There are three phases to deploy a PyTorch model to on-device: program preparation, runtime preparation, and program execution, as shown in the diagram below, with a number of user entry points.