wrong gradient type when dividing by integer

[CarloLucibello]

When performing division of a Float32 by an Int I get a Float64 gradient

julia> using Zygote

julia> gradient(x -> x/2, 1f0)
(0.5,)

@oxinabox @MikeInnes @sethaxen @DhairyaLGandhi any clue of when was this introduced?

This is causing FluxML/Flux.jl#1269 and possibly also FluxML/Flux.jl#1255

[CarloLucibello]

Is this the rule used? Should we replace it with

@scalar_rule x / y (one(x)/y, -((x / y) / y))

or something similar?

[sethaxen]

Yeah, that's the culprit:

julia> z, back = rrule(/, 1f0, 2)
(0.5f0, ChainRules.var"#575#/_pullback#149"{Float32,Int64}(1.0f0, 2))

julia> back(1)
(Zero(), 0.5, -0.25f0)

Your proposed change looks right to me. inv is just too problematic for promoting everything to Float64 or higher, and we should probably just remove it from all @scalar_rrules.

[mcabbott]

Also fixed in JuliaDiff/DiffRules.jl#46, but without a test it seems.

[DhairyaLGandhi]

+1 on removing dependency on inv