PyTorch 2.15 Enhances MPS Backend with FP8 and Custom Metal Kernels

1 reports archived1 independent sourcesupdated 10/20/2026, 08:00:00

Synthesis & Latest Updates

1 Sources Cross-Validated

PyTorch 2.15 significantly optimizes the MPS acceleration experience on Macs by introducing native FP8 support and refactoring linear layer and softmax decoding paths. Replacing generic MPSGraph with dedicated Metal kernels yields substantial performance gains for audio, video, and training workloads.

LATEST/PyTorch 2.15 significantly optimizes the MPS acceleration experience on Macs by introducing native FP8 support and refactoring linear layer and softmax decoding paths. Replacing generic MPSGraph with dedicated Metal kernels yields substantial performance gains for audio, video, and training workloads.

HEAT TRENDHourly heat curve

2 fully observed hours

Fewer than 3 fully observed hours; trend not plotted yet.

TIMELINECoverage timeline

Total 1 reports · Latest first

Hugging Face BlogT1·Oct 20, 08:0068 pts
PyTorch 2.15 Enhances MPS Backend with FP8 and Custom Metal Kernels
- MPS backend now officially supports FP8 precision for improved inference efficiency
- Language model decoding paths (linear layers/softmax) receive targeted acceleration
- Dedicated Metal kernels replace MPSGraph to optimize audio, video, and training workloads
View Related Reports & Analysis Original Source