PyTorch 2.0 also introduces torch.compile as the main API for wrapping models and returning a compiled model. This is a completely additive feature, helping to maintain backwards compatibility.