Overview

PyTorch 2.8, released on August 6, 2025, adds ONNX Attention-23 support and is the last version to support GPU on macOS Ventura.

Main Features

ONNX Attention-23

The ONNX export now supports the Attention-23 operator, making it easier to deploy Transformer models to third-party runtimes.

python 
import torch

model = torch.nn.TransformerEncoderLayer(
    d_model=512, nhead=8, batch_first=True
)
x = torch.randn(1, 10, 512)

# ONNX export with Attention-23
torch.onnx.export(model, x, 'transformer.onnx',
                  opset_version=23)

Last macOS Ventura GPU version

PyTorch 2.8 is the last version to support GPU acceleration (MPS) on macOS Ventura. Future versions will require macOS Sonoma or later.

Sources