Published on October 20, 2026
If you run models on a Mac or rely on MPS as your primary compute accelerator, there is plenty to look forward to in PyTorch 2.15! FP8 has officially landed on MPS, language model decoding gains faster execution paths through linear layers and softmax, and brand-new Metal kernels tackle some surprisingly expensive bottlenecks across audio, video, and training workloads. A quick introduction: we are the PyTorch team at Hugging Face. We contribute across PyTorch, with a key focus on the MPS backend to make the Mac a first-class platform for running and training PyTorch models. In early 2026, running even small models on MPS often meant navigating around unsupported operators, performance cliffs, and unexpected memory spikes. So, we ran real-world models from the Hugging Face Hub across language, vision, and video to pinpoint where MPS fell short, and resolved those issues upstream in PyTorch—increasingly by replacing MPSGraph with dedicated Metal kernels. Much of this post highlights the results of that work, accomplished alongside PyTorch maintainers and community contributors.
TL;DR
- FP8 comes to Mac: Storing
float8_e4m3fnvalues requires only half the space of BF16, with full support for scaled matrix multiplication alongside every other operator supported on CUDA and CPU! - MoE model decoding is 5–6.5× faster than in 2.14: Thanks to native grouped matmul with a dedicated decode path, fewer GPU-CPU synchronizations, and faster top-k routing, Qwen3.6-35B-A3B decode speed on an M3 Ultra surges from 9.4 tokens/s to 61.3 tokens/s.
- Every LLM gets faster per-token generation: Single-row linear layers now use dedicated matrix-vector kernels, and vocabulary softmax is roughly 8× faster.
- More linear algebra runs on the GPU: Added support for complex LU decomposition and triangular solve, faster inversion for small matrices, and faster QR decomposition.
- Audio, video, and training get love too: Convolutions, padding, dropout, interpolation backward passes, and reductions all gain faster execution paths. Wan2.2 video VAE decode peak memory plummeted from 270 GiB down to 8 GiB.
- Greater control from Python: Public MPS streams and more reliable event timing are now exposed.
- Updated requirements: MPS now requires macOS 15 or later; PyTorch 2.15 requires Python 3.11 or later.
- A sneak peek at 2.16: Attention fully running on Metal, further migrations from MPSGraph to Metal, and even faster LLM decoding.
What This Means for You
If you run models on a Mac using tools like transformers, diffusers, or ComfyUI, here is what changes in practice with 2.15:
