Ansible Role: Cuda

Installs NVIDIA CUDA, cuDNN, TensorRT (optionally) and TensorFlow.

Requirements

Ubuntu 22.04 and Python 3.10.

Dependancies

The jedimt.nvidia_repo role must be installed and run on the hosts to enable downloading packages for cudnn.

Role Variables

## Behavior control variables

# Controls if the system is rebooted to complete install
reboot_systems: false

# Specify if packages should be downloaded from the Internet or
# retrieved from a local HTTP/HTTPS link. If enabled, you must
# provide alternate URLs for the CUDA and TensorRT packages.
# NOTE: The package names must not be changed.

use_local_sources: false
local_cuda_pin_url: "http://linuxdev.tme.nebulon.com/install/cuda/cuda-ubuntu2204.pin"
local_cuda_package_url: "http://linuxdev.tme.nebulon.com/install/cuda/cuda-repo-ubuntu2204-11-8-local_11.8.0-520.61.05-1_amd64.deb"
local_tensorrt_package_url: "http://linuxdev.tme.nebulon.com/install/tensorrt/nv-tensorrt-local-repo-ubuntu2204-8.6.1-cuda-11.8_1.0-1_amd64.deb"

## Ubuntu distro and architecture - used in download URLs
cuda_distro: "ubuntu2204"
cuda_arch: "x86_64"

## CUDA variables
# URL prefix for developer.download.nvidia.com
# Helps shorten lines in playbook
nvidia_url: "https://developer.download.nvidia.com/compute"

# Semantic version string used in some URLs
cuda_semantic_version: "11.8.0"

# CUDA installer package version
cuda_package_version: "520.61.05"

# Base OS packages to install prior to NVIDIA components
base_apt_packages:
  - python3-pip

# CUDA packages
cuda_apt_packages:
  - cuda-{{ cuda_version }}

## cuDNN variables
cudnn_version: "8.9.6.50"
cudnn_cuda_version: "cuda11.8"

cudnn_apt_packages:
  - "libcudnn8={{ cudnn_version }}-1+{{ cudnn_cuda_version }}"
  - "libcudnn8-dev={{ cudnn_version }}-1+{{ cudnn_cuda_version }}"
  - "libcudnn8-samples={{ cudnn_version }}-1+{{ cudnn_cuda_version }}"
  - libfreeimage3
  - libfreeimage-dev

## TensorRT variables
# Should TensorRT be installed?
install_tensorrt: true

# Version of TensorRT to install
tensorrt_url_prefix: "https://developer.nvidia.com/downloads/compute/machine-learning/tensorrt/secure"
tensorrt_version: "8.6.1"

# APT packages for TensorRT
tensorrt_apt_packages:
  - tensorrt

## Tensorflow
# Python modules to install
tensorflow_python_modules:
  - tensorflow[and-cuda]==2.13.0

# Path to the Python virtual environment
tensorflow_venv_folder: '{{ ansible_env.HOME }}/venv_tensorflow'

# Python version for Python virtual environment
venv_python_version: '3.10'

Example Playbook

# ===========================================================================
# NVIDIA Cuda installation
# ===========================================================================
- name: Install and configure NVIDIA CUDA components
  hosts: servers
  tags: play_cuda_install
  gather_facts: true

  vars_files:
    # Ansible vault with all required passwords
    - "../../credentials.yml"

  roles:
    - { role: jedimt.nvidia_container_toolkit }
    - { role: jedimt.cuda,
        skip_validation: true,
        use_local_sources: true,
        reboot_systems: true,
        cuda_semantic_version: "11.8.0",
        cuda_package_version: "520.61.05",
        cudnn_version: "8.9.6.50",
        cudnn_cuda_version: "cuda11.8",
        tensorrt_version: "8.6.1",
        venv_python_version: "3.10",
        tensorflow_python_modules: ["tensorflow[and-cuda]==2.13.0"]
    }

License

MIT

Author Information

Aaron Patten aaronpatten@gmail.com