first commit

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.*"