Linux aarch64架构下scikit-learn导入失败问题求助

问题现象 #

在Linux aarch64架构的miniconda3虚拟环境中，通过pip install scikit-learn安装scikit-learn后，导入librosa时触发如下错误：

>>> import librosa
Traceback (most recent call last):
  File "/home/su/miniconda3/envs/acousticpdev/lib/python3.10/site-packages/sklearn/__check_build/__init__.py", line 48, in <module>
    from ._check_build import check_build  # noqa
ImportError: /home/su/miniconda3/envs/acousticpdev/lib/python3.10/site-packages/sklearn/__check_build/../../scikit_learn.libs/libgomp-d22c30c5.so.1.0.0: cannot allocate memory in static TLS block

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/su/miniconda3/envs/acousticpdev/lib/python3.10/site-packages/librosa/__init__.py", line 211, in <module>
    from . import decompose
  File "/home/su/miniconda3/envs/acousticpdev/lib/python3.10/site-packages/librosa/decompose.py", line 19, in <module>
    import sklearn.decomposition
  File "/home/su/miniconda3/envs/acousticpdev/lib/python3.10/site-packages/sklearn/__init__.py", line 81, in <module>
    from . import __check_build  # noqa: F401
  File "/home/su/miniconda3/envs/acousticpdev/lib/python3.10/site-packages/sklearn/__check_build/__init__.py", line 50, in <module>
    raise_build_error(e)
  File "/home/su/miniconda3/envs/acousticpdev/lib/python3.10/site-packages/sklearn/__check_build/__init__.py", line 31, in raise_build_error
    raise ImportError(
ImportError: /home/su/miniconda3/envs/acousticpdev/lib/python3.10/site-packages/sklearn/__check_build/../../scikit_learn.libs/libgomp-d22c30c5.so.1.0.0: cannot allocate memory in static TLS block
___________________________________________________________________________
Contents of /home/su/miniconda3/envs/acousticpdev/lib/python3.10/site-packages/sklearn/__check_build:
__pycache__               __init__.py               setup.py
_check_build.cpython-310-aarch64-linux-gnu.so
___________________________________________________________________________
It seems that scikit-learn has not been built correctly.

If you have installed scikit-learn from source, please do not forget
to build the package before using it: run `python setup.py install` or
`make` in the source directory.

If you have used an installer, please check that it is suited for your
Python version, your operating system and your platform.

尝试降级scikit-image到0.17.2后，立即导入librosa和scikit-learn正常，但退出Python后再次导入又出现相同错误。

环境背景 #

• 系统架构：Linux aarch64
• 虚拟环境：miniconda3（Python 3.10）
• 已安装关键包版本：scikit-learn 1.1.2、librosa 0.9.2、scikit-image 0.19.3
• 曾用pyenv创建过非conda环境，相关配置已写入~/.bashrc

解决方案 #

1. 改用conda安装scikit-learn（优先推荐） #

conda针对aarch64架构提供了预编译的优化包，能避免pip编译时的TLS参数问题：

# 卸载pip安装的scikit-learn
pip uninstall -y scikit-learn
# 用conda安装对应版本
conda install scikit-learn=1.1.2

2. 调整静态TLS内存分配环境变量 #

如果必须使用pip安装的scikit-learn，可通过环境变量调整TLS内存分配：

# 临时生效：启动Python前执行该命令
export LD_PRELOAD=/home/su/miniconda3/envs/acousticpdev/lib/python3.10/site-packages/scikit_learn.libs/libgomp-d22c30c5.so.1.0.0

# 永久生效：将命令添加到conda环境的激活脚本中
echo 'export LD_PRELOAD=/home/su/miniconda3/envs/acousticpdev/lib/python3.10/site-packages/scikit_learn.libs/libgomp-d22c30c5.so.1.0.0' >> /home/su/miniconda3/envs/acousticpdev/etc/conda/activate.d/env_vars.sh

或者调整TLS预留大小：

export MALLOC_TLS_PAD_OLD=0x4000

3. 隔离conda环境与pyenv的干扰 #

pyenv的PATH配置可能导致conda环境加载系统级的libgomp库，引发冲突：

# 编辑~/.bashrc，修改pyenv初始化逻辑，仅在非conda环境加载
sed -i 's/eval "$(pyenv init -)"/if [ -z "$CONDA_PREFIX" ]; then eval "$(pyenv init -)"; fi/' ~/.bashrc
# 重新加载配置
source ~/.bashrc
# 重新激活conda环境
conda activate acousticpdev

4. 重建干净的conda环境 #

若以上方法无效，彻底清理旧环境，新建纯净环境避免依赖冲突：

conda deactivate
conda env remove -n acousticpdev
# 创建新环境
conda create -n acousticpdev python=3.10
conda activate acousticpdev
# 优先安装conda维护的包，再用pip补充
conda install scikit-learn librosa numpy scipy
# 按需安装其他pip包
pip install tensorflow keras

问题原因 #

该错误是aarch64架构下，libgomp（OpenMP库）的静态TLS（线程本地存储）内存块分配不足导致的。pip安装的scikit-learn可能使用了未针对aarch64优化的编译参数，而conda预编译包则调整了相关参数；同时pyenv的环境变量可能导致conda环境加载了系统级的libgomp，引发库冲突。

内容的提问来源于stack exchange，提问作者Susan