Missing / incorrect impl of Base.parent for LogDensityProblemsAD.jl interface

[torfjelde]

LogDensityProblemsAD.jl implements Base.parent for ADGradientWrapper (https://github.com/tpapp/LogDensityProblemsAD.jl/blob/449e5661bc2667f7bef061e148a6ea5526cbb427/src/LogDensityProblemsAD.jl#L34) by just accessing the ℓ field, which is assumed to return the original logdensity passed to ADgradient.

IIRC that field is not the original logdensity but a Tapir.CoDual? Hence it might be worth overloading Base.parent to return the "correct" thing, or alternatively change what the ℓ field represents.

Lack of this will cause issues in applications where one wants to re-construct the gradient wrapper with a new logdensity programmatically, e.g. TuringLang/Turing.jl#2231 (and more specifically, TuringLang/Turing.jl#2231 (comment)).

[torfjelde]

Was about to make a PR myself, but figured I should first check if @willtebbutt is happy with this:)

[yebai]

The issue has also been noted in TuringLang/Turing.jl#2289 (comment). @torfjelde, feel free to create a PR.

[willtebbutt]

Ah, interesting. Is the ℓ field technically part of the public interface for ADGradientWrapper? This wasn't clear to me from the docs of the package, so I assumed that I had the freedom to choose my own field names / do what I like with the fields.

In my view, adding a method of Base.parent makes most sense. You should use Tapir.primal to extract the bit of the ℓ field that you're interested in.

[torfjelde]

Ah, interesting. Is the ℓ field technically part of the public interface for ADGradientWrapper? This wasn't clear to me from the docs of the package, so I assumed that I had the freedom to choose my own field names / do what I like with the fields.

Tbh, I don't know 😅 I read the src code as Base.parent being part of the API and the default to ℓ just being a, well, default 🤷 So in agreement with what you're saying: overload Base.parent seems to be the way to go

[willtebbutt]

Version 0.2.25 of Tapir.jl should resolve this. Could someone please check?

[yebai]

I triggered a rerun of TuringLang/Turing.jl#2289.

[yebai]

It seems there are additional places that need fix: https://github.com/TuringLang/Turing.jl/actions/runs/9988691426/job/27802772815?pr=2289#step:8:775

[willtebbutt]

I've commented directly on the PR associated to that run -- it looks more like a Turing.jl issue than a Tapir.jl issue to me, but I might be misinterpreting.

[torfjelde]

You are indeed correct @willtebbutt 👍

[yebai]

Closing this since the remaining issues with Turing are caused by separate reasons.