Michael Abbott

Can reproduce the errors shown. Note BTW that Molly only seems to work (for me) on 1.7, not on Julia 1.8. (But maybe it's an Apple M1 problem.) Without the...

Now I see. The CR rule accepts any BroadcastStyle, to handle tuples too, while the Zygote one restricts to AbstractArrayStyle. The cases where the CR rule is called all have...

This is not solved by changing the `rrule`'s signature to match the `@adjoint` rule, and reject `Broadcast.Style{Tuple}`. With https://github.com/JuliaDiff/ChainRules.jl/commit/6e383c10f8dfc0731be64a51bc3a33a6c7d21f5b you get the same error as with the `@opt_out`: ```julia julia>...

Do you know what the test errors are? Weird it doesn't print them in CI. Xref comments on testing this a https://github.com/FluxML/Zygote.jl/pull/1302 Edit: note also that https://github.com/JuliaDiff/ChainRules.jl/pull/671/files#r955087897 wants to change...

Yes, TensorOperations ought to be good at this. I believe it still decides the order of multiplication based only on the syntax, i.e. before knowing the sizes of the arrays....

I may have suggested that. But aren't we (in english) differentiating the function, not the array `x`? Maybe worth trying to line up with whatever no_gradient function we need too.

I think x is needed for Tracker to work, as it will never notice a function call which doesn't get a TrackedArray.

Hah I completely forgot that I already added that... and it's still there: https://github.com/FluxML/NNlib.jl/blob/0c8396e2f2707d4c223fb45348897eae28b62e2e/src/softmax.jl#L90-L91 It could just grow to `within_grad(x...)` or something? Maybe Tracker would like the use of it...

Can you explain more what you see this version which takes a function doing? Does it call the function, or infer something about it (like whether it returns a non-diff...

I guess I don't see what's gained by passing it more arguments. You can check `any(within_gradient, (x, y, z))`, or it could be spelled `within_grad(x...)`. You may only care about...