Broadcasting fails with CUDA: unsupported dynamic function invocation (call to overdub)

This code fails when using CUDADevice(). When using CPU() however, everything works just fine.
Michael Abbott identified this to be a problem with the broacasting, replacing these with map resolves the error. (slack thread)

using KernelAbstractions, CUDA
@kernel function ker(a)
	I = @index(Global, NTuple)
	a[I...] = round.(I .+ 0.5)[1]
end

f = ker(CUDADevice(), 256)
a = cu(rand(10, 10).*10)

event = f(a, ndrange=size(a))

ERROR: InvalidIRError: compiling kernel gpu_ker(Cassette.Context{nametype(CUDACtx),KernelAbstractions.CompilerMetadata{KernelAbstractions.NDIteration.DynamicSize,KernelAbstractions.NDIteration.DynamicCheck,Nothing,CartesianIndices{2,Tuple{Base.OneTo{Int64},Base.OneTo{Int64}}},KernelAbstractions.NDIteration.NDRange{2,KernelAbstractions.NDIteration.DynamicSize,KernelAbstractions.NDIteration.StaticSize{(256, 1)},CartesianIndices{2,Tuple{Base.OneTo{Int64},Base.OneTo{Int64}}},Nothing}},Nothing,KernelAbstractions.var"##PassType#253",Nothing,Cassette.DisableHooks}, typeof(gpu_ker), CuDeviceArray{Float32,2,CUDA.AS.Global}) resulted in invalid LLVM IR
Reason: unsupported dynamic function invocation (call to overdub)
Stacktrace:
 [1] copy at broadcast.jl:1023
 [2] materialize at broadcast.jl:820
 [3] macro expansion at Untitled-2:37
 [4] gpu_ker at C:\Users\jules\.julia\packages\KernelAbstractions\yw9SF\src\macros.jl:80
 [5] overdub at C:\Users\jules\.julia\packages\Cassette\158rp\src\overdub.jl:0
Reason: unsupported dynamic function invocation (call to overdub)
Stacktrace:
 [1] copy at broadcast.jl:1024
 [2] materialize at broadcast.jl:820
 [3] macro expansion at Untitled-2:37
 [4] gpu_ker at C:\Users\jules\.julia\packages\KernelAbstractions\yw9SF\src\macros.jl:80
 [5] overdub at C:\Users\jules\.julia\packages\Cassette\158rp\src\overdub.jl:0
Reason: unsupported dynamic function invocation (call to overdub)
Stacktrace:
 [1] macro expansion at Untitled-2:37
 [2] gpu_ker at C:\Users\jules\.julia\packages\KernelAbstractions\yw9SF\src\macros.jl:80
 [3] overdub at C:\Users\jules\.julia\packages\Cassette\158rp\src\overdub.jl:0
Reason: unsupported dynamic function invocation (call to overdub(overdub_context::Cassette.Context, overdub_arguments...) in Cassette at C:\Users\jules\.julia\packages\Cassette\158rp\src\overdub.jl:587)
Stacktrace:
 [1] axes at broadcast.jl:206
 [2] copy at broadcast.jl:1021
 [3] materialize at broadcast.jl:820
 [4] macro expansion at Untitled-2:37
 [5] gpu_ker at C:\Users\jules\.julia\packages\KernelAbstractions\yw9SF\src\macros.jl:80
 [6] overdub at C:\Users\jules\.julia\packages\Cassette\158rp\src\overdub.jl:0
Reason: unsupported dynamic function invocation (call to overdub)
Stacktrace:
 [1] copy at broadcast.jl:1022
 [2] materialize at broadcast.jl:820
 [3] macro expansion at Untitled-2:37
 [4] gpu_ker at C:\Users\jules\.julia\packages\KernelAbstractions\yw9SF\src\macros.jl:80
 [5] overdub at C:\Users\jules\.julia\packages\Cassette\158rp\src\overdub.jl:0
