Add qwen3.5-moe-0.87B-d0.8B.Q2_K full-v1 for Flexan/kshitijthakkar-qwen3.5-moe-0.87B-d0.8B-GGUF@a9b8adbec2cc

data/Flexan/kshitijthakkar-qwen3.5-moe-0.87B-d0.8B-GGUF/a9b8adbec2cc87479c772dac1944f313b4036c26/gguf/qwen3.5-moe-0.87B-d0.8B.Q2_K/full-v1/metadata.json

@@ -0,0 +1,28 @@
+{
+  "all_files": [
+    "qwen3.5-moe-0.87B-d0.8B.Q2_K.gguf"
+  ],
+  "all_layers": true,
+  "analysis_tool": "llama-moe-analyze",
+  "analyzer_id": "full-v1",
+  "command": {
+    "analyzer_id": "full-v1",
+    "context_size": 4096,
+    "token_count": 32
+  },
+  "created_at": "2026-04-11T04:07:32+00:00",
+  "distribution_id": "qwen3.5-moe-0.87B-d0.8B.Q2_K",
+  "file_hashes": {
+    "qwen3.5-moe-0.87B-d0.8B.Q2_K.gguf": "sha256:e8a84df1a50ce65cf80c2b55bba8c6e80f913679fdf9e9439f2c3b52ef3145d5"
+  },
+  "format": "gguf",
+  "primary_file": "qwen3.5-moe-0.87B-d0.8B.Q2_K.gguf",
+  "prompt_count": null,
+  "prompt_set": null,
+  "ranking_path": "ranking.csv",
+  "schema_version": 1,
+  "source_repo": "Flexan/kshitijthakkar-qwen3.5-moe-0.87B-d0.8B-GGUF",
+  "source_revision": "a9b8adbec2cc87479c772dac1944f313b4036c26",
+  "status": "complete",
+  "token_count": 32
+}

data/Flexan/kshitijthakkar-qwen3.5-moe-0.87B-d0.8B-GGUF/a9b8adbec2cc87479c772dac1944f313b4036c26/gguf/qwen3.5-moe-0.87B-d0.8B.Q2_K/full-v1/ranking.csv

@@ -0,0 +1,17 @@
+# MoE expert ranking by gate mass
+# Model: /Users/jdumay/.cache/huggingface/hub/models--Flexan--kshitijthakkar-qwen3.5-moe-0.87B-d0.8B-GGUF/snapshots/a9b8adbec2cc87479c772dac1944f313b4036c26/qwen3.5-moe-0.87B-d0.8B.Q2_K.gguf
+# Experts: 8 (top-2)
+# Prompts: 10 x 32 tokens
+# Layers logged: all
+# Total token-layer observations: 79905
+#
+# Format: expert_id,gate_mass,mass_pct,selection_count
+# Sorted by gate_mass descending (hottest first)
+0,9850.29,40.2628,66592
+5,2272.88,9.29034,4749
+2,2184.38,8.92858,4889
+7,2110.32,8.62588,4148
+3,2071.32,8.46647,3942
+6,2061.91,8.42799,4465
+1,1994.04,8.1506,4128
+4,1919.85,7.84735,3537

data/Flexan/kshitijthakkar-qwen3.5-moe-0.87B-d0.8B-GGUF/a9b8adbec2cc87479c772dac1944f313b4036c26/gguf/qwen3.5-moe-0.87B-d0.8B.Q2_K/full-v1/run.log

@@ -0,0 +1,302 @@
+$ /Users/jdumay/.codex/worktrees/4dc4/mesh-llm/llama.cpp/build/bin/llama-moe-analyze -m /Users/jdumay/.cache/huggingface/hub/models--Flexan--kshitijthakkar-qwen3.5-moe-0.87B-d0.8B-GGUF/snapshots/a9b8adbec2cc87479c772dac1944f313b4036c26/qwen3.5-moe-0.87B-d0.8B.Q2_K.gguf --all-layers --export-ranking /Users/jdumay/.cache/mesh-llm/moe-rankings/hf-Flexan--kshitijthakkar-qwen3.5-moe-0.87B-d0.8B-GGUF-a9b8adbec2cc87479c772dac1944f313b4036c26-qwen3.5-moe-0.87B-d0.8B.Q2_K.gguf.csv -n 32 -c 4096 -ngl 0
+
+[stdout]
+
+[stderr]
+ggml_metal_device_init: tensor API disabled for pre-M5 and pre-A19 devices
+ggml_metal_library_init: using embedded metal library
+ggml_metal_library_init: loaded in 0.031 sec
+ggml_metal_rsets_init: creating a residency set collection (keep_alive = 180 s)
+ggml_metal_device_init: GPU name:   MTL0
+ggml_metal_device_init: GPU family: MTLGPUFamilyApple8  (1008)
+ggml_metal_device_init: GPU family: MTLGPUFamilyCommon3 (3003)
+ggml_metal_device_init: GPU family: MTLGPUFamilyMetal4  (5002)
+ggml_metal_device_init: simdgroup reduction   = true
+ggml_metal_device_init: simdgroup matrix mul. = true
+ggml_metal_device_init: has unified memory    = true
+ggml_metal_device_init: has bfloat            = true
+ggml_metal_device_init: has tensor            = false
+ggml_metal_device_init: use residency sets    = true
+ggml_metal_device_init: use shared buffers    = true
+ggml_metal_device_init: recommendedMaxWorkingSetSize  = 19069.67 MB
+common_init_result: fitting params to device memory, for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on
+llama_params_fit_impl: projected to use 30 MiB of device memory vs. 18185 MiB of free device memory
+llama_params_fit_impl: will leave 18155 >= 1024 MiB of free device memory, no changes needed
