Request: Apple Silicon (Mac OS) Support

[aichr]

Summary

Is there any plan to support Apple Silicon (Mac OS) for running Utonia? We tried multiple approaches but hit fundamental CUDA dependencies.

What We Tried

1. Install via pip

2. Build from source

• Cloned spconv source from GitHub
• Build requires CUDA toolchain (nvcc, etc.)
• Not possible on Mac OS without CUDA

3. Mock dependencies

Created mock modules for , , to test if model loads:

• Model imports successfully
• Model loading fails because (from spconv) has no CPU implementation

Root Cause

The model fundamentally depends on:

1. spconv (sparse convolution) - CUDA-only, no Apple Silicon wheels
2. torch-scatter - CUDA-only, no Apple Silicon wheels
3. flash-attention - CUDA-only

These are not optional dependencies — they are core operations in the architecture.

Questions

1. Is Apple Silicon / MPS support on the roadmap?
2. Would you accept a PR to add CPU/MPS fallback for sparse convolutions?
3. Any suggestions for running on Mac OS?

Environment

• Mac mini with M4 chip (Apple Silicon)
• PyTorch 2.10 with MPS (Metal Performance Shaders) available
• macOS 14+

References

• Similar issue: Can't install on mac m1, please add support traveller59/spconv#600 ("Can't install on mac m1, please add support")
• Related: fairchem#1712 (Apple Silicon MPS support request)

Thanks for the great work on Utonia!

[Yujia-Zhang0913]

Hi, thanks for your detailed description.
However, the current architecture is indeed not suitable for CPU device. We also want to erase the inconvenience of SPConv in implementation. If there is any PR for running SPConv on CPU, we would be happy to see and accept that. And if we have any updates on implementation, I will comment on this issue.