@non_differentiable rand

[nmheim]

Really happy that ChainRules now has a @non_differentiable macro! Unfortunately it includes all rand methods, which is a bit too restrictive... there was a similar discussion about this in FluxML/Zygote.jl#234 .

MWE:

using Random
using Zygote

struct Normal
    μ
    σ
end

Random.rand(d::Normal) = randn()*d.σ + d.μ
f(a,b) = rand(Normal(a,b))
gs = Zygote.gradient(f,2,2) # returns (nothing,nothing) with ChainRules > 0.7.15

[willtebbutt]

Lol, yeah, we really shouldn't have done that 🤦 It was literally the motivating for a more flexible @non_differentiable / @nograd macro.

A PR modifying the implementation would be very welcome.

[nmheim]

Thanks for merging this so quickly!