af04ee9e7a
### What this PR does / why we need it?
Fix Qwen MoE accuracy bug in DP scenario.
Now the implentment of `FusedMoE` in vLLM use `All2AllManager` to
manager different all2all algorithm branch. And the default branch use
`Multicast` in `dispatch` phase and `all_reduce` in `combine` phase,
which are not implented in vLLM-Ascend. This leading to invoking into a
default implentment in `base_communicator`, with empty `dispatch` and
`combine` operations, thus causing the accuracy issue on it.
This pr is a temporary workaround, refacting all2all in vLLM-Ascend
could be a better way.
- vLLM version: v0.10.0
- vLLM main:
ad57f23f6a
---------
Signed-off-by: MengqingCao <cmq0113@163.com>
97 lines
3.8 KiB
Python
97 lines
3.8 KiB
Python
#
|
|
# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
|
|
#
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
# you may not use this file except in compliance with the License.
|
|
# You may obtain a copy of the License at
|
|
#
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
#
|
|
# Unless required by applicable law or agreed to in writing, software
|
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
# See the License for the specific language governing permissions and
|
|
# limitations under the License.
|
|
# This file is a part of the vllm-ascend project.
|
|
#
|
|
from typing import List, Optional
|
|
|
|
import torch
|
|
import torch.distributed as dist
|
|
from vllm.distributed.device_communicators.base_device_communicator import \
|
|
DeviceCommunicatorBase
|
|
from vllm.utils import logger
|
|
|
|
|
|
class NPUCommunicator(DeviceCommunicatorBase):
|
|
|
|
def __init__(self,
|
|
cpu_group: dist.ProcessGroup,
|
|
device: Optional[torch.device] = None,
|
|
device_group: Optional[dist.ProcessGroup] = None,
|
|
unique_name: str = ""):
|
|
super().__init__(cpu_group, device, device_group, unique_name)
|
|
# TODO(hz): Refer to CudaCommunicator's implementation to integrate PyHcclCommunicator
|
|
# init device according to rank
|
|
self.device = torch.npu.current_device()
|
|
|
|
if self.use_all2all:
|
|
from vllm.distributed.device_communicators.all2all import \
|
|
NaiveAll2AllManager
|
|
self.all2all_manager = NaiveAll2AllManager(self.cpu_group)
|
|
logger.info("Using naive all2all manager.")
|
|
|
|
def all_to_all(self,
|
|
input_: torch.Tensor,
|
|
scatter_dim: int = 0,
|
|
gather_dim: int = -1,
|
|
scatter_sizes: Optional[List[int]] = None,
|
|
gather_sizes: Optional[List[int]] = None) -> torch.Tensor:
|
|
|
|
if scatter_dim < 0:
|
|
scatter_dim += input_.dim()
|
|
if gather_dim < 0:
|
|
gather_dim += input_.dim()
|
|
|
|
if scatter_sizes is not None and gather_sizes is not None:
|
|
input_list = [
|
|
t.contiguous()
|
|
for t in torch.split(input_, scatter_sizes, scatter_dim)
|
|
]
|
|
output_list = []
|
|
tensor_shape_base = input_list[self.rank].size()
|
|
for i in range(self.world_size):
|
|
tensor_shape = list(tensor_shape_base)
|
|
tensor_shape[gather_dim] = gather_sizes[i]
|
|
output_list.append(
|
|
torch.empty(tensor_shape,
|
|
dtype=input_.dtype,
|
|
device=input_.device))
|
|
|
|
else:
|
|
input_list = [
|
|
t.contiguous() for t in torch.tensor_split(
|
|
input_, self.world_size, scatter_dim)
|
|
]
|
|
output_list = [
|
|
torch.empty_like(input_list[i]) for i in range(self.world_size)
|
|
]
|
|
|
|
dist.all_to_all(output_list, input_list, group=self.device_group)
|
|
output_tensor = torch.cat(output_list, dim=gather_dim).contiguous()
|
|
return output_tensor
|
|
|
|
# TODO: Add ut for dispatch and combine
|
|
def dispatch(
|
|
self, hidden_states: torch.Tensor,
|
|
router_logits: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
|
|
assert self.all2all_manager is not None
|
|
hidden_states, router_logits = self.all2all_manager.dispatch(
|
|
hidden_states, router_logits)
|
|
return hidden_states, router_logits
|
|
|
|
def combine(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
|
assert self.all2all_manager is not None
|
|
hidden_states = self.all2all_manager.combine(hidden_states)
|
|
return hidden_states
|