+llama_params_fit: successfully fit params to free device memory
+llama_params_fit: fitting params to free memory took 0.59 seconds
+llama_model_load_from_file_impl: using device MTL0 (Apple M2) (unknown id) - 18185 MiB free
+llama_model_loader: loaded meta data with 40 key-value pairs and 441 tensors from /Users/jdumay/.cache/huggingface/hub/models--Flexan--kshitijthakkar-qwen3.5-moe-0.87B-d0.8B-GGUF/snapshots/a9b8adbec2cc87479c772dac1944f313b4036c26/qwen3.5-moe-0.87B-d0.8B.Q2_K.gguf (version GGUF V3 (latest))
+llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
+llama_model_loader: - kv   0:                       general.architecture str              = qwen35moe
+llama_model_loader: - kv   1:                               general.type str              = model
+llama_model_loader: - kv   2:                               general.name str              = Kshitijthakkar Qwen3.5 Moe 0.87B d0.8B
+llama_model_loader: - kv   3:                           general.finetune str              = d0.8B
+llama_model_loader: - kv   4:                           general.basename str              = kshitijthakkar-qwen3.5-moe
+llama_model_loader: - kv   5:                         general.size_label str              = 0.87B
+llama_model_loader: - kv   6:                            general.license str              = apache-2.0
+llama_model_loader: - kv   7:                               general.tags arr[str,6]       = ["qwen3.5", "moe", "weight-transfer",...
+llama_model_loader: - kv   8:                      qwen35moe.block_count u32              = 24
+llama_model_loader: - kv   9:                   qwen35moe.context_length u32              = 262144
+llama_model_loader: - kv  10:                 qwen35moe.embedding_length u32              = 1024
+llama_model_loader: - kv  11:             qwen35moe.attention.head_count u32              = 8
+llama_model_loader: - kv  12:          qwen35moe.attention.head_count_kv u32              = 2
+llama_model_loader: - kv  13:          qwen35moe.rope.dimension_sections arr[i32,4]       = [86, 85, 85, 0]
+llama_model_loader: - kv  14:                   qwen35moe.rope.freq_base f32              = 10000000.000000
+llama_model_loader: - kv  15: qwen35moe.attention.layer_norm_rms_epsilon f32              = 0.000001
+llama_model_loader: - kv  16:                     qwen35moe.expert_count u32              = 8
+llama_model_loader: - kv  17:                qwen35moe.expert_used_count u32              = 2
+llama_model_loader: - kv  18:             qwen35moe.attention.key_length u32              = 256
+llama_model_loader: - kv  19:           qwen35moe.attention.value_length u32              = 256
+llama_model_loader: - kv  20:       qwen35moe.expert_feed_forward_length u32              = 400
+llama_model_loader: - kv  21: qwen35moe.expert_shared_feed_forward_length u32              = 400
+llama_model_loader: - kv  22:                  qwen35moe.ssm.conv_kernel u32              = 4
+llama_model_loader: - kv  23:                   qwen35moe.ssm.state_size u32              = 128
+llama_model_loader: - kv  24:                  qwen35moe.ssm.group_count u32              = 16
+llama_model_loader: - kv  25:               qwen35moe.ssm.time_step_rank u32              = 16
+llama_model_loader: - kv  26:                   qwen35moe.ssm.inner_size u32              = 2048
+llama_model_loader: - kv  27:          qwen35moe.full_attention_interval u32              = 4
+llama_model_loader: - kv  28:             qwen35moe.rope.dimension_count u32              = 64
+llama_model_loader: - kv  29:                       tokenizer.ggml.model str              = gpt2
+llama_model_loader: - kv  30:                         tokenizer.ggml.pre str              = qwen35
+llama_model_loader: - kv  31:                      tokenizer.ggml.tokens arr[str,248320]  = ["!", "\"", "#", "$", "%", "&", "'", ...
