fix(proxy): streaming loop yields once it has held the event loop 50 ms, clock read inline

litellm/proxy/proxy_server.py

@@ -10012,6 +10012,7 @@ def _pop_complete_sse_frame(buffer: str) -> tuple[str | None, str]:
 _STREAM_KEEPALIVE: Final = object()
 
 _KEEPALIVE_MIN_SECONDS: Final = 1.0
+_STREAM_LOOP_HOLD_SECONDS: Final = 0.05
 _KEEPALIVE_MAX_SECONDS: Final = 300.0
 _EMPTY_MAPPING: Final[Mapping[str, object]] = MappingProxyType({})
 _EMPTY_HEADERS: Final[Mapping[str, str]] = MappingProxyType({})
@@ -10287,10 +10288,13 @@ async def async_data_generator(
             else stream_iterator
         )
 
+        loop_released_at = time.monotonic()  # rebind-ok: last time this stream let other tasks run
         async for item in stream_source:
             # Unblock the event loop: a stream whose chunks are already buffered never awaits
             # between chunks, so without this yield one long stream holds the loop for its whole body
-            await asyncio.sleep(0)
+            if time.monotonic() - loop_released_at >= _STREAM_LOOP_HOLD_SECONDS:
+                await asyncio.sleep(0)
+                loop_released_at = time.monotonic()
             if item is _STREAM_KEEPALIVE:
                 yield ": ping\n\n"
                 continue

tests/unit/proxy/proxy_server/test_streaming_helpers.py

@@ -2095,13 +2095,17 @@ def test_fast_serialize_simple_model_response_stream_keeps_served_service_tier()
 @pytest.mark.asyncio
 async def test_async_data_generator_lets_other_tasks_run_during_a_stream_that_never_waits(monkeypatch):
     """A stream whose chunks are already buffered never suspends on its own, so
-    the generator has to hand the loop back itself: another task must get to
-    run before the stream is over."""
+    the generator has to hand the loop back itself once it has held the loop
+    long enough: another task must get to run before the stream is over, and
+    not before the hold budget is used up."""
     _patch_logging_flags(monkeypatch)
     order: list[str] = []  # mutable-ok: records the interleaving of the two tasks
+    clock: list[float] = [0.0]  # mutable-ok: fake monotonic clock, advanced 20 ms per chunk
+    monkeypatch.setattr(ps.time, "monotonic", lambda: clock[0])
 
     async def chunks():
         for index in range(6):
+            clock[0] += 0.02
             order.append(f"chunk{index}")
             yield _simple_chunk(content=str(index))
 
@@ -2118,3 +2122,4 @@ async def test_async_data_generator_lets_other_tasks_run_during_a_stream_that_ne
     await other
 
     assert "other" in order[:-1], order
+    assert order.index("other") >= 3, order