David Widmann

Which definitions? ChainRules? LogExpFunctions contains already derivatives for all functions defined in LogExpFunctions.

Nobody has tested it but they should be as they only involve basic functions or functions from LogExpFunctions for which rules are defined: https://github.com/JuliaStats/LogExpFunctions.jl/blob/master/src/chainrules.jl It might be more efficient though...

Mutation of the primal result of `softmax` leads to an incorrect pullback. The `copy` ensures that the pullback is always correct, regardless of downstream computations.

Can you try different older Zygote versions and check if it was introduced in some version? Unfortunately, in my experience the Zygote-ChainRules integration is quite unstable and it is not...

No clue, I saw so many mysterious bugs that I became tired of hunting them down and gave up on trying to understand all of them :smile: Based on the...

The main issue was fixed upstream, only some interesting issue due to `Zygote._zero` with `ChainRulesCore.NotImplemented` for matrices of Skellam variates is left: https://github.com/TuringLang/DistributionsAD.jl/pull/172

Great! This sounds definitely useful. It would be even better to add an adjoint for ChainRules instead of Zygote (ChainRules is the new way of defining forward and reverse mode...

Yeah, I've run into this issue before. But if it is only since the implementation > relies on running AD on the incomplete gamma function wouldn't it be even better...

It seems this is caused by a call of `cdf(Gamma(...), ...)` or something similar? Such calls are forwarded to `gammacdf` in StatsFuns. In StatsFuns >= 1.0.0 we use Julia implementations...

No, `MvNormalCanon` (which you are referring to) is not supported yet. Would be a valuable addition.