support varargs in @non_differentiable

[oxinabox]

follow up to #207
it would be useful to support varargs.

[gxyd]

Currently when I do:

jl> using ChainRules
jl> function h(a, x::Vararg{Any}; kwargs...)
        return a
    end
jl> @non_differentiable h(::Vararg{Any})
jl> rrule(h, 1, 2)[2](1)
(Zero(), DoesNotExist(), DoesNotExist())

Doesn't it mean that it does support Vararg's? Or am I missing something?

[mzgubic]

I can't reproduce your results, for me it errors:

julia> @non_differentiable h(::Vararg{Any})
ERROR: LoadError: @non_differentiable does not support Varargs like: ::Vararg{Any}
Stacktrace:
 [1] error(::String) at ./error.jl:33
 [2] @non_differentiable(::LineNumberNode, ::Module, ::Any) at /Users/mzgubic/Projects/ChainRules.jl/dev/ChainRulesCore/src/rule_definition_tools.jl:291
in expression starting at REPL[4]:1

What version of ChainRules and julia were you using?

Have you defined other methods of h in the REPL before?

[gxyd]

Sorry, you'll have to uncomment this block of code: https://github.com/JuliaDiff/ChainRulesCore.jl/blob/master/src/rule_definition_tools.jl#L290-L292

Have you defined other methods of h in the REPL before?

No, that's probably not the case.

[mzgubic]

Have you made any other changes?

I get:

julia> rrule(h, 1, 2)[2](1)
(Zero(), DoesNotExist())

And that is regardless of how many arguments I pass. Which I think makes sense, because it just treats the Vararg as if it was a normal argument (since it is not supported at present).

[gxyd]

Sorry, I think I made a few changes to the code at that point of time.