With this PR the whole procedure becomes more convenient for downstream packages: If the specified tag is not a ForwardDiff.Tag, the gradient computations will use ForwardDiff.Tag(tag, eltype(x)) instead. Thus the element type of x is taken into consideration automatically.
The ForwardDiff tags are by default dependent on the type of the input (see, e.g., https://github.com/JuliaDiff/ForwardDiff.jl/blob/e3670ce9055c66863f655d2bac2d6615c165d838/src/config.jl#L120 and https://www.stochasticlifestyle.com/improved-forwarddiff-jl-stacktraces-with-package-tags/). This is typically also the case for custom tags (see, e.g., again https://www.stochasticlifestyle.com/improved-forwarddiff-jl-stacktraces-with-package-tags/ and https://github.com/TuringLang/Turing.jl/blob/e32bb71accb4d71dfc5f0377984a00aa5d643c3a/src/Turing.jl#L30-L36).
Unfortunately, with the current setup in LogDensityProblemsAD one can only specify the exact
ForwardDiff.Tagand hence packages such as Turing have to compute it based on the type of the input and then forward it toADgradient(see https://github.com/TuringLang/Turing.jl/blob/e32bb71accb4d71dfc5f0377984a00aa5d643c3a/src/essential/ad.jl#L96).With this PR the whole procedure becomes more convenient for downstream packages: If the specified
tagis not aForwardDiff.Tag, the gradient computations will useForwardDiff.Tag(tag, eltype(x))instead. Thus the element type ofxis taken into consideration automatically.