first commit

dev/configs/bf16/deepseek-r1.yml

@@ -0,0 +1,43 @@
+model:
+  name: DeepSeek-R1
+  model_type: deepseek_v3
+
+paths:
+  checkpoint_dir: /home/mashun/vllm_project/checkpoints/DeepSeek-R1-0528-FP8
+  output_dir: /home/mashun/vllm_project/checkpoints/DeepSeek-R1-0528
+
+runtime:
+  num_workers: 8
+  mode: dequant # dequant | dequant+quant | quant (only support bf16)
+
+retry_files: ~
+
+quant:
+  version: 2
+  format:
+    TN: ~
+    NN: ~
+  batch_size: 64
+  block_size: 128
+  group_size: -1
+  symmetric: True
+  class:
+    int4: ~ 
+    int8: ~
+  ignore: ~
+   
+dequant:
+  ignore:
+    - "re:.*embed_tokens.*"
+    - "re:.*input_layernorm.*"
+    - "re:.*post_attention_layernorm.*"
+    - "re:.*kv_a_layernorm.*"
+    - "re:.*q_a_layernorm.*"
+    - "re:.*eh_proj.*"
+    - "re:.*enorm.*"
+    - "re:.*hnorm.*"
+    - "re:model.layers.\\d+.mlp.gate\\..*"
+    - "re:model.norm.*"
+    - "re:model.layers.\\d+.shared_head.norm.*"
+    - "re:.*lm_head.*"
+    - "re:model.layers.\\d+.shared_head.head.*"

dev/configs/w4a8_int/deepseek-r1.yml

@@ -0,0 +1,70 @@
+model:
+  name: DeepSeek-R1
+  model_type: deepseek_v3
+
+paths:
+  checkpoint_dir: /home/mashun/vllm_project/checkpoints/DeepSeek-R1-0528-FP8
+  output_dir: /home/mashun/vllm_project/checkpoints/DeepSeek-R1-0528-w4a8
+
+runtime:
+  num_workers: 16
+  mode: dequant+quant # dequant | dequant+quant | quant (only support bf16)
+
+retry_files: ~
+
+quant:
+  version: 2
+  format:
+    TN:
+      - "re:model.layers.\\d+.mlp.experts.*"
+      - "re:model.layers.\\d+.self_attn.*"
+    NN: ~
+  batch_size: 64
+  block_size: 128
+  group_size: -1
+  symmetric: True
+  class:
+    int4:
+      - "re:model.layers.\\d+.mlp.experts.\\d+.down_proj.*"
+      - "re:model.layers.\\d+.mlp.experts.\\d+.gate_proj.*"
+      - "re:model.layers.\\d+.mlp.experts.\\d+.up_proj.*"
+    int8: 
+      - "re:.*shared_experts.*"
+      - "re:model.layers.\\d+.self_attn.q_a_proj.*"
+      - "re:model.layers.\\d+.self_attn.q_b_proj.*"
+      - "re:model.layers.\\d+.self_attn.kv_a_proj_with_mqa.*"
+      - "re:model.layers.\\d+.self_attn.kv_b_proj.*"
+      - "re:model.layers.\\d+.self_attn.o_proj.*"
+  ignore: 
+    - "re:.*embed_tokens.*"
+    - "re:.*input_layernorm.*"
+    - "re:.*post_attention_layernorm.*"
+    - "re:.*kv_a_layernorm.*"
+    - "re:.*q_a_layernorm.*"
+    - "re:.*eh_proj.*"
+    - "re:.*enorm.*"
+    - "re:.*hnorm.*"
+    - "re:model.layers.\\d+.mlp.gate_proj.*"
+    - "re:model.layers.\\d+.mlp.up_proj.*"
+    - "re:model.layers.\\d+.mlp.down_proj.*"
+    - "re:model.layers.\\d+.mlp.gate\\..*"
+    - "re:model.norm.*"
+    - "re:model.layers.\\d+.shared_head.norm.*"
+    - "re:.*lm_head.*"
+    - "re:model.layers.\\d+.shared_head.head.*"
+   
+dequant:
+  ignore:
+    - "re:.*embed_tokens.*"
+    - "re:.*input_layernorm.*"
+    - "re:.*post_attention_layernorm.*"
+    - "re:.*kv_a_layernorm.*"
+    - "re:.*q_a_layernorm.*"
+    - "re:.*eh_proj.*"
+    - "re:.*enorm.*"
+    - "re:.*hnorm.*"
+    - "re:model.layers.\\d+.mlp.gate\\..*"
+    - "re:model.norm.*"
+    - "re:model.layers.\\d+.shared_head.norm.*"
+    - "re:.*lm_head.*"
+    - "re:model.layers.\\d+.shared_head.head.*"

dev/configs/w4a8_int/deepseek-v2.yml

@@ -0,0 +1,43 @@
+model:
+  name: DeepSeek-V2-Lite-Chat
+  model_type: deepseek_v2
+
+paths:
+  checkpoint_dir: /home/mashun/vllm_project/checkpoints/DeepSeek-V2-Lite-Chat
+  output_dir: /home/mashun/vllm_project/checkpoints/DeepSeek-V2-Lite-Chat-w4a8
+
+runtime:
+  num_workers: 4
+  mode: quant # dequant | dequant+quant | quant (only support bf16)
+
+retry_files: ~
+
+quant:
+  version: 2
+  format:
+    TN:
+      - "re:model.layers.\\d+.mlp.experts.*"
+      - "re:model.layers.\\d+.self_attn.*"
+    NN: ~
+  batch_size: 64
+  block_size: 128
+  group_size: -1
+  symmetric: True
+  class:
+    int4:
+      - "re:model.layers.\\d+.mlp.experts.*"
+    int8: 
+      - "re:model.layers.\\d+.self_attn.*"
+  ignore: 
+    - "re:model.embed_tokens.*"
+    - "re:model.norm.*"
+    - "re:lm_head.*"
+    - "re:.*layernorm.*"
+    - "re:model.layers.\\d+.mlp.gate\\..*"
+    - "re:model.layers.\\d+.mlp.shared_experts.*"
+    - "re:model.layers.\\d+.mlp.gate_proj.*"
+    - "re:model.layers.\\d+.mlp.up_proj.*"
+    - "re:model.layers.\\d+.mlp.down_proj.*"
+   
+dequant:
+  ignore: ~

dev/configs/w4a8_int/glm4_7.yml

@@ -0,0 +1,55 @@
+model:
+  name: GLM-4.7
+  model_type: glm4_moe
+
+paths:
+  checkpoint_dir: /home/mashun/vllm_project/checkpoints/GLM-4.7-FP8
+  output_dir: /home/mashun/vllm_project/checkpoints/GLM-4.7-w4a8
+
+runtime:
+  num_workers: 16
+  mode: dequant+quant # dequant | dequant+quant | quant (only support bf16)
+
+retry_files: ~
+
+quant:
+  version: 2
+  format: 
+    TN:
+      - "re:model.layers.\\d+.mlp.experts.\\d+.down_proj.*"
+      - "re:model.layers.\\d+.mlp.experts.\\d+.up_proj.*"
+      - "re:model.layers.\\d+.mlp.experts.\\d+.gate_proj.*"
+    NN: ~
+  batch_size: 64
+  block_size: 128
+  group_size: -1
+  symmetric: True
+  class:
+    int4:
+      - "re:model.layers.\\d+.mlp.experts.\\d+.down_proj.*"
+      - "re:model.layers.\\d+.mlp.experts.\\d+.up_proj.*"
+      - "re:model.layers.\\d+.mlp.experts.\\d+.gate_proj.*"
+    int8: ~
+  ignore: 
+    - "re:.*self_attn.*"
+    - "re:.*bias.*"
+    - "re:.*mlp.gate.*"
+    - "re:.*norm.*"
+    - "re:.*lm_head.*"
+    - "re:.*eh_proj.*"
+    - "re:.*embed_tokens.*"
+    - "re:.*shared_head.*"
+    - "re:model.layers.\\d+.mlp.shared_experts.*"
+    - "re:model.layers.\\d+.mlp.down_proj.*"
+    - "re:model.layers.\\d+.mlp.up_proj.*"
+    - "re:model.layers.\\d+.mlp.proj_proj.*"
+   
+dequant:
+  ignore:
+    - "re:.*bias"
+    - "re:.*mlp.gate\\..*"
+    - "re:.*norm.*"
+    - "lm_head.weight"
+    - "re:.*eh_proj.*"
+    - "re:.*embed_tokens.weight"
+    - "re:.*shared_head.*"

dev/configs/w4a8_int/glm4_7_flash.yml

@@ -0,0 +1,55 @@
+model:
+  name: GLM-4.7-Flash
+  model_type: glm4_moe
+
+paths:
+  checkpoint_dir: /home/mashun/vllm_project/checkpoints/GLM-4.7-Flash
+  output_dir: /home/mashun/vllm_project/checkpoints/GLM-4.7-Flash-w4a8
+
+runtime:
+  num_workers: 16
+  mode: quant # dequant | dequant+quant | quant (only support bf16)
+
+retry_files: ~
+
+quant:
+  version: 2
+  format: 
+    TN:
+      - "re:model.layers.\\d+.mlp.experts.\\d+.down_proj.*"
+      - "re:model.layers.\\d+.mlp.experts.\\d+.up_proj.*"
+      - "re:model.layers.\\d+.mlp.experts.\\d+.gate_proj.*"
+    NN: ~
+  batch_size: 64
+  block_size: 128
+  group_size: -1
+  symmetric: True
+  class:
+    int4:
+      - "re:model.layers.\\d+.mlp.experts.\\d+.down_proj.*"
+      - "re:model.layers.\\d+.mlp.experts.\\d+.up_proj.*"
+      - "re:model.layers.\\d+.mlp.experts.\\d+.gate_proj.*"
+    int8: ~
+  ignore: 
+    - "re:.*self_attn.*"
+    - "re:.*bias.*"
+    - "re:.*mlp.gate.*"
+    - "re:.*norm.*"
+    - "re:.*lm_head.*"
+    - "re:.*eh_proj.*"
+    - "re:.*embed_tokens.*"
+    - "re:.*shared_head.*"
+    - "re:model.layers.\\d+.mlp.shared_experts.*"
+    - "re:model.layers.\\d+.mlp.down_proj.*"
+    - "re:model.layers.\\d+.mlp.up_proj.*"
+    - "re:model.layers.\\d+.mlp.proj_proj.*"
+   
+dequant:
+  ignore:
+    - "re:.*bias"
+    - "re:.*mlp.gate\\..*"
+    - "re:.*norm.*"
+    - "lm_head.weight"
+    - "re:.*eh_proj.*"
+    - "re:.*embed_tokens.weight"
+    - "re:.*shared_head.*"

