Hugging Face’s PyTorch team reports MPS gains in PyTorch 2.15
🎧 Listen to this story
In a post, Hugging Face’s PyTorch team says PyTorch 2.15 adds FP8 to MPS, using half the space of BF16, along with new paths to speed up language models, mathematical operations, audio, video and training. The team reports that mixture-of-experts (MoE) models became 5–6.5× faster to decode; on an M3 Ultra, Qwen3.6-35B-A3B rose from 9.4 to 61.3 tokens/s, while softmax over the vocabulary became about 8× faster.
The team also says Wan2.2’s video VAE decoding now peaks at about 8 GiB of memory, down from 270 GiB, and that a fix improved Qwen-Image-2.1 image quality on MPS. PyTorch 2.15 requires macOS 15 or later and Python 3.11 or later.
Why it matters · editorial interpretation
The gains reported by Hugging Face’s team could make running language, image and video models more practical on Macs with MPS, especially for memory-constrained workloads. Requiring macOS 15 and Python 3.11 limits which environments can take advantage of these changes.