From 3fc28939099a3bad1bcd27e8eec20d9afe921879 Mon Sep 17 00:00:00 2001 From: crZhao Date: Mon, 17 Aug 2026 12:10:34 -0700 Subject: [PATCH] [Bugfix] Account for local DP workers in startup thread allocation (#52385) Signed-off-by: real-cpu --- tests/distributed/test_multiproc_executor.py | 47 ++++++++++++++++++++ vllm/v1/executor/multiproc_executor.py | 5 ++- 2 files changed, 51 insertions(+), 1 deletion(-) diff --git a/tests/distributed/test_multiproc_executor.py b/tests/distributed/test_multiproc_executor.py index 042349e4fb..599025d1c8 100644 --- a/tests/distributed/test_multiproc_executor.py +++ b/tests/distributed/test_multiproc_executor.py @@ -11,15 +11,62 @@ import multiprocessing import os import socket +import pytest + from tests.utils import multi_gpu_test from vllm.config import VllmConfig from vllm.engine.arg_utils import EngineArgs from vllm.v1.core.sched.output import SchedulerOutput +from vllm.v1.executor import multiproc_executor from vllm.v1.executor.multiproc_executor import MultiprocExecutor MODEL = "facebook/opt-125m" +@pytest.mark.parametrize( + ("local_world_size", "data_parallel_size_local", "expected_num_local_procs"), + [(1, 4, 4), (4, 1, 4), (2, 0, 2)], +) +def test_multiproc_executor_counts_all_local_dp_workers( + monkeypatch: pytest.MonkeyPatch, + local_world_size: int, + data_parallel_size_local: int, + expected_num_local_procs: int, +): + """All colocated DP workers share the node's startup CPU budget.""" + executor = object.__new__(MultiprocExecutor) + executor.world_size = local_world_size + executor.local_world_size = local_world_size + executor.parallel_config = type( + "ParallelConfig", + (), + {"data_parallel_size_local": data_parallel_size_local}, + )() + + monkeypatch.setattr( + executor, + "_get_parallel_sizes", + lambda: (local_world_size, 1, 1), + ) + + class StopExecutorInit(Exception): + pass + + def capture_num_local_procs(num_local_procs: int): + assert num_local_procs == expected_num_local_procs + raise StopExecutorInit + + monkeypatch.setattr( + multiproc_executor, + "set_multiprocessing_worker_envs", + capture_num_local_procs, + ) + + with pytest.raises(StopExecutorInit): + executor._init_executor() + executor._finalizer.detach() + + def create_vllm_config( tensor_parallel_size: int = 1, pipeline_parallel_size: int = 1, diff --git a/vllm/v1/executor/multiproc_executor.py b/vllm/v1/executor/multiproc_executor.py index df20a90d51..8f0f638deb 100644 --- a/vllm/v1/executor/multiproc_executor.py +++ b/vllm/v1/executor/multiproc_executor.py @@ -130,7 +130,10 @@ class MultiprocExecutor(Executor): f"_parallel_size ({pcp_size}). " ) - set_multiprocessing_worker_envs(self.local_world_size) + num_local_procs = self.local_world_size * max( + 1, self.parallel_config.data_parallel_size_local + ) + set_multiprocessing_worker_envs(num_local_procs) if aiter_requires_tcp_store(): distributed_init_method = get_distributed_init_method(