dev/configs/w4a8_int/glm5.yml

@@ -0,0 +1,56 @@
+model:
+  name: GLM-5
+  model_type: glm_moe_dsa
+
+paths:
+  checkpoint_dir: /mnt/data/models/ZhipuAI/GLM-5-FP8
+  output_dir: /mnt/data/models/ZhipuAI/GLM-5-w4a8
+
+runtime:
+  num_workers: 16
+  mode: dequant+quant # dequant | dequant+quant | quant (only support bf16)
+
+retry_files: ~
+
+quant:
+  version: 2
+  format: 
+    TN:
+      - "re:model.layers.\\d+.mlp.experts.\\d+.down_proj.*"
+      - "re:model.layers.\\d+.mlp.experts.\\d+.up_proj.*"
+      - "re:model.layers.\\d+.mlp.experts.\\d+.gate_proj.*"
+    NN: ~
+  batch_size: 64
+  block_size: 128
+  group_size: -1
+  symmetric: True
+  class:
+    int4:
+      - "re:model.layers.\\d+.mlp.experts.\\d+.down_proj.*"
+      - "re:model.layers.\\d+.mlp.experts.\\d+.up_proj.*"
+      - "re:model.layers.\\d+.mlp.experts.\\d+.gate_proj.*"
+    int8: ~
+  ignore: 
+    - "re:.*self_attn.*"
+    - "re:.*bias.*"
+    - "re:.*mlp.gate.*"
+    - "re:.*norm.*"
+    - "re:.*lm_head.*"
+    - "re:.*eh_proj.*"
+    - "re:.*embed_tokens.*"
+    - "re:.*shared_head.*"
+    - "re:model.layers.\\d+.mlp.shared_experts.*"
+    - "re:model.layers.\\d+.mlp.down_proj.*"
+    - "re:model.layers.\\d+.mlp.up_proj.*"
+    - "re:model.layers.\\d+.mlp.proj_proj.*"
+   
+dequant:
+  ignore:
+    - "re:.*self_attn.indexer.weights_proj.*"
+    - "re:.*bias"
+    - "re:.*mlp.gate\\..*"
+    - "re:.*norm.*"
+    - "lm_head.weight"
+    - "re:.*eh_proj.*"
+    - "re:.*embed_tokens.weight"
+    - "re:.*shared_head.*"

dev/configs/w4a8_int/glm5_nn_g64.yml

@@ -0,0 +1,53 @@
+model:
+  name: GLM-5
+  model_type: glm_moe_dsa
+
+paths:
+  checkpoint_dir: /mnt/data/models/ZhipuAI/GLM-5-FP8
+  output_dir: /mnt/data/models/ZhipuAI/GLM-5-w4a8-nn-g64
+
+runtime:
+  num_workers: 16
+  mode: dequant+quant # dequant | dequant+quant | quant (only support bf16)
+
+retry_files: ~
+
+quant:
+  version: 2
+  format:
+    TN: ~
+    NN: ~
+  batch_size: 64
+  block_size: 128
+  group_size: 64
+  symmetric: True
+  class:
+    int4:
+      - "re:model.layers.\\d+.mlp.experts.\\d+.down_proj.*"
+      - "re:model.layers.\\d+.mlp.experts.\\d+.up_proj.*"
+      - "re:model.layers.\\d+.mlp.experts.\\d+.gate_proj.*"
+    int8: ~
+  ignore: 
+    - "re:.*self_attn.*"
+    - "re:.*bias.*"
+    - "re:.*mlp.gate.*"
+    - "re:.*norm.*"
+    - "re:.*lm_head.*"
+    - "re:.*eh_proj.*"
+    - "re:.*embed_tokens.*"
+    - "re:.*shared_head.*"
+    - "re:model.layers.\\d+.mlp.shared_experts.*"
+    - "re:model.layers.\\d+.mlp.down_proj.*"
+    - "re:model.layers.\\d+.mlp.up_proj.*"
+    - "re:model.layers.\\d+.mlp.proj_proj.*"
+   
+dequant:
+  ignore:
+    - "re:.*self_attn.indexer.weights_proj.*"
+    - "re:.*bias"
+    - "re:.*mlp.gate\\..*"
+    - "re:.*norm.*"
+    - "lm_head.weight"
+    - "re:.*eh_proj.*"
+    - "re:.*embed_tokens.weight"
+    - "re:.*shared_head.*"

dev/configs/w4a8_int/kimik2.yml

@@ -0,0 +1,76 @@
+model:
+  name: Kimi-K2-Thinking
+  model_type: kimik2
+
+paths:
+  checkpoint_dir: /home/models/kimi/Kimi-K2.5
+  output_dir: /home/models/kimi/Kimi-K2.5-w4a8
+
+runtime:
+  num_workers: 16
+  mode: dequant+quant # dequant | dequant+quant | quant (only support bf16)
+
+retry_files: ~
+
+quant:
+  version: 2
+  format: 
+    TN:
+      - "re:model.layers.\\d+.mlp.experts.*"
+      - "re:.*shared_experts.*"
+      - "re:model.layers.\\d+.self_attn.q_a_proj.*"
+      - "re:model.layers.\\d+.self_attn.q_b_proj.*"
+      - "re:model.layers.\\d+.self_attn.kv_a_proj_with_mqa.*"
+      - "re:model.layers.\\d+.self_attn.kv_b_proj.*"
+      - "re:model.layers.\\d+.self_attn.o_proj.*"
+      - "re:model.layers.\\d+.mlp.shared_experts.*"
+      - "re:model.layers.\\d+.mlp.gate_proj.*"
+      - "re:model.layers.\\d+.mlp.down_proj.*"
+      - "re:model.layers.\\d+.mlp.up_proj.*"
+    NN: ~
+  batch_size: 64
+  block_size: 128
+  group_size: -1
+  symmetric: True
+  class:
+    int4:
+      - "re:model.layers.\\d+.mlp.experts.*"
+    int8:
+      - "re:.*shared_experts.*"
+      - "re:model.layers.\\d+.self_attn.q_a_proj.*"
+      - "re:model.layers.\\d+.self_attn.q_b_proj.*"
+      - "re:model.layers.\\d+.self_attn.kv_a_proj_with_mqa.*"
+      - "re:model.layers.\\d+.self_attn.kv_b_proj.*"
+      - "re:model.layers.\\d+.self_attn.o_proj.*"
+      - "re:model.layers.\\d+.mlp.shared_experts.*"
+      - "re:model.layers.\\d+.mlp.gate_proj.*"
+      - "re:model.layers.\\d+.mlp.down_proj.*"
+      - "re:model.layers.\\d+.mlp.up_proj.*"
+  ignore: 
+    - "re:model.layers.\\d+.self_attn.q_a_layernorm.*"
+    - "re:model.layers.\\d+.self_attn.kv_a_layernorm.*"
+    - "re:model.layers.\\d+.input_layernorm.*"
+    - "re:model.layers.\\d+.post_attention_layernorm.*"
+    - "re:model.layers.\\d+.mlp.gate.weight"
+    - "re:model.layers.\\d+.mlp.gate.e_score_correction_bias"
+    - "re:model.layers.\\d+.self_attn.rotary_emb.inv_freq"
+    - "re:model.embed_tokens.*"
+    - "re:model.norm.*"
+    - "re:lm_head.*"
+    
+
+dequant:
+  ignore:
+    - "re:lm_head.*"
+    - "re:.*self_attn.*"
+    - "re:.*shared_experts.*"
+    - "re:.*mlp\\.(gate|up|gate_up|down)_proj.*"
+    - "re:model.layers.\\d+.mlp.gate.e_score_correction_bias"
+    - "re:model.layers.\\d+.input_layernorm.weight"
+    - "re:model.layers.\\d+.post_attention_layernorm.weight"
+    - "re:model.layers.\\d+.mlp.gate.weight"
+    - "re:model.layers.\\d+.self_attn.rotary_emb.inv_freq"
+    - "re:model.norm.*"
+    - "re:model.layers.\\d+.self_attn.q_a_layernorm.*"
+    - "re:model.layers.\\d+.self_attn.kv_a_layernorm.*"
+    - "re:model.embed_tokens.*"

dev/configs/w4a8_int/kimik25.yml

@@ -0,0 +1,77 @@
+model:
+  name: Kimi-K2.5
+  model_type: kimik2
+
+paths:
+  checkpoint_dir: /home/models/kimi/Kimi-K2.5
+  output_dir: /home/models/kimi/Kimi-K2.5-w4a8
+
+runtime:
+  num_workers: 4
+  mode: dequant+quant # dequant | dequant+quant | quant (only support bf16)
+
+retry_files: ~
+
+quant:
+  version: 2
+  format: 
+    TN:
+      - "re:.*model.layers.\\d+.mlp.experts.*"
+      - "re:.*model.layers.\\d+.self_attn.q_a_proj.*"
+      - "re:.*model.layers.\\d+.self_attn.q_b_proj.*"
+      - "re:.*model.layers.\\d+.self_attn.kv_a_proj_with_mqa.*"
+      - "re:.*model.layers.\\d+.self_attn.kv_b_proj.*"
+      - "re:.*model.layers.\\d+.self_attn.o_proj.*"
+      - "re:.*model.layers.\\d+.mlp.shared_experts.*"
+      - "re:.*model.layers.\\d+.mlp.gate_proj.*"
+      - "re:.*model.layers.\\d+.mlp.down_proj.*"
+      - "re:.*model.layers.\\d+.mlp.up_proj.*"
+      - "re:.*shared_experts.*"
+    NN: ~
+  batch_size: 64
+  block_size: 128
+  group_size: -1
+  symmetric: True
+  class:
+    int4:
+      - "re:.*model.layers.\\d+.mlp.experts.*"
+    int8:
+      - "re:.*model.layers.\\d+.self_attn.q_a_proj.*"
+      - "re:.*model.layers.\\d+.self_attn.q_b_proj.*"
+      - "re:.*model.layers.\\d+.self_attn.kv_a_proj_with_mqa.*"
+      - "re:.*model.layers.\\d+.self_attn.kv_b_proj.*"
+      - "re:.*model.layers.\\d+.self_attn.o_proj.*"
+      - "re:.*model.layers.\\d+.mlp.shared_experts.*"
+      - "re:.*model.layers.\\d+.mlp.gate_proj.*"
+      - "re:.*model.layers.\\d+.mlp.down_proj.*"
+      - "re:.*model.layers.\\d+.mlp.up_proj.*"
+      - "re:.*shared_experts.*"
+  ignore: 
+    - "re:language_model.model.layers.\\d+.self_attn.q_a_layernorm.*"
+    - "re:language_model.model.layers.\\d+.self_attn.kv_a_layernorm.*"
+    - "re:language_model.model.layers.\\d+.input_layernorm.*"
+    - "re:language_model.model.layers.\\d+.post_attention_layernorm.*"
+    - "re:language_model.model.layers.\\d+.mlp.gate.weight"
+    - "re:language_model.model.layers.\\d+.mlp.gate.e_score_correction_bias"
+    - "re:language_model.model.layers.\\d+.self_attn.rotary_emb.inv_freq"
+    - "re:language_model.model.embed_tokens.*"
+    - "re:language_model.model.norm.*"
+    - "re:.*lm_head.*"
+    - "re:mm_projector.*"
+    - "re:vision_tower.*"
+    - "re:.*bias.*"
+    - "re:.*norm.*"
+    
+
+dequant:
+  ignore:
+    - "re:.*lm_head.*"
+    - "re:.*self_attn.*"
+    - "re:.*shared_experts.*"
+    - "re:.*mlp\\.(gate|up|gate_up|down)_proj.*"
+    - "re:.*bias.*"
+    - "re:.*norm.*"
+    - "re:language_model.model.layers.\\d+.mlp.gate.weight"
+    - "re:mm_projector.*"
+    - "re:vision_tower.*"
+    - "re:.*embed_tokens.*"

dev/configs/w4a8_int/minimax_m2_1.yml

@@ -0,0 +1,50 @@
+model:
+  name: MiniMax-M2.1
+  model_type: minimax_m2
+
+paths:
+  checkpoint_dir: /home/mashun/vllm_project/checkpoints/MiniMax-M2.1-FP8
+  output_dir: /home/mashun/vllm_project/checkpoints/MiniMax-M2.1-w4a8
+
+runtime:
+  num_workers: 16
+  mode: dequant+quant # dequant | dequant+quant | quant (only support bf16)
+
+retry_files: ~
+
+quant:
+  version: 2
+  format: 
+    TN:
+      - "re:model.layers.\\d+.block_sparse_moe.experts.*"
+    NN: ~
+  batch_size: 64
+  block_size: 128
+  group_size: -1
+  symmetric: True
+  class:
+    int4:
+      - "re:model.layers.\\d+.block_sparse_moe.experts.*"
+    int8: ~
+  ignore: 
+    - "model.embed_tokens.weight"
+    - "re:model.layers.\\d+.input_layernorm.weight"
+    - "re:model.layers.\\d+.self_attn.*"
+    - "re:model.layers.\\d+.post_attention_layernorm.weight"
+    - "re:model.layers.\\d+.block_sparse_moe.gate.weight"
+    - "re:model.layers.\\d+.block_sparse_moe.e_score_correction_bias"
+    - "model.norm.weight"
+    - "lm_head.weight"
+
+dequant:
+  ignore:
+    - "re:lm_head.*"
+    - "re:model.embed_tokens.*"
+    - "re:model.layers.\\d+.input_layernorm.*"
+    - "re:model.layers.\\d+.mlp.gate.*"
+    - "re:model.layers.\\d+.post_attention_layernorm.*"
+    - "re:model.layers.\\d+.self_attn.k_norm.*"
+    - "re:model.layers.\\d+.self_attn.q_norm.*"
+    - "re:model.norm.*"
+    - "re:model.layers.\\d+.block_sparse_moe.e_score_correction_bias"
+    - "re:model.layers.\\d+.block_sparse_moe.gate.weight"

dev/configs/w4a8_int/minimax_m2_5.yml

@@ -0,0 +1,50 @@
+model:
+  name: MiniMax-M2.5
+  model_type: minimax_m2
+
+paths:
+  checkpoint_dir: /home/mashun/vllm_project/checkpoints/MiniMax-M2.5
+  output_dir: /home/mashun/vllm_project/checkpoints/MiniMax-M2.5-w4a8
+
+runtime:
+  num_workers: 16
+  mode: dequant+quant # dequant | dequant+quant | quant (only support bf16)
+
+retry_files: ~
+
+quant:
+  version: 2
+  format: 
+    TN:
+      - "re:model.layers.\\d+.block_sparse_moe.experts.*"
+    NN: ~
+  batch_size: 64
+  block_size: 128
+  group_size: -1
+  symmetric: True
+  class:
+    int4:
+      - "re:model.layers.\\d+.block_sparse_moe.experts.*"
+    int8: ~
+  ignore: 
+    - "model.embed_tokens.weight"
+    - "re:model.layers.\\d+.input_layernorm.weight"
+    - "re:model.layers.\\d+.self_attn.*"
+    - "re:model.layers.\\d+.post_attention_layernorm.weight"
+    - "re:model.layers.\\d+.block_sparse_moe.gate.weight"
+    - "re:model.layers.\\d+.block_sparse_moe.e_score_correction_bias"
+    - "model.norm.weight"
+    - "lm_head.weight"
+
+dequant:
+  ignore:
+    - "re:lm_head.*"
+    - "re:model.embed_tokens.*"
+    - "re:model.layers.\\d+.input_layernorm.*"
+    - "re:model.layers.\\d+.mlp.gate.*"
+    - "re:model.layers.\\d+.post_attention_layernorm.*"
+    - "re:model.layers.\\d+.self_attn.k_norm.*"
+    - "re:model.layers.\\d+.self_attn.q_norm.*"
+    - "re:model.norm.*"
+    - "re:model.layers.\\d+.block_sparse_moe.e_score_correction_bias"
+    - "re:model.layers.\\d+.block_sparse_moe.gate.weight"

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.*"

dev/configs/w4a8_int/qwen3-30B-A3B.yml

@@ -0,0 +1,60 @@
+model:
+  name: Qwen3-30B-A3B-Instruct-2507
+  model_type: qwen3_moe
+
+paths:
+  checkpoint_dir: /home/mashun/vllm_project/checkpoints/Qwen3-30B-A3B-Instruct-2507
+  output_dir: /home/mashun/vllm_project/checkpoints/Qwen3-30B-A3B-Instruct-2507-w4a8
+
+runtime:
+  num_workers: 8
+  mode: quant # dequant | dequant+quant | quant (only support bf16)
+
+retry_files: ~
+
+quant:
+  version: 2
+  format:
+    TN:
+      - "re:model.layers.\\d+.mlp.experts.\\d+.down_proj"
+      - "re:model.layers.\\d+.mlp.experts.\\d+.gate_proj"
+      - "re:model.layers.\\d+.mlp.experts.\\d+.up_proj"
+      - "re:model.layers.\\d+.self_attn.q_proj.*"
+      - "re:model.layers.\\d+.self_attn.v_proj.*"
+      - "re:model.layers.\\d+.self_attn.o_proj.*"
+      - "re:model.layers.\\d+.self_attn.k_proj.*"
+    NN: ~
+  batch_size: 64
+  block_size: 128
+  group_size: -1
+  symmetric: True
+  class:
+    int4:
+      - "re:model.layers.\\d+.mlp.experts.\\d+.down_proj"
+      - "re:model.layers.\\d+.mlp.experts.\\d+.gate_proj"
+      - "re:model.layers.\\d+.mlp.experts.\\d+.up_proj"
+    int8:
+      - "re:model.layers.\\d+.self_attn.q_proj.*"
+      - "re:model.layers.\\d+.self_attn.v_proj.*"
+      - "re:model.layers.\\d+.self_attn.o_proj.*"
+      - "re:model.layers.\\d+.self_attn.k_proj.*"
+  ignore: 
+    - "re:lm_head.*"
+    - "re:model.norm.*"
+    - "re:model.embed_tokens.*"
+    - "re:model.layers.\\d+.input_layernorm.*"
+    - "re:model.layers.\\d+.mlp.gate.*"
+    - "re:model.layers.\\d+.post_attention_layernorm.*"
+    - "re:model.layers.\\d+.self_attn.k_norm.*"
+    - "re:model.layers.\\d+.self_attn.q_norm.*"
+
+dequant:
+  ignore:
+    - "re:lm_head.*"
+    - "re:model.embed_tokens.*"
+    - "re:model.layers.\\d+.input_layernorm.*"
+    - "re:model.layers.\\d+.mlp.gate.*"
+    - "re:model.layers.\\d+.post_attention_layernorm.*"
+    - "re:model.layers.\\d+.self_attn.k_norm.*"
+    - "re:model.layers.\\d+.self_attn.q_norm.*"
+    - "re:model.norm.*"

