CUDA.jl along cannot trigger automatic GPU backend selection

[chengchingwen]

julia> using CUDA, MLDataDevices

julia> MLDataDevices.Internal.get_gpu_device(force = true)
ERROR: DeviceSelectionException: No functional GPU device found!
Stacktrace:
 [1] get_gpu_device(; force::Bool)
   @ MLDataDevices.Internal ~/.julia/packages/MLDataDevices/uhCbD/src/internal.jl:93
 [2] top-level scope
   @ REPL[2]:1

julia> CUDA.functional()
true

It seems get_gpu_device would return CUDADevice successfully only when cuDNN is loaded.

[avik-pal]

Yes, that is an intentional choice, since a lot of the NNlib functions will fail without cuDNN loaded anyways

[chengchingwen]

It's kinda counterintuitive. NNlib's CUDA ext does not depend on cuDNN, while NNlib's cuDNN ext depends on CUDA. cuDNN also depends on CUDA. OTOH, not all ML libraries require cuDNN functionality to use nv gpu.