Valentin Churavy

> is this going towards implementing the handling on the Julia-side like I suggested above. Partially yes, it implements the reporting part. But unreachables are still left in the code....

> (e.g. by using a global variable for the status and returns from the call stack) We can't implement the "returns from a call stack" easily since we are not...

For future reference: ``` # https://www.aanda.org/articles/aa/full_html/2016/02/aa27339-15/aa27339-15.html function JVP_Finite_Diff(F,u,v) λ = 10e-6 δ = λ * (λ + norm(u, Inf)/norm(v,Inf)) (F(u + δ .* v) - F(u)) ./ δ end ```

"any undef thing" you mean a "Vector{Float64}(undef, 0)" would work?

Split the fix here out to #2654

I also used ``` function mynorm(X) acc = zero(eltype(X)) for x in X acc += x^2 end return sqrt(acc) end ``` as a sanitiy check and it agrees with `generic_norm2`...

The last remaining issue is that the handling of `norm(x) == 0.0` currently leads to `NaN` ``` dx = [1.0] x = [0.0] autodiff(Forward, LinearAlgebra.BLAS.nrm2, Duplicated(x, dx)) # NaN autodiff(Forward,...

> and your values normally (no units) So one of the reasons why units are interesting to me is that it makes for better input validation and output presentation. I...

In particular [email protected] (released a few days ago) comes with support for DynamicalQuantities and axis conversion. https://docs.makie.org/dev/explanations/dim-converts#Experimental-DynamicQuantities.jl-support https://github.com/MakieOrg/Makie.jl/pull/5280

The DQ in Makie is kinda nice. I can set an axis to days and after manually reattaching units to the time axis it automatically adjusts the axis