set VLLM_USE_PD_SPLIT=1

vllm/envs.py

@@ -1222,7 +1222,7 @@ environment_variables: dict[str, Callable[[], Any]] = {
 
     # vLLM will split prefill and decode, not mix up
     "VLLM_USE_PD_SPLIT":
-        lambda: (os.environ.get("VLLM_USE_PD_SPLIT", "False").lower() in
+        lambda: (os.environ.get("VLLM_USE_PD_SPLIT", "True").lower() in
                  ("true", "1")), 
 
     # vLLM will sync to avoid pp vmfault

vllm/v1/core/sched/scheduler.py

@@ -1051,14 +1051,15 @@ class Scheduler(SchedulerInterface):
     
     def schedule(self) -> SchedulerOutput:
         if envs.VLLM_USE_PD_SPLIT: 
-            return self.schedule_split_pd()
-        else:
-            if self.connector is not None:
-                return self.schedule_default()
-            if self.full_cuda_graph and self.use_mla and self.num_spec_tokens > 0 :
-                return self.schedule_split_pd()
+            if self.use_mla:
+                if self.full_cuda_graph and self.num_spec_tokens > 0:
+                    return self.schedule_split_pd() 
+                else:
+                    self.schedule_default()
             else:
-                return self.schedule_default()
+                return self.schedule_split_pd()
+        else:
+            return self.schedule_default()
 
     def _update_after_schedule(
         self,