dev/configs/w4a8_int/qwen3-32B.yml

@@ -0,0 +1,53 @@
+model:
+  name: Qwen3-32B
+  model_type: qwen3
+
+paths:
+  checkpoint_dir: /home/mashun/vllm_project/checkpoints/Qwen3-32B
+  output_dir: /home/mashun/vllm_project/checkpoints/Qwen3-32B-w4a8
+
+runtime:
+  num_workers: 16
+  mode: quant # dequant | dequant+quant | quant (only support bf16)
+
+retry_files: ~
+
+quant:
+  version: 2
+  format: 
+    TN: ~
+    NN: 
+      - "re:model.layers.\\d+.mlp.down_proj.*"
+      - "re:model.layers.\\d+.mlp.gate_proj.*"
+      - "re:model.layers.\\d+.mlp.up_proj.*"
+      - "re:model.layers.\\d+.self_attn.k_proj.*"
+      - "re:model.layers.\\d+.self_attn.o_proj.*"
+      - "re:model.layers.\\d+.self_attn.q_proj.*"
+      - "re:model.layers.\\d+.self_attn.v_proj.*"
+  batch_size: 64
+  block_size: 128
+  group_size: -1
+  symmetric: True
+  class:
+    int4:
+      - "re:model.layers.\\d+.mlp.down_proj.*"
+      - "re:model.layers.\\d+.mlp.gate_proj.*"
+      - "re:model.layers.\\d+.mlp.up_proj.*"
+      - "re:model.layers.\\d+.self_attn.k_proj.*"
+      - "re:model.layers.\\d+.self_attn.o_proj.*"
+      - "re:model.layers.\\d+.self_attn.q_proj.*"
+      - "re:model.layers.\\d+.self_attn.v_proj.*"
+    int8: ~
+  ignore: 
+    # - "re:model.layers.(?:[0-9]|[1-5][0-9]|6[0-2])(?:\\..*)?$"
+    - "re:model.layers.63.*"
+    - "re:lm_head.*"
+    - "re:model.embed_tokens.*"
+    - "re:model.layers.\\d+.input_layernorm.*"
+    - "re:model.layers.\\d+.post_attention_layernorm.*"
+    - "re:model.layers.\\d+.self_attn.k_norm.*"
+    - "re:model.layers.\\d+.self_attn.q_norm.*"
+    - "re:model.norm.*"
+
+dequant:
+  ignore: ~

dev/configs/w4a8_int/qwen3-VL-30B-A3B.yml

@@ -0,0 +1,51 @@
+model:
+  name: Qwen3-30B-A3B-Instruct-2507
+  model_type: qwen3_vl_moe
+
+paths:
+  checkpoint_dir: /home/mashun/vllm_project/checkpoints/Qwen3-VL-30B-A3B-Instruct
+  output_dir: /home/mashun/vllm_project/checkpoints/Qwen3-VL-30B-A3B-Instruct-w4a8
+
+runtime:
+  num_workers: 16
+  mode: quant # dequant | dequant+quant | quant (only support bf16)
+
+retry_files: ~
+
+quant:
+  version: 2
+  format: 
+    TN:
+      - "re:.*experts.*"
+    NN: ~
+  batch_size: 64
+  block_size: 128
+  group_size: -1
+  symmetric: True
+  class:
+    int4: 
+      - "re:.*experts.*"
+    int8: ~
+  ignore: 
+    - "re:lm_head.*"
+    - "re:.*embed_tokens.*"
+    - "re:.*input_layernorm.*"
+    - "re:.*post_attention_layernorm.*"
+    - "re:.*k_norm.*"
+    - "re:.*q_norm.*"
+    - "re:.*mlp\\.gate.*"
+    - "re:.*\\.norm.*"
+    - "re:.*visual.*"
+    - "re:.*k_proj.*"
+    - "re:.*o_proj.*"
+    - "re:.*v_proj.*"
+    - "re:.*q_proj.*"
+
+dequant:
+  ignore:
+    - "model.language_model.embed_tokens.weight"
+    - "lm_head.weight"
+    - "re:model.language_model.layers.\\d+.input_layernorm.weight"
+    - "re:model.language_model.layers.\\d+.post_attention_layernorm.weight"
+    - "re:model.language_model.norm.weight"
+    - "re:model.visual.*"

dev/configs/w4a8_int/qwen3-coder-30B-A3B.yml

@@ -0,0 +1,56 @@
+model:
+  name: Qwen3-Coder-30B-A3B-Instruct-FP8
+  model_type: qwen3_moe
+
+paths:
+  checkpoint_dir: /home/mashun/vllm_project/checkpoints/Qwen3-Coder-30B-A3B-Instruct-FP8
+  output_dir: /home/mashun/vllm_project/checkpoints/Qwen3-Coder-30B-A3B-Instruct-w4a8
+
+runtime:
+  num_workers: 8
+  mode: dequant+quant # dequant | dequant+quant | quant (only support bf16)
+
+retry_files: ~
+
+quant:
+  version: 2
+  format: 
+    TN:
+      - "re:model.layers.\\d+.mlp.experts.\\d+.down_proj"
+      - "re:model.layers.\\d+.mlp.experts.\\d+.gate_proj"
+      - "re:model.layers.\\d+.mlp.experts.\\d+.up_proj"
+    NN: ~
+  batch_size: 64
+  block_size: 128
+  group_size: -1
+  symmetric: True
+  class:
+    int4:
+      - "re:model.layers.\\d+.mlp.experts.\\d+.down_proj"
+      - "re:model.layers.\\d+.mlp.experts.\\d+.gate_proj"
+      - "re:model.layers.\\d+.mlp.experts.\\d+.up_proj"
+    int8: ~
+  ignore: 
+    - "re:lm_head.*"
+    - "re:model.norm.*"
+    - "re:model.embed_tokens.*"
+    - "re:model.layers.\\d+.input_layernorm.*"
+    - "re:model.layers.\\d+.mlp.gate.*"
+    - "re:model.layers.\\d+.post_attention_layernorm.*"
+    - "re:model.layers.\\d+.self_attn.k_norm.*"
+    - "re:model.layers.\\d+.self_attn.k_proj.*"
+    - "re:model.layers.\\d+.self_attn.o_proj.*"
+    - "re:model.layers.\\d+.self_attn.q_norm.*"
+    - "re:model.layers.\\d+.self_attn.q_proj.*"
+    - "re:model.layers.\\d+.self_attn.v_proj.*"
+
+dequant:
+  ignore:
+    - "re:lm_head.*"
+    - "re:model.embed_tokens.*"
+    - "re:model.layers.\\d+.input_layernorm.*"
+    - "re:model.layers.\\d+.mlp.gate.*"
+    - "re:model.layers.\\d+.post_attention_layernorm.*"
+    - "re:model.layers.\\d+.self_attn.k_norm.*"
+    - "re:model.layers.\\d+.self_attn.q_norm.*"
+    - "re:model.norm.*"

dev/configs/w4a8_int/qwen3-next.yml

@@ -0,0 +1,54 @@
+model:
+  name: Qwen3-Next-80B-A3B-Instruct
+  model_type: qwen3_moe
+
+paths:
+  checkpoint_dir: /data/model/Qwen3-Next-80B-A3B-Thinking
+  output_dir: /data/model/Qwen3-Next-80B-A3B-Thinking-w4a8
+
+runtime:
+  num_workers: 8
+  mode: quant # dequant | dequant+quant | quant (only support bf16)
+
+retry_files: ~
+
+quant:
+  version: 2
+  format: 
+    TN:
+      - "re:.*experts.*"
+    NN: ~
+  batch_size: 64
+  block_size: 128
+  group_size: -1
+  symmetric: True
+  class:
+    int4:
+      - "re:.*experts.*"
+    int8: ~
+  ignore:
+    - "re:.*linear_attn.*"
+    - "re:.*self_attn.*"
+    - "re:.*shared_expert.*"
+    - "re:.*linear_attn.in_proj_ba.*"
+    - "re:.*conv1d.*"
+    - "re:.*A_log.*"
+    - "re:.*shared_expert_gate.*"
+    - "re:.*bias.*"
+    - "re:.*norm.*"
+    - "re:.*mlp.gate.*"
+    - "re:.*embed_tokens.*"
+    - "re:.*lm_head.*"
+    - "re:.*mtp.*"
+
+dequant:
+  ignore: 
+    - "re:.*linear_attn.in_proj_ba*"
+    - "re:.*conv1d.*"
+    - "re:.*A_log.*"
+    - "re:.*shared_expert_gate.*"
+    - "re:.*bias.*"
+    - "re:.*norm.*"
+    - "re:.*mlp.gate.*"
+    - "re:.*embed_tokens.*"
+    - "re:.*lm_head.*"

dev/configs/w4a8_int/qwen3-omni-30B-A3B.yml

@@ -0,0 +1,58 @@
+model:
+  name: Qwen3-Coder-30B-A3B-Instruct-FP8
+  model_type: qwen3_moe
+
+paths:
+  checkpoint_dir: /home/mashun/vllm_project/checkpoints/Qwen3-Omni-30B-A3B-Thinking
+  output_dir: /home/mashun/vllm_project/checkpoints/Qwen3-Omni-30B-A3B-Thinking-w4a8
+
+runtime:
+  num_workers: 8
+  mode: quant # dequant | dequant+quant | quant (only support bf16)
+
+retry_files: ~
+
+quant:
+  version: 2
+  format: 
+    TN:
+      - "re:.*model.layers.\\d+.mlp.experts.\\d+.down_proj"
+      - "re:.*model.layers.\\d+.mlp.experts.\\d+.gate_proj"
+      - "re:.*model.layers.\\d+.mlp.experts.\\d+.up_proj"
+    NN: ~
+  batch_size: 64
+  block_size: 128
+  group_size: -1
+  symmetric: True
+  class:
+    int4:
+      - "re:.*model.layers.\\d+.mlp.experts.\\d+.down_proj"
+      - "re:.*model.layers.\\d+.mlp.experts.\\d+.gate_proj"
+      - "re:.*model.layers.\\d+.mlp.experts.\\d+.up_proj"
+    int8: ~
+  ignore: 
+    - "re:.*lm_head.*"
+    - "re:.*model.norm.*"
+    - "re:.*model.embed_tokens.*"
+    - "re:.*model.layers.\\d+.input_layernorm.*"
+    - "re:.*model.layers.\\d+.mlp.gate.*"
+    - "re:.*model.layers.\\d+.post_attention_layernorm.*"
+    - "re:.*model.layers.\\d+.self_attn.k_norm.*"
+    - "re:.*model.layers.\\d+.self_attn.k_proj.*"
+    - "re:.*model.layers.\\d+.self_attn.o_proj.*"
+    - "re:.*model.layers.\\d+.self_attn.q_norm.*"
+    - "re:.*model.layers.\\d+.self_attn.q_proj.*"
+    - "re:.*model.layers.\\d+.self_attn.v_proj.*"
+    - "re:.*audio_tower.*"
+    - "re:.*visual.*"
+
+dequant:
+  ignore:
+    - "re:lm_head.*"
+    - "re:model.embed_tokens.*"
+    - "re:model.layers.\\d+.input_layernorm.*"
+    - "re:model.layers.\\d+.mlp.gate.*"
+    - "re:model.layers.\\d+.post_attention_layernorm.*"
+    - "re:model.layers.\\d+.self_attn.k_norm.*"
+    - "re:model.layers.\\d+.self_attn.q_norm.*"
+    - "re:model.norm.*"

