Terjemahan disediakan oleh mesin penerjemah. Jika konten terjemahan yang diberikan bertentangan dengan versi bahasa Inggris aslinya, utamakan versi bahasa Inggris.

Prasyarat

aws sagemaker create-cluster \
  --cluster-name my-mig-cluster \
  --orchestrator 'Eks={ClusterArn=arn:aws:eks:region:account:cluster/cluster-name}' \
  --instance-groups '{
    "InstanceGroupName": "gpu-group",
    "InstanceType": "ml.p4d.24xlarge",
    "InstanceCount": 1,
    "LifeCycleConfig": {
       "SourceS3Uri": "s3://my-bucket",
       "OnCreate": "on_create_script.sh"
    },
    "KubernetesConfig": {
       "Labels": {
          "nvidia.com/mig.config": "all-1g.5gb"
       }
    },
    "ExecutionRole": "arn:aws:iam::account:role/execution-role",
    "ThreadsPerCore": 1
  }' \
  --vpc-config '{
     "SecurityGroupIds": ["sg-12345"],
     "Subnets": ["subnet-12345"]
  }' \
  --node-provisioning-mode Continuous

{
  "ClusterName": "my-mig-cluster",
  "InstanceGroups": [
    {
      "InstanceGroupName": "gpu-group",
      "InstanceType": "ml.p4d.24xlarge",
      "InstanceCount": 1,
      "KubernetesConfig": {
        "Labels": {
          "nvidia.com/mig.config": "all-2g.10gb"
        }
      },
      "ExecutionRole": "arn:aws:iam::account:role/execution-role"
    }
  ],
  "Orchestrator": {
    "Eks": {
      "ClusterArn": "arn:aws:eks:region:account:cluster/cluster-name"
    }
  },
  "NodeProvisioningMode": "Continuous"
}

helm install gpuo helm_chart/HyperPodHelmChart/charts/gpu-operator \
-f helm_chart/HyperPodHelmChart/charts/gpu-operator/regional-values/values-{$AWS_REGION}.yaml \
-n kube-system

kubectl get pods -n kube-system | grep -E "(gpu-operator|nvidia-)"

helm upgrade dependencies helm_chart/HyperPodHelmChart \
--set nvidia-device-plugin.devicePlugin.enabled=false \
-n kube-system

kubectl describe nodes | grep "nvidia.com/gpu"

# List all nodes in the instance group
kubectl get nodes -l sagemaker.amazonaws.com/instance-group-name=INSTANCE_GROUP_NAME

# Example:
kubectl get nodes -l sagemaker.amazonaws.com/instance-group-name=p4d-group

# Cordon a single node
kubectl cordon NODE_NAME

# Example:
kubectl cordon hyperpod-i-014a41a7001adca60

# Drain the node (ignore DaemonSets and evict pods)
kubectl drain NODE_NAME \
  --ignore-daemonsets \
  --delete-emptydir-data \
  --force \
  --grace-period=300

# Example:
kubectl drain hyperpod-i-014a41a7001adca60 \
  --ignore-daemonsets \
  --delete-emptydir-data \
  --force \
  --grace-period=300

# Check for any remaining pods outside system namespaces
kubectl get pods --all-namespaces --field-selector spec.nodeName=NODE_NAME \
  | grep -v "kube-system" \
  | grep -v "cert-manager" \
  | grep -v "kubeflow" \
  | grep -v "hyperpod-inference-system" \
  | grep -v "kube-public" \
  | grep -v "mpi-operator" \
  | grep -v "gpu-operator" \
  | grep -v "aws-hyperpod" \
  | grep -v "jupyter-k8s-system" \
  | grep -v "hyperpod-observability" \
  | grep -v "kueue-system" \
  | grep -v "keda"

# Example:
kubectl get pods --all-namespaces --field-selector spec.nodeName=hyperpod-i-014a41a7001adca60 \
  | grep -v "kube-system" \
  | grep -v "cert-manager" \
  | grep -v "kubeflow" \
  | grep -v "hyperpod-inference-system" \
  | grep -v "kube-public" \
  | grep -v "mpi-operator" \
  | grep -v "gpu-operator" \
  | grep -v "aws-hyperpod" \
  | grep -v "jupyter-k8s-system" \
  | grep -v "hyperpod-observability" \
  | grep -v "kueue-system" \
  | grep -v "keda"

# Check node status - should show "SchedulingDisabled"
kubectl get nodes -l sagemaker.amazonaws.com/instance-group-name=INSTANCE_GROUP_NAME

aws sagemaker update-cluster \
  --cluster-name my-mig-cluster \
  --instance-groups '{
    "InstanceGroupName": "gpu-group",
    "InstanceType": "ml.p4d.24xlarge",
    "InstanceCount": 1,
    "KubernetesConfig": {
       "Labels": {
          "nvidia.com/mig.config": "all-3g.20gb"
       }
    },
    "ExecutionRole": "arn:aws:iam::account:role/execution-role"
  }'

# Update kubeconfig
aws eks update-kubeconfig --name your-eks-cluster --region us-east-2

# Check MIG labels
kubectl get node NODE_NAME -o=jsonpath='{.metadata.labels}' | grep mig

# Check available MIG resources
kubectl describe node NODE_NAME | grep -A 10 "Allocatable:"

# Check GPU Operator status
kubectl get pods -n gpu-operator-resources

# View MIG configuration
kubectl exec -n gpu-operator-resources nvidia-driver-XXXXX -- nvidia-smi mig -lgi

# Check device plugin configuration
kubectl logs -n gpu-operator-resources nvidia-device-plugin-XXXXX

# Monitor node events
kubectl get events --field-selector involvedObject.name=NODE_NAME