Carlo Lucibello

What should we do here? I think we should take the performance hit for the correctness. > So the main concern wrt. dependencies here is not CUDA (because there's already...

Yes, we could do that, not ideal but seems the best alternative

@jgreener64 do you want to implement this plan?

I'm not aware of any recent change for the broadcast infrastructure. Could you perform the same benchmarks on some previous Zygote release? If performance is the same (which is still...

@mcabbott @willtebbutt so we want to do all this by v0.6, migrating stuff to packages and ChainRules if package owners are not willing to carry the extra dependence (which could...

NNlib and LoopVectorization excision are already on master, so we could just do whatever is left to do for SpecialFunctions and tag v0.6, without the additional complexity of having to...

I guess the most disturbing is 5., shouldn't return ```julia [1] => (parent = [1],) # this is xt :(Main.x) => [1] [1] => [1] # this is x ```...

For a user define struct we have ```julia julia> struct A; x; end julia> x = rand(2); a = A(x); julia> Base.sum(a) = sum(a.x) julia> gradient(() -> sum(a), Params([x])).grads IdDict{Any,...

Seems hard to not consider last example in https://github.com/FluxML/Zygote.jl/issues/991#issuecomment-864375988 a bug. I don't even know precisely why it happens, probably when we hit an `AbstractArray{ gradient(x -> sum(a), x) (nothing,)...

repeat adjoints are still in Zygote https://github.com/FluxML/Zygote.jl/blob/956cbcf3c572c0eb09c146189bb38b1b434634ff/src/lib/array.jl#L130 not sure why they were ported to ChainRules, probably just lack of time