Install CUDA and repacke the NVIDIA kernel modules

install CUDA.md

RHEL 7

download the CUDA RPMs:

ssh -f -N -o ControlMaster=auto -D 1080 cmsusr.cms 
export ALL_PROXY=socks5://localhost:1080
curl http://developer.download.nvidia.com/compute/cuda/10.1/Prod/local_installers/cuda-repo-rhel7-10-1-local-10.1.243-418.87.00-1.0-1.x86_64.rpm -o cuda-repo-rhel7-10-1-local-10.1.243-418.87.00-1.0-1.x86_64.rpm

unpack the individual RPMs under /var/cuda-repo-10-1-local-10.1.243-418.87.00/:

sudo rpm -i cuda-repo-rhel7-10-1-local-10.1.243-418.87.00-1.0-1.x86_64.rpm

install the packages:

sudo yum clean all
sudo yum -y install kernel-devel rpm-build
sudo yum -y install nvidia-driver-latest-dkms
sudo yum -y install cuda

Note that this will also build the kernel modules:

/lib/modules/3.10.0-957.1.3.el7.x86_64/extra/nvidia-drm.ko.xz
/lib/modules/3.10.0-957.1.3.el7.x86_64/extra/nvidia.ko.xz
/lib/modules/3.10.0-957.1.3.el7.x86_64/extra/nvidia-modeset.ko.xz
/lib/modules/3.10.0-957.1.3.el7.x86_64/extra/nvidia-uvm.ko.xz

repackage the kernel modules

Rather than building the kernel modules on every machine, we can build them once, and package them in an RPM that provides the same capabilities as nvidia-driver-latest-dkms-418.87.00-1.el7.x86_64.rpm:

~fwyzard/cuda/package_nvidia_modules.sh

This will work for a single kernel at a time (so the kernel should be among the dependencies ?), but it might be good enough for the production environment.

installing the repackaged modules

# install the precompiled kernel drivers
sudo cp ~fwyzard/cuda/nvidia-driver-latest-3.10.0-957.1.3.el7.x86_64-418.87.00-1.el7.x86_64.rpm /var/cuda-repo-10-1-local-10.1.243-418.87.00/
sudo createrepo --update /var/cuda-repo-10-1-local-10.1.243-418.87.00/
sudo yum install nvidia-driver-latest-3.10.0-957.1.3.el7.x86_64-418.87.00

# prevent GCC from being installed as a dependency of NVCC
sudo rpm -i ~fwyzard/cuda/gcc-c++-dummy-4.8.5-1.el7.x86_64.rpm

# install CUDA
sudo yum install cuda