Hugging Face details PyTorch 2.15's MPS backend improvements
What happened
On 20 October 2026, a Hugging Face blog post detailed improvements to PyTorch 2.15's MPS backend: added support for FP8 tensors and F.scaled_mm, and MoE model decoding 5–6.5 times faster than in 2.14. The post cites Qwen3.6-35B-A3B decoding on an M3 Ultra rising from 9.4 tokens/s to 61.3 tokens/s. The information currently appears only in this single first-hand Hugging Face blog post, with no other source adding to or corroborating it.
Written by AI from the reports · updated 4 h ago
Timeline
Follow the reports to see the event from each side.
- Hugging Face BlogMPS in PyTorch 2.15
If you run models on your Mac or use MPS as your primary acceleration tool for computing, there's plenty to look forward to in PyTorch 2.15! FP8 arrives on MPS, language-model decoding gets faster paths through linear layers and softmax and new Metal kernels tackle some surprisingly expensive parts of audio, video and training workloads.
Heat of this event
Current heat 6·Peak in the comparable range 7（Oct 20 23:00）·Change over 24 hours in the comparable range –
The trend compares only the same accounts observed without a break, so its range can be narrower than the current heat count. Move the pointer or tap the chart to see each hour; with a keyboard, use the left and right arrow keys.