[Bugfix][Core] Allow AuxOutput reset after abort without running requests (#59060)

Signed-off-by: aoshen02 <aoshen@inferact.ai>
Co-authored-by: OpenAI Codex <codex@openai.com>

tests/v1/core/test_scheduler.py

@@ -1562,13 +1562,7 @@ def test_scheduler_reset_prefix_cache():
     assert not scheduler.reset_prefix_cache()
     scheduler.aux_output_connector.reset.assert_not_called()
 
-    with pytest.raises(RuntimeError, match=r"pause\(mode='keep'\)"):
-        scheduler.reset_prefix_cache(reset_running_requests=True)
-
-    # pause(mode="keep") also waits for scheduled model outputs to drain.
-    scheduler.set_pause_state(PauseState.PAUSED_ALL)
-    with pytest.raises(RuntimeError, match="model output is in flight"):
-        scheduler.reset_prefix_cache(reset_running_requests=True)
+    # Pause completes pending model outputs before the caller resets the scheduler.
     for request in requests:
         request.num_in_flight_tokens = 0
 

vllm/v1/core/sched/scheduler.py

@@ -2781,16 +2781,6 @@ class Scheduler(SchedulerInterface):
         Otherwise, this method will only reset the KV prefix cache when there
         is no running requests taking KV cache.
         """
-        if reset_running_requests and self.aux_output_connector is not None:
-            if self._pause_state != PauseState.PAUSED_ALL:
-                raise RuntimeError(
-                    "AuxOutput Connector only supports resetting running requests "
-                    "after pause(mode='keep')."
-                )
-            if any(request.num_in_flight_tokens for request in self.requests.values()):
-                raise RuntimeError(
-                    "AuxOutput Connector cannot reset while model output is in flight."
-                )
         if reset_running_requests:
             # For logging.
             timestamp = time.monotonic()