Move AD Rules to SciMLSensitivity

[prbzrg]

DiffEqFlux.jl/src/DiffEqFlux.jl

Lines 24 to 40 in e32422d

	# ForwardDiff integration
	ZygoteRules.@adjoint function ForwardDiff.Dual{T}(x, ẋ::Tuple) where T
	@assert length(ẋ) == 1
	ForwardDiff.Dual{T}(x, ẋ), ḋ -> (ḋ.partials[1], (ḋ.value,))
	end
	ZygoteRules.@adjoint ZygoteRules.literal_getproperty(d::ForwardDiff.Dual{T}, ::Val{:partials}) where T =
	d.partials, ṗ -> (ForwardDiff.Dual{T}(ṗ[1], 0),)
	ZygoteRules.@adjoint ZygoteRules.literal_getproperty(d::ForwardDiff.Dual{T}, ::Val{:value}) where T =
	d.value, ẋ -> (ForwardDiff.Dual{T}(0, ẋ),)
	ZygoteRules.@adjoint ZygoteRules.literal_getproperty(A::Tridiagonal, ::Val{:dl}) = A.dl,y -> Tridiagonal(dl,zeros(length(d)),zeros(length(du)),)
	ZygoteRules.@adjoint ZygoteRules.literal_getproperty(A::Tridiagonal, ::Val{:d}) = A.d,y -> Tridiagonal(zeros(length(dl)),d,zeros(length(du)),)
	ZygoteRules.@adjoint ZygoteRules.literal_getproperty(A::Tridiagonal, ::Val{:du}) = A.dl,y -> Tridiagonal(zeros(length(dl)),zeros(length(d),du),)
	ZygoteRules.@adjoint Tridiagonal(dl, d, du) = Tridiagonal(dl, d, du), p̄ -> (diag(p̄[2:end,1:end-1]),diag(p̄),diag(p̄[1:end-1,2:end]))

Related to SciML/SciMLSensitivity.jl#582

[ChrisRackauckas]

Those rules should just be removed or upstreamed. @oxinabox @DhairyaLGandhi are any of those required these days?

[prbzrg]

What's update on this?

[ChrisRackauckas]

Tridiagonal rules moved with JuliaDiff/ChainRules.jl#758

[ChrisRackauckas]

Oh, these all were removed awhile ago.