Michael Abbott

> the lower cutoff x find_worst(deriv_telu_fast, deriv_telu_exact, -0.1:0.0001f0:1) # with abs(x) < 0.01 (302129306388, 0.009999997221166262) julia> find_worst(deriv_telu_fast, deriv_telu_exact, -0.1:0.0001f0:1) # with abs(x) < sqrt(eps(T)) (3, -0.04700000133889262) ``` I wonder a...

This is #2290, for which there are [two PRs](https://github.com/FluxML/Flux.jl/pulls?q=is%3Apr+is%3Aopen+Lecun+), which seem to have fallen through the cracks. #2311 looks OK, needs tests.

~~The code you sketch looks like it'll return a 3-array `z[i,j,b]`, but I think you intend to sum i and j? If so,~~ sorry, I misread. There is nothing built...

I'm somewhat surprised by these performance numbers, and recommend checking them on other systems, other sizes. To add something like this to NNlib, the minimum standard is that it: *...

Seems fine. Is it possible to add a test on CI somehow, perhaps in https://github.com/FluxML/NNlib.jl/blob/master/test/ext_cuda/scatter.jl ?