PyTorch 2.15 release candidate adds FP8 tensors, native grouped matmul and faster LLM decoding kernels on Apple MPS backend

PyTorch 2.15, available as a release candidate and not a final release, improves the MPS backend that runs models on Apple silicon. It adds FP8 tensors (torch.float8_e4m3fn) and F.scaled_mm, a native Metal grouped matmul for mixture-of-experts layers, and a faster single-row decoding path for linear layers. It also adds more linear algebra operations and Metal kernels, public MPS streams, and a fix for Qwen-Image-2.1 VAE encoder corruption. The reported figures are Wan2.2 VAE decode peak memory falling from about 270 GiB to about 8 GiB, MoE decoding 5-6.5x faster than 2.14 on an M3 Ultra with torch.compile (Qwen3.6-35B-A3B from 9.4 to 61.3 tokens/s), and a 1.23x geometric-mean speedup on linear layers across 96 BF16 shapes. MPS now requires macOS 15 or later, and PyTorch 2.15 requires Python 3.11 or later.

People running LLMs, MoE models or video generation locally on Macs, including ComfyUI users, stand to gain the most, and the memory reduction could make some video models usable on machines that could not run them before. The minimum requirements are a real cost: anyone on macOS 14 or older, or Python 3.10 or older, cannot take these gains without upgrading. The effect is limited to one backend on one hardware family, so it does not change the wider picture for training or serving at scale. Mac users should treat the gains as promising but unproven until they test their own models.

The claim that this makes the Mac a "first-class place" to run and train PyTorch models is framing, not something the numbers show. The speedups come from the authors' own tests, mostly on a top-end M3 Ultra with torch.compile and a favourable MoE case, so typical machines and dense models will likely see far smaller gains (the 1.23x linear-layer figure is closer to the realistic case). The 270 GiB to 8 GiB drop is also a best case for one model, and the 270 GiB baseline suggests a pathological earlier behaviour rather than a normal workload.

1. 01Whether the final 2.15 release ships on schedule and the MPS changes survive the release candidate stage without regressions or reverts
2. 02Independent benchmarks on M-series chips below the M3 Ultra (such as M2 or M4 Pro and Max) and on dense models, to see whether the 5-6.5x MoE decoding gain holds outside the authors' setup
3. 03Reports from ComfyUI, Hugging Face Diffusers and Transformers users on Wan2.2 and Qwen-Image memory use and any breakage caused by the macOS 15 and Python 3.11 requirements