Gradient of function containing copyto! fails about zygote.jl HOT 5 OPEN

You could express this without array mutation, by e.g. using a Ref (or just returning y, ideally). I'd like to make examples like this work, but it may not happen all that soon.

from zygote.jl.

Hmm... But then it is not so clear to me how one can avoid allocating a new y each time the function is run. I mean, without array mutation.

from zygote.jl.

Now that Zygote integrates with ChainRules, the best way to handle this would be defining a custom rrule. See https://juliadiff.org/ChainRulesCore.jl/stable/writing_good_rules.html#Patterns-that-need-rules-in-[Zygote.jl](https://github.com/FluxML/Zygote.jl) for more.

from zygote.jl.

Since it's not solved and a custom rrule isn't going to scale well long term, I'm going to keep this open till we have a good answer to non allocating or in place versions of passing gradients along

from zygote.jl.

I'd rather close this in favour of the couple other issues we have about copyto! or mutational then, no point in redundancy.

from zygote.jl.