update VLLM_USE_PD_SPLIT=0

vllm/envs.py

@@ -1157,7 +1157,7 @@ environment_variables: dict[str, Callable[[], Any]] = {
                  ("true", "1")),
     # vLLM will split prefill and decode, not mix up
     "VLLM_USE_PD_SPLIT":
-        lambda: (os.environ.get("VLLM_USE_PD_SPLIT", "True").lower() in
+        lambda: (os.environ.get("VLLM_USE_PD_SPLIT", "False").lower() in
                  ("true", "1")), 
     # vLLM will sync to avoid pp vmfault
     "VLLM_USE_PP_SYNC":

vllm/model_executor/model_loader/utils.py

@@ -253,8 +253,9 @@ def get_model_architecture(
                     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'
+            else:
                 if not envs.is_set("VLLM_USE_PD_SPLIT"):
-                    os.environ['VLLM_USE_PD_SPLIT'] = '0'
+                    os.environ['VLLM_USE_PD_SPLIT'] = '1'
                 
                 
             if os.getenv('GEMM_PAD') != '1': 
@@ -271,8 +272,9 @@ def get_model_architecture(
                     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'
+            else:
                 if not envs.is_set("VLLM_USE_PD_SPLIT"):
-                    os.environ['VLLM_USE_PD_SPLIT'] = '0'
+                    os.environ['VLLM_USE_PD_SPLIT'] = '1'
                     
         # awq相关配置
         try: