Michael Abbott

From reading the log... error is from this utility function: https://github.com/FluxML/Flux.jl/blob/0e36af98f6fc5b7f3c95fe819a02172cfaaaf777/test/test_utils.jl#L38-L44 Seems to be called with 1st argument Zygote gradient, and 2nd argument Enzyme gradient: https://github.com/FluxML/Flux.jl/blob/0e36af98f6fc5b7f3c95fe819a02172cfaaaf777/test/test_utils.jl#L109 And Zygote will often...

Yes, that's right. Sorry I stopped before the conclusion! Somehow the function for doing approximate equality on nested structs needs to be made smart enough to handle the difference between...

Sorry I don't have bandwidth to think about this. If anyone else does, trying to simplify gradient tests would be very welcome. Zygote may be on the way out, so...

We could also revive #572, closed it was thought that SciML printing problems were solved in Base. Can you comment on the comparison?

Would someone care to answer this, and related comments above? > It seems quite an obscure internal detail, which should never change any results. So perhaps need not be a...

Defaulting to this thing (which really still ought to be called `HashTag` not `SmallTag`) seems fine. Mainly targeting 1.11+ seems fine by me too. If there really has to be...

> person doing the autodiff should have control Can you explain what you mean, who are these multiple people? You are envisaging that libraries which call ForwardDiff may want to...

It's not going in circles. I'm against clutter, especially user-facing clutter. If this is safe for everyone, then I don't understand the desire for every user-facing function to gain a...

I am keen not to drive the code in circles. Switching from the keyword to Preferences is some work, and if there are solid arguments against it, better to figure...

We can be civil and discuss pros and cons of all options. Ideally before coding each of them in sequence. I'm against unnecessary user-facing clutter, but open to persuasion about...