# Hugging Face — MPS in PyTorch 2.15

- Company: Hugging Face (huggingface.co)
- Announced: 2026-10-20
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://huggingface.co/blog/mps-pytorch-2-15
- Record: https://forck.live/items/25135-mps-in-pytorch-2-15
- Subject: Platform
- Models affected: Qwen3.6-35B-A3B, LFM2.5-8B-A1B, Wan2.2, Qwen-Image-2.1

Hugging Face's PyTorch team says PyTorch 2.15 brings FP8 tensors, faster language-model decoding and new Metal kernels to the MPS backend for Apple silicon Macs. Mixture-of-experts models decode 5–6.5× faster than in 2.14: on an M3 Ultra with torch.compile, Qwen3.6-35B-A3B goes from 9.4 to 61.3 tokens/s and LFM2.5-8B-A1B from 23.6 to 123.6 tokens/s. Dense models gain from matrix-vector kernels wired into F.linear, a 1.23× geometric-mean speedup across 96 BF16 layer shapes, and softmax over the vocabulary is about 8× faster. FP8 adds torch.float8_e4m3fn tensors at one byte per value, half the storage of FP16 or BF16, multiplied on the GPU with F.scaled_mm using tensorwise or rowwise scaling. A padding fix repairs the Qwen-Image-2.1 VAE encoder, which silently corrupted images on MPS in 2.14; an encode-decode round trip now matches CPU at 29.95 dB PSNR, up from 10.06 dB. Wan2.2's tiled VAE decode now peaks at about 8 GiB of driver memory instead of 270 GiB, and MPS requires macOS 15 or later while PyTorch 2.15 requires Python 3.11 or later.

## Evidence

Verbatim from https://huggingface.co/blog/mps-pytorch-2-15:

> Mixture-of-experts models decode 5–6.5× faster than in 2.14: on an M3 Ultra with torch.compile , Qwen3.6-35B-A3B goes from 9.4 to 61.3 tokens/s and LFM2.5-8B-A1B from 23.6 to 123.6 tokens/s.

---

Record: https://forck.live/items/25135-mps-in-pytorch-2-15
Catalogue: https://forck.live/llms.txt
Current issue: https://forck.live/feed.md