This project contains an up-to-date K3S image to run NVIDIA workloads on Kubernetes following the K3d official documentation.
The native K3S image is based on Alpine but the NVIDIA container runtime is not supported on Alpine. Therefore, we need to build a custom image based on Ubuntu that supports it.
First, you need to install the NVIDIA Container Toolkit on your machine.
Also make sure that you have zstd installed, since the last versions of K3S use it to generate a embeedded tarball.
Run the build.sh script, which takes the K3S git tag as argument, by default it uses: release-1.21.
The script performs the following steps:
- pulls K3S
- builds K3S
- build the custom K3S Docker image
The resulting image is tagged as k3s-gpu:<version tag>. The version tag is the git tag but the '+' sign is replaced with a '-'.
#!/bin/bash
set -e
cd $(dirname $0)
K3S_TAG="${1:-release-1.21}"
IMAGE_TAG="${K3S_TAG/+/-}"
if [ -d k3s ]; then
rm -rf k3s
fi
git clone --depth 1 https://github.com/rancher/k3s.git -b $K3S_TAG
cd k3s
mkdir -p build/data && ./scripts/download && go generate
SKIP_VALIDATE=true make
cd ..
unzstd k3s/build/out/data.tar.zst
docker build -t k3s-gpu:$IMAGE_TAG .You can run a container based on the new image with Docker:
docker run --name k3s-gpu -d --privileged --gpus all k3s-gpu:release-1.21Deploy a test pod:
docker cp cuda-vector-add.yaml k3s-gpu:/cuda-vector-add.yaml
docker exec k3s-gpu kubectl apply -f /cuda-vector-add.yaml
docker exec k3s-gpu kubectl logs cuda-vector-addTou can use the image with k3d:
k3d cluster create --no-lb --image k3s-gpu:release-1.21 --gpus allDeploy a test pod:
kubectl apply -f cuda-vector-add.yaml
kubectl logs cuda-vector-addMost of the work in the repo is based on the official K3d documentation and the following articles: