Add AutoReactant{AutoEnzyme}

[wsmoses]

It's not really an AD backend, but this can help make things easier for folks to use throughout the ecosystem.

I'm not quite sure this is the right place/abstraction though, alternatively we can add bool flags to turn on reactant/etc as applicable throughout

cc @avik-pal

[gdalle]

I understand where you're coming from but I'm not sure this is the right abstraction either. As you said, Reactant is not an AD backend, and this package is for AD backends.
Is there any reason to use Reactant with another AD than Enzyme? If not, maybe a parameter inside AutoEnzyme would make more sense?

[wsmoses]

The argument for a distinct thing is that there are reactant-specific attributes (like device/sharding, not yet added here) that are relevant to reactant compilation, but not enzyme.

It also can be relevant for other AD tools, but that's presently unsupported, so it's best to keep the assurances limited until then.

[gdalle]

Can you give an example of where this would simplify things? Lux's API?

[wsmoses]

Lux already has a better way of calling from the trainapi overall.

However, in places, like for example, some turing/sciml/etc packages where adtype is passed through a long train of configs before being selected on -- this means Reactant can be tried without having to rewrite all the intermediate calls. It still might be better to do that rewriting to enable fusion of AD and the outside context, but this at least makes it easy to use/try out.

[gdalle]

In that case it might be time to revive my initial attempt JuliaDiff/DifferentiationInterface.jl#325 for the general case.

I'm not fundamentally opposed to the principle of the current PR (haven't reviewed yed) but there is a documentation challenge to convey what Reactant is, how it is not an AD backend but it can instead combine with AD backends

[wsmoses]

In that case it might be time to revive my initial attempt JuliaDiff/DifferentiationInterface.jl#325 for the general case.

Yeah go for, I think that is just waiting for someone to bring it over the line