Remove AsyncLLMEngine busy loop, shield background task (#1059)

tests/async_engine/test_request_tracker.py

@@ -4,10 +4,25 @@ from vllm.engine.async_llm_engine import RequestTracker
 from vllm.outputs import RequestOutput
 
 
+class DummyEvent:
+
+    def __init__(self):
+        self._flag = False
+
+    def set(self):
+        self._flag = True
+
+    def clear(self):
+        self._flag = False
+
+
 def test_request_tracker():
     tracker = RequestTracker()
+    tracker.new_requests_event = DummyEvent()
     stream_1 = tracker.add_request("1")
+    assert tracker.new_requests_event._flag
     new, finished = tracker.get_new_and_finished_requests()
+    assert not tracker.new_requests_event._flag
     assert len(new) == 1
     assert new[0]["request_id"] == "1"
     assert not finished
@@ -15,7 +30,9 @@ def test_request_tracker():
 
     stream_2 = tracker.add_request("2")
     stream_3 = tracker.add_request("3")
+    assert tracker.new_requests_event._flag
     new, finished = tracker.get_new_and_finished_requests()
+    assert not tracker.new_requests_event._flag
     assert len(new) == 2
     assert new[0]["request_id"] == "2"
     assert new[1]["request_id"] == "3"
@@ -26,6 +43,7 @@ def test_request_tracker():
     # request_ids must be unique
     with pytest.raises(KeyError):
         tracker.add_request("1")
+    assert not tracker.new_requests_event._flag
 
     tracker.abort_request("1")
     new, finished = tracker.get_new_and_finished_requests()
@@ -36,6 +54,7 @@ def test_request_tracker():
 
     stream_4 = tracker.add_request("4")
     tracker.abort_request("4")
+    assert tracker.new_requests_event._flag
     new, finished = tracker.get_new_and_finished_requests()
     assert len(finished) == 1
     assert "4" in finished
@@ -43,9 +62,11 @@ def test_request_tracker():
     assert stream_4.finished
 
     stream_5 = tracker.add_request("5")
+    assert tracker.new_requests_event._flag
     tracker.process_request_output(
         RequestOutput("2", "output", [], [], finished=True))
     new, finished = tracker.get_new_and_finished_requests()
+    assert not tracker.new_requests_event._flag
     assert len(finished) == 1
     assert "2" in finished
     assert len(new) == 1