How general/efficient are these AD systems ?

New ops in tensorflow seem to be oriented towards forward-mode AD rather than reverse-mode (for which one needs a pull back op on a dual).