first commit

dev/configs/w4a8_int/mistral-3-large.yml

@@ -0,0 +1,66 @@
+model:
+  name: Mistral-Large-3-675B-Instruct-2512
+  model_type: mistral
+
+paths:
+  checkpoint_dir: /home/mashun/vllm_project/checkpoints/Mistral-Large-3-675B-Instruct-2512
+  output_dir: /home/mashun/vllm_project/checkpoints/Mistral-Large-3-675B-Instruct-2512-BF16
+
+runtime:
+  num_workers: 16
+  mode: dequant # dequant | dequant+quant | quant (only support bf16)
+
+retry_files: ~
+
+quant:
+  version: 2
+  format: 
+    TN:
+      - "re:layers.\\d+.experts.*"
+      - "re:layers.\\d+.shared_experts.*"
+    NN: ~
+  batch_size: 64
+  block_size: 128
+  group_size: -1
+  symmetric: True
+  class:
+    int4:
+      - "re:layers.\\d+.experts.*"
+    int8: 
+      - "re:layers.\\d+.shared_experts.*"
+  ignore: 
+    - "re:^vision_language_adapter.*"
+    - "re:tok_embeddings.*"
+    - "re:pre_mm_projector_norm.*"
+    - "re:patch_merger.merging_layer.*"
+    - "output.weight"
+    - "norm.weight"
+    - "re:layers.\\d+.attention.kv_a_norm.*"
+    - "re:layers.\\d+.attention.q_a_norm.*"
+    - "re:layers.\\d+.attention.wkv_a_with_mqa.*"
+    - "re:layers.\\d+.attention.wq_a.*"
+    - "re:layers.\\d+.attention.wkv_b.*"
+    - "re:layers.\\d+.attention.wo.*"
+    - "re:layers.\\d+.attention.wq_b.*"
+    - "re:layers.\\d+.attention_norm.*"
+    - "re:layers.\\d+.ffn_norm.*"
+    - "re:layers.\\d+.gate.*"
+    - "re:layers.\\d+.feed_forward.*"
+    - "re:vision_encoder.*"
+
+dequant:
+  ignore:
+    - "re:vision_language_adapter.*"
+    - "re:tok_embeddings.*"
+    - "re:pre_mm_projector_norm.*"
+    - "re:patch_merger.merging_layer.*"
+    - "output.weight"
+    - "norm.weight"
+    - "re:layers.\\d+.attention.kv_a_norm.*"
+    - "re:layers.\\d+.attention.q_a_norm.*"
+    - "re:layers.\\d+.attention.wkv_a_with_mqa.*"
+    - "re:layers.\\d+.attention.wq_a.*"
+    - "re:layers.\\d+.attention_norm.*"
+    - "re:layers.\\d+.ffn_norm.*"
+    - "re:layers.\\d+.gate.*"
+    - "re:vision_encoder.*"