David Widmann

One could keep an explicit list of supported and tested distributions in DistributionsAD, and only define it for those. However, I think a proper fix would be to change the...

> Is this something that the Distributions folks might be receptive to do you think? I don't know but I assume they might be fine with it. I mean, it...

It seems the error arises from the use of LoopVectorization in your implementation of `logsumexp` which tries to call `typemin` which is not defined by ReverseDiff for `::Type{

Possibly also related: https://github.com/TuringLang/Bijectors.jl/issues/134

I assume it is not supported since Distributions only contains types and methods for uni-, multi-, and matrixvariate distributions. Therefore `filldist` (and `arraydist`) only support vectors and matrices of univariate...

Have you benchmarked the dependencies? I suggest doing that before duplicating any implementation. E.g., ExprTools is a very lightweight package that contains basically only `splitdef` and `combinedef` from MacroTools, and...

> What is the opinion on developers of this package about splitting the gradient code from the AD integration code into separate packages? By gradient code, do you mean (for...

> Users that just want to obtain a particular gradient associated with a distribution can use the specialized gradient code however they want (e.g. in a highly optimized implementation that...

IMO ideally DistributionsAD (which basically only contains type piracy :smile:) shouldn't exist at all. Of course, it's very understandable that nobody wants Distributions to depend on a bunch of AD...

This is part of https://github.com/TuringLang/DistributionsAD.jl/pull/95.