The new PyTorch 2.0 promises to accelerate ML training and development, while being backward-compatible with existing PyTorch code.