[BugFix][TTS] R1.1-R1.4: turn silent async-chunk failures into visible ones (#6033)
Signed-off-by: Yueqian Lin <linyueqian@outlook.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>
This commit is contained in:
1 parent
baba7d1e2e
commit
611c083587
10 files changed
+750
-32
No files matched your search
@@ -411,6 +411,51 @@ async def test_async_chunk_prewarm_failure_reports_failing_downstream_stage(orch
|
||||
orchestrator_fixture.thread.join(timeout=5)
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_async_chunk_prewarm_without_prompt_token_ids_fails_only_that_request(
|
||||
orchestrator_factory,
|
||||
) -> None:
|
||||
"""R1.3 of #4855: an async-chunk request whose stage-0 prompt carries no
|
||||
token ids cannot prewarm anything, so it must fail with a client error
|
||||
instead of stalling.
|
||||
|
||||
Non-fatal is the point: ``fatal=True`` reads as "the engine died" and the
|
||||
entrypoints act on it (``OmniLLM.generate`` raises out of its batch loop,
|
||||
the serving layer calls ``terminate_if_errored``). One client's embeds-only
|
||||
prompt must not reach other requests.
|
||||
"""
|
||||
stage0 = FakeStageClient(stage_type="llm", final_output=False)
|
||||
stage1 = FakeStageClient(stage_type="llm", final_output=True)
|
||||
orchestrator_fixture = orchestrator_factory([stage0, stage1], async_chunk=True)
|
||||
|
||||
try:
|
||||
await _enqueue_add_request(
|
||||
orchestrator_fixture,
|
||||
request_id="req-embeds",
|
||||
# No prompt_token_ids: what an embeds-only prompt looks like here.
|
||||
prompt=SimpleNamespace(request_id="req-embeds"),
|
||||
original_prompt={"prompt": "embeds only"},
|
||||
sampling_params_list=[_sampling_params(), _sampling_params()],
|
||||
final_stage_id=1,
|
||||
)
|
||||
|
||||
error_msg = await _wait_for_error_message(orchestrator_fixture, request_id="req-embeds")
|
||||
assert error_msg.fatal is False
|
||||
assert error_msg.status_code == 400
|
||||
assert error_msg.error_type == "BadRequestError"
|
||||
assert error_msg.stage_id == 0
|
||||
assert "prompt_token_ids" in error_msg.error
|
||||
|
||||
# Nothing was prewarmed downstream, and the request is gone.
|
||||
assert stage1.add_request_calls == []
|
||||
await _wait_for(lambda: "req-embeds" not in orchestrator_fixture.orchestrator.request_states)
|
||||
assert orchestrator_fixture.thread.is_alive()
|
||||
finally:
|
||||
if orchestrator_fixture.thread.is_alive():
|
||||
orchestrator_fixture.request_sync_q.put_nowait(ShutdownRequestMessage())
|
||||
orchestrator_fixture.thread.join(timeout=5)
|
||||
|
||||
|
||||
# ───────── Direct unit tests for the fault-isolation helpers ─────────
|
||||
|
||||
|
||||
|
||||
@@ -256,12 +256,13 @@ async def test_async_prewarm_skips_outgoing_only_stage() -> None:
|
||||
duplex_identity=SimpleNamespace(),
|
||||
)
|
||||
|
||||
await orchestrator._prewarm_async_chunk_stages(
|
||||
prewarmed = await orchestrator._prewarm_async_chunk_stages(
|
||||
"req-prewarm",
|
||||
SimpleNamespace(prompt_token_ids=[1, 2], resumable=True),
|
||||
req_state,
|
||||
)
|
||||
|
||||
assert prewarmed is True
|
||||
assert stage1.submitted == []
|
||||
assert len(stage2.submitted) == 1
|
||||
assert 1 not in req_state.stage_submit_ts
|
||||
@@ -277,6 +278,7 @@ async def test_async_prewarm_skips_outgoing_only_stage() -> None:
|
||||
@pytest.mark.asyncio
|
||||
async def test_duplex_prewarm_runs_after_first_stage0_submission() -> None:
|
||||
port, stage_pools, request_states, prewarm, submission = _duplex_stage_port_submission()
|
||||
prewarm.return_value = True
|
||||
|
||||
result = await port.submit(submission)
|
||||
|
||||
@@ -285,6 +287,30 @@ async def test_duplex_prewarm_runs_after_first_stage0_submission() -> None:
|
||||
prewarm.assert_awaited_once_with("req-duplex", ANY, request_states["req-duplex"])
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_duplex_submit_bails_out_when_prewarm_failed_the_request() -> None:
|
||||
"""A failed prewarm has already aborted the request and popped its state.
|
||||
|
||||
Returning a success result here would hand the control plane a replica for a
|
||||
request that no longer exists, and the trailing bookkeeping would re-register
|
||||
a running counter the cleanup just released -- a leak that never decrements.
|
||||
"""
|
||||
port, stage_pools, request_states, prewarm, submission = _duplex_stage_port_submission()
|
||||
counter = MagicMock()
|
||||
port._running_counter = counter
|
||||
prewarm.return_value = False
|
||||
request_state = request_states["req-duplex"]
|
||||
|
||||
with pytest.raises(RuntimeError, match="prewarm failed"):
|
||||
await port.submit(submission)
|
||||
|
||||
prewarm.assert_awaited_once()
|
||||
assert request_state.duplex_stage_fences == {}
|
||||
assert request_state.stage_submit_ts == {}
|
||||
assert request_state.running_counter_registered is False
|
||||
counter.increment.assert_not_called()
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_async_route_forwards_to_outgoing_only_stage() -> None:
|
||||
orchestrator = object.__new__(Orchestrator)
|
||||
|
||||
Reference in new issue
Block a user