修复awq模型的VLLM_USE_LIGHTOP_MOE_SUM_MUL_ADD设置位置

vllm/model_executor/model_loader/utils.py

@@ -232,6 +232,11 @@ def get_model_architecture(
                                 'ChatGLMModel', 'Glm4ForCausalLM', 'ChatGLMForConditionalGeneration', 'BaichuanForCausalLM', 'BloomForCausalLM', 'TeleChat2ForCausalLM', 'MixtralForCausalLM', 'FalconForCausalLM',
                                 'MedusaModel', 'MLPSpeculatorPreTrainedModel', 'DeepseekV2ForCausalLM', 'DeepseekV3ForCausalLM', 'DeepSeekMTPModel']  
     if any(arch in architectures for arch in support_nn_architectures): 
+        #针对使用dtype为fp16的情况的量化默认关闭"VLLM_USE_LIGHTOP_MOE_SUM_MUL_ADD"
+        if model_config.quantization in {"awq", "awq_marlin", "moe_wna16"}:
+            if not envs.is_set("VLLM_USE_LIGHTOP_MOE_SUM_MUL_ADD"):
+                os.environ['VLLM_USE_LIGHTOP_MOE_SUM_MUL_ADD'] = '0'      
+        
         if not envs.VLLM_USE_NN:
             if os.getenv('LLAMA_NN') != '0': 
                 if (architectures == ['QWenLMHeadModel'] or architectures == ['ChatGLMModel'] ) and visions != []:
@@ -287,11 +292,7 @@ def get_model_architecture(
             if os.getenv('FA_PAD') != '1': 
                 os.environ['FA_PAD'] = '0'
         else:
-            if architectures in [['DeepseekV3ForCausalLM'], ['DeepSeekMTPModel']]:
-                #针对使用dtype为fp16的情况的量化默认关闭"VLLM_USE_LIGHTOP_MOE_SUM_MUL_ADD"
-                if model_config.quantization in {"awq", "awq_marlin", "moe_wna16"}:
-                    if not envs.is_set("VLLM_USE_LIGHTOP_MOE_SUM_MUL_ADD"):
-                        os.environ['VLLM_USE_LIGHTOP_MOE_SUM_MUL_ADD'] = '0'                
+            if architectures in [['DeepseekV3ForCausalLM'], ['DeepSeekMTPModel']]:          
                 if not envs.is_set("VLLM_USE_LIGHTOP"):
                     os.environ['VLLM_USE_LIGHTOP'] = '1'
                 if not envs.is_set("VLLM_USE_LIGHTOP_MOE_SUM_MUL_ADD"):