SVD is broken in Zygote

[MikeInnes]

e.g.

gradient(x -> sum(svd(x).S), rand(10,10))

fails due to Zygote's NamedTuples not being supported by the rrule for this function.

Would be nice to have a type alias for Composite/NamedTuple so that rrules can be generic across both. Perhaps for now we can just remove the type restriction, since no other adjoint for SVD is possible anyway.

[nickrobinson251]

Is this fixed by removing the restriction to Composite here?

ChainRules.jl/src/rulesets/LinearAlgebra/factorization.jl

Line 10 in 3a969a2

function svd_pullback(Ȳ::Composite{<:SVD})

If so, I think removing that restriction is fine (likewise for the cholesky rrule).

Although i think we may also want to revisit adopting Composite in Zygote.jl FluxML/Zygote.jl#603

[MikeInnes]

Yes, I believe that would do the job. I agree we should talk about that but obviously would love to fix the regression before having the design discussion.

[oxinabox]

we don't need to support NamedTuple we need to support Composite{Any}