How can be openMM-ML v1.5 installed with CUDA 12.2?

[qinghualiao]

Hello,

Recently, I asked our cluster administrator to install openMM/openMM-ML,
and we found that there is a conflict of dependency regarding CUDA version.

The highest CUDA version supported by the operating system is 12.2.
We can install openMM/openMM-ML, pytorch, mace-torch and related programs with CUDA 12.0.
But the version of openMM-ML is 1.2, if we install 1.5, then CUDA 12.9 will be used for the installation,
which is not supported by the operating system (upgrading CUDA in the system is one option, but not for now).

We are wondering there is a way to install openMM-ML with CUDA 12.0 or 12.2. Thanks very much!

All the best,
Qinghua

[peastman]

The problem is that recent versions of PyTorch are built against newer versions of CUDA than that. I assume you must have installed an older version? OpenMM-ML is a pure Python package, so it doesn't really care what version of CUDA you have, but it uses OpenMM-Torch, which is a binary package providing the interface to PyTorch. The easiest approach is probably to install both of them from source, building OpenMM-Torch against your installed PyTorch.

This will get easier in the near future. The next version of OpenMM-ML will remove the need for OpenMM-Torch, so you'll just be dealing with a single pure Python package. It requires a new feature of OpenMM 8.5, which we just released a beta of. The official release will probably be sometime next month.

[qinghualiao]

Thanks very much, looking forward to both of the new releases of openMM and openMM-ML.

All the best,
Qinghua

[Mert-csrg]

@qinghualiao and I had the same problem (maybe it's because we work in the same group) and I was able to get a working stack with openmm-ml 1.5 by building the CUDA-dependent packages from source. So, anyway, maybe the following may be useful for others who have the same problem.

Working stack in my build:

• OpenMM 8.4.0 (source, CUDA enabled)
• openmm-ml 1.5 (source)
• openmm-torch v1.5.1 (source)
• PyTorch v2.9.1 (source, built against CUDA 12.2)
• CUDA base image: nvidia/cuda:12.2.2-devel-ubuntu22.04
• Tested on H100 with NVIDIA driver 535.86.10

Dockerfile:
Dockerfile.openmm.cuda122.source

Base env file:
environment-openmm-linux-amd64-gpu-source-base.yml

Built image:
openmm-ml:cuda122-source

Build command:

docker buildx build \
  --platform linux/amd64 \
  -f Dockerfile.openmm.cuda122.source \
  -t openmm-ml:cuda122-source \
  --load \
  .