CUDABackend throws an error for empty array

[sshin23]

When KA is used with CUDABackend, it throws an error when the array is empty. MWE here:

using KernelAbstractions, CUDA

@kernel function add1(I)
    i = @index(Global)
    I[i] += 1
end

a = zeros(0)
add1(CPU())(a; ndrange = length(a)) # ok

a = CUDA.zeros(0)
add1(CUDABackend())(a; ndrange = length(a)) # throws an error

This is the stack trace:

ERROR: DivideError: integer division error
Stacktrace:
  [1] div
    @ ./int.jl:295 [inlined]
  [2] div
    @ ./div.jl:310 [inlined]
  [3] div
    @ ./div.jl:355 [inlined]
  [4] fld
    @ ./div.jl:321 [inlined]
  [5] mod
    @ ./int.jl:287 [inlined]
  [6] JuliaGPU/KernelAbstractions.jl#4
    @ ~/.julia/packages/KernelAbstractions/zPAn3/src/nditeration.jl:121 [inlined]
  [7] ntuple
    @ ./ntuple.jl:48 [inlined]
  [8] partition
    @ ~/.julia/packages/KernelAbstractions/zPAn3/src/nditeration.jl:119 [inlined]
  [9] partition
    @ ~/.julia/packages/KernelAbstractions/zPAn3/src/KernelAbstractions.jl:628 [inlined]
 [10] launch_config
    @ ~/git/CUDA.jl/src/CUDAKernels.jl:88 [inlined]
 [11] (::KernelAbstractions.Kernel{…})(args::CuArray{…}; ndrange::Int64, workgroupsize::Nothing)
    @ CUDA.CUDAKernels ~/git/CUDA.jl/src/CUDAKernels.jl:108
 [12] top-level scope
    @ REPL[10]:12
 [13] top-level scope
    @ ~/git/CUDA.jl/src/initialization.jl:206
Some type information was truncated. Use `show(err)` to see complete types.

Currently, I'm on julia v1.10.2 and

[052768ef] CUDA v5.1.2
[63c18a36] KernelAbstractions v0.9.18

[maleadt]

@vchuravy Why the transfer? None of the back-ends handle this, even KA's own CPU backend, so maybe partition should deal with this?