install-nvidia-cuda

Install pre-compiled or dkms Nvidia drivers and cuda-toolkit on:

• RHEL8
• RHEL9
• Ubuntu 22.04
• Ubuntu 24.04
• Debian 12

Other distros in these families may also work.

Based on the instructions from:

• https://docs.nvidia.com/datacenter/tesla/driver-installation-guide/
• https://docs.nvidia.com/cuda/cuda-installation-guide-linux/

Requirements

See Nvidia docs for system requirements.
Drivers and tool-kit can take up 10Gb+ on some systems.

Role Variables

• driver_version - Default propriety.
  open for the latest open drivers.
  propriety for the latest propriety drivers.
  Or another value from https://docs.nvidia.com/datacenter/tesla/driver-installation-guide/index.html#precompiled-streams.
  Open drivers require a 'Turing' (RTX 20xx, 16xx, etc.) or newer Nvidia GPU.
• allow_autoremove - Default false. If autoremove should run when packages are installed or removed.
• driver_packages - Default true. Install (or remove) Nvidia driver packages.
• cuda_packages - Default true. Install (or remove) the CUDA toolkit packages.

Dependencies

• community.general.dnf_config_manager module

Example Playbook

Can test the ouput of lspci to only run on nodes with Nvidia GPU hardware. The propriety, pre-compiled drivers should work with older, pre-Turing, GPUs if required.

---
- name: Install CUDA on nodes with GPUs
  hosts: all
  become: true
  gather_facts: true
  pre_tasks:
  - name: Check which nodes have NVIDIA GPUs
    register: lspci_result
    changed_when: false
    check_mode: false
    ansible.builtin.command:
      cmd: "lspci"

  tasks:
  - name: Install CUDA
    vars:
      driver_version: propriety
    when: lspci_result['stdout'] | regex_search('vga.*nvidia', ignorecase=true, multiline=false)
    ansible.builtin.include_role:
      name: install_nvidia_cuda

License

BSD-3-Clause