modelscope / modelscope/FunASR
[安装] 昇腾 NPU 环境下 pip install funasr 可能替换 torch 导致 torch_npu 失效,建议提供对 NPU 友好的安装方式
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 20.4k
- Forks
- 2k
- Avg merge
- 4h 55m
- Merged PRs (30d)
- 169
Description
环境
- funasr:1.3.14
- torch:2.10.0+cpu / torch_npu:2.10.0(昇腾 Atlas 300I Pro / 310P3,CANN 9.1.0-beta.1)
问题描述
昇腾 NPU 环境要求 torch 使用 CPU 版本,并与 torch_npu 严格配对
(如 torch 2.10.0+cpu + torch_npu 2.10.0)。默认依赖解析安装 funasr 时,
可能拉取/升级通用(CUDA)版 torch,静默破坏这一配对——之后 import torch_npu
会因 C++ 扩展不匹配而失败。
我们目前的规避方式(openEuler 24.03,容器内):
pip install funasr --no-deps -i <国内镜像>
pip install torch_complex kaldiio omegaconf librosa kaldi-native-fbank
editdistance jieba zhconv tgt umap-learn ... # 手动逐个补齐
这种做法容易出错:不少运行时依赖(kaldiio、omegaconf、torch_complex、
umap-learn、praat-parselmouth 等)很容易遗漏,只在 import 时才报
ModuleNotFoundError。
建议(任一均可)
- 提供依赖分组,如 pip install "funasr[npu]"(torch 视为环境自带,不强制版本);
- 或将依赖声明放宽到"已安装的 torch 不被替换"(避免触发升级的上限约束);
- 或补充文档《在昇腾 NPU 上安装》,列出 --no-deps 方式所需的完整运行时依赖清单。
如有需要,我们愿意贡献在 310P3 上实测验证过的完整依赖清单。
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
Start by inspecting FunASR's declared runtime dependencies and existing installation documentation. Reproduce installation with a preinstalled CPU torch and torch_npu pair, then verify that the pair is not replaced and all runtime imports work. Done means an NPU-friendly installation path or dependency group is documented and validated, including the dependencies currently listed in the issue.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python, pytorch
- Domain
- tooling
- Issue type
- Feature
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Active
- Clarity
- Mostly clear
- Newbie friendliness
- 55/100