Installation Woes

[JacksonBurns]

Hello,

I tried installing minimol according to the indicated instructions (from pip) and had this issue:

OSError: [...]/_version_cuda.so: undefined symbol: _ZN5torch3jit17parseSchemaOrNameERKNSt7__cxx1112basic_stringIcSt11char_traitsIcESaIcEEE

which is caused by incompatible torch and torch-geometric versions.

After much debugging, I got things up and running with this series of commands:

conda env create --name minimol 'python==3.10.*'
pip install 'torch==2.6.0' 'numpy<2'
pip install pyg_lib torch_scatter torch_sparse torch_cluster torch_spline_conv -f https://data.pyg.org/whl/torch-2.6.0+cu124.html
pip install minimol 'torch==2.6.0' 'numpy<2' 'scipy==1.12.*' # force minimol to respect the installed version of pytorch and numpy

The version of torch suggested in the README isn't listed on the pytorch geometric install page with a version of CUDA that I use, so I bumped it.

Similar issues extend to the Google colab. On trying to execute the notebook it eventually crashes from an error related to the numpy version being >1 when 1 is expected

[callumm-graphcore]

Hi Jackson, what version of CUDA are you using?

[JacksonBurns]

@callumm-graphcore 12.4

[CurtisColwell]

Thanks @JacksonBurns! I spent too much time trying to get this installed until I finally came across this. Still works despite my CUDA being at 13.0

[JacksonBurns]

Glad it worked! Turns out most of the install hinges on the CUDA version, which dictates which version of torch and torch geomtric one can install. Here's my latest install script:

conda create --name minimol 'python==3.10.*' --yes
conda activate minimol
pip install torch==2.8.0 torchvision==0.23.0 torchaudio==2.8.0 --index-url https://download.pytorch.org/whl/cu129 'numpy<2'
pip install pyg_lib torch_scatter torch_sparse torch_cluster torch_spline_conv -f https://data.pyg.org/whl/torch-2.8.0+cu129.html
pip install minimol 'torch==2.8.0' 'numpy<2' 'scipy==1.12.*' # force minimol to respect the installed version of pytorch and numpy

Most important us just to find a version of torch-geomtric that supports your cuda version, install the corresponding torch version, and ensure that the numpy and scipy limitations are respected.