Failure to import openmmtorch

[eva-not]

I'm trying to install openmm-torch in a fresh conda environment (Ubuntu 20.04 LTS with CUDA 12.0). When I use the default installation command
conda install -c conda-forge openmm-torch
and try to import openmmtorch, I get the following error:

Traceback (most recent call last): File "<stdin>", line 1, in <module> File "/home/eva/anaconda3/envs/openmm-torch/lib/python3.11/site-packages/openmmtorch.py", line 15, in <module> import _openmmtorch ImportError: libtorch_cuda.so: cannot open shared object file: No such file or directory

From the output of conda list, I can see that the installed pytorch is the CPU version cpu_generic_py311h495eaba_0, which probably explains the incompatibility.

If I install openmm-torch and pytorch simultaneously with
mamba install pytorch torchvision torchaudio pytorch-cuda=11.8 openmm-torch -c pytorch -c nvidia -c conda-forge
and try to import openmmtorch, I get this other error:
Traceback (most recent call last): File "<stdin>", line 1, in <module> File "/home/eva/anaconda3/envs/openmm-torch/lib/python3.10/site-packages/openmmtorch.py", line 15, in <module> import _openmmtorch ImportError: /home/eva/anaconda3/envs/openmm-torch/lib/python3.10/site-packages/../../libOpenMMTorch.so: undefined symbol: _ZN5torch3jit4loadERKNSt7__cxx1112basic_stringIcSt11char_traitsIcESaIcEEEN3c108optionalINS9_6DeviceEEEb

Is there a way to overcome this? Any help would be greatly appreciated!

[peastman]

You need to remove the extra channels-c pytorch -c nvidia. The builds of PyTorch from the different channels aren't compatible with each other.

[sef43]

Hi @eva-not

The first import error sounds like #94

Can you try setting:
export CONDA_OVERRIDE_CUDA=12.0
before installing the package from conda-forge

[eva-not]

Thank you both! I managed to install it with:
conda install pytorch=2=*cuda* -c conda-forge
conda install openmm-torch -c conda-forge

and importing works fine now.