PyTorch 2.15 release candidate adds FP8 support for MPS and speeds up model execution
The PyTorch 2.15 release candidate expands MPS support on Mac computers to include FP8. The team at Hugging Face reports up to 6.5 times faster MoE model decoding and a reduction in peak memory usage during video decoding from 270 to approximately 8 GiB.
The PyTorch 2.15 release candidate expands MPS support on Mac computers with Apple silicon. It adds the torch.float8_e4m3fn data type and matrix multiplication using F.scaled_mm directly on the GPU. Values in this FP8 format occupy one byte, half the space required by FP16 or BF16; this refers to the size of stored values, not the total memory usage of the model.
According to the team at Hugging Face, the new operation implementations make MoE model decoding approximately 5–6.5 times faster than in PyTorch 2.14. On M3 Ultra with torch.compile, the speed of Qwen3.6-35B-A3B increased from 9.4 to 61.3 tokens per second, and that of LFM2.5-8B-A1B from 23.6 to 123.6 tokens per second. The changes include processing groups of matrices directly on the GPU and fewer synchronizations between the GPU and CPU.
The team also reports a reduction in peak driver memory usage during tiled video decoding using the VAE in Wan2.2 from 270 to approximately 8 GiB. The release also fixes a bug in which the VAE encoder in Qwen-Image-2.1 on MPS corrupted images and disrupted image editing, for example in ComfyUI.
A release candidate is available to try, and the source recommends installing it in a separate environment. It requires a Mac with Apple silicon, macOS 15 or later, and Python 3.11 or later.
Why it matters
Users of local models gain the option to store supported values in FP8 with half the memory requirements of BF16. The fixes addressing peak memory usage and image corruption also have practical significance for video generation and image editing. The measured speeds apply to specific models on M3 Ultra with torch.compile.
Two audiences, two different impacts
What this means
For individuals
If you run local language models on a Mac, the new operations may reduce the wait for generated text. To try them, however, you need at least macOS 15 and Python 3.11.
For a business
For businesses running image workflows on MPS, the fix for image corruption in Qwen-Image-2.1 is significant. It affects the correctness of image-editing results, not just processing speed.
ProcessesCheck the original
Event sources
clearly official source · 1 publisher, 0 independent. We count feeds from the same owner only once.