Hugging Face Blog·· 2026-10-20AI score58
PyTorch 2.15 brings FP8 and Metal kernels to MPS, Wan2.2 video VAE decoding VRAM reduced to 8 GiB
MPS in PyTorch 2.15
AI briefing
Hugging Face's PyTorch team released PyTorch 2.15, adding FP8 tensors, native grouped matmul, and multiple Metal kernels to the MPS backend for Mac.
Open original huggingface.co