[Report] Issues Installing PyEPRI on Apple Silicon and How I Resolved Them

[Yohei-Kudo]

I would like to report an issue I encountered when installing PyEPRI on an Apple Silicon Mac, along with the workaround that resolved it. I hope this helps other users and contributes to improving the package.

---

Environment

• MacBook Pro (Apple Silicon, M2 Max)
• RAM: 32 GB
• OS: macOS 15.5
• Python version: 3.11
• PyTorch backend: torch("cpu")
• Environment manager: Started with pip (venv), later switched to conda
• PyEPRI installation: GitHub clone → pip install -e .

---

Issue Summary

Initially, I created a virtual environment using pip and installed PyEPRI with:

git clone https://github.com/remy-abergel/pyepri.git
cd pyepri
pip install -e .

When running the TV-regularized imaging example from the documentation, Python crashed during the optimization loop-not a Python exception, but a full interpreter crash resulting in a macOS crash report.

Details:

• Crash occurred inside the optimization step.
• Backend: Pytorch CPU.
• Reinstalling (fresh clone + pip install) produced the same crash.
• The behavior and crash report suggested an OpenMP runtime conflict.

---

Attempts to Fix the Issue

1. Reinstalling PyEPRI in a fresh pip virtual environment

-> Crash persisted

2. Switching to a conda environment and installing PyEPRI with --no-deps

-> This eliminated the OpenMP conflict and resolved the crash.

Using conda to manage libraries such as NumPy, finufft, and PyTorch avoided the OpenMP runtime clash that occurred in the pip environment.
Additionally, installing PyEPRI with --no-deps prevented pip from overwriting conda-managed dependencies.

---

Final Working Installation Commands

The following set of commands produced a stable environment where PyEPRI runs correctly on Apple Silicon:

conda create -n pyepri python=3.11
conda activate pyepri 

Install core scientific packages

conda install -c conda-forge numpy=1.26 scipy matplotlib pyvista psutil jupyter

Install finufft

conda install -c conda-forge finufft

Install PyTorch

conda install -c pytorch torchvision torchaudio

Install PyEPRI without modifying dependency versions

pip install --no-deps -e . 

With this setup:

• The TV-regularized imaging example runs without crashing
• No OpenMP-related issues occur
• The optimization loop completes successfully

---

Conclusion

PyEPRI can run successfully on Apple Silicon, but pip-based installations may trigger OpenMP runtime conflicts, leading to interpreter crashes during optimization.
Using conda and installing PyEPRI with --no-deps fully resolved the issue in my environment.

I hope this report is helpful for other users.

[remy-abergel]

Thank you very much for sharing this report and the helpful workaround. I’ll see if I can isolate the issue and get pip install working in a standard venv for Mac users.

I’ve also received reports from another user about issues with using PyVista on macOS for 3D visualization. Are you able to use the displayers.isosurf3d function properly in your conda environment? (Note that this function is currently only available on the main branch of GitHub.)

Anyway, thank you again for your very helpful contribution.

[Yohei-Kudo]

Thank you very much for your message.

Using the environment I created based on the report, I tested displayers.isosurf3d() and it worked correctly.
This suggests that PyVista is functioning properly as well, at least in this conda environment.

[remy-abergel]

That's great, thank you :-)