MPS in PyTorch 2.15
PyTorch 2.15 adds FP8 tensor support on Apple silicon Macs and speeds up language-model decoding, especially for mixture-of-experts models, with more GPU linear algebra and operations moved to dedicated Metal kernels. It requires macOS 15 or later and Python 3.11 or later; 2.16 will run attention entirely on Metal.
What the article says
In a post dated 20 October 2026, the Hugging Face PyTorch team presents progress on the MPS backend, which runs PyTorch on Apple Silicon Mac GPUs, with version 2.15 (still a release candidate). The team tests real Hub models, fixes issues upstream and increasingly replaces MPSGraph with dedicated Metal kernels.
New items include FP8 support with F.scaled_mm, faster LLM decoding (MoE layers up to 5 to 6.5 times faster than in 2.14, softmax about 8 times faster), added linear algebra and convolution operations, and MPS streams controllable from Python. On an M3 Ultra, Qwen3.6-35B-A3B goes from 9.4 to 61.3 tokens/s. Wan2.2 VAE decoding fits in about 8 GiB instead of more than 270 GiB, and a bug that corrupted images in ComfyUI is fixed. macOS 15 and Python 3.11 are the minimum, and the measurements are the team's own.
Read the official article AnnuIA profile
Summary written by AnnuIA in its own words from the official page of Hugging Face. The content and rights belong to the publisher: please read the original article.