improve chunked prefill performance

[Bugfix] Fix vllm-project#7592 vllm 0.5.4 enable_chunked_prefill throughput is slightly lower than 0.5.3~0.5.0. (vllm-project#7874)

vllm/core/scheduler.py

@@ -883,16 +883,21 @@ def _schedule_chunked_prefill(self):
 
         # Update waiting requests.
         self.waiting.extendleft(running_scheduled.preempted)
+
         # Update new running requests.
-        self.running.extend([s.seq_group for s in prefills.seq_groups])
-        self.running.extend(
-            [s.seq_group for s in running_scheduled.decode_seq_groups])
-        self.running.extend(
-            [s.seq_group for s in running_scheduled.prefill_seq_groups])
+        # By default, vLLM scheduler prioritizes prefills.
+        # Once chunked prefill is enabled,
+        # the policy is changed to prioritize decode requests.
         self.running.extend(
             [s.seq_group for s in swapped_in.decode_seq_groups])
         self.running.extend(
             [s.seq_group for s in swapped_in.prefill_seq_groups])
+        self.running.extend(
+            [s.seq_group for s in running_scheduled.decode_seq_groups])
+        self.running.extend(
+            [s.seq_group for s in running_scheduled.prefill_seq_groups])
+        self.running.extend([s.seq_group for s in prefills.seq_groups])
+
         # Update swapped requests.
         self.swapped.extend(running_scheduled.swapped_out)
         return SchedulerOutputs(