theaicatalog
Hugging Face ยท Open source

PyTorch 2.15 adds FP8 and faster local LLM decoding on Macs

Hugging Face's PyTorch team says PyTorch 2.15 adds FP8 support on Macs through the MPS backend. Mixture-of-experts models decode 5 to 6.5 times faster than in 2.14, with Qwen3.6-35B-A3B rising from 9.4 to 61.3 tokens per second on an M3 Ultra with torch.compile. Softmax over the vocabulary is about 8 times faster, and Wan2.2's video VAE decode peaks at about 8 GiB of memory instead of 270 GiB. The release requires macOS 15 or later and Python 3.11 or later.