Nvidia-docker: Docker build issues.

[R1234A]

OS - Ubuntu(16.04)-host
GPU - GTX 1060(6GB)
GPU architecture - Pascal(I guess)

I am trying to build caffe in docker container. But getting errors and have some doubts regarding dockers.
1)My host machine is taking by default 384.130 but according to https://github.com/NVIDIA/nvidia-docker/wiki/CUDA I have to downgrade to 375 version according to the link chart in my host machine. Is this right way ?

2)Then CUDA and CUDNN is not to be installed in host machine. Is it correct?

3)Inside docker container no nvidia drivers is to be installed. Am I right?

4)I have tried 1st step and 2nd step but when creating container for caffe with GPU support giving error at running test of caffe as : "cuda driver version is insufficient for cuda runtime version docker". What is the solution??

Please help for with all answers, because I am bit confused with these versions of host and docker.(with GPU)

[3XX0]

1) 384.130 is fine if you want to run CUDA <= 9.0
2) Yes
3) Yes
4) This error is coming from your Docker image, not nvidia-docker. You can check that CUDA is working properly from within an image by running a sample

[R1234A]

I have to install cudnn inside container seprately??

[flx42]

No, cuDNN is included in some images, look at the tags we provide: https://hub.docker.com/r/nvidia/cuda/

[R1234A]

Which image should I use runtime or devel of CUDA-9.0 to build caffe with CUDA support and CUDNN support inside docker??

[flx42]

9.0-cudnn7-devel