Installation dependency conflicts with modern Python/PyTorch versions

[ozakary]

Hello,

I'm trying to install and run Eigenn on a system with an NVIDIA GPU (RTX 1000 Ada with CUDA 12.4), but I'm encountering several dependency conflicts that prevent successful installation following the README instructions.

Environment:

• OS: Ubuntu
• GPU: NVIDIA RTX 1000 Ada with CUDA 12.4
• Python: 3.9 (also tried 3.8)

Here are the steps I followed:

conda create -n eigenn_conda_env python=3.9
conda activate eigenn_conda_env
conda install pytorch -c pytorch
conda install pyg -c pyg -c conda-forge
conda install pytorch-lightning==1.5.2 torchmetrics==0.6.0 lightning-bolts -c conda-forge
pip install e3nn
git clone https://github.com/mjwen/eigenn.git
cd eigenn
pip install -e .
pip install wandb
pip install loguru

Error:

When running python train_atomic.py, I encounter:

ModuleNotFoundError: No module named 'torch._six'

This occurs because newer versions of PyTorch have removed the _six module, but lightning-bolts still depends on it.

Additional Errors Encountered

1- After trying various workarounds:

When using specific older PyTorch versions (1.10.0):

AttributeError: module 'distutils' has no attribute 'version'

2- After patching the distutils issue:

TypeError: Descriptors cannot be created directly.
If this call came from a _pb2.py file, your generated code is out of date and must be regenerated with protoc >= 3.19.0.

3- After fixing protobuf issues:

ModuleNotFoundError: No module named 'torch_sparse'

Attempted Solutions

I've tried numerous approaches:

• Installing specific PyTorch versions (1.10.0, 1.10.1)
• Adding torch-sparse and other PyG dependencies explicitly:

pip install torch-sparse torch-scatter torch-cluster torch-spline-conv -f https://data.pyg.org/whl/torch-1.10.1+cu113.html

• Downgrading protobuf:

pip install protobuf==3.20.3

• Patching the distutils issue
• Using both Python 3.8 and 3.9
• Installing NumPy 1.24.3 to avoid NumPy 2.x compatibility issues:

pip install numpy==1.24.3

Each solution addresses one problem but introduces new conflicts, suggesting the dependency ecosystem has evolved significantly since this project was last updated.

Request

Could you please provide updated installation instructions with specific, compatible versions of all dependencies?

This would be immensely helpful not just for me but for anyone else trying to use Eigenn with modern system configurations.

Thank you for your assistance and for developing this valuable tool.

[mjwen]

Hi @ozakary, thanks for reporting this!

This repo is a bit old, and we have actually reimplemented and simplified it, see: https://github.com/wengroup/matten

And here are example scripts for training both atomic tensors (e.g. NMR) and structural tensors (e.g. elastic): https://github.com/wengroup/matten/tree/main/scripts