[Rebase] Rebase to vllm 0.26.0 (#5443)

Signed-off-by: tzhouam <tzhouam@connect.ust.hk>
Signed-off-by: tzhouam <tzhouam@gmail.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Co-authored-by: tzhouam <tzhouam@gmail.com>
This commit is contained in:
authored and GitHub committed 2026-07-27 23:35:49 +08:00
1 parent fb89ab43ea
commit cbab1d4d8f
41 files changed
+334 -183

No files matched your search

@@ -73,6 +73,9 @@ def _run_resumable_segment_stop(
sched._update_request_with_output.side_effect = stop_request
sched._handle_stopped_request.return_value = session_finished
# vLLM 0.26 returns (kv_xfer_params, ec_xfer_params); an unconfigured
# MagicMock iterates empty and fails to unpack at the call site.
sched._free_request.return_value = (None, None)
sched.chunk_transfer_adapter = None
sched.running = [session]
sched.waiting_for_transfer_free = set()
@@ -137,6 +140,23 @@ def test_resumable_session_terminal_is_not_marked_as_segment_boundary() -> None:
assert output.is_segment_finished is False
def test_update_from_output_settles_in_flight_tokens() -> None:
"""vLLM 0.26: schedule() increments num_in_flight_tokens per scheduled
token; update_from_output must decrement it symmetrically. If the
decrement is dropped the counter grows monotonically and both readers
(allocate_slots, _connector_finished) clamp
max(0, num_computed_tokens - num_in_flight_tokens) to zero forever,
silently freezing sliding-window block freeing.
"""
session = _make_request()
session.status = RequestStatus.RUNNING
session.num_in_flight_tokens = 1 # as left by schedule() for this step
_run_resumable_segment_stop(session)
assert session.num_in_flight_tokens == 0
def test_running_decode_step_without_inter_stage_payload_does_not_raise() -> None:
"""A decode step that neither stops nor carries an inter-stage payload.
@@ -76,8 +76,9 @@ def test_finish_requests_purges_finished_request_from_running(
)
def fake_finish_requests(self, request_ids, finished_status):
# super() returning the abort tuple but leaving self.running untouched
return [(finished.request_id, 0)]
# vLLM 0.26 super() returns the finished Request objects but here
# leaves self.running untouched to simulate the missed removal.
return [finished]
monkeypatch.setattr(scheduler_mod.VLLMScheduler, "finish_requests", fake_finish_requests)
@@ -88,7 +89,7 @@ def test_finish_requests_purges_finished_request_from_running(
)
assert scheduler.running == []
assert result == [(finished.request_id, 0)]
assert result == [finished]
@pytest.mark.parametrize(("scheduler_cls", "scheduler_mod"), _SCHEDULER_PARAMS)
@@ -111,7 +112,7 @@ def test_finish_requests_purges_untracked_request_from_running(
)
def fake_finish_requests(self, request_ids, finished_status):
return [(untracked.request_id, 0)]
return [untracked]
monkeypatch.setattr(scheduler_mod.VLLMScheduler, "finish_requests", fake_finish_requests)
@@ -122,7 +123,7 @@ def test_finish_requests_purges_untracked_request_from_running(
)
assert scheduler.running == []
assert result == [(untracked.request_id, 0)]
assert result == [untracked]
@pytest.mark.parametrize(("scheduler_cls", "scheduler_mod"), _SCHEDULER_PARAMS)
@@ -136,7 +137,7 @@ def test_finish_requests_leaves_healthy_running_intact(
is conservative and won't aggressively drop healthy entries.
"""
alive = _StubRequest("req-alive", RequestStatus.RUNNING)
leaving_id = "req-leaving"
leaving = _StubRequest("req-leaving", RequestStatus.FINISHED_STOPPED)
scheduler = _make_scheduler(
scheduler_cls,
requests={alive.request_id: alive},
@@ -145,13 +146,13 @@ def test_finish_requests_leaves_healthy_running_intact(
)
def fake_finish_requests(self, request_ids, finished_status):
return [(leaving_id, 0)]
return [leaving]
monkeypatch.setattr(scheduler_mod.VLLMScheduler, "finish_requests", fake_finish_requests)
scheduler_cls.finish_requests(
scheduler,
[leaving_id],
[leaving.request_id],
RequestStatus.FINISHED_STOPPED,
)
@@ -53,7 +53,11 @@ def test_finish_requests_cleans_input_coordinator_for_finished_ids(
def fake_finish_requests(self, request_ids, finished_status):
assert request_ids == ["req-a", "req-b"]
assert finished_status == RequestStatus.FINISHED_STOPPED
return [("req-a", 0), ("req-b", 1)]
# vLLM 0.26 returns the finished Request objects, not (req_id, client_index) tuples.
return [
SimpleNamespace(request_id="req-a", client_index=0),
SimpleNamespace(request_id="req-b", client_index=1),
]
monkeypatch.setattr(scheduler_mod.VLLMScheduler, "finish_requests", fake_finish_requests)
@@ -63,7 +67,7 @@ def test_finish_requests_cleans_input_coordinator_for_finished_ids(
RequestStatus.FINISHED_STOPPED,
)
assert result == [("req-a", 0), ("req-b", 1)]
assert [(r.request_id, r.client_index) for r in result] == [("req-a", 0), ("req-b", 1)]
assert coordinator.freed == ["req-a", "req-b"]
@@ -82,7 +86,7 @@ def test_ar_free_request_cleans_input_coordinator_on_normal_free() -> None:
result = OmniARScheduler._free_request(scheduler, DummyRequest())
assert result is None
assert result == (None, None)
assert coordinator.freed == ["req-free"]
assert scheduler._omits_kv_transfer_cache == {}
assert scheduler._new_prompt_len_snapshot == {}