Mike J Innes

Both MathLink and WSTP refer to the implementation rather than the purpose of the package (which made some sense when Mathematica.jl existed as a separate package on top of this,...

I really just think docs strings would be better here. It's a case of KISS until there's a clear need for any more complexity.

That's a good use case, although even then, wasn't #14 meant to implement something fairly general and language-agnostic? It seems better to have the same default for all languages if...

Thanks a lot for those links, that's really useful. To paraphrase the discussion there, it sounds like the most important reason is that C++ needs abort semantics where execution ends...

Excellent, thanks for the additional insight. I agree that traps are fatal at the task level, and signal a bug somewhere. > [traps] are indication of a bug not in...

> For a language without undefined behavior, a trap would typically signify a code generator bug. Ok I see, that clarifies things. Thanks again all for taking the time to...

Yes, I think we're on the same page, and I'm glad to see that throwing variants are on the table. FWIW, some languages may also want to turn traps into...

No, there isn't currently, at least not directly. The best way to do this is instead to create a normal-argument wrapper around the function and differentiate that: ```julia _foo(a, b)...

Is the issue that the gradient for `normalise` ends up incorrect because the gradient for `std` is incorrect (with respect to the kwarg `mean`)? If so that's a Zygote bug...

Zygote can actually get gradients of kwargs, so it's not an issue there as such; the problem is in [this definition](https://github.com/FluxML/Zygote.jl/blob/0e25138a3c3c88600de2a4ae2999f5a5d4676f02/src/lib/array.jl#L224) for `var` (along with `std` below) which assumes the...