Overview

This doc is a quick how-to, in order to use ansible to configure A100 GPUs for MIG usage.

Refer to MIG-Kubernetes - CloudKB - Confluence (atlassian.net) for more information.

 Instructions

1. Spin up an instance with an A100

2. Do the following steps from an existing machine with Ansible / Public keys:

   git clone https://github.com/NVIDIA/deepops.git cd deepops # Configure env, this will prompt for sudo - feel free to inspect script ./scripts/setup.sh

3. Update the following in the config/inventory file to add a single GPU host:

   [all] gpu01 ansible_host=172.16.x.y # Bottom of the file [all:vars] ansible_user=ubuntu # or rocky

4. Select a profile in config/group_vars/all.yml based on the available profiles in config/nvidia-mig-config.yml:

   mig_manager_profile: all-1g.10gb # To deploy 7x10gb slots

   (If required, per host profiles can be set in the config/inventory instead:)

   [all] gpu01 ansible_host=172.16.x.y mig_manager_profile=all-1g.10gb gpu02 ansible_host=172.16.x.y mig_manager_profile=all-1g.20gb

5. Deploy the GPU Drivers and MIG with the selected profile:

   source /opt/deepops/env/bin/activate ansible-playbook playbooks/nvidia-software/nvidia-cuda.yml ansible-playbook playbooks/nvidia-software/nvidia-mig.yml --become

 Related articles

MIG-Kubernetes - CloudKB - Confluence (atlassian.net)