Prerequisites Setup and configuration Configure KV caching and intelligent routing for improved performance Deploy your model from Amazon S3, Amazon FSx, or Hugging Face Hub Verify the status of your deployment Manage your deployment

Deploy models from Amazon S3, Amazon FSx, or Hugging Face Hub using kubectl

The following steps show you how to deploy models stored on Amazon S3, Amazon FSx, or Hugging Face Hub to a Amazon SageMaker HyperPod cluster using kubectl.

The following instructions contain code cells and commands designed to run in a terminal. Ensure you have configured your environment with AWS credentials before executing these commands.

Prerequisites

Before you begin, verify that you've:

Set up inference capabilities on your Amazon SageMaker HyperPod clusters. For more information, see Setting up your HyperPod clusters for model deployment.
Installed kubectl utility in your terminal.

Setup and configuration

Replace all placeholder values with your actual resource identifiers.

Select your Region in your environment.
```
export REGION=<region>
```
Initialize your cluster name. This identifies the HyperPod cluster where your model will be deployed.

Note
Check with your cluster admin to ensure permissions are granted for this role or user. You can run !aws sts get-caller-identity --query "Arn" to check which role or user you are using in your terminal.
```
# Specify your hyperpod cluster name here
HYPERPOD_CLUSTER_NAME="<Hyperpod_cluster_name>"

# NOTE: For sample deployment, we use g5.24xlarge for Llama 3.1 8B model which has sufficient memory and GPU
instance_type="ml.g5.24xlarge"
```
Initialize your cluster namespace. Your cluster admin should've already created a hyperpod-inference service account in your namespace.
```
cluster_namespace="<namespace>"
```

Create a CRD using one of the following options:

Using Amazon FSx as the model source

Set up a SageMaker endpoint name.


export SAGEMAKER_ENDPOINT_NAME="llama-fsx"

Configure the Amazon FSx file system ID to be used.
```
export FSX_FILE_SYSTEM_ID="fs-1234abcd"
```

The following is an example yaml file for creating an endpoint with Amazon FSx and a Llama model.

Note

For clusters with GPU partitioning enabled, replace nvidia.com/gpu with the appropriate MIG resource name such as nvidia.com/mig-1g.10gb. For more information, see Task Submission with MIG.

cat <<EOF> deploy_fsx_cluster_inference.yaml
---
apiVersion: inference.sagemaker.aws.amazon.com/v1
kind: InferenceEndpointConfig
metadata:
  name: $SAGEMAKER_ENDPOINT_NAME
  namespace: $CLUSTER_NAMESPACE
spec:
  modelName: Llama-3.1-8B-Instruct
  instanceType: ml.g5.24xlarge
  invocationEndpoint: v1/chat/completions
  replicas: 2
  modelSourceConfig:
    fsxStorage:
      fileSystemId: $FSX_FILE_SYSTEM_ID
    modelLocation: Llama-3.1-8B-Instruct
    modelSourceType: fsx
  worker:
    image: vllm/vllm-openai:v0.10.1
    modelInvocationPort:
      containerPort: 8000
      name: http
    modelVolumeMount:
      mountPath: /opt/ml/model
      name: model-weights
    resources:
      limits:
        nvidia.com/gpu: 4
      requests:
        cpu: 30000m
        memory: 100Gi
        nvidia.com/gpu: 4
    args:
      - "--model"
      - "/opt/ml/model"
      - "--port"
      - "8000"
      - "--tensor-parallel-size"
      - "4"
      - "--served-model-name"
      - "Llama-3.1-8B-Instruct"
    environmentVariables:
      - name: VLLM_REQUEST_TIMEOUT
        value: "600"
EOF

Using Amazon S3 as the model source

Set up a SageMaker endpoint name.


export SAGEMAKER_ENDPOINT_NAME="llama-s3"

Configure the Amazon S3 bucket location where the model is located.
```
export S3_MODEL_LOCATION="<your-s3-bucket-name>"
```

The following is an example yaml file for creating an endpoint with Amazon S3 and a Llama model using vLLM as the inference runtime.

