Error with installing openmmtorch and torch_scatter/torch_geometric

[AdeeshKolluru]

I want to install openmm-torch and pyg. I'm wondering if there are any known versions of these packages that work together.

Simple test

conda create -n test python=3.12
conda activate test
conda install openmm-torch -c conda-forge

this installs torch2.5.1.post303 with cuda version 11.8 on my system.

now when I try to install other packages through pip (conda isn't able to find a matching version)

pip install pyg_lib torch_scatter torch_sparse torch_cluster torch_spline_conv -f https://data.pyg.org/whl/torch-2.5.1+cu118.html

its unable to install these packages correctly

>>> import torch_scatter
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "*/lib/python3.12/site-packages/torch_scatter/__init__.py", line 16, in <module>
    torch.ops.load_library(spec.origin)
  File "*/lib/python3.12/site-packages/torch/_ops.py", line 1350, in load_library
    ctypes.CDLL(path)
  File "*/lib/python3.12/ctypes/__init__.py", line 379, in __init__
    self._handle = _dlopen(self._name, mode)
                   ^^^^^^^^^^^^^^^^^^^^^^^^^
OSError: */lib/python3.12/site-packages/torch_scatter/_version_cuda.so: undefined symbol: _ZN5torch3jit17parseSchemaOrNameERKSsb

It does seem like TorchMDNet uses all these packages but by default their environment.yml doesn't install openmm-torch. It would be great to know if there's any way around this.

Thanks!

[peastman]

Mixing packages from pip and conda doesn't usually work. They use different versions of the C++ ABI, so they can't link to each other.

I'm trying to create a pip package for OpenMM-Torch, but I'm currently blocked by a PyTorch bug.