Merge branch 'v0.11.0-dev-yql-12.24' into 'v0.11.0-dev'

在v0.11中暂不支持w4a16的moe算子scale和zero合并操作

See merge request dcutoolkit/deeplearing/vllm!318

vllm/model_executor/layers/quantization/moe_wna16.py

@@ -195,6 +195,8 @@ class MoeWNA16Method(FusedMoEMethodBase):
                  moe: "FusedMoEConfig") -> None:
         super().__init__(moe)
         self.quant_config = quant_config
+        # if os.getenv('AWQ_MOE_SZ') == None:
+        #     os.environ['AWQ_MOE_SZ'] = '1'
         self.use_w4a16_moe_sz = os.environ.get('AWQ_MOE_SZ') == '1'
         self.use_w4a16_cuda = 0
         self.use_moe_lmslim = 0

vllm/model_executor/model_loader/utils.py

@@ -244,8 +244,8 @@ def _get_model_architecture(
                     
         # awq相关配置
         try:
-            if os.getenv('AWQ_MOE_SZ') == None:
-                os.environ['AWQ_MOE_SZ'] = '1'
+            # if os.getenv('AWQ_MOE_SZ') == None:
+            #     os.environ['AWQ_MOE_SZ'] = '1'
             if os.getenv('AWQ_PAD') == None and (torch.cuda.get_device_properties(torch.cuda.current_device()).multi_processor_count == 120):
                 os.environ['AWQ_PAD'] = '1'
         except Exception as e: