Build with mkldnn support alongside cuda

[shunr]

Comment:

Hi,

I'm currently working on torch code that should run on both cpu and cuda backends. Specifically, I'm using bfloat16 which requires the mkldnn backend to be performant on cpu. Is it possible to ship the cuda build with mkldnn turned on?

pytorch-cpu-feedstock/recipe/build.sh

Lines 121 to 123 in 9e99e03

	# Even though cudnn is used for CUDA builds, it's good to enable
	# for MKLDNN for CUDA builds when CUDA builds are used on a machine
	# with no NVIDIA GPUs. However compilation fails with mkldnn and cuda enabled.

The above comment seems to suggest that USE_MKLDNN=1 should in theory be usable with cuda, but some compilation error led to a temporary workaround.

Thanks.

[mgorny]

FYI, I've been able to enable it successfully after applying the fixes from #291.