xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-05-16 04:02:18 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Misc]allow disable pynccl (#25421)

Browse Source 
Signed-off-by: Lu Fang <fanglu@fb.com>
Co-authored-by: Lucia (Lu) Fang <fanglu@meta.com>
This commit is contained in:
Lucia Fang 2025-09-30 23:04:13 -07:00 committed by GitHub
parent 2a69ab4899
commit f48b6a03ba
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
3 changed files with 12 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
vllm/distributed/device_communicators/cuda_communicator.py
View File

@ -147,6 +147,10 @@ class CudaCommunicator(DeviceCommunicatorBase):
assert out is not None assert out is not None
return out return out
pynccl_comm = self.pynccl_comm pynccl_comm = self.pynccl_comm
if pynccl_comm is None or pynccl_comm.disabled:
out = input_.clone()
torch.distributed.all_reduce(out, group=self.device_group)
return out
assert pynccl_comm is not None assert pynccl_comm is not None
out = pynccl_comm.all_reduce(input_) out = pynccl_comm.all_reduce(input_)
if out is None: if out is None:

3
vllm/distributed/device_communicators/pynccl.py
View File

@ -8,6 +8,7 @@ import torch
import torch.distributed as dist import torch.distributed as dist
from torch.distributed import ProcessGroup, ReduceOp from torch.distributed import ProcessGroup, ReduceOp
import vllm.envs as envs
from vllm.distributed.device_communicators.pynccl_wrapper import ( from vllm.distributed.device_communicators.pynccl_wrapper import (
NCCLLibrary, buffer_type, cudaStream_t, ncclComm_t, ncclDataTypeEnum, NCCLLibrary, buffer_type, cudaStream_t, ncclComm_t, ncclDataTypeEnum,
ncclRedOpTypeEnum, ncclUniqueId) ncclRedOpTypeEnum, ncclUniqueId)
@ -83,7 +84,7 @@ class PyNcclCommunicator:
self.group = group self.group = group
# if world_size == 1, no need to create communicator # if world_size == 1, no need to create communicator
if self.world_size == 1: if self.world_size == 1 or envs.VLLM_DISABLE_PYNCCL:
self.available = False self.available = False
self.disabled = True self.disabled = True
return return

6
vllm/envs.py
View File

@ -98,6 +98,7 @@ if TYPE_CHECKING:
VLLM_SKIP_P2P_CHECK: bool = False VLLM_SKIP_P2P_CHECK: bool = False
VLLM_DISABLED_KERNELS: list[str] = [] VLLM_DISABLED_KERNELS: list[str] = []
VLLM_DISABLE_NCCL_FOR_DP_SYNCHRONIZATION: bool = False VLLM_DISABLE_NCCL_FOR_DP_SYNCHRONIZATION: bool = False
VLLM_DISABLE_PYNCCL: bool = False
VLLM_USE_V1: bool = True VLLM_USE_V1: bool = True
VLLM_ROCM_USE_AITER: bool = False VLLM_ROCM_USE_AITER: bool = False
VLLM_ROCM_USE_AITER_PAGED_ATTN: bool = False VLLM_ROCM_USE_AITER_PAGED_ATTN: bool = False
@ -897,6 +898,11 @@ environment_variables: dict[str, Callable[[], Any]] = {
(os.getenv("VLLM_DISABLE_NCCL_FOR_DP_SYNCHRONIZATION", "False").lower() in (os.getenv("VLLM_DISABLE_NCCL_FOR_DP_SYNCHRONIZATION", "False").lower() in
("true", "1")), ("true", "1")),
# Disable pynccl (using torch.distributed instead)
"VLLM_DISABLE_PYNCCL":
lambda:
(os.getenv("VLLM_DISABLE_PYNCCL", "False").lower() in ("true", "1")),
# If set, use the V1 code path. # If set, use the V1 code path.
"VLLM_USE_V1": "VLLM_USE_V1":
lambda: bool(int(os.getenv("VLLM_USE_V1", "1"))), lambda: bool(int(os.getenv("VLLM_USE_V1", "1"))),