PyTorch MPS changes kernels—and fixes image output
The PyTorch 2.15 article covers Metal execution paths, FP8 support and a correctness fix worth checking before throughput.
F.linear connects to specialized matrix-vector kernels for eligible decoding inputs in the MPS changes described for PyTorch 2.15. That is a specific execution-path improvement, rather than a blanket performance promise for every model.
Writing on October 20, 2026, Irakli Salia and Khushi Agrawal of Hugging Face's PyTorch team also describe FP8 support, native grouped matrix multiplication and operations moving from MPSGraph to dedicated Metal kernels.
Correctness is the first change I would check. The authors report that the previous padding path silently corrupted images in the Qwen-Image-2.1 VAE encoder, disrupting image editing. The new path fixes that workflow. They also connect the padding change to lower driver-memory usage during Wan2.2 video decoding.
For an upgrade evaluation, I would rerun a known input and inspect its output before comparing speed. The documented prerequisites are Apple silicon, macOS 15 or later and Python 3.11 or later. Keep the release boundary clear: the article provides a release-candidate installation path and discusses subsequent-version work separately.
Source: Irakli Salia and Khushi Agrawal, Hugging Face PyTorch team, MPS in PyTorch 2.15.