AD failure where Zygote succeeds

[marius311]

MWE:

julia> Zygote.gradient(nt -> nt.x^2, (x=1.,))
((x = 2.0,),)

julia> AD.gradient(AD.ZygoteBackend(), nt -> nt.x^2, (x=1.,))
ERROR: MethodError: no method matching adjoint(::ChainRulesCore.Tangent{NamedTuple{(:x,), Tuple{Float64}}, NamedTuple{(:x,), Tuple{Float64}}})

Closest candidates are:
  adjoint(::Union{QR, LinearAlgebra.QRCompactWY, QRPivoted}) at ~/.julia/juliaup/julia-1.7.3+0.x64/share/julia/stdlib/v1.7/LinearAlgebra/src/qr.jl:509
  adjoint(::Union{Cholesky, CholeskyPivoted}) at ~/.julia/juliaup/julia-1.7.3+0.x64/share/julia/stdlib/v1.7/LinearAlgebra/src/cholesky.jl:538
  adjoint(::Hessenberg) at ~/.julia/juliaup/julia-1.7.3+0.x64/share/julia/stdlib/v1.7/LinearAlgebra/src/hessenberg.jl:423
  ...
Stacktrace:
 [1] _broadcast_getindex_evalf
   @ ./broadcast.jl:670 [inlined]
 [2] _broadcast_getindex
   @ ./broadcast.jl:643 [inlined]
 [3] (::Base.Broadcast.var"#29#30"{Base.Broadcast.Broadcasted{Base.Broadcast.Style{Tuple}, Nothing, typeof(adjoint), Tuple{Tuple{ChainRulesCore.Tangent{NamedTuple{(:x,), Tuple{Float64}}, NamedTuple{(:x,), Tuple{Float64}}}}}}})(k::Int64)
   @ Base.Broadcast ./broadcast.jl:1075
 [4] ntuple
   @ ./ntuple.jl:48 [inlined]
 [5] copy
   @ ./broadcast.jl:1075 [inlined]
 [6] materialize(bc::Base.Broadcast.Broadcasted{Base.Broadcast.Style{Tuple}, Nothing, typeof(adjoint), Tuple{Tuple{ChainRulesCore.Tangent{NamedTuple{(:x,), Tuple{Float64}}, NamedTuple{(:x,), Tuple{Float64}}}}}})
   @ Base.Broadcast ./broadcast.jl:860
 [7] jacobian(ab::AbstractDifferentiation.ReverseRuleConfigBackend{Zygote.ZygoteRuleConfig{Zygote.Context{false}}}, f::Function, xs::NamedTuple{(:x,), Tuple{Float64}})
   @ AbstractDifferentiation ~/.julia/packages/AbstractDifferentiation/o62DE/src/AbstractDifferentiation.jl:591
 [8] gradient(ab::AbstractDifferentiation.ReverseRuleConfigBackend{Zygote.ZygoteRuleConfig{Zygote.Context{false}}}, f::Function, xs::NamedTuple{(:x,), Tuple{Float64}})
   @ AbstractDifferentiation ~/.julia/packages/AbstractDifferentiation/o62DE/src/AbstractDifferentiation.jl:48
 [9] top-level scope
   @ REPL[6]:1

Does this package expect all arguments be scalars/vectors that have an adjoint defined? Eg this works:

AD.gradient(AD.ZygoteBackend(), nt -> nt.x^2, ComponentVector(x=1.,))

(I'm hoping not because it would definitely hurt usability to not be able to use Zygote's full capability, where that is not a requirement)

AbstractDifferentiation v0.4.3
Zygote v0.6.44

[ToucheSir]

For some reason, gradient(::ReverseRuleConfigBackend, ...) falls back to jacobian instead of calling https://github.com/JuliaDiff/AbstractDifferentiation.jl/blob/v0.4.3/src/ruleconfig.jl#L10 directly. That seems somewhat concerning, is there a reason the two functions haven't been hooked up? In the meantime, I suppose you could call pullback_function yourself.

[devmotion]

This causes also #57, I noticed a while ago (#57 (comment)).

[ToucheSir]

Since there was a request for naming issues related to performance or awkward API in #88 (comment), wanted to bump this since I think it covers both. It also affects more than Zygote (as seen in the many linked issues) and even has (stalled?) PRs trying to address it.

[devmotion]

Fixed by #128. On the master branch I get

julia> Zygote.gradient(nt -> nt.x^2, (x=1.,))
((x = 2.0,),)

julia> AD.gradient(AD.ZygoteBackend(), nt -> nt.x^2, (x=1.,))
((x = 2.0,),)