AI
PyTorch 2.15 makes Apple silicon more practical for local AI workloads
PyTorch 2.15 substantially improves its MPS backend for Apple silicon. The release adds FP8 tensor support, dedicated Metal kernels, public MPS streams and faster operations for language-model decoding, linear algebra, video and training. Reported gains include much faster mixture-of-experts decoding, an eightfold softmax improvement in one test and a dramatic reduction in Wan2.2 video-decoding memory use. Results come from the PyTorch team’s own benchmarks and require macOS 15 or later.
Editorial summary · Publisher article
What matters
- MoE decoding improved from 9.4 to 61.3 tokens per second for one Qwen model on an M3 Ultra.
- Wan2.2 video VAE decoding reportedly fell from about 270 GiB to 8 GiB of peak driver memory.
- FP8 tensors use half the storage of BF16 and can be multiplied directly on the MPS GPU.
Context & caveats
- The reported performance figures are workload- and hardware-specific; PyTorch 2.16 work remains partly in development.
Why this made the edition
Detailed publisher analysis reports concrete PyTorch 2.15 changes, benchmarks, compatibility requirements, correctness fixes and planned follow-up work for Mac-based AI workloads.
Source material