+llama_model_loader: - kv  32:                  tokenizer.ggml.token_type arr[i32,248320]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
+llama_model_loader: - kv  33:                      tokenizer.ggml.merges arr[str,247587]  = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
+llama_model_loader: - kv  34:                tokenizer.ggml.eos_token_id u32              = 248046
+llama_model_loader: - kv  35:            tokenizer.ggml.padding_token_id u32              = 248044
+llama_model_loader: - kv  36:               tokenizer.ggml.add_bos_token bool             = false
+llama_model_loader: - kv  37:                    tokenizer.chat_template str              = {%- set image_count = namespace(value...
+llama_model_loader: - kv  38:               general.quantization_version u32              = 2
+llama_model_loader: - kv  39:                          general.file_type u32              = 10
+llama_model_loader: - type  f32:  181 tensors
+llama_model_loader: - type  f16:   48 tensors
+llama_model_loader: - type q8_0:   12 tensors
+llama_model_loader: - type q2_K:  193 tensors
+llama_model_loader: - type q5_K:    6 tensors
+llama_model_loader: - type q6_K:    1 tensors
+print_info: file format = GGUF V3 (latest)
+print_info: file type   = Q2_K - Medium
+print_info: file size   = 587.11 MiB (4.89 BPW) 
+load: 0 unused tokens
+load: printing all EOG tokens:
+load:   - 248044 ('<|endoftext|>')
+load:   - 248046 ('<|im_end|>')
+load:   - 248063 ('<|fim_pad|>')
+load:   - 248064 ('<|repo_name|>')
+load:   - 248065 ('<|file_sep|>')
+load: special tokens cache size = 33
+load: token to piece cache size = 1.7581 MB
+print_info: arch                  = qwen35moe
+print_info: vocab_only            = 0
+print_info: no_alloc              = 0
+print_info: n_ctx_train           = 262144
+print_info: n_embd                = 1024
+print_info: n_embd_inp            = 1024
+print_info: n_layer               = 24
+print_info: n_head                = 8
+print_info: n_head_kv             = 2
+print_info: n_rot                 = 64
+print_info: n_swa                 = 0
+print_info: is_swa_any            = 0
+print_info: n_embd_head_k         = 256
+print_info: n_embd_head_v         = 256
+print_info: n_gqa                 = 4
+print_info: n_embd_k_gqa          = 512
+print_info: n_embd_v_gqa          = 512
+print_info: f_norm_eps            = 0.0e+00
+print_info: f_norm_rms_eps        = 1.0e-06
+print_info: f_clamp_kqv           = 0.0e+00
+print_info: f_max_alibi_bias      = 0.0e+00
+print_info: f_logit_scale         = 0.0e+00
+print_info: f_attn_scale          = 0.0e+00
+print_info: n_ff                  = 0
+print_info: n_expert              = 8
+print_info: n_expert_used         = 2
+print_info: n_expert_groups       = 0
+print_info: n_group_used          = 0
+print_info: causal attn           = 1
+print_info: pooling type          = -1
+print_info: rope type             = 40
+print_info: rope scaling          = linear
+print_info: freq_base_train       = 10000000.0
+print_info: freq_scale_train      = 1
+print_info: n_ctx_orig_yarn       = 262144
+print_info: rope_yarn_log_mul     = 0.0000
+print_info: rope_finetuned        = unknown
+print_info: mrope sections        = [86, 85, 85, 0]
+print_info: ssm_d_conv            = 4
+print_info: ssm_d_inner           = 2048
+print_info: ssm_d_state           = 128
+print_info: ssm_dt_rank           = 16
+print_info: ssm_n_group           = 16
+print_info: ssm_dt_b_c_rms        = 0
+print_info: model type            = ?B
+print_info: model params          = 1.01 B
+print_info: general.name          = Kshitijthakkar Qwen3.5 Moe 0.87B d0.8B
+print_info: vocab type            = BPE
+print_info: n_vocab               = 248320
+print_info: n_merges              = 247587
+print_info: BOS token             = 11 ','
+print_info: EOS token             = 248046 '<|im_end|>'
+print_info: EOT token             = 248046 '<|im_end|>'
+print_info: PAD token             = 248044 '<|endoftext|>'
+print_info: LF token              = 198 'Ċ'
+print_info: FIM PRE token         = 248060 '<|fim_prefix|>'
+print_info: FIM SUF token         = 248062 '<|fim_suffix|>'
+print_info: FIM MID token         = 248061 '<|fim_middle|>'
+print_info: FIM PAD token         = 248063 '<|fim_pad|>'
+print_info: FIM REP token         = 248064 '<|repo_name|>'
+print_info: FIM SEP token         = 248065 '<|file_sep|>'
+print_info: EOG token             = 248044 '<|endoftext|>'
+print_info: EOG token             = 248046 '<|im_end|>'
+print_info: EOG token             = 248063 '<|fim_pad|>'
+print_info: EOG token             = 248064 '<|repo_name|>'
+print_info: EOG token             = 248065 '<|file_sep|>'
+print_info: max token length      = 256
+load_tensors: loading model tensors, this can take a while... (mmap = true, direct_io = false)
+load_tensors: offloading 0 repeating layers to GPU
+load_tensors: offloaded 0/25 layers to GPU
+load_tensors:   CPU_Mapped model buffer size =   388.18 MiB
+load_tensors:   CPU_REPACK model buffer size =   213.55 MiB
+......................................................
+common_init_result: added <|endoftext|> logit bias = -inf
+common_init_result: added <|im_end|> logit bias = -inf
+common_init_result: added <|fim_pad|> logit bias = -inf
+common_init_result: added <|repo_name|> logit bias = -inf
+common_init_result: added <|file_sep|> logit bias = -inf
+llama_context: constructing llama_context
+llama_context: n_seq_max     = 1
+llama_context: n_ctx         = 4096
+llama_context: n_ctx_seq     = 4096
+llama_context: n_batch       = 2048
+llama_context: n_ubatch      = 512
+llama_context: causal_attn   = 1
+llama_context: flash_attn    = auto
+llama_context: kv_unified    = false
+llama_context: freq_base     = 10000000.0
+llama_context: freq_scale    = 1
+llama_context: n_ctx_seq (4096) < n_ctx_train (262144) -- the full capacity of the model will not be utilized
+ggml_metal_init: allocating
+ggml_metal_init: found device: Apple M2
+ggml_metal_init: picking default device: Apple M2
+ggml_metal_init: use fusion         = true
+ggml_metal_init: use concurrency    = true
+ggml_metal_init: use graph optimize = true
+llama_context:        CPU  output buffer size =     0.95 MiB
+llama_kv_cache:        CPU KV buffer size =    48.00 MiB
+llama_kv_cache: size =   48.00 MiB (  4096 cells,   6 layers,  1/1 seqs), K (f16):   24.00 MiB, V (f16):   24.00 MiB
+llama_kv_cache: attn_rot_k = 0
+llama_kv_cache: attn_rot_v = 0
+llama_memory_recurrent:        CPU RS buffer size =    19.27 MiB
+llama_memory_recurrent: size =   19.27 MiB (     1 cells,  24 layers,  1 seqs), R (f32):    1.27 MiB, S (f32):   18.00 MiB
+sched_reserve: reserving ...
+sched_reserve: Flash Attention was auto, set to enabled
+sched_reserve: resolving fused Gated Delta Net support:
+sched_reserve: fused Gated Delta Net (autoregressive) enabled
+sched_reserve: fused Gated Delta Net (chunked) enabled
+sched_reserve:       MTL0 compute buffer size =    30.06 MiB
+sched_reserve:        CPU compute buffer size =   489.00 MiB
+sched_reserve: graph nodes  = 1953
+sched_reserve: graph splits = 427 (with bs=512), 1 (with bs=1)
+sched_reserve: reserve took 12.60 ms, sched copies = 1
+
+=== MoE Expert Routing Analysis ===
+Model experts: 8, used per token: 2
+Logging ALL MoE layers (--all-layers)
+Will export expert ranking to: /Users/jdumay/.cache/mesh-llm/moe-rankings/hf-Flexan--kshitijthakkar-qwen3.5-moe-0.87B-d0.8B-GGUF-a9b8adbec2cc87479c772dac1944f313b4036c26-qwen3.5-moe-0.87B-d0.8B.Q2_K.gguf.csv
+Running 10 prompts, generating 32 tokens each
+Logging first 9999 MoE layers per eval
+
+Prompt 1/10: <|im_start|>user
+Write a Python function to find the nth Fib...
+  collected 1584 layer snapshots (total: 1584)
+Prompt 2/10: <|im_start|>user
+Write a Rust function that reads a CSV file...
