GPU-enabled hosted runners

Tier: Premium, Ultimate Offering: GitLab.com

GitLab provides GPU-enabled hosted runners to accelerate heavy compute workloads for ModelOps or HPC such as the training or deployment of Large Language Models (LLMs) as part of ModelOps workloads.

GitLab provides GPU-enabled runners only on Linux. For more information about how these runners work, see Hosted runners on Linux

Machine types available for GPU-enabled runners

The following machine types are available for GPU-enabled runners on Linux x86-64.

Runner Tag vCPUs Memory Storage GPU GPU Memory
saas-linux-medium-amd64-gpu-standard 4 15 GB 50 GB 1 Nvidia Tesla T4 (or similar) 16 GB

Container images with GPU drivers

As with GitLab hosted runners on Linux, your job runs in an isolated virtual machine (VM) with a bring-your-own-image policy. GitLab mounts the GPU from the host VM into your isolated environment. To use the GPU, you must use a Docker image with the GPU driver installed. For Nvidia GPUs, you can use their CUDA Toolkit.

Example .gitlab-ci.yml file

In the following example of the .gitlab-ci.yml file, the Nvidia CUDA base Ubuntu image is used. In the script: section, you install Python. 

gpu-job:
  stage: build
  tags:
    - saas-linux-medium-amd64-gpu-standard
  image: nvcr.io/nvidia/cuda:12.1.1-base-ubuntu22.04
  script:
    - apt-get update
    - apt-get install -y python3.10
    - python3.10 --version

If you don’t want to install larger libraries such as Tensorflow or XGBoost each time you run a job, you can create your own image with all the required components pre-installed. Watch this demo to learn how to leverage GPU-enabled hosted runners to train an XGBoost model:

Video demonstration of GitLab GPU-enabled hosted runners: Train XGboost models with GitLab.