Build a custom Hadron image with NVIDIA drivers

This NVIDIA Dockerfile builds a final Hadron-based image with NVIDIA drivers/tools already included. Users use it by selecting:

• Hadron version (via a build arg)
• NVIDIA driver version (via a build arg)
• an output image name (their registry/repo)

The produced output is the final layered image target (commonly named hadron-extension), which is based on Hadron and includes the NVIDIA payload.

Build arguments

The Dockerfile supports:

• HADRON_VERSION (or tag)
  The Hadron version to build against and to use as the final base image.
• NVIDIA_VERSION
  The NVIDIA driver/userspace version to include.
• JOBS
  Parallelism for compilation.

info

The Dockerfile uses the same HADRON_VERSION for both hadron-toolchain and hadron. This ensures kernel module compatibility.

Build the final “Hadron + NVIDIA” image

From the directory containing the Dockerfile:

HADRON_VERSION="v0.0.4"          # set to the Hadron release/tag you want
NVIDIA_VERSION="580.126.20"      # set as needed
IMAGE="my-registry.example.com/myteam/hadron-nvidia:${HADRON_VERSION}-${NVIDIA_VERSION}"

docker build \
  -f Dockerfile.nvidia \
  --build-arg HADRON_VERSION="${HADRON_VERSION}" \
  --build-arg NVIDIA_VERSION="${NVIDIA_VERSION}" \
  --build-arg JOBS="$(nproc)" \
  --target hadron-extension \
  -t "${IMAGE}" \
  .

Push to your registry

docker push "${IMAGE}"

Users can now consume your custom Hadron image like any other image to build artifacts using Auroraboot or to run K8s workloads with NVIDIA support

tip

Quick validation on your image:

docker run --rm -it "${IMAGE}" nvidia-smi

Actual GPU access depends on how/where it runs, but this confirms the image contains the NVIDIA tooling and runtime wiring expected by your system.