PyTorch 2.15 boosts MPS performance with FP8 and faster LLM decoding

pytorch 2.15 brings significant improvements to the mps backend, including fp8 support, which reduces memory usage by half. language-model decoding sees a 5-6.5x speedup, with qwen3.6-35b-a3b achieving 61.3 tokens/s on an m3 ultra. the update also introduces optimized metal kernels for audio, video, and training workloads, such as faster convolutions, padding, and reductions. additionally, the mps backend now requires macos 15 or later and python 3.11 or later. these changes benefit mac users running llms and other ai models, enabling faster and more efficient processing.