[Bugfix] Terminate streaming sessions on ABORT/ERROR and prevent zombie sessions (#58942)

Signed-off-by: David Cheung <d7cheung@gmail.com>
Signed-off-by: Nick Hill <nickhill123@gmail.com>
Co-authored-by: Benedikt Falk <bfalk@nvidia.com>
Co-authored-by: Nick Hill <nickhill123@gmail.com>
This commit is contained in:
David Cheung
2026-10-15 13:32:19 +00:00
committed by GitHub
co-authored by Benedikt Falk Nick Hill
parent 766c837bd5
commit 1463501553
6 changed files with 213 additions and 11 deletions
@@ -210,6 +210,7 @@ def _make_dplb_client(num_engines: int = 3, client_count: int = 1) -> DPLBAsyncM
client = object.__new__(DPLBAsyncMPClient)
client.client_count = client_count
client.reqs_in_flight = {}
client.streaming_reqs = set()
client.engine_inflight = Counter()
client.core_engines = [bytes([i, 0]) for i in range(num_engines)]
client.lb_engines = [[0, 0, 0.0] for _ in range(num_engines)]
@@ -407,6 +408,42 @@ def test_dplb_finished_requests_release_inflight():
assert req.request_id not in client.reqs_in_flight
@pytest.mark.parametrize("final_sent", [False, True])
def test_dplb_streaming_session_abort_goes_to_all_engines(final_sent: bool):
"""A streaming-input chunk in flight when its session finished may
re-create it on any engine, so an open session's abort reaches every
engine. Other aborts go only to the routed engine."""
client = _make_dplb_client(num_engines=3)
client.resources = SimpleNamespace(engine_dead=False)
sent: dict[bytes, list[str]] = {}
async def abort_requests(request_ids, engine):
sent[engine] = request_ids
client._abort_requests = abort_requests
chunk = make_request(SamplingParams(max_tokens=1))
chunk.resumable = True
client.get_core_engine_for_request(chunk)
if final_sent:
chunk.resumable = False
client.get_core_engine_for_request(chunk)
# The engine finished the session, dropping its routing entry.
outputs = EngineCoreOutputs(finished_requests={chunk.request_id})
asyncio.run(DPLBAsyncMPClient.process_engine_outputs(client, outputs))
client.reqs_in_flight["live"] = client.core_engines[1]
asyncio.run(client.abort_requests_async([chunk.request_id, "live"]))
if final_sent:
expected = {client.core_engines[1]: ["live"]}
else:
expected = {engine: [chunk.request_id] for engine in client.core_engines}
expected[client.core_engines[1]].append("live")
assert sent == expected
assert not client.streaming_reqs
@pytest.mark.parametrize(
("effective_size", "other_size"),
[(None, None), (4224, 4224), (4224, 1056), (4224, None)],
+84
View File
@@ -1642,6 +1642,90 @@ async def test_cumulative_output_collector_n():
assert third[0].text == "c"
def _make_streaming_input_request(
request_id: str, prompt_token_ids: list[int]
) -> EngineCoreRequest:
"""One chunk of a streaming-input (resumable) request."""
return EngineCoreRequest(
request_id=request_id,
prompt_token_ids=prompt_token_ids,
mm_features=None,
sampling_params=SamplingParams(
detokenize=False, output_kind=RequestOutputKind.DELTA
),
pooling_params=None,
arrival_time=0,
lora_request=None,
cache_salt=None,
data_parallel_rank=None,
resumable=True,
external_req_id=f"{request_id}-ext",
)
def _add_streaming_request_with_queued_chunk(
processor: OutputProcessor, request_id: str
) -> RequestOutputCollector:
queue = RequestOutputCollector(RequestOutputKind.DELTA, request_id=request_id)
processor.add_request(
_make_streaming_input_request(request_id, [1, 2, 3]), None, queue=queue
)
processor.add_request(_make_streaming_input_request(request_id, [7, 8]), None)
return queue
def test_streaming_input_stop_is_chunk_boundary():
"""STOP is a per-chunk boundary: finished is forced False, the queued
chunk is applied, and the request state stays alive."""
