Crie e configure um HyperPod cluster com o escalonamento automático do Karpenter

Carregue a configuração do seu ambiente e extraia valores das CloudFormation pilhas.

source .env
SUBNET1=$(cfn-output $VPC_STACK_NAME PrivateSubnet1)
SUBNET2=$(cfn-output $VPC_STACK_NAME PrivateSubnet2)
SUBNET3=$(cfn-output $VPC_STACK_NAME PrivateSubnet3)
SECURITY_GROUP=$(cfn-output $VPC_STACK_NAME NoIngressSecurityGroup)
EKS_CLUSTER_ARN=$(cfn-output $EKS_STACK_NAME ClusterArn)
EXECUTION_ROLE=$(cfn-output $SAGEMAKER_STACK_NAME ExecutionRole)
SERVICE_ROLE=$(cfn-output $SAGEMAKER_STACK_NAME ServiceRole)
BUCKET_NAME=$(cfn-output $SAGEMAKER_STACK_NAME Bucket)
HP_CLUSTER_NAME="hyperpod-eks-test-$(date +%s)"
EKS_CLUSTER_NAME=$(cfn-output $EKS_STACK_NAME ClusterName)
HP_CLUSTER_ROLE=$(cfn-output $SAGEMAKER_STACK_NAME ClusterRole)

Faça upload do script de inicialização do nó em seu bucket do Amazon S3.

aws s3 cp lifecyclescripts/on_create_noop.sh s3://$BUCKET_NAME

Crie um arquivo de configuração de cluster com suas variáveis de ambiente.

cat > cluster_config.json << EOF
{
    "ClusterName": "$HP_CLUSTER_NAME",
    "InstanceGroups": [
        {
            "InstanceCount": 1,
            "InstanceGroupName": "system",
            "InstanceType": "ml.c5.xlarge",
            "LifeCycleConfig": {
                "SourceS3Uri": "s3://$BUCKET_NAME",
                "OnCreate": "on_create_noop.sh"
            },
            "ExecutionRole": "$EXECUTION_ROLE"
        },
        {
            "InstanceCount": 0,
            "InstanceGroupName": "auto-c5-az1",
            "InstanceType": "ml.c5.xlarge",
            "LifeCycleConfig": {
                "SourceS3Uri": "s3://$BUCKET_NAME",
                "OnCreate": "on_create_noop.sh"
            },
            "ExecutionRole": "$EXECUTION_ROLE"
        },
        {
            "InstanceCount": 0,
            "InstanceGroupName": "auto-c5-4xaz2",
            "InstanceType": "ml.c5.4xlarge",
            "LifeCycleConfig": {
                "SourceS3Uri": "s3://$BUCKET_NAME",
                "OnCreate": "on_create_noop.sh"
            },
            "ExecutionRole": "$EXECUTION_ROLE",
            "OverrideVpcConfig": {
                "SecurityGroupIds": [
                    "$SECURITY_GROUP"
                ],
                "Subnets": [
                    "$SUBNET2"
                ]
            }
        },
        {
            "InstanceCount": 0,
            "InstanceGroupName": "auto-g5-az3",
            "InstanceType": "ml.g5.xlarge",
            "LifeCycleConfig": {
                "SourceS3Uri": "s3://$BUCKET_NAME",
                "OnCreate": "on_create_noop.sh"
            },
            "ExecutionRole": "$EXECUTION_ROLE",
            "OverrideVpcConfig": {
                "SecurityGroupIds": [
                    "$SECURITY_GROUP"
                ],
                "Subnets": [
                    "$SUBNET3"
                ]
            }
        }
    ],
    "VpcConfig": {
        "SecurityGroupIds": [
            "$SECURITY_GROUP"
        ],
        "Subnets": [
            "$SUBNET1"
        ]
    },
    "Orchestrator": {
        "Eks": {
            "ClusterArn": "$EKS_CLUSTER_ARN"
        }
    },
    "ClusterRole": "$HP_CLUSTER_ROLE",
    "AutoScaling": {
        "Mode": "Enable",
        "AutoScalerType": "Karpenter"
    },
    "NodeProvisioningMode": "Continuous"
}
EOF

Execute o comando a seguir para criar seu HyperPod cluster.

aws sagemaker create-cluster --cli-input-json file://./cluster_config.json

O processo de criação do cluster leva aproximadamente 20 minutos. Monitore o status do cluster até que ambos ClusterStatus e AutoScaling .Status sejam exibidos InService.

Salve o ARN do cluster para operações subsequentes.

HP_CLUSTER_ARN=$(aws sagemaker describe-cluster --cluster-name $HP_CLUSTER_NAME \
   --output text --query ClusterArn)

Execute o comando a seguir para habilitar o ajuste de escala automático baseado no Karpenter em qualquer cluster preexistente que tenha o modo de provisionamento contínuo de nós.

aws sagemaker update-cluster \
    --cluster-name $HP_CLUSTER_NAME \
    --auto-scaling Mode=Enable,AutoScalerType=Karpenter \
    --cluster-role $HP_CLUSTER_ROLE

Verifique se o Karpenter foi habilitado com êxito:

aws sagemaker describe-cluster --cluster-name $HP_CLUSTER_NAME --query 'AutoScaling'

Saída esperada:

{
    "Mode": "Enable",
    "AutoScalerType": "Karpenter",
    "Status": "InService"
}