This example shows how to run SkyPilot on AWS SageMaker HyperPod with EKS.

• An existing SageMaker HyperPod with EKS (or you can create one with AWS doc)
• SkyPilot installed: installation doc

pip install skypilot-nightly[kubernetes]

1. Get your EKS cluster name and region from the SageMaker HyperPod console.

aws eks update-kubeconfig --name $EKS_CLUSTER_NAME --region $EKS_REGION

2. Connect SkyPilot to the cluster.

sky check k8s

If you are using it with SkyPilot for the first time, you may see a hint to create GPU labels for your nodes. Follow the instructions to create the labels.

sky gpus label

Note: This command currently only supports NVIDIA GPUs.

sky gpus list --cloud k8s

sky launch -c dev --gpus A10G

This launches a SkyPilot cluster with a single A10G GPU, and you can use it as a interactive dev environment by sshing into it or connecting to it with VSCode.

ssh dev

sky launch -c train train.yaml

This will launch a distributed training job with 2 nodes on the HyperPod cluster. You can check the logs with:

sky logs train

To terminate the SkyPilot cluster, you can run:

sky down train

You can start many jobs to run them in parallel. See this doc for more details.

for i in {1..10}; do
    sky jobs launch -n job-$i train.yaml --async
done

This will launch 10 distributed training jobs on the HyperPod cluster, with automatic recovery. You can check the status of all jobs with:

sky jobs queue

Or, find the jobs in a dashboard:

sky dashboard