dev/configs/w4a8_int/qwen3-vl-32B.yml

@@ -0,0 +1,56 @@
+model:
+  name: Qwen3-VL-32B-Instruct
+  model_type: qwen3_vl
+
+paths:
+  checkpoint_dir: /home/mashun/vllm_project/checkpoints/Qwen3-VL-32B-Instruct
+  output_dir: /home/mashun/vllm_project/checkpoints/Qwen3-VL-32B-Instruct-w4a8
+
+runtime:
+  num_workers: 8
+  mode: quant # dequant | dequant+quant | quant (only support bf16)
+
+retry_files: ~
+
+quant:
+  version: 2
+  format:
+    TN: ~
+    NN:
+      - "re:model.language_model.layers.\\d+.self_attn.v_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.q_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.k_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.o_proj.*"
+      - "re:model.language_model.layers.\\d+.mlp.up_proj.*"
+      - "re:model.language_model.layers.\\d+.mlp.gate_proj.*"
+      - "re:model.language_model.layers.\\d+.mlp.down_proj.*"
+  batch_size: 64
+  block_size: 128
+  group_size: -1
+  symmetric: True
+  class:
+    int4:
+      - "re:model.language_model.layers.\\d+.self_attn.v_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.q_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.k_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.o_proj.*"
+      - "re:model.language_model.layers.\\d+.mlp.up_proj.*"
+      - "re:model.language_model.layers.\\d+.mlp.gate_proj.*"
+      - "re:model.language_model.layers.\\d+.mlp.down_proj.*"
+    int8: ~
+  ignore: 
+    - "re:.*visual.*"
+    - "re:.*norm.*"
+    - "re:lm_head.*"
+    - "re:.*embed_tokens.*"
+
+dequant:
+  ignore:
+    - "re:lm_head.*"
+    - "re:model.embed_tokens.*"
+    - "re:model.layers.\\d+.input_layernorm.*"
+    - "re:model.layers.\\d+.mlp.gate.*"
+    - "re:model.layers.\\d+.post_attention_layernorm.*"
+    - "re:model.layers.\\d+.self_attn.k_norm.*"
+    - "re:model.layers.\\d+.self_attn.q_norm.*"
+    - "re:model.norm.*"

dev/configs/w4a8_int/qwen35-122B-A10B.yml

@@ -0,0 +1,73 @@
+model:
+  name: Qwen3.5-122B-A10B
+  model_type: qwen3_5_moe
+
+paths:
+  checkpoint_dir: /home/mashun/vllm_project/checkpoints/Qwen3.5-122B-A10B
+  output_dir: /home/mashun/vllm_project/checkpoints/Qwen3.5-122B-A10B-w4a8
+
+runtime:
+  num_workers: 16
+  mode: quant # dequant | dequant+quant | quant (only support bf16)
+
+retry_files: ~
+
+quant:
+  version: 2
+  format: 
+    TN:
+      - "re:model.language_model.layers.\\d+.mlp.experts.gate_up_proj.*"
+      - "re:model.language_model.layers.\\d+.mlp.experts.down_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.o_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.q_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.k_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.v_proj.*"
+    NN: ~
+  batch_size: 64
+  block_size: 128
+  group_size: -1
+  symmetric: True
+  class:
+    int4: 
+      - "re:model.language_model.layers.\\d+.mlp.experts.gate_up_proj.*"
+      - "re:model.language_model.layers.\\d+.mlp.experts.down_proj.*"
+    int8:
+      - "re:model.language_model.layers.\\d+.self_attn.o_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.q_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.k_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.v_proj.*"
+  ignore: 
+    - "re:lm_head.*"
+    - "re:.*embed_tokens.*"
+    - "re:.*visual.*"
+    - "re:.*mtp.*"
+    - "re:.*norm.*"
+    - "re:.*linear_attn.*"
+    - "re:.*shared_expert.*"
+    - "re:model.language_model.layers.\\d+.mlp.gate\\..*"
+    - "re:.*bias.*"
+
+dequant:
+  ignore:
+    - "re:.*mtp.fc.*"
+    - "re:.*mtp.layers.\\d+.mlp.gate.*"
+    - "re:.*mtp.layers.\\d+.mlp.shared_expert_gate.*"
+    - "re:.*visual.*"
+    - "re:.*shared_expert_gate.*"
+    - "re:.model.language_model.layers.\\d+.mlp.gate.*"
+    - "re:.*conv1d.*"
+    - "re:.*in_proj_a.*"
+    - "re:.*in_proj_b.*"
+    - "re:.*lm_head.*"
+    - "re:.*embed_tokens.*"
+    - "re:.*bias.*"
+    - "re:.*A_log.*"
+    - "re:.*model.language_model.layers.\\d+.linear_attn.norm.*"
+    - "re:.*input_layernorm.*"
+    - "re:.*post_attention_layernorm.*"
+    - "re:.*mtp.pre_fc_norm_embedding.*"
+    - "re:.*q_norm.*"
+    - "re:.*k_norm.*"
+    - "re:.*mtp.pre_fc_norm_hidden.*"
+    - "re:.*model.language_model.norm.*"
+    - "re:.*mtp.norm.*"

dev/configs/w4a8_int/qwen35-27B.yml

@@ -0,0 +1,50 @@
+model:
+  name: Qwen3.5-27B
+  model_type: qwen3_5
+
+paths:
+  checkpoint_dir: /home/mashun/vllm_project/checkpoints/Qwen3.5-27B
+  output_dir: /home/mashun/vllm_project/checkpoints/Qwen3.5-27B-w4a8
+
+runtime:
+  num_workers: 8
+  mode: quant # dequant | dequant+quant | quant (only support bf16)
+
+retry_files: ~
+
+quant:
+  version: 2
+  format:
+    TN: ~
+    NN:
+      - "re:model.language_model.layers.\\d+.mlp.down_proj.*"
+      - "re:model.language_model.layers.\\d+.mlp.gate_proj.*"
+      - "re:model.language_model.layers.\\d+.mlp.up_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.k_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.q_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.o_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.v_proj.*"
+  batch_size: 64
+  block_size: 128
+  group_size: -1
+  symmetric: True
+  class:
+    int4:
+      - "re:model.language_model.layers.\\d+.mlp.down_proj.*"
+      - "re:model.language_model.layers.\\d+.mlp.gate_proj.*"
+      - "re:model.language_model.layers.\\d+.mlp.up_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.k_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.q_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.o_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.v_proj.*"
+    int8: ~
+  ignore: 
+    - "re:lm_head.*"
+    - "re:.*visual.*"
+    - "re:.*mtp.*"
+    - "re:.*linear_attn.*"
+    - "re:.*norm.*"
+    - "re:.*embed_tokens.*"
+
+dequant:
+  ignore: ~

dev/configs/w4a8_int/qwen35-35B-A3B.yml

@@ -0,0 +1,50 @@
+model:
+  name: Qwen3.5-35B-A3B
+  model_type: qwen3_5_moe
+
+paths:
+  checkpoint_dir: /home/mashun/vllm_project/checkpoints/Qwen3.5-35B-A3B
+  output_dir: /home/mashun/vllm_project/checkpoints/Qwen3.5-35B-A3B-w4a8
+
+runtime:
+  num_workers: 16
+  mode: quant # dequant | dequant+quant | quant (only support bf16)
+
+retry_files: ~
+
+quant:
+  version: 2
+  format: 
+    TN:
+      - "re:model.language_model.layers.\\d+.mlp.experts.gate_up_proj.*"
+      - "re:model.language_model.layers.\\d+.mlp.experts.down_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.o_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.q_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.k_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.v_proj.*"
+    NN: ~
+  batch_size: 64
+  block_size: 128
+  group_size: -1
+  symmetric: True
+  class:
+    int4: 
+      - "re:model.language_model.layers.\\d+.mlp.experts.gate_up_proj.*"
+      - "re:model.language_model.layers.\\d+.mlp.experts.down_proj.*"
+    int8:
+      - "re:model.language_model.layers.\\d+.self_attn.o_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.q_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.k_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.v_proj.*"
+  ignore: 
+    - "re:lm_head.*"
+    - "re:.*embed_tokens.*"
+    - "re:.*visual.*"
+    - "re:.*mtp.*"
+    - "re:.*norm.*"
+    - "re:.*linear_attn.*"
+    - "re:model.language_model.layers.\\d+.mlp.gate\\..*"
+    - "re:.*shared_expert.*"
+
+dequant:
+  ignore: ~

dev/configs/w4a8_int/qwen35-397B-A17B.yml

@@ -0,0 +1,75 @@
+model:
+  name: Qwen3.5-397B-A17B
+  model_type: qwen3_5_moe
+
+paths:
+  checkpoint_dir: /home/mashun/vllm_project/checkpoints/Qwen3.5-397B-A17B-FP8
+  output_dir: /home/mashun/vllm_project/checkpoints/Qwen3.5-397B-A17B-w4a8
+
+runtime:
+  num_workers: 16
+  mode: dequant+quant # dequant | dequant+quant | quant (only support bf16)
+
+retry_files: ~
+
+quant:
+  version: 2
+  format: 
+    TN:
+      - "re:model.language_model.layers.\\d+.mlp.experts.\\d+.gate_proj.*"
+      - "re:model.language_model.layers.\\d+.mlp.experts.\\d+.up_proj.*"
+      - "re:model.language_model.layers.\\d+.mlp.experts.\\d+.down_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.o_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.q_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.k_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.v_proj.*"
+    NN: ~
+  batch_size: 64
+  block_size: 128
+  group_size: -1
+  symmetric: True
+  class:
+    int4: 
+      - "re:model.language_model.layers.\\d+.mlp.experts.\\d+.gate_proj.*"
+      - "re:model.language_model.layers.\\d+.mlp.experts.\\d+.up_proj.*"
+      - "re:model.language_model.layers.\\d+.mlp.experts.\\d+.down_proj.*"
+    int8:
+      - "re:model.language_model.layers.\\d+.self_attn.o_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.q_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.k_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.v_proj.*"
+  ignore: 
+    - "re:lm_head.*"
+    - "re:.*embed_tokens.*"
+    - "re:.*visual.*"
+    - "re:.*mtp.*"
+    - "re:.*norm.*"
+    - "re:.*linear_attn.*"
+    - "re:.*shared_expert.*"
+    - "re:model.language_model.layers.\\d+.mlp.gate\\..*"
+    - "re:.*bias.*"
+
+dequant:
+  ignore:
+    - "re:.*mtp.fc.*"
+    - "re:.*mtp.layers.\\d+.mlp.gate.*"
+    - "re:.*mtp.layers.\\d+.mlp.shared_expert_gate.*"
+    - "re:.*visual.*"
+    - "re:.*shared_expert_gate.*"
+    - "re:.model.language_model.layers.\\d+.mlp.gate.*"
+    - "re:.*conv1d.*"
+    - "re:.*in_proj_a.*"
+    - "re:.*in_proj_b.*"
+    - "re:.*lm_head.*"
+    - "re:.*embed_tokens.*"
+    - "re:.*bias.*"
+    - "re:.*A_log.*"
+    - "re:.*model.language_model.layers.\\d+.linear_attn.norm.*"
+    - "re:.*input_layernorm.*"
+    - "re:.*post_attention_layernorm.*"
+    - "re:.*mtp.pre_fc_norm_embedding.*"
+    - "re:.*q_norm.*"
+    - "re:.*k_norm.*"
+    - "re:.*mtp.pre_fc_norm_hidden.*"
+    - "re:.*model.language_model.norm.*"
+    - "re:.*mtp.norm.*"

