Skip to content

[Bugfix] fix xpu communicator #13368

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Feb 17, 2025
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
54 changes: 54 additions & 0 deletions vllm/distributed/device_communicators/xpu_communicator.py
Original file line number Diff line number Diff line change
@@ -0,0 +1,54 @@
# SPDX-License-Identifier: Apache-2.0

from typing import Optional

import torch
import torch.distributed as dist
from torch.distributed import ProcessGroup

from .base_device_communicator import DeviceCommunicatorBase


class XpuCommunicator(DeviceCommunicatorBase):

def __init__(self,
cpu_group: ProcessGroup,
device: Optional[torch.device] = None,
device_group: Optional[ProcessGroup] = None,
unique_name: str = ""):
super().__init__(cpu_group, device, device_group, unique_name)

def all_reduce(self, input_) -> torch.Tensor:
dist.all_reduce(input_, group=self.device_group)
return input_

def gather(self,
input_: torch.Tensor,
dst: int = 0,
dim: int = -1) -> Optional[torch.Tensor]:
assert -input_.dim() <= dim < input_.dim(), (
f"Invalid dim ({dim}) for input tensor with shape {input_.size()}")
if dim < 0:
# Convert negative dim to positive.
dim += input_.dim()
# For xpu path, gather doesn't work properly together with ray
# cluster so we use all_gather instead for now.
input_size = input_.size()
# Allocate output tensor.
output_tensor = torch.empty((self.world_size, ) + input_size,
dtype=input_.dtype,
device=input_.device)
# All-gather.
dist.all_gather_into_tensor(output_tensor,
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@youkaichao #13208 breaks XPU path since here we use all_gather_into_tensor as workaround. Please help take a look at this quick fix. Thanks.

input_,
group=self.device_group)
if self.rank_in_group == dst:
# Reshape
output_tensor = output_tensor.movedim(0, dim)
output_tensor = output_tensor.reshape(input_size[:dim] +
(self.world_size *
input_size[dim], ) +
input_size[dim + 1:])
else:
output_tensor = None
return output_tensor
4 changes: 4 additions & 0 deletions vllm/platforms/xpu.py
Original file line number Diff line number Diff line change
Expand Up @@ -135,3 +135,7 @@ def device_support_bf16(cls) -> bool:
logger.warning("Unknown device name %s, always use float16",
device_name)
return False

@classmethod
def get_device_communicator_cls(cls) -> str:
return "vllm.distributed.device_communicators.xpu_communicator.XpuCommunicator" # noqa