Blackwell GPU Support: Alternative Dockerfile using NVIDIA PyTorch Container

[eidamo]

Description

I've successfully built and run OpenFold-3 inference on Blackwell GPUs (compute capability 12.0) using an alternative Dockerfile approach.

My Approach

Instead of building from nvidia/cuda:12.1.1-cudnn8-devel-ubuntu22.04, I used:

FROM nvcr.io/nvidia/pytorch:25.02-py3

This provides native Blackwell support out of the box.

Testing

• ✅ Successfully ran inference on Blackwell GPU
• ❌ Have not tested training
• ❓ Uncertain about backward compatibility with older GPU architectures

Question for Maintainers

Would you be interested in:

1. Option A: An alternative Dockerfile.blackwell for users with Blackwell GPUs?
2. Option B: Modifying the main Dockerfile to support newer architectures?
3. Option C: Documentation on building for Blackwell GPUs separately?

I'm happy to contribute whichever approach you prefer. I can share my complete Dockerfile if helpful.

Environment

• GPU: NVIDIA RTX PRO 6000 Blackwell (compute capability 12.0)
• Base Image: nvcr.io/nvidia/pytorch:25.02-py3
• Tested: Inference only

[jnwei]

Hi @eidamo Thank you so much for reaching out.

The Blackwell Dockerfile sounds like it could be very helpful, I would be happy to review a PR for it. I think your Option A proposal of a separate Dockerfile would make the most sense.

Just to give a sense of the roadmap with Docker: We plan to create a conda-forge and bioconda recipe soon, after which we plan to update the docker files to use the conda recipes. After that, we can work towards having a regular release of OpenFold on Dockerhub or another release platform.

I am sure that the Blackwell Dockerfile you provide will be a good starting point for users on Blackwell GPUs!

[eidamo]

Hi @jnwei Thanks for inviting me to contribute. I've opened PR #26 for Blackwell support.

Let me know if you have questions about the approach.