[BugFix] get_and_reset only when scheduler outputs are not empty (#6266)

vllm/engine/async_llm_engine.py

@@ -225,11 +225,11 @@ class _AsyncLLMEngine(LLMEngine):
         """
         seq_group_metadata_list, scheduler_outputs = self.scheduler[
             virtual_engine].schedule()
-        finished_requests_ids = self.scheduler[
-            virtual_engine].get_and_reset_finished_requests_ids()
 
         if not scheduler_outputs.is_empty():
             # Execute the model.
+            finished_requests_ids = self.scheduler[
+                virtual_engine].get_and_reset_finished_requests_ids()
             execute_model_req = ExecuteModelRequest(
                 seq_group_metadata_list=seq_group_metadata_list,
                 blocks_to_swap_in=scheduler_outputs.blocks_to_swap_in,

vllm/engine/llm_engine.py

@@ -871,10 +871,10 @@ class LLMEngine:
                 "as performance will be severely degraded otherwise.")
         seq_group_metadata_list, scheduler_outputs = self.scheduler[
             0].schedule()
-        finished_requests_ids = self.scheduler[
-            0].get_and_reset_finished_requests_ids()
 
         if not scheduler_outputs.is_empty():
+            finished_requests_ids = self.scheduler[
+                0].get_and_reset_finished_requests_ids()
             execute_model_req = ExecuteModelRequest(
                 seq_group_metadata_list=seq_group_metadata_list,
                 blocks_to_swap_in=scheduler_outputs.blocks_to_swap_in,