Loading
Loading articles
By Hugging Face Team
AI summary
PyTorch 2.15 introduces significant performance improvements for the MPS (Metal Performance Shaders) backend on macOS, focusing on migrating operations from MPSGraph to dedicated Metal kernels. Key enhancements include native FP8 support, optimized decoding paths for mixture-of-experts (MoE) and dense LLMs, and drastically reduced memory usage for video VAE decoding. These updates streamline local model inference and training workflows by addressing performance bottlenecks in linear layers, softmax, and linear algebra operations.
Key points