[Core][Distributed] use cpu/gloo to initialize pynccl (#4248)

This commit is contained in:
youkaichao
2024-04-23 18:32:19 -07:00
committed by GitHub
parent 79a268c4ab
commit 91f50a6fe2
5 changed files with 93 additions and 71 deletions
@@ -2,7 +2,7 @@ import contextlib
from typing import Optional
import torch
from torch.distributed import ReduceOp
from torch.distributed import ProcessGroup, ReduceOp
from vllm.logger import init_logger
@@ -37,17 +37,11 @@ def set_pynccl_stream(stream: torch.cuda.Stream):
pass
def init_process_group(world_size: int,
rank: int,
init_method: str,
local_rank: int = -1) -> None:
def init_process_group(group: Optional[ProcessGroup] = None) -> None:
assert not is_initialized()
global comm
logger.info(f"vLLM is using nccl=={ncclGetVersion()}")
comm = NCCLCommunicator(init_method=init_method,
world_size=world_size,
local_rank=local_rank,
rank=rank)
comm = NCCLCommunicator(group=group)
def all_reduce(input_: torch.Tensor, op=ReduceOp.SUM) -> None: