Support non-Array AbstractVectors for AutoEnzyme hess and fgh!

[wsmoses]

Found this while doing the second order scimlsensitivity testing

In OptimizationEnzymeExt.jl (

lines 157–165
), the Hessian (hess) and gradient-Hessian (fgh!) builders construct the shadow vectors using plain Array(r) and zeros(eltype(x), length(x)):

julia
# In OptimizationEnzymeExt.jl (CURRENT CODE)
vdθ  = Tuple((Array(r) for r in eachrow(I(length(x)) * one(eltype(x)))))  # NTuple{N, Vector}
bθ   = zeros(eltype(x), length(x))                                        # Vector
vdbθ = Tuple(zeros(eltype(x), length(x)) for i in eachindex(x))            # NTuple{N, Vector}
When x (or θ) is a custom container type like ComponentVector, θ has type ComponentVector{...}, but its shadows in vdθ, bθ, and vdbθ are plain Vectors.

Enzyme requires BatchDuplicated(primal, shadow) to have the exact same type for both primal and shadow elements (T and NTuple{N, T}). Passing Vector shadows for a ComponentVector primal causes Enzyme.BatchDuplicated to fail with a MethodError.

[ChrisRackauckas]

This isn't correct, it's coercing into an array which for example would not be the correct behavior on GPUArrays.

[wsmoses]

this is the existing behavior here:

        vdθ = Tuple((Array(r) for r in eachrow(I(length(x)) * one(eltype(x)))))

already does the conversion [we just need to match this for both primal+shadow]

[ChrisRackauckas]

well then that's the incorrect part that should get fixed.

[wsmoses]

how? you cannot guarantee the primal/shadow are the same type [and index offsets] in a nice way that I can think of offhand otherwise (ofc longer term you can make a helper to adapt the array type, but that's a follow up)

[ChrisRackauckas]

make_zero, Adapt.jl, ArrayInterface.restructure

[wsmoses]

I don't think any of those work here, but I'll let you continue it

[ChrisRackauckas]

Should test other things like GPUArrays, StaticArrays, ComponentArrays, etc.

[wsmoses]

I'll let you take over this PR then

[ChrisRackauckas]

that's fine.

[wsmoses]

also for context iirc this was needed for the second order ode tests in scimlsensitivity