float(::Dual) disagrees with Base

[mcabbott]

I was surprised by this:

using ForwardDiff: Dual, value
x1 = Dual.(1:4, 2:5, 3:6)
value.(float.(x1)) # Vector{Float16}
float.(value.(x1)) # Vector{Float64}

While one could argue that Julia should have different promotion / widening rules, it seems hard to argue that Dual types should disagree. This example caught me when trying to wrap a function which only accepts Float32 & Float64, as seems fairly common for C libraries.

Edit -- xref #419, about AbstractFloat.

And #362, about Int -> Float16. Example from here is ForwardDiff.derivative(first∘sincos, 1) isa Float16

[stevengj]

It seems like this line

ForwardDiff.jl/src/dual.jl

Lines 383 to 384 in 6b393f4

	Base.float(d::Dual{T,V,N}) where {T,V,N} = convert(Dual{T,promote_type(V, Float16),N}, d)
	Base.AbstractFloat(d::Dual{T,V,N}) where {T,V,N} = convert(Dual{T,promote_type(V, Float16),N}, d)

should just use float(V) rather than promote_type(V, Float16)