Kyurae Kim

I've now added the pre-packaged location-scale family. Overall, to the user, the basic interface looks like the following: ```julia μ_y, σ_y = 1.0, 1.0 μ_z, Σ_z = [1.0, 2.0], [1.0...

@devmotion That's unfortunate... One thing I noted for `AdvancedVI.jl` is that we can't reuse all of the nice machinery behind `LogDensityProblemsAD`, and rolling our own AD interface definitely does not...

> What's the reasoning behind moving away from the current design here? > > In particular I mean, why don't we keep the algorithms and stuff like > > ```julia...

Hi @torfjelde , I've added the interface for stateful objectives. Like `Optimisers.jl`, it's supposed to be passed around as a tuple. For ADVI, I think only the control variates will...

I think most of the things are there except for convergence diagnostics/early termination stuff, which I think isn't very urgent.

Hi @yebai, One of the failing tests is a [critical bug](https://github.com/JuliaDiff/ReverseDiff.jl/issues/235) on `ReverseDiff`'s side. I could nudge the test so that it doesn't fail anymore.

@yebai Those are coming from the fp32 tests against a full-rank Gaussian posterior. I think those are not critical; Maybe I'll just remove fp32 tests.

@yebai I think the x86 error is a `ReTest` bug. I found that the [failing function](https://github.com/JuliaTesting/ReTest.jl/blob/4831b0fc23f2897bbeb999de9bdd14e968653199/src/ReTest.jl#L304) takes an `Int64` but, the [caller](https://github.com/JuliaTesting/ReTest.jl/blob/4831b0fc23f2897bbeb999de9bdd14e968653199/src/ReTest.jl#L1488) shoves in an integer literal, which is an...

I'll also remove the tests for `FullMonteCarlo` since it's useless.

Currently in the [works](https://github.com/FluxML/Zygote.jl/issues/1448) of enabling the inference tests for `Zygote`, which is currently broken for `Stacked` bijectors