outdated requirements, how to install on cuda 12.x and python 3.11 or later ?

[schloegl]

env/SE3nv.yml
contains these dependencies

dependencies:
  - python=3.9
  - pytorch=1.9
  - cudatoolkit=11.1
  - dgl-cuda11.1
  - ...

However, this page [1] says, that L40s would require cuda12.1 or later. Moreover, python/3.9 has reached end-of-life [2]. So I'm wondering what the recommended procedure for installing rfdiffusion on a new software stack (e.python 3.11, cuda 12.x) would be.

Cheer,
Alois

[1] https://forums.developer.nvidia.com/t/minimum-required-cuda-version-by-gpu/276955
[2] https://devguide.python.org/versions/

[tydingcw]

I created a conda environment that works for L40s on my cluster using cuda 11.8:

conda env create -f env/RFdiff_accre.yml
conda activate SE3nv
conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia
conda install torchdata==0.9.0 -c pytorch -c conda-forge
conda install pandas
conda install pydantic
conda install -c dglteam/label/th23_cu118 dgl

cat ~/git_repos/RFdiffusion/env/RFdiff_accre.yml
name: SE3nv
channels:
  - defaults
  - conda-forge
  - pytorch
  - dglteam
  - nvidia
dependencies:
  - python=3.9
  - pytorch=1.9
  - torchaudio
  - torchvision
  - cudatoolkit=11.8
  - pip
  - pip:
    - hydra-core
    - pyrsistent