Habana AI Driver Container

This repository contains everything needed to build a container image with a precompiled Habana AI driver. It requires to know the kernel version, since it builds and installs the driver package for a given kernel.

The image creation uses a multi-stage approach that leverages the driver-toolkit container image to build the kernel module, then install the module in the final image, along with kernel module management packages.

Manual build of the container image

Below is an example for building a driver image for the version 1.6.0-439 of the Habana AI driver and the version 4.18.0-372.19.1.el8_6 of the kernel.

export ARCH="x86_64"
export HABANA_VERSION="1.6.0-439"
export KERNEL_VERSION="4.18.0-372.19.1.el8_6"

We are using ubi8/ubi-minimal as the base image for our driver container. More precisely, we're using version 8.6 to match the kernel version. For a better traceability, we store the base image digest in the driver container image labels. So, we need to retrieve the SHA256 digest of the ubi8/ubi-minimal:8.6 image. For that, we use curl to get the image manifest headers that contains the digest in the redirect URL.

export RHEL_VERSION="8.6"
export BASE_DIGEST=$(curl -sI \
    -H "Accept: application/vnd.docker.distribution.manifest.v2+json" \
    "https://registry.access.redhat.com/v2/ubi8/ubi-minimal/manifests/${RHEL_VERSION}" \
    | grep -i "^Location: " | rev | cut -d '/' -f 1 | rev)

We also need the Git digest of the commit used by Habana AI to build the driver RPM. We can extract it from the DKMS configuration file shipped within the RPM. We use rpm2cpio and cpio to read the dkms.conf file.

export HABANA_GIT_SHA=$(rpm2cpio \
    https://vault.habana.ai/artifactory/rhel/8/${RHEL_VERSION}/habanalabs-${HABANA_VERSION}.el8.noarch.rpm \
    | cpio -i --to-stdout ./usr/src/habanalabs-${HABANA_VERSION}/dkms.conf 2>/dev/null \
    | grep "KMD_LAST_GIT_SHA=" | cut -d "=" -f 2)

With all these variables set, we can now proceed to building the driver container image with podman build.

podman build \
    --build-arg ARCH=${ARCH} \
    --build-arg BASE_DIGEST=${BASE_DIGEST} \
    --build-arg HABANA_VERSION=${HABANA_VERSION} \
    --build-arg HABANA_GIT_SHA=${HABANA_GIT_SHA} \
    --build-arg KERNEL_VERSION=${KERNEL_VERSION} \
    --build-arg RHEL_VERSION=${RHEL_VERSION} \
    --tag ghcr.io/fabiendupont/habana-ai-driver:${HABANA_VERSION}-${KERNEL_VERSION}.${ARCH} \
    --file Dockerfile .

The resulting container image is much smaller than the driver-toolkit image, with only 143 MB. We could get a smaller image with ubi8/ubi-micro base image, but that would require more hacks to install the kernel module management commands.

For that image to be usable in our OpenShift clusters, we simply push it to ghcr.io.

podman login ghcr.io
podman push ghcr.io/fabiendupont/habana-ai-driver:${HABANA_VERSION}-${KERNEL_VERSION}.${ARCH}

Maintain a library of Habana AI driver images

TODO