+  collected 1584 layer snapshots (total: 3168)
+Prompt 3/10: <|im_start|>user
+Explain how a B-tree index works in a datab...
+  collected 1584 layer snapshots (total: 4752)
+Prompt 4/10: <|im_start|>user
+If all roses are flowers and some flowers f...
+  collected 1584 layer snapshots (total: 6336)
+Prompt 5/10: <|im_start|>user
+A train travels 120km in 2 hours. It then s...
+  collected 1584 layer snapshots (total: 7920)
+Prompt 6/10: <|im_start|>user
+Hello! What's the best way to learn a new l...
+  collected 1584 layer snapshots (total: 9504)
+Prompt 7/10: <|im_start|>user
+Tell me a joke about programmers.<|im_end|>...
+  collected 1584 layer snapshots (total: 11088)
+Prompt 8/10: <|im_start|>user
+Summarize the key differences between TCP a...
+  collected 1584 layer snapshots (total: 12672)
+Prompt 9/10: <|im_start|>user
+Translate 'The weather is beautiful today' ...
+  collected 1584 layer snapshots (total: 14256)
+Prompt 10/10: <|im_start|>user
+List 5 healthy breakfast options with brief...
+  collected 1584 layer snapshots (total: 15840)
+
+=== Expert Popularity (gate mass, summed across all tokens & logged layers) ===
+Total tokens × layers: 79905
+
+Top 20 experts by gate mass:
+  Expert    Mass          Mass%     Selected  
+  0         9850.2938     40.26     66592     
+  5         2272.8815     9.29      4749      
+  2         2184.3772     8.93      4889      
+  7         2110.3206     8.63      4148      
+  3         2071.3219     8.47      3942      
+  6         2061.9076     8.43      4465      
+  1         1994.0433     8.15      4128      
+  4         1919.8542     7.85      3537      
+
+Concentration:
+  Top   4 experts: 67.1% of total gate mass
+  Top   8 experts: 100.0% of total gate mass
+
+Exported expert ranking to: /Users/jdumay/.cache/mesh-llm/moe-rankings/hf-Flexan--kshitijthakkar-qwen3.5-moe-0.87B-d0.8B-GGUF-a9b8adbec2cc87479c772dac1944f313b4036c26-qwen3.5-moe-0.87B-d0.8B.Q2_K.gguf.csv
+Use with moe-split: --group-map <file generated from this ranking>
+
+
+=== Group Masking Analysis (best-group capture ratio) ===
+For each group count, what fraction of the unrestricted top-2 mass
+is captured by the best single group?
+
+  Groups    Replicas  Exp/Grp     Mean        P25         P50         P5        
+  2         0         4           0.990       1.000       1.000       0.916     
+  2         1         4           0.993       1.000       1.000       0.942     
+  2         2         4           0.995       1.000       1.000       0.967     
+  2         4         4           0.998       1.000       1.000       1.000     
+
+  4         0         2           0.977       1.000       1.000       0.825     
+  4         1         2           0.986       1.000       1.000       0.876     
+  4         2         2           0.992       1.000       1.000       0.926     
+  4         4         2           0.997       1.000       1.000       1.000     
+
+  8         0         1           0.908       1.000       1.000       0.516     
+  8         1         1           0.974       1.000       1.000       0.792     
+  8         2         1           0.986       1.000       1.000       0.870     
+  8         4         1           0.997       1.000       1.000       1.000     
+
+=== Interpretation ===
+Mean close to 1.0 = masking barely hurts (best group captures most of top-k mass)
+Mean < 0.7 = significant quality risk from group restriction
+P5 close to 1.0 = even worst-case tokens are OK
+P5 < 0.5 = some tokens will be badly served by any single group
+
+=== Phase 1b: Masked Generation Quality (logprob comparison) ===
+Testing 4 groups (2 experts/group) vs baseline (all 8 experts)
+Using first 5 prompts, generating 32 tokens each
+
+  Group 0 (experts 0-1 + 2 hot replicas): avg logprob delta = -0.2655
+  Group 1 (experts 2-3 + 2 hot replicas): avg logprob delta = -0.0722
+  Group 2 (experts 4-5 + 2 hot replicas): avg logprob delta = -0.0818
+  Group 3 (experts 6-7 + 2 hot replicas): avg logprob delta = -0.0192
+
+=== Interpretation (logprob delta) ===
+Delta near 0.0 = masking barely affects generation quality
+Delta < -0.1 = noticeable quality loss
+Delta < -0.5 = significant degradation
+
+ggml_metal_free: deallocating