[安装] 在昇腾 NPU 环境下,pip install funasr 可能会替换掉 torch,导致 torch_npu 失效,建议提供对 NPU 友好的安装方式
作者: Liuzi0816创建于 2026年9月15日更新于 2026年9月16日
Environment
- funasr: 1.3.14
- torch: 2.10.0 + cpu / torch_npu: 2.10.0 (Ascend NPU Atlas 300I Pro / 310P3, CANN 9.1.0-beta.1)
Problem Description
The Ascend NPU environment requires torch to be used with the CPU version and must be strictly paired with torch_npu (such as torch 2.10.0 + cpu + torch_npu 2.10.0). By default, when installing funasr with no dependencies, it may pull or upgrade the general (CUDA) version of torch, silently breaking this pairing — subsequent import of torch_npu will fail due to C++ extension mismatch.
内容来源: modelscope/FunASR