Understand this story
On October 20, 2026, PyTorch 2.15 added support for torch.float8_e4m3fn tensors on the MPS backend. Reporting from the same source also mentions that this version makes MoE model decoding 5–6.5 times faster than 2.14: on an M3 Ultra with torch.compile, Qwen3.6-35B-A3B improved from 9.4 tokens/s to 61.3 tokens/s, and LFM2.5-8B-A1B improved from 23.6 tokens/s to 123.6 tokens/s. The report states that Wan2.2's tiled VAE video decoding peak driver memory dropped from about 270 GiB to 8 GiB; the image corruption issue in Qwen-Image-2.1's VAE encoding on MPS was also fixed with a padding change, with encode-decode round-trip PSNR rising from 10.06 dB to 29.95 dB. This version requires macOS 15 or later and Python 3.11 or later. There are currently no subsequent independent developments.
AI-generated from public reports
Latest development · Report
PyTorch 2.15 Brings FP8 and Faster MoE Decoding to MPSDevelopments
1 confirmed developments
- Source:Hugging Face
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Oct 20
Hugging FaceFirst-partySelected
PyTorch 2.15 Brings FP8 and Faster MoE Decoding to MPS
PyTorch 2.15 adds FP8 support to the MPS backend and makes MoE model decoding 5–6.5 times faster than 2.14: on an M3 Ultra with torch.compile, Qwen3.6-35B-A3B goes from 9.4 tokens/s to 61.3 tokens/s, and LFM2.5-8B-A1B goes from 23.6 to 123.6 tokens/s. For Wan2.2's tiled VAE video decoding, peak driver memory drops from about 270 GiB to 8 GiB, and the image corruption issue with Qwen-Image-2.1's VAE encoding on MPS is also fixed with a padding change, raising the encode-decode round-trip PSNR from 10.06 dB to 29.95 dB. This version requires macOS 15 or later and Python 3.11 or later.