use split_group for pytorch process group creation (#41980)

Signed-off-by: Tushar Jain <tushar00jain@users.noreply.github.com>
Co-authored-by: Tushar Jain <tushar00jain@users.noreply.github.com>
This commit is contained in:
Tushar Jain
2026-06-04 14:36:07 -04:00
committed by GitHub
co-authored by Tushar Jain
parent a947f7a420
commit 38fd2405f3
11 changed files with 528 additions and 51 deletions
@@ -10,6 +10,7 @@ import torch
from torch.multiprocessing import spawn # pyright: ignore[reportPrivateImportUsage]
from typing_extensions import ParamSpec
import vllm.envs as envs
from vllm.config import VllmConfig, set_current_vllm_config
from vllm.distributed import (
cleanup_dist_env_and_memory,
@@ -60,7 +61,15 @@ def _set_vllm_config(
tensor_model_parallel_size=vllm_config.parallel_config.tensor_parallel_size,
pipeline_model_parallel_size=vllm_config.parallel_config.pipeline_parallel_size,
)
cpu_group = torch.distributed.new_group(list(range(world_size)), backend="gloo")
if envs.VLLM_DISTRIBUTED_USE_SPLIT_GROUP:
cpu_group = torch.distributed.split_group(
split_ranks=[list(range(world_size))],
group_desc="moe_test_cpu",
)
else:
cpu_group = torch.distributed.new_group(
list(range(world_size)), backend="gloo"
)
return cpu_group