Simone Carlo Surace

I don't think so.

+1 for merging this

I'm really looking forward to this feature. The convenience of being able to test locally with the most recent development version will be nice.

Would it be appropriate to add some overloads for `Diagonal{AbstractGPUArray}` analogous to those for `cholesky` I added [here](https://github.com/JuliaGPU/GPUArrays.jl/pull/384)?

Actually, for diagonal matrices we can just use broadcasting syntax. I'll open a PR.

This is a sure sign that I don't actually understand how Zygote works. Where in the call stack would it matter whether or not `MyPDMat` is a subtype of `AbstractMatrix`?...

Making it a subtype of `AbstractMatrix` indeed makes it fail: ```julia struct MyOtherPDMat{T

> The usual approach for fixing these errors is defining an rrule or a projector with CR, as Will discussed in the linked PR. Which is the function that needs...

Thanks, this sounds great! Any tips on how to find out efficiently which rrule is problematic?

I wasn't able to figure out which rrule to opt out of. However, I came up with an rrule that allows me to differentiate through the `rrule` interface directly. I...