup model) took 7.79 s
+(EngineCore pid=235) INFO 07-08 00:11:03 [vllm.py:1057] Asynchronous scheduling is enabled.
+(EngineCore pid=235) WARNING 07-08 00:11:03 [vllm.py:1113] Enforce eager set, disabling torch.compile and CUDAGraphs. This is equivalent to setting -cc.mode=none -cc.cudagraph_mode=none
+(EngineCore pid=235) WARNING 07-08 00:11:03 [vllm.py:1142] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored.
+(EngineCore pid=235) INFO 07-08 00:11:03 [kernel.py:280] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native'])
+(EngineCore pid=235) INFO 07-08 00:11:03 [vllm.py:1320] Cudagraph is disabled under eager mode
+(EngineCore pid=235) INFO 07-08 00:11:03 [compilation.py:310] Enabled custom fusions: norm_quant, act_quant, allreduce_rms
+
+
+wrote_prefill_ref {"dtype": "torch.float32", "path": "/root/kld/glm52_current_bf16_returnlogits_ref_20260708T000424Z/ref/logits_0.safetensors", "shape": [2047, 154880], "window_idx": 0}
+collect_prefill_ref_done {"elapsed_sec": 5.831637144088745, "output_dir": "/root/kld/glm52_current_bf16_returnlogits_ref_20260708T000424Z/ref", "windows": 1}
+(EngineCore pid=235) INFO 07-08 00:11:08 [core.py:1212] [shutdown] EngineCore: trigger received signal=SIGTERM
+(EngineCore pid=235) INFO 07-08 00:11:08 [core.py:1331] [shutdown] EngineCore: start mode=abort timeout=0s
+(EngineCore pid=235) INFO 07-08 00:11:08 [core.py:1362] [shutdown] EngineCore: request processing complete; starting resource teardown
+(EngineCore pid=235) INFO 07-08 00:11:08 [core.py:1225] [shutdown] EngineCore: exiting busy loop
+(Worker_TP0 pid=277) INFO 07-08 00:11:08 [multiproc_executor.py:790] Parent process exited, terminating worker queues
+(EngineCore pid=235) INFO 07-08 00:11:08 [multiproc_executor.py:426] [shutdown] Executor: waiting for worker exit count=16
+WARNING 07-08 00:11:13 [utils.py:620] [shutdown] Process manager: force killing remaining processes count=1
+(EngineCore pid=235) WARNING 07-08 00:11:13 [multiproc_executor.py:438] [shutdown] Executor: workers still running after grace period; sending SIGTERM count=16
+/usr/lib/python3.12/multiprocessing/resource_tracker.py:254: UserWarning: resource_tracker: There appear to be 1 leaked shared_memory objects to clean up at shutdown
+  warnings.warn('resource_tracker: There appear to be %d '

reference-logits/logits_0.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:87f992a689c054a0548a4b3863da6c809f9239beacd5786d0401e45904fec063
+size 1268157528

reference-logits/manifest.json

@@ -0,0 +1,36 @@
+{
+  "context_length": 2048,
+  "elapsed_sec": 5.831637144088745,
+  "max_windows": 1,
+  "model": "/root/.cache/huggingface/hub/models--zai-org--GLM-5.2/snapshots/4d67f66cc64d3219133b767c253b2ad1425c6c88",
+  "stride": 512,
+  "token_first16": [
+    284,
+    8396,
+    425,
+    10960,
+    465,
+    284,
+    14721,
+    8396,
+    425,
+    10960,
+    465,
+    374,
+    458,
+    6364,
+    4531,
+    1154
+  ],
+  "windows": [
+    {
+      "dtype": "torch.float32",
+      "path": "/root/kld/glm52_current_bf16_returnlogits_ref_20260708T000424Z/ref/logits_0.safetensors",
+      "shape": [
+        2047,
+        154880
+      ],
+      "window_idx": 0
+    }
+  ]
+}