mirror of
https://github.com/vllm-project/vllm-ascend.git
synced 2025-10-20 21:53:54 +08:00
c3d1a3782a
This is the first step to support trl vllm serve on Ascend NPU https://github.com/vllm-project/vllm-ascend/issues/459. This PR can work properly only when https://github.com/vllm-project/vllm/pull/16464 is merged into vLLM. --------- Signed-off-by: hzji210@gmail.com <hzji210@gmail.com>
86 lines
2.8 KiB
Python
86 lines
2.8 KiB
Python
#
|
|
# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
|
|
# This file is a part of the vllm-ascend project.
|
|
# Adapted from vllm-project/vllm/vllm/worker/worker.py
|
|
# Copyright 2023 The vLLM team.
|
|
#
|
|
# 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.
|
|
#
|
|
import torch
|
|
import torch_npu # noqa: F401
|
|
from vllm.logger import logger
|
|
|
|
import vllm_ascend.envs as envs
|
|
|
|
|
|
def try_register_lib(lib_name: str, lib_info: str = ""):
|
|
import importlib
|
|
import importlib.util
|
|
try:
|
|
module_spec = importlib.util.find_spec(lib_name)
|
|
if module_spec is not None:
|
|
importlib.import_module(lib_name)
|
|
if lib_info:
|
|
logger.info(lib_info)
|
|
except Exception:
|
|
pass
|
|
|
|
|
|
def find_hccl_library() -> str:
|
|
"""
|
|
We either use the library file specified by the `HCCL_SO_PATH`
|
|
environment variable, or we find the library file brought by PyTorch.
|
|
After importing `torch`, `libhccl.so` can be
|
|
found by `ctypes` automatically.
|
|
"""
|
|
so_file = envs.HCCL_SO_PATH
|
|
|
|
# manually load the hccl library
|
|
if so_file:
|
|
logger.info("Found hccl from environment variable HCCL_SO_PATH=%s",
|
|
so_file)
|
|
else:
|
|
if torch.version.cann is not None:
|
|
so_file = "libhccl.so"
|
|
else:
|
|
raise ValueError("HCCL only supports Ascend NPU backends.")
|
|
logger.info("Found hccl from library %s", so_file)
|
|
return so_file
|
|
|
|
|
|
_current_stream = None
|
|
|
|
|
|
def current_stream() -> torch.npu.Stream:
|
|
"""
|
|
replace `torch.npu.current_stream()` with `vllm.utils.current_stream()`.
|
|
it turns out that `torch.npu.current_stream()` is quite expensive,
|
|
as it will construct a new stream object at each call.
|
|
here we patch `torch.npu.set_stream` to keep track of the current stream
|
|
directly, so that we can avoid calling `torch.npu.current_stream()`.
|
|
|
|
"""
|
|
global _current_stream
|
|
if _current_stream is None:
|
|
# when this function is called before any stream is set,
|
|
# we return the default stream.
|
|
_current_stream = torch.npu.current_stream()
|
|
return _current_stream
|
|
|
|
|
|
def adapt_patch(is_global_patch: bool = False):
|
|
if is_global_patch:
|
|
from vllm_ascend.patch import platform # noqa: F401
|
|
else:
|
|
from vllm_ascend.patch import worker # noqa: F401
|