Run Docker containers with NVIDIA GPUs on Deadline Cloud workers

The docker_nvidia_container_toolkit host configuration script installs Docker and the NVIDIA Container Toolkit on Linux service-managed fleet workers, which lets jobs run GPU containers directly with docker run. Many GPU workloads, such as ComfyUI and Stable Diffusion inference servers, ship as container images and require Docker with GPU passthrough. Containers also provide a clean way to package complex dependency stacks (CUDA, Python, application-specific libraries) without polluting the host or conflicting with other jobs on the fleet.

The script does the following:

1. Installs Docker through dnf and starts the service.

2. Adds job-user to the docker group so that jobs can run containers without sudo.

3. Installs the NVIDIA Container Toolkit from the official repository.

4. Configures the Docker daemon to use the NVIDIA runtime.

5. Generates the CDI (Container Device Interface) spec for GPU access.

6. Restarts Docker and verifies the setup.

The fleet AMI must have NVIDIA GPU drivers already installed, which Deadline Cloud GPU AMIs provide, and the fleet must use a GPU instance type such as g6.xlarge, g6e.xlarge, or p4d.24xlarge.

Once the host configuration has run, jobs can launch GPU containers. The following example runs a container with GPU access, host networking, and the standard Deadline Cloud usage-based license environment variables passed through:

docker run --rm \
  --runtime=nvidia \
  --gpus all \
  --network host \
  -e ADSKFLEX_LICENSE_FILE \
  -e FLEXLM_TIMEOUT \
  -e foundry_LICENSE \
  -e PIXAR_LICENSE_FILE \
  -e g_licenseServerRLM \
  -e redshift_LICENSE \
  -e SESI_LMHOST \
  -e VRAY_AUTH_CLIENT_FILE_PATH \
  -e VRAY_AUTH_CLIENT_SETTINGS \
  your-image:latest