first commit

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: ~