When discussing the ML Compute (Apple ML library) integration, I made this description of some of the more complex use cases for the autograd.
This might be interesting to other to looks at, link here.
Feel free to ask any question here or if anything is missing!