PyTorch 2.15 MPS Backend Released: FP8 Comes to Mac, MoE Decoding Speedup

According to an article published on the Hugging Face Blog on October 20, 2026, the Hugging Face PyTorch team introduced major updates to the MPS backend in PyTorch 2.15. The update centers on two points: first, new support for the torch.float8_e4m3fn FP8 data type, bringing FP8 capability to the Mac platform; second, the addition of the F.scaled_mm operator. In the article title, the team noted that with these updates, MoE (Mixture of Experts) model decoding speed can improve by 5 to 6.5x. This release, published by the Hugging Face PyTorch team as the primary source, marks a further expansion of PyTorch's low-precision compute support on Apple Silicon's MPS backend. No additional details from other media or official channels have been seen, and no claims contradicting the report have emerged.

AI overview of reporting · Updated 3 h前