Stacktrace:
 [1] check_ir(::GPUCompiler.CompilerJob{GPUCompiler.PTXCompilerTarget,CUDA.CUDACompilerParams}, ::LLVM.Module) at C:\Users\jules\.julia\packages\GPUCompiler\4e9CU\src\validation.jl:123
 [2] macro expansion at C:\Users\jules\.julia\packages\GPUCompiler\4e9CU\src\driver.jl:241 [inlined]
 [3] macro expansion at C:\Users\jules\.julia\packages\TimerOutputs\dVnaw\src\TimerOutput.jl:206 [inlined]
 [4] codegen(::Symbol, ::GPUCompiler.CompilerJob{GPUCompiler.PTXCompilerTarget,CUDA.CUDACompilerParams}; libraries::Bool, deferred_codegen::Bool, optimize::Bool, strip::Bool, validate::Bool, only_entry::Bool) at C:\Users\jules\.julia\packages\GPUCompiler\4e9CU\src\driver.jl:239
 [5] compile(::Symbol, ::GPUCompiler.CompilerJob{GPUCompiler.PTXCompilerTarget,CUDA.CUDACompilerParams}; libraries::Bool, deferred_codegen::Bool, optimize::Bool, strip::Bool, validate::Bool, only_entry::Bool) at C:\Users\jules\.julia\packages\GPUCompiler\4e9CU\src\driver.jl:39
 [6] compile at C:\Users\jules\.julia\packages\GPUCompiler\4e9CU\src\driver.jl:35 [inlined]
 [7] _cufunction(::GPUCompiler.FunctionSpec{typeof(Cassette.overdub),Tuple{Cassette.Context{nametype(CUDACtx),KernelAbstractions.CompilerMetadata{KernelAbstractions.NDIteration.DynamicSize,KernelAbstractions.NDIteration.DynamicCheck,Nothing,CartesianIndices{2,Tuple{Base.OneTo{Int64},Base.OneTo{Int64}}},KernelAbstractions.NDIteration.NDRange{2,KernelAbstractions.NDIteration.DynamicSize,KernelAbstractions.NDIteration.StaticSize{(256, 1)},CartesianIndices{2,Tuple{Base.OneTo{Int64},Base.OneTo{Int64}}},Nothing}},Nothing,KernelAbstractions.var"##PassType#253",Nothing,Cassette.DisableHooks},typeof(gpu_ker),CuDeviceArray{Float32,2,CUDA.AS.Global}}}; kwargs::Base.Iterators.Pairs{Symbol,Int64,Tuple{Symbol},NamedTuple{(:maxthreads,),Tuple{Int64}}}) at C:\Users\jules\.julia\packages\CUDA\h38pe\src\compiler\execution.jl:308
 [8] #87 at C:\Users\jules\.julia\packages\GPUCompiler\4e9CU\src\cache.jl:21 [inlined]
 [9] get!(::GPUCompiler.var"#87#88"{Base.Iterators.Pairs{Symbol,Int64,Tuple{Symbol},NamedTuple{(:maxthreads,),Tuple{Int64}}},typeof(CUDA._cufunction),GPUCompiler.FunctionSpec{typeof(Cassette.overdub),Tuple{Cassette.Context{nametype(CUDACtx),KernelAbstractions.CompilerMetadata{KernelAbstractions.NDIteration.DynamicSize,KernelAbstractions.NDIteration.DynamicCheck,Nothing,CartesianIndices{2,Tuple{Base.OneTo{Int64},Base.OneTo{Int64}}},KernelAbstractions.NDIteration.NDRange{2,KernelAbstractions.NDIteration.DynamicSize,KernelAbstractions.NDIteration.StaticSize{(256, 1)},CartesianIndices{2,Tuple{Base.OneTo{Int64},Base.OneTo{Int64}}},Nothing}},Nothing,KernelAbstractions.var"##PassType#253",Nothing,Cassette.DisableHooks},typeof(gpu_ker),CuDeviceArray{Float32,2,CUDA.AS.Global}}}}, ::Dict{UInt64,Any}, ::UInt64) at .\dict.jl:452
 [10] macro expansion at .\lock.jl:183 [inlined]
