PyTorch 2.15 RC speeds up local AI workloads on Macs
The PyTorch 2.15 release candidate adds FP8 support, faster language-model decoding, and major video-memory improvements to the MPS backend on Apple Silicon.
AT A GLANCE
Key points
- FP8 tensors on MPS can use half the storage of BF16
- Dedicated Metal kernels accelerate language-model decoding paths
- Peak driver memory in the Wan 2.2 VAE test fell from about 270 GiB to 8 GiB
- The release candidate requires macOS 15+ and Python 3.11+
What is changing?
Hugging Face’s PyTorch team detailed the Apple Metal Performance Shaders improvements in the PyTorch 2.15 release candidate on October 20. The update supports FP8 tensors and scaled matrix multiplication on MPS, allowing eligible weights to use half the storage of BF16.
For language models, dedicated Metal kernels accelerate linear layers, mixture-of-experts routing, and softmax during sampling. In the team’s M3 Ultra tests, decoding throughput for selected MoE models rose substantially compared with 2.14. Peak driver memory during the tiled VAE decode for Wan 2.2 also fell from about 270 GiB to 8 GiB.
The release requires macOS 15 or later and Python 3.11 or later. The installation command shared in the article points to the 2.15.0 release candidate in the test repository, so these results should not be presented as final stable-release performance.
Explore the original sources
Source date: 20 October 2026.
Published by AI Radar: 20 Oct 2026, 18:12 (Istanbul, UTC+3).
FOLLOW THE TOPIC
Topics and tags
Share your discovery.
Send this story to a friend.
Reader comments
What do you think about this development? Share your experience or question. Comments are published after editorial approval.
Your name/nickname and comment become public if approved. Read: Privacy notice.
Loading comments…