fix gradient algebra on gpu

[CarloLucibello]

fix #971

I could not fix the gradient + Dict case, if someone has any idea on how to that it would be appreciated, otherwise will just leave the broken test there

[darsnack]

Does

Base.issetequal(x::Params, y::Base.KeySet) = issetequal(x.params, y)

work? What was the error with AbstractSet (may not want to define it so broadly anyways)?

[CarloLucibello]

What was the error with AbstractSet (may not want to define it so broadly anyways)?

turns AbstractSet works indeed, the problem was in the dictionary creation, this gives an error

w = cu(rand(2))
Dict(w => w)

CuArrays are not hashable and IdDict should be used instead.

I currently overload issetequal(::Params, ::AbstractSet) for generality, but I'm never sure when those kind of general method overloads would lead to ambiguities in the wild. Better go with the stricter Base.KeySet? Although that seems too limited, I think I prefer AbstractSet

[CarloLucibello]

Nightly fail unrelated and fixed in FluxML/IRTools.jl#86

[darsnack]

LGTM!

[CarloLucibello]

bors r+

[bors]

Build succeeded:

• buildkite/zygote-dot-jl