gpu-installer

nvidia gpu installer

the main idea is installing all nvidia drivers files in one direcotry, stolen from ubuntu nvidia-driver-installer from google.
But add features/fix issues below:

1. add centos support.
2. fix Installer freezes node issue.
3. fix machine-rebooting issue: we just need to run this installer only once, not k8s daemonset.

requirements

1. nvidia gpu card inserted
2. docker installed/started(refer centos/ubuntu to get docker) and can pull image(refer this for local proxy case)
3. by now centos 7.X and ubuntu 16.04+ are supported
4. network is reachable

run

the default install directory is /var/IEF/nvidia, don't modify this script if you use the IEF service.

# fix the old wrong installed scripts, if wrong output of running nvidia-smi -L
bash nvidia-gpu-installer.sh fix

# check usage
bash nvidia-gpu-installer.sh -h

# simple run with default version 384.111, default download url https://us.download.nvidia.com/tesla/384.111/NVIDIA-Linux-x86_64-384.111.run
bash nvidia-gpu-installer.sh install

# OR: try to uninstall existing drivers without asking when installing
bash nvidia-gpu-installer.sh install -y

# OR: run with driver version 396.44, default download url https://us.download.nvidia.com/tesla/396.44/NVIDIA-Linux-x86_64-396.44.run
NVIDIA_DRIVER_VERSION=396.44 bash nvidia-gpu-installer.sh install

# OR: run with driver version 396.44 and download url http://cn.download.nvidia.com/tesla/396.44/NVIDIA-Linux-x86_64-396.44.run
NVIDIA_DRIVER_VERSION=396.44 NVIDIA_DRIVER_DOWNLOAD_URL=http://cn.download.nvidia.com/tesla/396.44/NVIDIA-Linux-x86_64-396.44.run bash nvidia-gpu-installer.sh install

# OR: run it under local proxy
export http_proxy=http://10.90.2.2:808
export https_proxy=http://10.90.2.2:808
bash nvidia-gpu-installer.sh install

# clean drivers if you don't need it
bash nvidia-gpu-installer.sh clean

FAQ

1. How to check gpu drivers are installed successfully?

# check that nvidia/nvidia-uvm ko are loaded
lsmod |grep -e nvidia -e nvidia-uvm

# check that device files are created
ls /dev/nvidiactl /dev/nvidia-uvm /dev/nvidia?

# check by nvidia-smi
/var/IEF/nvidia/bin/nvidia-smi

2. After reboot, gpu drivers are not loaded?

# check the loader script status
systemctl status nvidia-drivers-loader
if : found "Invalid module format"; then
  echo "your kernel version may change since last installation time, please switch to the old kernel version!!"
fi

3. APT reported the "E: Failed to fetch ... Hash Sum mismatch" error
   Try to run this script again, it maybe work