dev/configs/w4a8_int/step3-vl-10B.yml

@@ -0,0 +1,55 @@
+model:
+  name: Step3-VL-10B
+  model_type: qwen3
+
+paths:
+  checkpoint_dir: /home/mashun/vllm_project/checkpoints/Step3-VL-10B
+  output_dir: /home/mashun/vllm_project/checkpoints/Step3-VL-10B-w4a8
+
+runtime:
+  num_workers: 8
+  mode: quant # dequant | dequant+quant | quant (only support bf16)
+
+retry_files: ~
+
+quant:
+  version: 2
+  format: 
+    TN: ~
+    NN: 
+      - "re:model.layers.\\d+.mlp.down_proj.*"
+      - "re:model.layers.\\d+.mlp.gate_proj.*"
+      - "re:model.layers.\\d+.mlp.up_proj.*"
+      - "re:model.layers.\\d+.self_attn.k_proj.*"
+      - "re:model.layers.\\d+.self_attn.o_proj.*"
+      - "re:model.layers.\\d+.self_attn.q_proj.*"
+      - "re:model.layers.\\d+.self_attn.v_proj.*"
+  batch_size: 64
+  block_size: 128
+  group_size: -1
+  symmetric: True
+  class:
+    int4:
+      - "re:model.layers.\\d+.mlp.down_proj.*"
+      - "re:model.layers.\\d+.mlp.gate_proj.*"
+      - "re:model.layers.\\d+.mlp.up_proj.*"
+      - "re:model.layers.\\d+.self_attn.k_proj.*"
+      - "re:model.layers.\\d+.self_attn.o_proj.*"
+      - "re:model.layers.\\d+.self_attn.q_proj.*"
+      - "re:model.layers.\\d+.self_attn.v_proj.*"
+    int8: ~
+  ignore: 
+    # - "re:model.layers.(?:[0-9]|[1-5][0-9]|6[0-2])(?:\\..*)?$"
+    - "re:model.layers.46.*"
+    - "re:lm_head.*"
+    - "re:model.embed_tokens.*"
+    - "re:model.layers.\\d+.input_layernorm.*"
+    - "re:model.layers.\\d+.post_attention_layernorm.*"
+    - "re:model.layers.\\d+.self_attn.k_norm.*"
+    - "re:model.layers.\\d+.self_attn.q_norm.*"
+    - "re:model.norm.*"
+    - "re:vision_model.*"
+    - "re:vit_large_projector.*"
+
+dequant:
+  ignore: ~

dev/configs/w4a8_int/step35-flash.yml

@@ -0,0 +1,62 @@
+model:
+  name: Step-3.5-Flash
+  model_type: step3p5
+
+paths:
+  checkpoint_dir: /home/mashun/vllm_project/checkpoints/Step-3.5-Flash
+  output_dir: /home/mashun/vllm_project/checkpoints/Step-3.5-Flash-w4a8
+
+runtime:
+  num_workers: 16
+  mode: quant # dequant | dequant+quant | quant (only support bf16)
+
+retry_files: ~
+
+quant:
+  version: 2
+  format:
+    TN:
+      - "re:.*moe.down_proj.*"
+      - "re:.*moe.gate_proj.*"
+      - "re:.*moe.up_proj.*"
+      - "re:.*self_attn.g_proj.*"
+      - "re:.*self_attn.k_proj.*"
+      - "re:.*self_attn.o_proj.*"
+      - "re:.*self_attn.q_proj.*"
+      - "re:.*self_attn.v_proj.*"
+      - "re:.*share_expert.*"
+      - "re:.*mlp.down_proj.*"
+      - "re:.*mlp.up_proj.*"
+      - "re:.*mlp.gate_proj.*"
+    NN: ~
+  batch_size: 64
+  block_size: 128
+  group_size: -1
+  symmetric: True
+  class:
+    int4:
+      - "re:.*moe.down_proj.*"
+      - "re:.*moe.gate_proj.*"
+      - "re:.*moe.up_proj.*"
+    int8: 
+      - "re:.*self_attn.g_proj.*"
+      - "re:.*self_attn.k_proj.*"
+      - "re:.*self_attn.o_proj.*"
+      - "re:.*self_attn.q_proj.*"
+      - "re:.*self_attn.v_proj.*"
+      - "re:.*mlp.down_proj.*"
+      - "re:.*mlp.up_proj.*"
+      - "re:.*mlp.gate_proj.*"
+  ignore: 
+    - "re:.*norm.*"
+    - "re:.*bias.*"
+    - "re:.*moe.gate\\..*"
+    - "re:.*eh_proj.*"
+    - "re:.*transformer.*"
+    - "re:.*lm_head.*"
+    - "re:.*embed_tokens.*"
+    - "re:.*share_expert.*"
+
+   
+dequant:
+  ignore: ~

dev/configs/w8a8_int/qwen3-30B-A3B.yml

@@ -0,0 +1,60 @@
+model:
+  name: Qwen3-30B-A3B-Instruct-2507
+  model_type: qwen3_moe
+
+paths:
+  checkpoint_dir: /home/mashun/vllm_project/checkpoints/Qwen3-30B-A3B-Instruct-2507
+  output_dir: /home/mashun/vllm_project/checkpoints/Qwen3-30B-A3B-Instruct-2507-w8a8
+
+runtime:
+  num_workers: 8
+  mode: quant # dequant | dequant+quant | quant (only support bf16)
+
+retry_files: ~
+
+quant:
+  version: 2
+  format:
+    TN:
+      - "re:model.layers.\\d+.mlp.experts.\\d+.down_proj"
+      - "re:model.layers.\\d+.mlp.experts.\\d+.gate_proj"
+      - "re:model.layers.\\d+.mlp.experts.\\d+.up_proj"
+      - "re:model.layers.\\d+.self_attn.q_proj.*"
+      - "re:model.layers.\\d+.self_attn.v_proj.*"
+      - "re:model.layers.\\d+.self_attn.o_proj.*"
+      - "re:model.layers.\\d+.self_attn.k_proj.*"
+    NN: ~
+  batch_size: 64
+  block_size: 128
+  group_size: -1
+  symmetric: True
+  class:
+    int4: ~
+    int8:
+      - "re:model.layers.\\d+.mlp.experts.\\d+.down_proj"
+      - "re:model.layers.\\d+.mlp.experts.\\d+.gate_proj"
+      - "re:model.layers.\\d+.mlp.experts.\\d+.up_proj"
+      - "re:model.layers.\\d+.self_attn.q_proj.*"
+      - "re:model.layers.\\d+.self_attn.v_proj.*"
+      - "re:model.layers.\\d+.self_attn.o_proj.*"
+      - "re:model.layers.\\d+.self_attn.k_proj.*"
+  ignore: 
+    - "re:lm_head.*"
+    - "re:model.norm.*"
+    - "re:model.embed_tokens.*"
+    - "re:model.layers.\\d+.input_layernorm.*"
+    - "re:model.layers.\\d+.mlp.gate.*"
+    - "re:model.layers.\\d+.post_attention_layernorm.*"
+    - "re:model.layers.\\d+.self_attn.k_norm.*"
+    - "re:model.layers.\\d+.self_attn.q_norm.*"
+
+dequant:
+  ignore:
+    - "re:lm_head.*"
+    - "re:model.embed_tokens.*"
+    - "re:model.layers.\\d+.input_layernorm.*"
+    - "re:model.layers.\\d+.mlp.gate.*"
+    - "re:model.layers.\\d+.post_attention_layernorm.*"
+    - "re:model.layers.\\d+.self_attn.k_norm.*"
+    - "re:model.layers.\\d+.self_attn.q_norm.*"
+    - "re:model.norm.*"

dev/configs/w8a8_int/qwen3-32B.yml

@@ -0,0 +1,53 @@
+model:
+  name: Qwen3-32B
+  model_type: qwen3
+
+paths:
+  checkpoint_dir: /home/mashun/vllm_project/checkpoints/Qwen3-32B
+  output_dir: /home/mashun/vllm_project/checkpoints/Qwen3-32B-w4a8
+
+runtime:
+  num_workers: 16
+  mode: quant # dequant | dequant+quant | quant (only support bf16)
+
+retry_files: ~
+
+quant:
+  version: 2
+  format: 
+    TN: 
+      - "re:model.layers.\\d+.mlp.down_proj.*"
+      - "re:model.layers.\\d+.mlp.gate_proj.*"
+      - "re:model.layers.\\d+.mlp.up_proj.*"
+      - "re:model.layers.\\d+.self_attn.k_proj.*"
+      - "re:model.layers.\\d+.self_attn.o_proj.*"
+      - "re:model.layers.\\d+.self_attn.q_proj.*"
+      - "re:model.layers.\\d+.self_attn.v_proj.*"
+    NN: ~
+  batch_size: 64
+  block_size: 128
+  group_size: -1
+  symmetric: True
+  class:
+    int4: ~
+    int8:
+      - "re:model.layers.\\d+.mlp.down_proj.*"
+      - "re:model.layers.\\d+.mlp.gate_proj.*"
+      - "re:model.layers.\\d+.mlp.up_proj.*"
+      - "re:model.layers.\\d+.self_attn.k_proj.*"
+      - "re:model.layers.\\d+.self_attn.o_proj.*"
+      - "re:model.layers.\\d+.self_attn.q_proj.*"
+      - "re:model.layers.\\d+.self_attn.v_proj.*"
+  ignore: 
+    # - "re:model.layers.(?:[0-9]|[1-5][0-9]|6[0-2])(?:\\..*)?$"
+    # - "re:model.layers.63.*"
+    - "re:lm_head.*"
+    - "re:model.embed_tokens.*"
+    - "re:model.layers.\\d+.input_layernorm.*"
+    - "re:model.layers.\\d+.post_attention_layernorm.*"
+    - "re:model.layers.\\d+.self_attn.k_norm.*"
+    - "re:model.layers.\\d+.self_attn.q_norm.*"
+    - "re:model.norm.*"
+
+dequant:
+  ignore: ~

