Michael Abbott

Oh right, I guess `NotImplemented` must have roughly the rules I imagined. However, my original example is trickier, since the pullback never gets called. ``` julia> function ChainRulesCore.rrule(::typeof(fill!), A::Vector, x::Number)...

I don't know the internals, maybe this could be changed? It calls the `rrule` on the forward pass, but not the pullback, I think because what's returned by `fill!` isn't...

This seems fine. How many other functions will need this? `cbrt` is currently a scalar rule. Powers are their own messy thing.

Functions like `inv`, `log` etc. are a slightly different class to `sqrt`, since the primal is infinite. The motivating case for sqrt is I think something like `f(x) = sqrt(x^2...

Xref https://github.com/JuliaDiff/ChainRules.jl/issues/383: The rules for repeat with `inner` etc. were just copied over from Zygote, and are very slow, too. My suggestion there was to re-write them using broadcasting, which...

Maybe this can be closed since this test is running? However, it seems to fail on every run, e.g. recently here: https://github.com/JuliaDiff/ChainRules.jl/runs/4903197151?check_suite_focus=true ``` 518 Got exception outside of a @test...

That's not good. I have never seen this before. Could you post what `versioninfo()` says? And perhaps exactly what version of DiffRules this is with, like `] st DiffRules` although...

Does this still happen? I'm sorry I can't reproduce this, and am not sure what to suggest. If so, does it happen with `using CUDA, DiffRules`? The CUDA.jl people may...

There is nothing built-in to deal with n-dimensional arrays. ~~But you can do something like this:~~ (No, in fact you can't) ```julia @generated function spatial_grad_square_ND(arr) out, add = [], []...

Oh I'm sorry, I thought I'd tried this but I messed up. There are indeed closures. It might be possible to make this work with https://github.com/thautwarm/GeneralizedGenerated.jl, however right now it...