require cudnn takes 10 minutes on a Volta with 1 GPU (Cuda 9, cudnn 7.1)

[ajhool]

The following command takes 10 minutes:

    #mycode.lua
    print('params.backend == cudnn, require cudnn')
    require 'cudnn'
    print('cudnn required')

require nn takes a split second.

And am building it using docker:

FROM nvidia/cuda:9.0-cudnn7-devel-ubuntu16.04 as base
...
RUN git clone https://github.com/soumith/cudnn.torch.git -b R7 && cd cudnn.torch && \
    luarocks make cudnn-scm-1.rockspec

#385 also talks about a slow loading time due to gpu initialization. Does setting the # of visible GPUs to zero affect runtime?