first commit

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: ~