// the find
mrdbourke/pytorch-apple-silicon
Setup PyTorch on Mac/Apple Silicon plus a few benchmarks.
A README-driven walkthrough for getting PyTorch running with GPU acceleration (MPS) on Apple Silicon Macs, written by the mrdbourke/ZTM crowd. It's two Jupyter notebooks plus setup instructions, not a library — aimed at people new to conda/ML tooling on Mac rather than anyone already comfortable with the ecosystem.
The two-track structure (short version for people comfortable with a terminal, long version with conda/package-manager explanations for beginners) is genuinely useful and well thought out. It includes real benchmark numbers (CPU vs MPS on TinyVGG/CIFAR10 at two image sizes) instead of just asserting the GPU path is faster. The verification snippet (checking torch.backends.mps.is_built()/is_available()) is exactly the sanity check someone needs after following unfamiliar setup steps.
Last pushed June 2023 and pins Python 3.8 and PyTorch 1.12 — both well behind current PyTorch (MPS support has moved a lot since beta) and Python versions, so following it literally will fight you on a modern Mac. Benchmarks are from a single M1 Pro run in May 2022 on one toy model; there's nothing for M2/M3/M4 or any workload beyond a tiny CNN, so the performance claims don't generalize. It's not a tool you install and use — no package, no CLI, no tests — just a README and two notebooks, so there's nothing to maintain compatibility going forward. Apple's and PyTorch's own docs on MPS have likely caught up to and surpassed this as the authoritative source.