The release adds MPS support for torch.float8_e4m3fn tensors and scaled matrix multiplication, and introduces Metal implementations for operations including grouped matrix multiplication and softmax. In the authors’ reported tests, Qwen3.6-35B-A3B decoding on an M3 Ultra with torch.compile increased from 9.4 to 61.3 tokens per second compared with PyTorch 2.14. For Wan2.2 tiled VAE decoding, reported peak total driver memory fell from about 270 GiB to 8 GiB. These are workload-specific measurements, not general guarantees. The article says PyTorch 2.15 requires Python 3.11 or later and MPS requires macOS 15 or later. Source: https://huggingface.co/blog/mps-pytorch-2-15
MPS in PyTorch 2.15
In one minute
Hugging Face’s PyTorch team describes MPS backend changes in PyTorch 2.15 for Apple silicon.