[Bugfix] Account for local DP workers in startup thread allocation (#52385)
Signed-off-by: real-cpu <zhaochenrui757@gmail.com>
This commit is contained in:
@@ -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,
|
||||
|
||||
Reference in New Issue
Block a user