Mixed Activity of Float, Vector in reverse mode

[ptiede]

On Enzyme 0.11.7 (it seems main is broken right now?), I am having some gradients drop when I try to take a derivative w.r.t. a NamedTuple.

A MWE example is

f(p) = p.b^2 + sum(abs2, p.c)
p = (b = 2.0, c = ones(2))
dp = (b = 0.0, c = zeros(2))
autodiff(Reverse, f, Duplicated(p, dp))
@show dp
# dp = (b = 0.0, c = [2.0, 2.0])
# correct gradient (b = 4.0, c=[2.0, 2.0])

My guess is that this is because the NamedTuple has Mixed activity. If I do the following

@noinline fref(p) = f(p[])
pref = Ref(p)
dpref = Ref(dp)
autodiff(Reverse, fref, Duplicated(pref, dpref))

the gradients are correct.

[wsmoses]

Oh yeah the issue here is mixedactivity precisely. Marking it duplicated won't compute the derivative of any in-place values (since we can't update the float in db in place, passing it as a duplicated cannot let us change it).

This is hopefully described in the docs now here: https://enzymead.github.io/Enzyme.jl/dev/#Mixed-Activity

Active variables are used for immutable variables (like Float64), whereas Duplicated variables are used for mutable variables (like Vector{Float64}). Speciically, since Active variables are immutable, functions with Active inputs will return the adjoint of that variable. In contrast Duplicated variables will have their derivatives +='d in place.

This error indicates that you have a type, like Tuple{Float, Vector{Float64}} that has immutable components and mutable components. Therefore neither Active nor Duplicated can be used for this type.

Please feel free to open a PR if you think that needs more (or perhaps we could issue a warning in reverse mode?)

[ptiede]

Ok excellent, I was hoping I understood the docs. Is there a way that Enzyme could potentially alert the user to this during compilation? The docs suggest I should see an error Type T has mixed internal activity type which I don't see on the main branch.