set VLLM_USE_LIGHTOP_FILL_MOE_ALIGN=0

vllm/attention/layer.py

@@ -536,4 +536,4 @@ direct_register_custom_op(
     mutates_args=["output"],
     fake_impl=unified_attention_with_output_fake,
     dispatch_key=current_platform.dispatch_key,
-)
+)
\ No newline at end of file

vllm/model_executor/model_loader/utils.py

@@ -251,8 +251,8 @@ def get_model_architecture(
                     os.environ['VLLM_USE_LIGHTOP_MOE_SUM_MUL_ADD'] = '1'
                 if not envs.is_set("VLLM_USE_OPT_CAT"):
                     os.environ['VLLM_USE_OPT_CAT'] = '1'
-                if not envs.is_set("VLLM_USE_LIGHTOP_FILL_MOE_ALIGN"):
-                    os.environ['VLLM_USE_LIGHTOP_FILL_MOE_ALIGN'] = '1'
+                # if not envs.is_set("VLLM_USE_LIGHTOP_FILL_MOE_ALIGN"):
+                #     os.environ['VLLM_USE_LIGHTOP_FILL_MOE_ALIGN'] = '1'
                 
             if os.getenv('GEMM_PAD') != '1': 
                 os.environ['GEMM_PAD'] = '0'
@@ -266,8 +266,8 @@ def get_model_architecture(
                     os.environ['VLLM_USE_LIGHTOP_MOE_SUM_MUL_ADD'] = '1'
                 if not envs.is_set("VLLM_USE_OPT_CAT"):
                     os.environ['VLLM_USE_OPT_CAT'] = '1'
-                if not envs.is_set("VLLM_USE_LIGHTOP_FILL_MOE_ALIGN"):
-                    os.environ['VLLM_USE_LIGHTOP_FILL_MOE_ALIGN'] = '1'
+                # if not envs.is_set("VLLM_USE_LIGHTOP_FILL_MOE_ALIGN"):
+                #     os.environ['VLLM_USE_LIGHTOP_FILL_MOE_ALIGN'] = '1'
                     
         # awq相关配置
         try: