[BugFix] Use file:// rendezvous for single-node executors to eliminate startup port races (#50999)

Signed-off-by: Parsa Mazaheri <p_mazaheri@apple.com>
Signed-off-by: aoshen02 <aoshen02@users.noreply.github.com>
Signed-off-by: aoshen02 <aoshen@inferact.ai>
Co-authored-by: Parsa Mazaheri <p_mazaheri@apple.com>
Co-authored-by: aoshen02 <aoshen02@users.noreply.github.com>
Co-authored-by: Claude <noreply@anthropic.com>
This commit is contained in:
aoshen02
2026-08-09 22:00:04 -04:00
committed by GitHub
co-authored by Parsa Mazaheri aoshen02 Claude
parent 0820125ae9
commit 11ba93f364
6 changed files with 142 additions and 13 deletions
+2 -2
View File
@@ -12,7 +12,7 @@ import torch.distributed as dist
import vllm.envs as envs
from vllm.logger import init_logger
from vllm.platforms import current_platform
from vllm.utils.network_utils import get_distributed_init_method, get_ip, get_open_port
from vllm.utils.network_utils import get_file_store_init_method
from vllm.v1.core.sched.output import GrammarOutput, SchedulerOutput
from vllm.v1.executor.abstract import Executor
from vllm.v1.executor.vllm_net_devices import set_worker_net_device
@@ -70,7 +70,7 @@ class UniProcExecutor(Executor):
def _distributed_args(self) -> tuple[str, int, int]:
"""Return (distributed_init_method, rank, local_rank)."""
distributed_init_method = get_distributed_init_method(get_ip(), get_open_port())
distributed_init_method = get_file_store_init_method()
# set local rank as the device index if specified
device_info = self.vllm_config.device_config.device.__str__().split(":")
local_rank = int(device_info[1]) if len(device_info) > 1 else 0