In DI, the pullback functions use split reverse mode by default, so they disregard the mode field of AutoEnzyme (because I figure no one is going to pass a ReverseSplitWithPrimal anyway). But still I'd like to respect some of the attributes of the ReverseMode that the user provides.
Would you welcome a PR with functions that translate a ReverseMode to the corresponding ReverseModeSplit, setting reasonable defaults for the remaining type parameters?
In DI, the
pullback
functions use split reverse mode by default, so they disregard themode
field ofAutoEnzyme
(because I figure no one is going to pass aReverseSplitWithPrimal
anyway). But still I'd like to respect some of the attributes of theReverseMode
that the user provides. Would you welcome a PR with functions that translate aReverseMode
to the correspondingReverseModeSplit
, setting reasonable defaults for the remaining type parameters?