EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[dist] fix communicator patch (#58)

Browse Source 
### What this PR does / why we need it?
fix communicator patch so parallel could work.
see #52

Signed-off-by: MengqingCao <cmq0113@163.com>
This commit is contained in:
Mengqing Cao
2025-02-14 10:45:49 +08:00
committed by GitHub
parent e264987af2
commit b88443b6c6
4 changed files with 7 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
setup.py
View File

@@ -95,7 +95,7 @@ setup(
"Topic :: Scientific/Engineering :: Artificial Intelligence", "Topic :: Scientific/Engineering :: Artificial Intelligence",
"Topic :: Scientific/Engineering :: Information Analysis", "Topic :: Scientific/Engineering :: Information Analysis",
], ],
packages=find_packages(exclude=("docs", "examples", "tests*", "patch")), packages=find_packages(exclude=("docs", "examples", "tests*")),
python_requires=">=3.9", python_requires=">=3.9",
install_requires=get_requirements(), install_requires=get_requirements(),
extras_require={}, extras_require={},

6
vllm_ascend/patch/patch_commnicator.py
View File

@@ -19,11 +19,11 @@
# https://github.com/vllm-project/vllm/pull/11324. # https://github.com/vllm-project/vllm/pull/11324.
import torch import torch
from vllm.distributed.parallel_state import GroupCoordinator import vllm
from vllm.utils import resolve_obj_by_qualname from vllm.utils import resolve_obj_by_qualname
class GroupCoordinatorPatch(GroupCoordinator): class GroupCoordinatorPatch(vllm.distributed.parallel_state.GroupCoordinator):
def __init__(self, *args, **kwargs): def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs) super().__init__(*args, **kwargs)
@@ -66,4 +66,4 @@ class GroupCoordinatorPatch(GroupCoordinator):
return self.communicator.all_gather(input_, dim) return self.communicator.all_gather(input_, dim)
GroupCoordinator = GroupCoordinatorPatch vllm.distributed.parallel_state.GroupCoordinator = GroupCoordinatorPatch

3
vllm_ascend/platform.py
View File

@@ -88,9 +88,8 @@ class NPUPlatform(Platform):
@classmethod @classmethod
def check_and_update_config(cls, vllm_config: VllmConfig) -> None: def check_and_update_config(cls, vllm_config: VllmConfig) -> None:
# Register ops and patch when setup. # Register ops when setup.
from vllm_ascend import ops # noqa: F401 from vllm_ascend import ops # noqa: F401
from vllm_ascend import patch # noqa: F401
parallel_config = vllm_config.parallel_config parallel_config = vllm_config.parallel_config
if parallel_config.worker_cls == "auto": if parallel_config.worker_cls == "auto":

2
vllm_ascend/worker.py
View File

@@ -457,6 +457,8 @@ def init_worker_distributed_environment(
backend: str = "hccl") -> None: backend: str = "hccl") -> None:
"""Initialize the distributed environment.""" """Initialize the distributed environment."""
set_custom_all_reduce(not parallel_config.disable_custom_all_reduce) set_custom_all_reduce(not parallel_config.disable_custom_all_reduce)
# register communicator patch before init dist env
from vllm_ascend import patch # noqa: F401
init_distributed_environment(parallel_config.world_size, rank, init_distributed_environment(parallel_config.world_size, rank,
distributed_init_method, local_rank, backend) distributed_init_method, local_rank, backend)