无CUDA版PyTorch运行项目遇mkl-service及numpy DLL加载错误求助

数据老陈

2026-7-15

解决无CUDA版PyTorch环境下numpy DLL加载失败及mkl-service导入警告问题

报错信息

D:\software\anaconda\envs\pytorch\lib\site-packages\numpy_init_.py:138: UserWarning: mkl-service package failed to import, therefore Intel(R) MKL initialization ensuring its correct out-of-the box operation under condition when Gnu OpenMP had already been loaded by Python process is not assured. Please install mkl-service package...
Traceback (most recent call last):
File "D:\software\anaconda\envs\pytorch\lib\site-packages\numpy\core_init_.py", line 22, in
from . import multiarray
File "D:\software\anaconda\envs\pytorch\lib\site-packages\numpy\core\multiarray.py", line 12, in
from . import overrides
File "D:\software\anaconda\envs\pytorch\lib\site-packages\numpy\core\overrides.py", line 7, in
from numpy.core._multiarray_umath import (
ImportError: DLL load failed: 找不到指定的模块。

解决方案

彻底重装numpy与mkl-service（推荐）
基于Anaconda环境的依赖特性，用conda执行以下操作：

# 强制卸载现有冲突包
conda remove --force numpy mkl-service
# 安装与PyTorch兼容的指定版本numpy（以1.21.x为例，可根据PyTorch版本微调）
conda install numpy=1.21.* mkl-service

配置系统环境变量
将Anaconda虚拟环境的Library\bin路径（如D:\software\anaconda\envs\pytorch\Library\bin）添加到系统环境变量Path的最前端，确保系统优先加载当前环境的DLL文件。
切换至OpenBLAS版本numpy
若MKL相关问题持续无法解决，可安装不带MKL依赖的numpy：
```
conda install numpy nomkl
```
验证修复结果
打开Python终端，执行以下命令确认无报错：
```
import numpy
import torch
print(numpy.__version__, torch.__version__)
```