William Moses

I mean an any and real should also be duplicated.... Essentially anything that isn't a floating point register, integer, or structure containing just those should not be active.

Can you paste whatever you see in a log/backtrace? I'm not able to reproduce it from your code snippet on main. ```julia julia> using Enzyme julia> julia> function aux(f, μ)...

Adding in a couple of corrections to above test case, can you also confirm this is the bug you see? ```julia using Enzyme, CUDA using Random using Adapt gpu(x) =...

Haven't yet gone through it all so there may be some surprises, but I believe the answer is just API/integration. Essentially because LLVM presently permits only one architecture in a...

@LeilaGhaffari per your other Github issue, one what timeline would support for this be needed? To do some expectation setting / figure out prioritization, enabling support for custom base types...

@vchuravy how would you get the augmented fwd as distinct from gradient in this case?

Okay, I'm okay with this. The other two things we may want to include in this: * whether or not to return the original return. * whether or not to...

I think the issue here is that the type is integer rather than floating point, what happens if you do 2.0 and 3.0? The justification for integers being that integers...

Also wait @oschulz, your example with floats isn't what I'd expect (see my use below). What version are you using? ``` julia> autodiff(x -> x.a * x.b, Active((a = 2.0,...

The other reason why integers are inactive by default, for example, is that most of the time you really don't want to differentiate with respect to them. E.g. they are...