Flux with Julia master fails with IRCode error

Specific Domains Machine Learning
1

Using the sample code from the documentation for Flux Optimizers, with Julia 1.5.2 the code works (apart from a warning about slow compilation) but fails with master on

julia> grads = gradient(() -> loss(x, y), θ)
ERROR: MethodError: no method matching Core.Compiler.IRCode(::Vector{Any}, ::Vector{Any}, ::Vector{Int32}, ::Vector{UInt8}, ::Core.Compiler.CFG, ::Vector{Core.LineInfoNode}, ::Vector{Any}, ::Vector{Any}, ::Vector{Any})
Stacktrace:
 [1] Core.Compiler.IRCode(ir::IRTools.Inner.IR)
   @ IRTools.Inner.Wrap ~/.julia/packages/IRTools/GVPoj/src/ir/wrap.jl:55
 [2] update!(ci::Core.CodeInfo, ir::IRTools.Inner.IR)
   @ IRTools.Inner ~/.julia/packages/IRTools/GVPoj/src/reflection/utils.jl:143

As the code hierarchy of Julia grows it gets harder and harder to know where the basic issue lies (Flux, Zygote, …) so I am not sure where this might be reported as an issue, if an issue it is.