Replace AbstractDifferentiation with DifferentiationInterface

[gdalle]

Fix #19

I wasn't able to run the tests because Soss hasn't been updated in 2 years, and has compat conflicts with DifferentiationInterface. The first one I ran into is FillArrays having released a v1 which Soss didn't know about.

I guess we may need to ask @csherrer to take a look at the CompatHelper PRs?

[marius311]

Many thanks, looking. I think we could drop Soss, as much as I like how fast/clean it is.

[gdalle]

I don't want to be the reason for that though

[marius311]

Before even Soss which could be dropped, on Julia 1.9 I run into

https://github.com/gdalle/DifferentiationInterface.jl/blob/0f7957d4c636722b0629096f0d4fc9e5009dced3/DifferentiationInterface/Project.toml#L55

https://github.com/JuliaDiff/FiniteDiff.jl/blob/27f31db7e4192e086a07c3fafe6394fa05214bc6/Project.toml#L30

which makes DI + FiniteDiff only installable on Julia 1.10+. But I don't see an easy way to drop FiniteDiff (which is a transitive dep of something I def need, Optim -> NLSolverBase -> FiniteDiff).

So it would seem this PR has to make MuseInference 1.10+ unless Im mistaken? (which I don't necessarily want to do)

[gdalle]

Since I started developing DI fairly recently, the compat bounds reflect the latest versions at the time, which are more recent than what you have. And the SciML ecosystem (including FiniteDiff) has been dropping Julia 1.6 faster than other packages.
In this instance, the compat bound v2.23.1 is important because DI caught a bad bug in FiniteDiff, which was then fixed in JuliaDiff/FiniteDiff.jl#186, the fix being part of https://github.com/JuliaDiff/FiniteDiff.jl/releases/tag/v2.23.1.

[marius311]

That makes sense. Maybe this is a rabbit hole you dont feel like going down, and it could well not be the only thing, but perhaps below was too aggressive of a julia version bump? Couldnt find mention anywhere of why its needed. This isnt a MuseInference thing, anything DI+FiniteDiff in the same project is going to have to be 1.10+.

JuliaDiff/FiniteDiff.jl@8ee5d7c

[gdalle]

I don't maintain FiniteDiff, despite it being in the JuliaDiff orgnaization it is mostly a SciML thing, and SciML has been dropping 1.6 like a hot potato. The right person to ask about this would be @ChrisRackauckas

[gdalle]

In any case, 1.10 will be the next LTS as soon as 1.11 comes out, so 1.6 should be droppable everywhere in a matter of months, if not weeks

[marius311]

Note this is about 1.9, not just 1.6, but in any case, I'll weight if I'm ready to drop 1.9.

[gdalle]

Hey @marius311, just checking in on my old PRs. Do you still want to explore the switch to DI?