OpenFold3 Docker Deployment for DGX Spark (ARM64/Blackwell)

Accelerated Computing DGX Spark / GB10 User Forum DGX Spark / GB10 Projects
1

Hi everyone,

I’ve created a Docker deployment for running OpenFold3 protein structure prediction on the DGX Spark (GB10). Since there’s no official ARM64 image available, I built one from scratch and solved several compatibility issues along the way.

What it solves:

  • Triton + Blackwell: Uses

    triton-nightly

    to support

    sm_121

    kernels

  • DeepSpeed JIT errors: Patches the

    compute_121 not recognized

    NVCC issue by mapping to

    compute_120

  • Pre-compiled kernels: Build script bakes JIT-compiled CUDA ops into the image (no 3-minute cold start)

Benchmarks on DGX Spark:

Example Inference Time
Ubiquitin (76 residues) 9s
Protein homomer 8s
Protein multimer 2m 11s
Protein-ligand 3m 16s

GitHub: https://github.com/adrian-greenneuron/openfold3-DGX-Spark

The repo includes Dockerfile, build scripts, DeepSpeed patches, and full documentation. Happy to answer questions or take PRs!

2

Thanks for sharing your project. This is exactly what the forum is meant to do and to inspire others in the community.