MPS in PyTorch 2.15
- ID
- 37350
- Status
- summarized
- Published
- 20 Oct 2026, 8:00 AM
- Fetched
- 20 Oct 2026, 8:28 PM
- Provider
- Hugging Face Blog
- Category
- developer-ai
- Original URL
- https://huggingface.co/blog/mps-pytorch-2-15
- Source URL
- https://huggingface.co/blog/feed.xml
Summary
- Score
- 7.5
- Created
- 20 Oct 2026, 8:29 PM
- Tags
- Audience
- developersai_ml_learnersvibe_codersai_agent_users
What happened
The Hugging Face PyTorch team detailed what lands in PyTorch 2.15's MPS backend for Apple Silicon: FP8 (float8_e4m3fn) storage and scaled matmul, a native grouped matmul with a dedicated MoE decoding path, specialized matrix-vector kernels for single-row linear layers, and a softmax over the vocabulary that is about 8x faster. Measured on an M3 Ultra with torch.compile, Qwen3.6-35B-A3B goes from 9.4 to 61.3 tokens/s and LFM2.5-8B-A1B from 23.6 to 123.6 tokens/s versus 2.14, while Wan2.2's tiled VAE video decode drops from a 270 GiB peak to 8 GiB. The release also exposes public MPS streams, adds complex LU and triangular solves plus faster small-matrix inverses and QR on GPU, and raises requirements to macOS 15+ and Python 3.11+; 2.16 is slated to move attention fully onto Metal.
Why it matters
If you run local models on a Mac (transformers, diffusers, ComfyUI), the upgrade gate is the decision: PyTorch 2.15 needs macOS 15+ and Python 3.11+, so anyone on macOS 14 or Python 3.10 stays on 2.14 and gets none of the 5-6.5x MoE decode gain. For those who can upgrade, the Wan2.2 VAE number (270 GiB to 8 GiB peak) is the concrete one: video decode that previously needed a datacenter GPU now fits in unified memory, which changes whether local video generation is viable for you at all.
Discussion angle
The token/s numbers are the authors' own benchmarks on an M3 Ultra with torch.compile — worth running the same Qwen3.6-35B-A3B or LFM2.5-8B-A1B test on your own Mac before and after the 2.15 upgrade, and asking whether a ~6x local decode speedup makes running agent workloads on your laptop cheaper than renting a GPU.