Models & Labs
PyTorch 2.15 brings FP8 and MoE speedups to Mac
Hugging Face Bloghigh confidence
Why it matters
- →FP8 support on MPS allows Mac users to run larger models within limited memory constraints.
- →MoE decoding optimizations deliver massive speedups, making local inference on Apple Silicon viable for complex architectures.
- →Memory footprint reductions in video workloads resolve previously impossible hardware limitations.
PyTorch 2.15 introduces significant performance and feature upgrades for the Metal Performance Shaders (MPS) backend, targeting Apple Silicon users. Key additions include native FP8 tensor support via F.scaled_mm, which halves memory usage compared to BF16, and optimized decoding paths for Mixture-of-Experts models that achieve up to 6.5x faster token generation on M3 Ultra hardware. The update also resolves critical memory leaks in video processing, reducing peak VRAM consumption for Wan2.2 VAE decoding from 270 GiB to 8 GiB. Compatibility now requires macOS 15 and Python 3.11 or later.
Read originalThe story around this
Earlier coverage that leads up to this article, and what followed. Lines connect each piece to the closest one after it, converging here.
More from Hugging Face Blog
More in Models & Labs
Models & Labsmodels
llama.cpp b12029 adds CUDA 13 and Snapdragon support
This release quietly cements llama.cpp as the universal inference runtime by finally supporting NVIDIA's latest CUDA 13 stack alongside AMD's ROCm 10.0. The inclusion of native builds for Linux arm64 with Snapdragon chips, covering CPU, Adreno GPU, and Hexagon NPU, signals a serious push into mobile edge deployment. While Apple Silicon KleidiAI is disabled in this specific build, the broader hardware coverage means developers no longer need to compile from source for most modern accelerators. It’s less about new features and more about removing friction for anyone running local models on bleeding-edge or diverse hardware.
llama.cpp Releases
© The Verge AIModels & Labsother
