PyTorch 2.15 brings FP8 and faster decoding to Mac

PyTorch 2.15 replaces MPSGraph paths for F.linear and grouped matmul with native Metal kernels, adds FP8 quantization on MPS and native grouped matmul for MoE decode. MoE decoding is 5–6.5x faster, video VAE memory drops from 270 GiB to 8 GiB, and FP8 roughly halves parameter memory. It requires macOS 15+ and Python 3.11+.

• MoE decoding on Mac is 5–6.5x faster
• Video VAE memory drops from 270 GiB to 8 GiB
• FP8 roughly halves parameter memory
• Requires macOS 15+ and Python 3.11+