[Core] Bound draft-token RPC waits by the execute-model timeout (#58779)

Signed-off-by: VS Chandra Mourya <219748331+vschandramourya@users.noreply.github.com>
Co-authored-by: VS Chandra Mourya <219748331+vschandramourya@users.noreply.github.com>
This commit is contained in:
Chandra Mourya
2026-09-26 00:40:22 +00:00
committed by GitHub
co-authored by VS Chandra Mourya
parent fdcce47e9b
commit 8a2364605c
2 changed files with 32 additions and 2 deletions
+29 -1
View File
@@ -2,12 +2,14 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import weakref
from collections import deque
from types import SimpleNamespace
from typing import Any
import pytest
from vllm.v1.executor.multiproc_executor import WorkerProc
from vllm.v1.executor.multiproc_executor import MultiprocExecutor, WorkerProc
from vllm.v1.outputs import DraftTokenIds
class _ExitWorkerLoop(RuntimeError):
@@ -65,3 +67,29 @@ def test_execute_worker_rpc_returns_worker_exception():
assert len(outputs) == 1
assert isinstance(outputs[0], RuntimeError)
assert str(outputs[0]) == "test error"
@pytest.mark.parametrize("stalled", [False, True])
def test_take_draft_token_ids_uses_execute_model_timeout(monkeypatch, stalled):
"""A wedged draft-token readback must not block EngineCore forever."""
monkeypatch.setenv("VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS", "7")
draft = DraftTokenIds(["req"], [[1, 2, 3]])
def dequeue(*, timeout=None):
assert timeout is not None and 0 < timeout <= 7
if stalled:
raise TimeoutError
return WorkerProc.ResponseStatus.SUCCESS, draft
executor: Any = MultiprocExecutor.__new__(MultiprocExecutor)
executor.is_failed = False
executor.output_rank = 1
executor.futures_queue = deque()
executor.rpc_broadcast_mq = SimpleNamespace(enqueue=lambda payload: None)
executor.response_mqs = [None, SimpleNamespace(dequeue=dequeue)]
if stalled:
with pytest.raises(TimeoutError, match="take_draft_token_ids timed out"):
executor.take_draft_token_ids()
else:
assert executor.take_draft_token_ids() is draft
+3 -1
View File
@@ -369,7 +369,9 @@ class MultiprocExecutor(Executor):
def take_draft_token_ids(self) -> DraftTokenIds | None:
# OPTIMIZATION: Get output only from a single worker (output_rank)
return self.collective_rpc(
"take_draft_token_ids", unique_reply_rank=self.output_rank
"take_draft_token_ids",
unique_reply_rank=self.output_rank,
timeout=envs.VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS,
)
def collective_rpc( # type: ignore[override]