pyg_lib installation fails inside Docker container due to incompatible torch / torch.cuda versions

[hknewton]

Hi,

Thanks for the great work on ptm-mamba! We’re excited to reproduce and build on this work.

We encountered an issue when trying to install pyg_lib inside the official Docker container provided for ptm-mamba. The installation fails due to a mismatch between the versions of torch and torch.cuda bundled in the container and the versions supported by pyg_lib.

1. Details
   Docker container: listed in the repo

PyTorch version: torch==2.2

torch.version.cuda: 12.3
Installed via pip with instructions provided in repo.

2. Error
   The error message indicates that there is no compatible build of pyg_lib for torch==2.2 with torch.cuda==12.3:

ERROR: Could not find a version that satisfies the requirement pyg_lib
ERROR: No matching distribution found for pyg_lib

This seems to be due to the fact that pyg_lib currently provides prebuilt wheels only up to CUDA 11.8 or 12.1 for specific PyTorch versions. See: https://github.com/pyg-team/pyg-lib#installation

3. Our Suggested Solutions
   Please provide a Docker container variant with torch and torch.cuda versions compatible with pyg_lib (e.g., torch==2.1, torch.cuda==11.8)

Or, please document the incompatibility and suggest a workaround or base image for those needing to use PyG-based components

Please let us know if there is a recommended approach to install pyg_lib with this container or if we should downgrade torch in our own forked environment.

Thanks again for your work on this project!

[pengzhangzhi]

Given that:

• Your PyTorch version is 2.2.0
• Your CUDA version is 12.3 (torch.version.cuda == '12.3')

And based on the supported pyg-lib wheels:

• cu123 (CUDA 12.3) is not available
• The closest available CUDA wheels are for cu121 (CUDA 12.1) and cu124 (CUDA 12.4)

Therefore, there is no prebuilt pyg-lib wheel available for torch==2.2.0 with CUDA 12.3.

---

**Recommended Install Command **

12.3 Cuda is compatible with CUDA 12.1, U can use:

pip install pyg-lib -f https://data.pyg.org/whl/torch-2.2.0+cu121.html

You can check the installed CUDA version used by PyTorch with:

import torch
print(torch.__version__, torch.version.cuda)

---

Safer Option: Downgrade to CUDA 12.1 Runtime in Docker

Switch to an image like:

FROM nvidia/cuda:12.1.1-cudnn8-devel-ubuntu22.04

Then your install command will work:

pip install torch==2.2.0 torchvision==0.17.0 torchaudio==2.2.0 \
  --index-url https://download.pytorch.org/whl/cu121

pip install pyg-lib -f https://data.pyg.org/whl/torch-2.2.0+cu121.html