 [11] check_cache(::typeof(CUDA._cufunction), ::GPUCompiler.FunctionSpec{typeof(Cassette.overdub),Tuple{Cassette.Context{nametype(CUDACtx),KernelAbstractions.CompilerMetadata{KernelAbstractions.NDIteration.DynamicSize,KernelAbstractions.NDIteration.DynamicCheck,Nothing,CartesianIndices{2,Tuple{Base.OneTo{Int64},Base.OneTo{Int64}}},KernelAbstractions.NDIteration.NDRange{2,KernelAbstractions.NDIteration.DynamicSize,KernelAbstractions.NDIteration.StaticSize{(256, 1)},CartesianIndices{2,Tuple{Base.OneTo{Int64},Base.OneTo{Int64}}},Nothing}},Nothing,KernelAbstractions.var"##PassType#253",Nothing,Cassette.DisableHooks},typeof(gpu_ker),CuDeviceArray{Float32,2,CUDA.AS.Global}}}, ::UInt64; kwargs::Base.Iterators.Pairs{Symbol,Int64,Tuple{Symbol},NamedTuple{(:maxthreads,),Tuple{Int64}}}) at C:\Users\jules\.julia\packages\GPUCompiler\4e9CU\src\cache.jl:19
 [12] + at .\int.jl:53 [inlined]
 [13] hash_64_64 at .\hashing.jl:35 [inlined]
 [14] hash_uint64 at .\hashing.jl:62 [inlined]
 [15] hx at .\float.jl:568 [inlined]
 [16] hash at .\float.jl:571 [inlined]
 [17] cached_compilation(::typeof(CUDA._cufunction), ::GPUCompiler.FunctionSpec{typeof(Cassette.overdub),Tuple{Cassette.Context{nametype(CUDACtx),KernelAbstractions.CompilerMetadata{KernelAbstractions.NDIteration.DynamicSize,KernelAbstractions.NDIteration.DynamicCheck,Nothing,CartesianIndices{2,Tuple{Base.OneTo{Int64},Base.OneTo{Int64}}},KernelAbstractions.NDIteration.NDRange{2,KernelAbstractions.NDIteration.DynamicSize,KernelAbstractions.NDIteration.StaticSize{(256, 1)},CartesianIndices{2,Tuple{Base.OneTo{Int64},Base.OneTo{Int64}}},Nothing}},Nothing,KernelAbstractions.var"##PassType#253",Nothing,Cassette.DisableHooks},typeof(gpu_ker),CuDeviceArray{Float32,2,CUDA.AS.Global}}}, ::UInt64; kwargs::Base.Iterators.Pairs{Symbol,Int64,Tuple{Symbol},NamedTuple{(:maxthreads,),Tuple{Int64}}}) at C:\Users\jules\.julia\packages\GPUCompiler\4e9CU\src\cache.jl:0
 [18] cufunction(::Function, ::Type{T} where T; name::String, kwargs::Base.Iterators.Pairs{Symbol,Int64,Tuple{Symbol},NamedTuple{(:maxthreads,),Tuple{Int64}}}) at C:\Users\jules\.julia\packages\CUDA\h38pe\src\compiler\execution.jl:296
 [19] macro expansion at C:\Users\jules\.julia\packages\CUDA\h38pe\src\compiler\execution.jl:108 [inlined]
 [20] (::KernelAbstractions.Kernel{CUDADevice,KernelAbstractions.NDIteration.StaticSize{(256,)},KernelAbstractions.NDIteration.DynamicSize,typeof(gpu_ker)})(::CuArray{Float32,2,Nothing}; ndrange::Tuple{Int64,Int64}, dependencies::Nothing, workgroupsize::Nothing, progress::Function) at C:\Users\jules\.julia\packages\KernelAbstractions\yw9SF\src\backends\cuda.jl:211
 [21] top-level scope at REPL[2]:1

Is this a bug, or is this just something KernelAbstractions doesn't support?
Thanks

Start with the minimal @kernel example using CUDADevice() and compare its broadcast expression with the map version mentioned in the report. Read the CUDA backend entry point in src/backends/cuda.jl and the kernel expansion referenced through src/macros.jl; done means the example compiles and runs on CUDA without the InvalidIRError while retaining the CPU behavior.