David Widmann

I went through the CI logs once more and I think https://github.com/FluxML/Zygote.jl/actions/runs/6631670834/job/18015683033?pr=1466#step:6:818 might be problematic: There `(nothing,)` seems to be correct - at least I assumed the return value of...

I guess my assumption about the return type of `gradient` was wrong (but this is not documented anywhere, is it?): https://github.com/FluxML/Zygote.jl/blob/f7551273080569316f5127cc6a537b62ef50f592/src/compiler/interface.jl#L98

Yeah, I think such conventions should be documented more clearly and followed more consistently. IMO it would be reasonable to demand that `gradient` always returns a tuple with gradients for...

I assume they can't be removed right now since the AbstractFFT PR is not sufficient - downstream packages have to implement the adjoint functionality first. Otherwise the rules will just...

Indeed, tests fail currently since e.g. `AbstractFFTs.ProjectionStyle` is not defined for FFTW plans.

IMO ideally these rules would not be a part of Zygote at all (not even using weak dependencies) but in Distances, using a weak dependency on ChainRulesCore.

Ah, it seems that this is the intention of the issue? It was not clear to me from the comment above but the title seems to indicate that? In that...

Adding things to Distances requires the approval of thr maintainers of Distances, not of Zygote. So to me this seems to be mainly a Distances issue? Only removing rules from...

> Maybe the extension in Distances.jl could depend on Zygote being loaded? Hopefully not. ChainRulesCore is a tiny (weak) dependency whereas Zygote would be huge. > I wonder if an...

FYI I opened a PR to get feedback from maintainers of Distances: https://github.com/JuliaStats/Distances.jl/pull/246