MPS in PyTorch 2.15

PyTorch 2.15 introduces significant performance improvements for the MPS backend on Apple Silicon Macs. Key updates include the addition of FP8 data types for reduced memory usage and faster matrix multiplication for language model decoding. Mixture-of-Experts (MoE) models experience a 5-6.5x speedup in token generation due to a new native grouped matmul implementation and other optimizations. The release also delivers faster linear algebra operations, a drastically reduced memory footprint for video decoding, and more direct control over the GPU via MPS streams.