ansible-role-cuda

Installs CUDA

Tested with Tesla P100, K80, Tesla M40, CentOS7, Ubuntu 16.04, Cuda 7.5 and 8.0

Optionally also installs cuda_init which initializes the GPUs during boot.

Requirements

Outbound access to http://developer.download.nvidia.com/compute/cuda/repos/

Role Variables

gpu: False
cuda_packages:
 - cuda
cuda_restart_node_on_install: True
cuda_init: True
cuda_bash_profile: True

• gpu: True is needed. Without it this role does nothing.
• cuda_packages: List that can be updated to include more packages that are installed after nvidia cuda repo is installed, or to a specific cuda version like cuda_packages: [ "cuda=10.0.130-1" ]
• cuda_init: Installs a bash script that is executed via systemd
• cuda_gpu_name0: "/dev/nvidia0" # set this to the device ansible looks for. If it does not exist then if cuda_init is True then it will run the cuda_init.sh script
• cuda_restart_node_on_install: restarts the system when packages are installed or updated

Example Playbook

playbook.yml:

- name: install a cuda with a specific version
  hosts: deep_learning
  vars:
    - cuda_packages:
        - cuda=10.0.130-1
  roles:
    - fgci.cuda

inventory:

[deep_learning]
host1.example gpu=True

Example Errors

This error means you are not using a supported OS (like Ubuntu 18.10 which does not have a cuda URL)

   "msg": "No file was found when using with_first_found. Use the 'skip: true' option to allow this task to be skipped if no files are found"

Tests in playbook

Tests are using this one: https://github.com/lae/ansible-role-travis-lxc

License

MIT

Author Information