processor = OutputProcessor(None, log_stats=False)
queue = _add_streaming_request_with_queued_chunk(processor, "stream-req")
processed = processor.process_outputs(
[
EngineCoreOutput(
request_id="stream-req",
new_token_ids=[4],
finish_reason=FinishReason.STOP,
)
]
)
assert processed.reqs_to_abort == []
assert processor.get_num_unfinished_requests() == 1
output = queue.get_nowait()
assert output is not None
assert output.finished is False
assert output.outputs[0].finish_reason == "stop"
req_state = processor.request_states["stream-req"]
assert req_state.prompt_token_ids is not None
assert req_state.prompt_token_ids[-2:] == [7, 8]
@pytest.mark.parametrize("finish_reason", [FinishReason.ABORT, FinishReason.ERROR])
def test_streaming_input_abort_error_is_terminal(finish_reason: FinishReason):
"""ABORT/ERROR is terminal, not a chunk boundary: finished stays True,
the state is freed, and an abort is routed back to the engine core."""
processor = OutputProcessor(None, log_stats=False)
queue = _add_streaming_request_with_queued_chunk(processor, "stream-req")
processed = processor.process_outputs(
[
EngineCoreOutput(
request_id="stream-req",
new_token_ids=[],
finish_reason=finish_reason,
)
]
)
assert processor.get_num_unfinished_requests() == 0
assert processed.reqs_to_abort == ["stream-req"]
output = queue.get_nowait()
assert output is not None
assert output.finished is True
assert output.outputs[0].finish_reason == str(finish_reason)
@pytest.mark.parametrize("runner", ["generate", "pooling"])
@pytest.mark.parametrize("abort_by", ["internal", "external"])
def test_abort_requests(runner: str, abort_by: str, dummy_test_vectors):
@@ -3,6 +3,7 @@
import asyncio
from collections.abc import AsyncGenerator
from unittest import mock
from unittest.mock import AsyncMock, MagicMock
import pytest
@@ -198,3 +199,52 @@ async def test_empty_input_stream_finishes_without_engine_request():
assert queue.get_nowait() is STREAM_FINISHED
llm._add_request.assert_not_awaited()
def _make_streaming_llm(num_chunks: int, has_request: bool = True):
llm = AsyncLLM.__new__(AsyncLLM)
llm._validate_streaming_input_sampling_params = MagicMock()
llm.get_supported_tasks = AsyncMock(return_value=("generate",))
llm._run_output_handler = MagicMock()
llm._add_request = AsyncMock()
llm.model_config = MagicMock()
llm.output_processor = MagicMock()
llm.output_processor.has_request.return_value = has_request
# The first processed request is the final (end-of-stream) request.
reqs = [MagicMock(prompt_embeds=None) for _ in range(num_chunks + 1)]
reqs[0].request_id = "r1-int"
llm.input_processor = MagicMock()
llm.input_processor.process_inputs.side_effect = reqs
return llm, reqs
async def _stream_chunks(llm: AsyncLLM, num_chunks: int) -> None:
async def chunks():
for i in range(num_chunks):
yield StreamingInput(prompt=f"frame-{i}")
with mock.patch(
"vllm.v1.engine.async_llm.extract_prompt_components",
return_value=("", None, None),
):
queue = await llm._add_streaming_input_request(
"r1", chunks(), SamplingParams(max_tokens=4)
)
task = queue._input_stream_task
assert task is not None
await task
@pytest.mark.asyncio
@pytest.mark.parametrize("session_live", [True, False])
async def test_handle_inputs_stops_after_session_finished(session_live: bool):
"""Once the session's request state is gone (aborted or errored), no
further chunks or final request are sent: they would recreate a request
state that is never cleaned up."""
llm, reqs = _make_streaming_llm(num_chunks=2, has_request=session_live)
await _stream_chunks(llm, num_chunks=2)
sent = [call.args[0] for call in llm._add_request.await_args_list]
# Both chunks then the final request, or only the first chunk.
assert sent == ([reqs[1], reqs[2], reqs[0]] if session_live else [reqs[1]])