[CI/Build] Use file rendezvous for local distributed tests (#51652)

Signed-off-by: yu-xin-c <2182712990@qq.com>
This commit is contained in:
Stellar鱼
2026-08-12 10:48:12 +08:00
committed by GitHub
parent 20727e2841
commit 3fa89225bc
3 changed files with 31 additions and 12 deletions
@@ -29,7 +29,7 @@ from vllm.distributed.parallel_state import (
initialize_model_parallel,
)
from vllm.platforms import current_platform
from vllm.utils.network_utils import get_open_port
from vllm.utils.network_utils import get_file_store_init_method
from vllm.utils.system_utils import update_environment_variables
from vllm.utils.torch_utils import set_random_seed
@@ -266,7 +266,7 @@ def test_async_tp_pass_replace(
dynamic: bool,
):
num_processes = 2
master_port = str(get_open_port())
distributed_init_method = get_file_store_init_method()
def run_torch_spawn(fn, nprocs):
# need to use torch.mp.spawn otherwise will have problems with
@@ -281,7 +281,7 @@ def test_async_tp_pass_replace(
hidden_size,
dtype,
dynamic,
master_port,
distributed_init_method,
),
nprocs=nprocs,
)
@@ -314,7 +314,7 @@ def async_tp_pass_on_test_model(
hidden_size: int,
dtype: torch.dtype,
dynamic: bool,
master_port: str = "0",
distributed_init_method: str,
):
set_random_seed(0)
@@ -328,13 +328,16 @@ def async_tp_pass_on_test_model(
"RANK": str(local_rank),
"LOCAL_RANK": str(local_rank),
"WORLD_SIZE": str(world_size),
"MASTER_ADDR": "localhost",
"MASTER_PORT": master_port,
}
)
# initialize distributed
init_distributed_environment()
init_distributed_environment(
world_size=world_size,
rank=local_rank,
distributed_init_method=distributed_init_method,
local_rank=local_rank,
)
# configure vllm config for SequenceParallelismPass
vllm_config = VllmConfig()
+13 -5
View File
@@ -17,7 +17,7 @@ import torch.distributed as dist
import vllm.envs as envs
from vllm.config.parallel import ParallelConfig
from vllm.utils.network_utils import get_open_port
from vllm.utils.network_utils import get_file_store_init_method
from vllm.utils.system_utils import update_environment_variables
mp.set_start_method("spawn", force=True)
@@ -76,7 +76,7 @@ def _assert_packed_a2a_close(
def _distributed_run(fn, world_size: int, extra_env: dict[str, str]) -> None:
port = str(get_open_port())
distributed_init_method = get_file_store_init_method()
processes: list[mp.Process] = []
for rank in range(world_size):
env = {
@@ -84,8 +84,7 @@ def _distributed_run(fn, world_size: int, extra_env: dict[str, str]) -> None:
"LOCAL_RANK": str(rank),
"WORLD_SIZE": str(world_size),
"LOCAL_WORLD_SIZE": str(world_size),
"MASTER_ADDR": "localhost",
"MASTER_PORT": port,
"DISTRIBUTED_INIT_METHOD": distributed_init_method,
**extra_env,
}
process = mp.Process(target=fn, args=(env,))
@@ -483,14 +482,23 @@ class TestPackedA2AKernels:
def _distributed_packed_a2a_worker(env: dict[str, str]) -> None:
update_environment_variables(env)
local_rank = int(env["LOCAL_RANK"])
world_size = int(env["WORLD_SIZE"])
torch.accelerator.set_device_index(local_rank)
if envs.VLLM_DISTRIBUTED_USE_SPLIT_GROUP:
dist.init_process_group(
backend="cpu:gloo,cuda:nccl",
device_id=torch.device(f"cuda:{local_rank}"),
init_method=env["DISTRIBUTED_INIT_METHOD"],
rank=local_rank,
world_size=world_size,
)
else:
dist.init_process_group(backend="nccl")
dist.init_process_group(
backend="nccl",
init_method=env["DISTRIBUTED_INIT_METHOD"],
rank=local_rank,
world_size=world_size,
)
use_workspace = env.get("USE_WORKSPACE") == "1"
if use_workspace:
from vllm.v1.worker.workspace import init_workspace_manager
+8
View File
@@ -8,6 +8,7 @@ import zmq
from vllm.utils import network_utils
from vllm.utils.network_utils import (
get_file_store_init_method,
get_open_port,
get_open_ports_list,
get_tcp_uri,
@@ -19,6 +20,13 @@ from vllm.utils.network_utils import (
)
def test_get_file_store_init_method_is_unique():
init_methods = {get_file_store_init_method() for _ in range(2)}
assert len(init_methods) == 2
assert all(method.startswith("file://") for method in init_methods)
def _call_with_timeout(func, timeout: float = 10.0):
"""Run func in a daemon thread so a livelock regression fails the test
quickly instead of hanging the suite."""