MPS in PyTorch 2.15

Why it matters

The PyTorch team at Hugging Face has published a detailed account of its work on the MPS backend, the component that lets PyTorch models run on Apple silicon Macs. The team explains that running models on MPS earlier in the year meant contending with unsupported operations, performance cliffs, and unexpected memory growth, so they now test real models from the Hugging Face Hub across language, image, and video tasks, then fix shortcomings upstream in PyTorch, increasingly by replacing MPSGraph with dedicated Metal kernels. The latest release brings FP8 tensors to the Mac, allowing weights to be stored in half the space of BF16 with scaled matrix multiplication support. Mixture-of-experts models decode dramatically faster thanks to a native grouped matmul on Metal, fewer GPU-CPU synchronizations, and faster routing. Dense language models also gain speed from specialized matrix-vector kernels and a much faster softmax during sampling. Video generation memory use drops sharply for Wan2.2's VAE, and a padding fix resolves silent image corruption in Qwen-Image-2.1 editing workflows. The backend now requires a newer macOS and Python version.