dev/configs/w8a8_int/qwen3-VL-30B-A3B.yml

@@ -0,0 +1,51 @@
+model:
+  name: Qwen3-30B-A3B-Instruct-2507
+  model_type: qwen3_vl_moe
+
+paths:
+  checkpoint_dir: /home/mashun/vllm_project/checkpoints/Qwen3-VL-30B-A3B-Instruct
+  output_dir: /home/mashun/vllm_project/checkpoints/Qwen3-VL-30B-A3B-Instruct-w4a8
+
+runtime:
+  num_workers: 16
+  mode: quant # dequant | dequant+quant | quant (only support bf16)
+
+retry_files: ~
+
+quant:
+  version: 2
+  format: 
+    TN:
+      - "re:.*experts.*"
+    NN: ~
+  batch_size: 64
+  block_size: 128
+  group_size: -1
+  symmetric: True
+  class:
+    int4: ~
+    int8:
+      - "re:.*experts.*"
+  ignore: 
+    - "re:lm_head.*"
+    - "re:.*embed_tokens.*"
+    - "re:.*input_layernorm.*"
+    - "re:.*post_attention_layernorm.*"
+    - "re:.*k_norm.*"
+    - "re:.*q_norm.*"
+    - "re:.*mlp\\.gate.*"
+    - "re:.*\\.norm.*"
+    - "re:.*visual.*"
+    - "re:.*k_proj.*"
+    - "re:.*o_proj.*"
+    - "re:.*v_proj.*"
+    - "re:.*q_proj.*"
+
+dequant:
+  ignore:
+    - "model.language_model.embed_tokens.weight"
+    - "lm_head.weight"
+    - "re:model.language_model.layers.\\d+.input_layernorm.weight"
+    - "re:model.language_model.layers.\\d+.post_attention_layernorm.weight"
+    - "re:model.language_model.norm.weight"
+    - "re:model.visual.*"

dev/configs/w8a8_int/qwen3-coder-30B-A3B.yml

@@ -0,0 +1,56 @@
+model:
+  name: Qwen3-Coder-30B-A3B-Instruct-FP8
+  model_type: qwen3_moe
+
+paths:
+  checkpoint_dir: /home/mashun/vllm_project/checkpoints/Qwen3-Coder-30B-A3B-Instruct-FP8
+  output_dir: /home/mashun/vllm_project/checkpoints/Qwen3-Coder-30B-A3B-Instruct-w4a8
+
+runtime:
+  num_workers: 8
+  mode: dequant+quant # dequant | dequant+quant | quant (only support bf16)
+
+retry_files: ~
+
+quant:
+  version: 2
+  format: 
+    TN:
+      - "re:model.layers.\\d+.mlp.experts.\\d+.down_proj"
+      - "re:model.layers.\\d+.mlp.experts.\\d+.gate_proj"
+      - "re:model.layers.\\d+.mlp.experts.\\d+.up_proj"
+    NN: ~
+  batch_size: 64
+  block_size: 128
+  group_size: -1
+  symmetric: True
+  class:
+    int4: ~
+    int8:
+      - "re:model.layers.\\d+.mlp.experts.\\d+.down_proj"
+      - "re:model.layers.\\d+.mlp.experts.\\d+.gate_proj"
+      - "re:model.layers.\\d+.mlp.experts.\\d+.up_proj"
+  ignore: 
+    - "re:lm_head.*"
+    - "re:model.norm.*"
+    - "re:model.embed_tokens.*"
+    - "re:model.layers.\\d+.input_layernorm.*"
+    - "re:model.layers.\\d+.mlp.gate.*"
+    - "re:model.layers.\\d+.post_attention_layernorm.*"
+    - "re:model.layers.\\d+.self_attn.k_norm.*"
+    - "re:model.layers.\\d+.self_attn.k_proj.*"
+    - "re:model.layers.\\d+.self_attn.o_proj.*"
+    - "re:model.layers.\\d+.self_attn.q_norm.*"
+    - "re:model.layers.\\d+.self_attn.q_proj.*"
+    - "re:model.layers.\\d+.self_attn.v_proj.*"
+
+dequant:
+  ignore:
+    - "re:lm_head.*"
+    - "re:model.embed_tokens.*"
+    - "re:model.layers.\\d+.input_layernorm.*"
+    - "re:model.layers.\\d+.mlp.gate.*"
+    - "re:model.layers.\\d+.post_attention_layernorm.*"
+    - "re:model.layers.\\d+.self_attn.k_norm.*"
+    - "re:model.layers.\\d+.self_attn.q_norm.*"
+    - "re:model.norm.*"

dev/configs/w8a8_int/qwen3-next.yml

@@ -0,0 +1,52 @@
+model:
+  name: Qwen3-Next-80B-A3B-Instruct
+  model_type: qwen3_moe
+
+paths:
+  checkpoint_dir: /home/mashun/vllm_project/checkpoints/Qwen3-Next-80B-A3B-Instruct
+  output_dir: /home/mashun/vllm_project/checkpoints/Qwen3-Next-80B-A3B-Instruct-w8a8
+
+runtime:
+  num_workers: 16
+  mode: dequant # dequant | dequant+quant | quant (only support bf16)
+
+quant:
+  version: 2
+  format: 
+    TN:
+      - "re:.*experts.*"
+    NN: ~
+  batch_size: 64
+  block_size: 128
+  group_size: -1
+  symmetric: True
+  class:
+    int4: ~
+    int8:
+      - "re:.*experts.*"
+  ignore:
+    - "re:.*linear_attn.*"
+    - "re:.*self_attn.*"
+    - "re:.*shared_expert.*"
+    - "re:.*linear_attn.in_proj_ba.*"
+    - "re:.*conv1d.*"
+    - "re:.*A_log.*"
+    - "re:.*shared_expert_gate.*"
+    - "re:.*bias.*"
+    - "re:.*norm.*"
+    - "re:.*mlp.gate.*"
+    - "re:.*embed_tokens.*"
+    - "re:.*lm_head.*"
+    - "re:.*mtp.*"
+
+dequant:
+  ignore: 
+    - "re:.*linear_attn.in_proj_ba*"
+    - "re:.*conv1d.*"
+    - "re:.*A_log.*"
+    - "re:.*shared_expert_gate.*"
+    - "re:.*bias.*"
+    - "re:.*norm.*"
+    - "re:.*mlp.gate.*"
+    - "re:.*embed_tokens.*"
+    - "re:.*lm_head.*"

dev/configs/w8a8_int/qwen3-omni-30B-A3B.yml

@@ -0,0 +1,58 @@
+model:
+  name: Qwen3-Coder-30B-A3B-Instruct-FP8
+  model_type: qwen3_moe
+
+paths:
+  checkpoint_dir: /home/mashun/vllm_project/checkpoints/Qwen3-Omni-30B-A3B-Thinking
+  output_dir: /home/mashun/vllm_project/checkpoints/Qwen3-Omni-30B-A3B-Thinking-w4a8
+
+runtime:
+  num_workers: 8
+  mode: quant # dequant | dequant+quant | quant (only support bf16)
+
+retry_files: ~
+
+quant:
+  version: 2
+  format: 
+    TN:
+      - "re:.*model.layers.\\d+.mlp.experts.\\d+.down_proj"
+      - "re:.*model.layers.\\d+.mlp.experts.\\d+.gate_proj"
+      - "re:.*model.layers.\\d+.mlp.experts.\\d+.up_proj"
+    NN: ~
+  batch_size: 64
+  block_size: 128
+  group_size: -1
+  symmetric: True
+  class:
+    int4: ~
+    int8:
+      - "re:.*model.layers.\\d+.mlp.experts.\\d+.down_proj"
+      - "re:.*model.layers.\\d+.mlp.experts.\\d+.gate_proj"
+      - "re:.*model.layers.\\d+.mlp.experts.\\d+.up_proj"
+  ignore: 
+    - "re:.*lm_head.*"
+    - "re:.*model.norm.*"
+    - "re:.*model.embed_tokens.*"
+    - "re:.*model.layers.\\d+.input_layernorm.*"
+    - "re:.*model.layers.\\d+.mlp.gate.*"
+    - "re:.*model.layers.\\d+.post_attention_layernorm.*"
+    - "re:.*model.layers.\\d+.self_attn.k_norm.*"
+    - "re:.*model.layers.\\d+.self_attn.k_proj.*"
+    - "re:.*model.layers.\\d+.self_attn.o_proj.*"
+    - "re:.*model.layers.\\d+.self_attn.q_norm.*"
+    - "re:.*model.layers.\\d+.self_attn.q_proj.*"
+    - "re:.*model.layers.\\d+.self_attn.v_proj.*"
+    - "re:.*audio_tower.*"
+    - "re:.*visual.*"
+
+dequant:
+  ignore:
+    - "re:lm_head.*"
+    - "re:model.embed_tokens.*"
+    - "re:model.layers.\\d+.input_layernorm.*"
+    - "re:model.layers.\\d+.mlp.gate.*"
+    - "re:model.layers.\\d+.post_attention_layernorm.*"
+    - "re:model.layers.\\d+.self_attn.k_norm.*"
+    - "re:model.layers.\\d+.self_attn.q_norm.*"
+    - "re:model.norm.*"

