The development

What happened

Hugging Face’s PyTorch contributors describe Apple Silicon MPS improvements in the 2.15 release candidate for running and training language, image and video AI models locally on Macs. The 20 October post covers FP8, targeted Metal kernels and correctness fixes; it is not a final general-availability announcement.

Key facts

  • MPS requires an Apple Silicon Mac with macOS 15 or later; PyTorch 2.15 requires Python 3.11 or later.

  • One-byte FP8 tensors use half BF16’s storage and support scaled matrix multiplication. Selected linear, grouped-matrix and softmax operations gain dedicated Metal paths.

  • In the contributors’ Wan2.2 tiled video-decoder test, peak driver memory fell from about 270 GiB to 8 GiB, not a universal memory reduction.

  • Padding changes address Qwen-Image-2.1 VAE image corruption. Public MPS streams and additional GPU linear algebra are also included; 2.16 previews are separate future work.

The context

Why it matters

Memory and correctness fixes may make previously impractical local workflows usable. Performance remains workload-dependent; test actual models against this release candidate instead of assuming advertised gains apply to all Macs.