I think this is more exploiting XLA to speedup autograd than for deep learning. You would generally use tensorflow for actual training, I've never encountered a

I think this is more exploiting XLA to speedup autograd than for deep learning. You would generally use tensorflow for actual training, I've never encountered any situations where autograd had to be used during training.