dev/configs/w8a8_int/qwen3-vl-32B.yml

@@ -0,0 +1,56 @@
+model:
+  name: Qwen3-VL-32B-Instruct
+  model_type: qwen3_vl
+
+paths:
+  checkpoint_dir: /home/mashun/vllm_project/checkpoints/Qwen3-VL-32B-Instruct
+  output_dir: /home/mashun/vllm_project/checkpoints/Qwen3-VL-32B-Instruct-w4a8
+
+runtime:
+  num_workers: 8
+  mode: quant # dequant | dequant+quant | quant (only support bf16)
+
+retry_files: ~
+
+quant:
+  version: 2
+  format:
+    TN:
+      - "re:model.language_model.layers.\\d+.self_attn.v_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.q_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.k_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.o_proj.*"
+      - "re:model.language_model.layers.\\d+.mlp.up_proj.*"
+      - "re:model.language_model.layers.\\d+.mlp.gate_proj.*"
+      - "re:model.language_model.layers.\\d+.mlp.down_proj.*"
+    NN: ~
+  batch_size: 64
+  block_size: 128
+  group_size: -1
+  symmetric: True
+  class:
+    int4: ~
+    int8:
+      - "re:model.language_model.layers.\\d+.self_attn.v_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.q_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.k_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.o_proj.*"
+      - "re:model.language_model.layers.\\d+.mlp.up_proj.*"
+      - "re:model.language_model.layers.\\d+.mlp.gate_proj.*"
+      - "re:model.language_model.layers.\\d+.mlp.down_proj.*"
+  ignore: 
+    - "re:.*visual.*"
+    - "re:.*norm.*"
+    - "re:lm_head.*"
+    - "re:.*embed_tokens.*"
+
+dequant:
+  ignore:
+    - "re:lm_head.*"
+    - "re:model.embed_tokens.*"
+    - "re:model.layers.\\d+.input_layernorm.*"
+    - "re:model.layers.\\d+.mlp.gate.*"
+    - "re:model.layers.\\d+.post_attention_layernorm.*"
+    - "re:model.layers.\\d+.self_attn.k_norm.*"
+    - "re:model.layers.\\d+.self_attn.q_norm.*"
+    - "re:model.norm.*"

dev/configs/w8a8_int/qwen35-27B.yml

@@ -0,0 +1,50 @@
+model:
+  name: Qwen3.5-27B
+  model_type: qwen3_5
+
+paths:
+  checkpoint_dir: /home/mashun/vllm_project/checkpoints/Qwen3.5-27B
+  output_dir: /home/mashun/vllm_project/checkpoints/Qwen3.5-27B-w8a8
+
+runtime:
+  num_workers: 8
+  mode: quant # dequant | dequant+quant | quant (only support bf16)
+
+retry_files: ~
+
+quant:
+  version: 2
+  format:
+    TN:
+      - "re:model.language_model.layers.\\d+.mlp.down_proj.*"
+      - "re:model.language_model.layers.\\d+.mlp.gate_proj.*"
+      - "re:model.language_model.layers.\\d+.mlp.up_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.k_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.q_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.o_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.v_proj.*"
+    NN: ~
+  batch_size: 64
+  block_size: 128
+  group_size: -1
+  symmetric: True
+  class:
+    int4: ~
+    int8:
+      - "re:model.language_model.layers.\\d+.mlp.down_proj.*"
+      - "re:model.language_model.layers.\\d+.mlp.gate_proj.*"
+      - "re:model.language_model.layers.\\d+.mlp.up_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.k_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.q_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.o_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.v_proj.*"
+  ignore: 
+    - "re:lm_head.*"
+    - "re:.*visual.*"
+    - "re:.*mtp.*"
+    - "re:.*linear_attn.*"
+    - "re:.*norm.*"
+    - "re:.*embed_tokens.*"
+
+dequant:
+  ignore: ~

dev/configs/w8a8_int/qwen35-35B-A3B.yml

@@ -0,0 +1,50 @@
+model:
+  name: Qwen3.5-35B-A3B
+  model_type: qwen3_5_moe
+
+paths:
+  checkpoint_dir: /home/mashun/vllm_project/checkpoints/Qwen3.5-35B-A3B
+  output_dir: /home/mashun/vllm_project/checkpoints/Qwen3.5-35B-A3B-w8a8
+
+runtime:
+  num_workers: 16
+  mode: quant # dequant | dequant+quant | quant (only support bf16)
+
+retry_files: ~
+
+quant:
+  version: 2
+  format: 
+    TN:
+      - "re:model.language_model.layers.\\d+.mlp.experts.gate_up_proj.*"
+      - "re:model.language_model.layers.\\d+.mlp.experts.down_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.o_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.q_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.k_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.v_proj.*"
+    NN: ~
+  batch_size: 64
+  block_size: 128
+  group_size: -1
+  symmetric: True
+  class:
+    int4: ~
+    int8:
+      - "re:model.language_model.layers.\\d+.mlp.experts.gate_up_proj.*"
+      - "re:model.language_model.layers.\\d+.mlp.experts.down_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.o_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.q_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.k_proj.*"
+      - "re:model.language_model.layers.\\d+.self_attn.v_proj.*"
+  ignore: 
+    - "re:lm_head.*"
+    - "re:.*embed_tokens.*"
+    - "re:.*visual.*"
+    - "re:.*mtp.*"
+    - "re:.*norm.*"
+    - "re:.*linear_attn.*"
+    - "re:model.language_model.layers.\\d+.mlp.gate\\..*"
+    - "re:.*shared_expert.*"
+
+dequant:
+  ignore: ~

dev/configs/w8a8_int/step3-vl-10B.yml

@@ -0,0 +1,55 @@
+model:
+  name: Step3-VL-10B
+  model_type: qwen3
+
+paths:
+  checkpoint_dir: /home/mashun/vllm_project/checkpoints/Step3-VL-10B
+  output_dir: /home/mashun/vllm_project/checkpoints/Step3-VL-10B-w4a8
+
+runtime:
+  num_workers: 8
+  mode: quant # dequant | dequant+quant | quant (only support bf16)
+
+retry_files: ~
+
+quant:
+  version: 2
+  format: 
+    TN:
+      - "re:model.layers.\\d+.mlp.down_proj.*"
+      - "re:model.layers.\\d+.mlp.gate_proj.*"
+      - "re:model.layers.\\d+.mlp.up_proj.*"
+      - "re:model.layers.\\d+.self_attn.k_proj.*"
+      - "re:model.layers.\\d+.self_attn.o_proj.*"
+      - "re:model.layers.\\d+.self_attn.q_proj.*"
+      - "re:model.layers.\\d+.self_attn.v_proj.*"
+    NN: ~
+  batch_size: 64
+  block_size: 128
+  group_size: -1
+  symmetric: True
+  class:
+    int4: ~
+    int8:
+      - "re:model.layers.\\d+.mlp.down_proj.*"
+      - "re:model.layers.\\d+.mlp.gate_proj.*"
+      - "re:model.layers.\\d+.mlp.up_proj.*"
+      - "re:model.layers.\\d+.self_attn.k_proj.*"
+      - "re:model.layers.\\d+.self_attn.o_proj.*"
+      - "re:model.layers.\\d+.self_attn.q_proj.*"
+      - "re:model.layers.\\d+.self_attn.v_proj.*"
+  ignore: 
+    # - "re:model.layers.(?:[0-9]|[1-5][0-9]|6[0-2])(?:\\..*)?$"
+    # - "re:model.layers.46.*"
+    - "re:lm_head.*"
+    - "re:model.embed_tokens.*"
+    - "re:model.layers.\\d+.input_layernorm.*"
+    - "re:model.layers.\\d+.post_attention_layernorm.*"
+    - "re:model.layers.\\d+.self_attn.k_norm.*"
+    - "re:model.layers.\\d+.self_attn.q_norm.*"
+    - "re:model.norm.*"
+    - "re:vision_model.*"
+    - "re:vit_large_projector.*"
+
+dequant:
+  ignore: ~

dev/configs/w8a8_int/step35-flash.yml

@@ -0,0 +1,62 @@
+model:
+  name: Step-3.5-Flash
+  model_type: step3p5
+
+paths:
+  checkpoint_dir: /home/mashun/vllm_project/checkpoints/Step-3.5-Flash
+  output_dir: /home/mashun/vllm_project/checkpoints/Step-3.5-Flash-w8a8
+
+runtime:
+  num_workers: 16
+  mode: quant # dequant | dequant+quant | quant (only support bf16)
+
+retry_files: ~
+
+quant:
+  version: 2
+  format:
+    TN:
+      - "re:.*moe.down_proj.*"
+      - "re:.*moe.gate_proj.*"
+      - "re:.*moe.up_proj.*"
+      - "re:.*self_attn.g_proj.*"
+      - "re:.*self_attn.k_proj.*"
+      - "re:.*self_attn.o_proj.*"
+      - "re:.*self_attn.q_proj.*"
+      - "re:.*self_attn.v_proj.*"
+      - "re:.*share_expert.*"
+      - "re:.*mlp.down_proj.*"
+      - "re:.*mlp.up_proj.*"
+      - "re:.*mlp.gate_proj.*"
+    NN: ~
+  batch_size: 64
+  block_size: 128
+  group_size: -1
+  symmetric: True
+  class:
+    int4: ~
+    int8: 
+      - "re:.*moe.down_proj.*"
+      - "re:.*moe.gate_proj.*"
+      - "re:.*moe.up_proj.*"
+      - "re:.*self_attn.g_proj.*"
+      - "re:.*self_attn.k_proj.*"
+      - "re:.*self_attn.o_proj.*"
+      - "re:.*self_attn.q_proj.*"
+      - "re:.*self_attn.v_proj.*"
+      - "re:.*mlp.down_proj.*"
+      - "re:.*mlp.up_proj.*"
+      - "re:.*mlp.gate_proj.*"
+  ignore: 
+    - "re:.*norm.*"
+    - "re:.*bias.*"
+    - "re:.*moe.gate\\..*"
+    - "re:.*eh_proj.*"
+    - "re:.*transformer.*"
+    - "re:.*lm_head.*"
+    - "re:.*embed_tokens.*"
+    - "re:.*share_expert.*"
+
+   
+dequant:
+  ignore: ~