Jupyter Notebook导入PyTorch报错：dlopen无法加载更多静态TLS对象

I’ve run into this exact annoying issue before—PyTorch imports fine in the regular terminal but throws that static TLS error in Jupyter. Let me break down why this happens and share the fixes that worked for me:

What’s going on here? #

Static TLS (Thread Local Storage) is a limited resource in Linux systems. Jupyter Notebook loads a bunch of background libraries (like for UI, kernel management, etc.) before you even run your code, and some of these might be using up the available static TLS slots. When you try to import PyTorch later, there’s no room left for its static TLS objects, hence the error. The regular terminal doesn’t load all those extra libraries, so PyTorch has enough slots to work with.

Solutions to try #

• Force PyTorch to load first via LD_PRELOAD
  This trick makes PyTorch’s core library load before any other code grabs static TLS slots. First, find where your libtorch.so is located (adjust the path to your conda/pip environment):

  find /path/to/your/python/env -name libtorch.so
  

  Then start Jupyter with this environment variable set:

  LD_PRELOAD=/path/to/libtorch.so jupyter notebook
  

  This ensures PyTorch gets the TLS slots it needs before any other libraries take them.

• Change the import order in your notebook
  If you’re importing other heavy libraries (like OpenCV, TensorFlow, or scikit-learn) before PyTorch, swap the order. Import torch first thing in your notebook cell, then load the other libraries. This way PyTorch reserves the TLS slots early, before other libraries use them up.

• Create a clean, dedicated environment
  Sometimes conflicting libraries in your existing environment are the culprit. Spin up a new conda environment with only PyTorch and Jupyter:

  conda create -n torch_jupyter python=3.9 pytorch torchvision torchaudio cpuonly -c pytorch
  conda activate torch_jupyter
  pip install jupyter
  

  Then launch Jupyter from this clean environment—no conflicting libraries means less chance of TLS slot exhaustion.

• Upgrade PyTorch or switch to JupyterLab
  Newer PyTorch versions have made improvements to reduce static TLS usage. Try upgrading with:

  pip install --upgrade torch torchvision torchaudio
  

  Alternatively, JupyterLab has a different initialization sequence than classic Jupyter Notebook, which might avoid loading as many TLS-heavy libraries upfront. Install it with pip install jupyterlab and give it a shot.

内容的提问来源于stack exchange，提问作者Ioannis Nasios