[Community] ARM64/Blackwell Docker Deployment for DGX Spark

[adrian-greenneuron]

Hi OpenFold team,

I wanted to share my experience getting OpenFold3 running on the NVIDIA DGX Spark (Grace Blackwell GB10, ARM64). I couldn't find any existing examples of this setup, so I'm documenting what worked for me.

What I Built

A Docker deployment with the following fixes:

1. DeepSpeed compute_121 error — NVCC doesn't recognize compute_121, so I patched op_builder/builder.py to map it to compute_120
2. Triton kernels — Required triton-nightly to support sm_121
3. JIT compilation overhead — Used a docker commit workflow to bake pre-compiled kernels into the image (reduces cold start from ~3 min to ~9 sec)

Benchmarks

Example	Inference Time
Ubiquitin (76 residues)	9s
Protein-ligand (MCL1)	3m 16s

Repo

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

---

Questions for the community:

• Has anyone else gotten OpenFold3 running on ARM64 or Blackwell GPUs?
• Are there better approaches to the DeepSpeed/Triton compatibility issues I encountered?
• Would it be useful to upstream any of these fixes?

I hope this helps others working with similar hardware. Feedback and suggestions are very welcome!

[jandom]

Oh I love the docker commit trick to compile these, excellent!!

also, I've seen our Blackwell docker image but I wasn't sure arm64 would work... that's great!

All in all, fantastic and thank your for this contribution 🙏

[jandom]

@eidamo this seems both similar and subtly different to what you proposed in #26, any comments on this build strategy?

[qurat-ul-ain95]

@adrian-greenneuron I have OF3 working on ARM64 (gracehopper GPUs) but without deepspeed.

[jandom]

We now have access to a DGX box to test this!

[jandom]

Okay, @adrian-greenneuron I can confirm that our current recipe for Blackwell doesn't really work. Trying your solution now.