Martin Smit

> p.s. notice that this PR needs to be re0opened due to the huge amount of code changes that have happened in the meantime. I think we should keep it...

Strangely, when I run `JuliaFormatter.format()` I get a parsing error in `src/core/model_standard.jl`. I'm using the latest version of JuliaFormatter 1.0.42, on Julia 1.9.4. The parsing error is reported to be...

When I read the docs and code for the functions in question, it seems rather that the functions I originally referred to do not do what I think they did....

Thanks for the reply. Forgive me for not understanding fully, do you think you could expand a little on how `@grad` could be used in combination with `AD` to make...

Could you give me a starting point that I could expand on? I'm not too familiar with `AbstractDifferentiation` but I'd love to build a usable MVP of this idea.

I had a quick read of the above links as well as the AbstractDifferentiation [PR about ReverseDiff](https://github.com/JuliaDiff/AbstractDifferentiation.jl/pull/29). I see that it's a relatively difficult problem to solve at such a...

Base on my (limited) understanding of the problem I think the answer is no. That said, Mohamed may have a better idea to deal with it. Maybe Julia can do...

After a call with Mohamed, I think the implementation we decided to try out will address `scan` quite nicely. I'll write up what we discussed here so it's public. For...

My previous comment was discussing the compiled tape in an uncompiled tape case, but the uncompiled tape in a compiled tape is easier to address. I'm leaving this comment as...

Ok, I've got a draft implementation for defining cached sub-tapes: ```julia import AbstractDifferentiation as AD using ReverseDiff using ReverseDiff: @grad, compile, GradientTape import AbstractDifferentiation: primal_value, pullback_function, value_and_pullback_function struct CachedReverseDiffBackend{F,T}