Home/Events/MPS in PyTorch 2.15

MPS in PyTorch 2.15

Confirmed

Confidence

90%

Impact: 80%

Updated 1h ago

Consensus Brief

PyTorch 2.15 introduces significant enhancements for Mac users utilizing the MPS backend, including the addition of FP8 support, faster language-model decoding, and improved performance for audio and video workloads. The update aims to make Mac a more viable platform for running and training PyTorch models, addressing previous limitations in model support and performance.

Sourced from

Hugging FaceT1 Hugging FaceT1+1 more

Primary: Hugging Face

What Changed Since Last Update

1h ago

New official source added: Hugging Face published an update on Tue, 20 Oc ("MPS in PyTorch 2.15").

Claim Ledger

4 claims tracked across sources

Confirmed Fact

FP8 arrives on MPS, allowing float8_e4m3fn values to be stored in half the space of BF16.

Confirmed Fact

MoE models decode 5–6.5× faster than in 2.14 on an M3 Ultra.

Confirmed Fact

Softmax over the vocabulary is about 8× faster in PyTorch 2.15.

Confirmed Fact

MPS now requires macOS 15 or later and Python 3.11 or later.

Role-Based Impact Analysis

Source Timeline

2 sources corroborating

MPS in PyTorch 2.15

Hugging Face·12h ago

MPS in PyTorch 2.15

Hugging Face·12h ago