Note

For clusters with GPU partitioning enabled, replace nvidia.com/gpu with the appropriate MIG resource name such as nvidia.com/mig-1g.10gb. For more information, see Task Submission with MIG.

cat <<EOF> deploy_s3_inference.yaml
---
apiVersion: inference.sagemaker.aws.amazon.com/v1
kind: InferenceEndpointConfig
metadata:
  name: $SAGEMAKER_ENDPOINT_NAME
  namespace: $CLUSTER_NAMESPACE
spec:
  modelName: Llama-3.1-8B-Instruct
  instanceType: ml.g5.24xlarge
  invocationEndpoint: v1/chat/completions
  replicas: 2
  modelSourceConfig:
    modelSourceType: s3
    s3Storage:
      bucketName: $S3_MODEL_LOCATION
      region: $REGION
    modelLocation: Llama-3.1-8B-Instruct
    prefetchEnabled: true
  worker:
    image: vllm/vllm-openai:v0.10.1
    modelInvocationPort:
      containerPort: 8000
      name: http
    modelVolumeMount:
      name: model-weights
      mountPath: /opt/ml/model
    resources:
      limits:
        nvidia.com/gpu: 4
      requests:
        cpu: 30000m
        memory: 100Gi
        nvidia.com/gpu: 4
    args:
      - "--model"
      - "/opt/ml/model"
      - "--port"
      - "8000"
      - "--tensor-parallel-size"
      - "4"
      - "--served-model-name"
      - "Llama-3.1-8B-Instruct"
    environmentVariables:
      - name: VLLM_REQUEST_TIMEOUT
        value: "600"
EOF

Using Hugging Face Hub as the model source

Create a Kubernetes Secret containing your Hugging Face API token. This token is required for gated models and recommended for all downloads. You can generate a token at huggingface.co/settings/tokens.

Important
Deploying models from Hugging Face Hub requires outbound internet access from your cluster nodes to Hugging Face domains including *.huggingface.co and *.hf.co. Ensure that your VPC networking configuration (NAT gateway, security groups, and network ACLs) allows HTTPS egress to these domains. Without internet access, the model download will fail.

Note
For production environments, we recommend using Amazon S3 or Amazon FSx as the model source instead of Hugging Face Hub. With Amazon S3 and Amazon FSx, model artifacts are stored within your AWS account, eliminating the dependency on external internet connectivity and providing more predictable deployment times. Hugging Face Hub is best suited for development, experimentation, and quick prototyping where direct access to the Hugging Face model repository is convenient.
```
kubectl create secret generic hf-token-secret \
  --from-literal=token=hf_YOUR_TOKEN_HERE \
  -n $CLUSTER_NAMESPACE
```

Set up a SageMaker endpoint name.


export SAGEMAKER_ENDPOINT_NAME="mistral7b-hf"

The following is an example YAML file for deploying a Mistral 7B model from Hugging Face Hub using vLLM as the inference runtime. With prefetchEnabled: true, the operator uses an init container to download the model before the inference container starts.

Note

For clusters with GPU partitioning enabled, replace nvidia.com/gpu with the appropriate MIG resource name such as nvidia.com/mig-1g.10gb. For more information, see Task Submission with MIG.

cat <<EOF> deploy_hf_inference.yaml
---
apiVersion: inference.sagemaker.aws.amazon.com/v1
kind: InferenceEndpointConfig
metadata:
  name: $SAGEMAKER_ENDPOINT_NAME
  namespace: $CLUSTER_NAMESPACE
spec:
  modelName: mistral-7b
  modelSourceConfig:
    modelSourceType: huggingface
    prefetchEnabled: true
    huggingFaceModel:
      modelId: "mistralai/Mistral-7B-Instruct-v0.3"
      tokenSecretRef:
        name: hf-token-secret
        key: token
  instanceType: "ml.g5.24xlarge"
  invocationEndpoint: v1/chat/completions
  worker:
    image: "vllm/vllm-openai:v0.10.1"
    modelInvocationPort:
      containerPort: 8000
      name: http
    modelVolumeMount:
      name: model-weights
      mountPath: /opt/ml/model
    resources:
      requests:
        nvidia.com/gpu: "4"
        memory: "96Gi"
        cpu: "16"
      limits:
        nvidia.com/gpu: "4"
        memory: "96Gi"
        cpu: "16"
    args:
      - "--model"
      - "/opt/ml/model"
      - "--port"
      - "8000"
      - "--tensor-parallel-size"
      - "4"
      - "--served-model-name"
      - "mistralai/Mistral-7B-Instruct-v0.3"
    environmentVariables:
      - name: VLLM_REQUEST_TIMEOUT
        value: "600"
EOF

The key Hugging Face configuration fields are:
- modelSourceType (required) — Set to huggingface.
- huggingFaceModel.modelId (required) — The Hugging Face Hub model identifier in org/model format (for example, mistralai/Mistral-7B-Instruct-v0.3).
- huggingFaceModel.commitSHA (optional) — A 40-character Git commit SHA to pin a specific model version. If omitted, defaults to the main branch.
- huggingFaceModel.tokenSecretRef (optional) — Reference to a Kubernetes Secret containing your Hugging Face API token. Required for gated models. The token is only used during model download and is not exposed to the inference container.
- prefetchEnabled (optional) — When true, an init container downloads the model before the inference container starts. When false, the inference runtime (vLLM, TGI, SGLang) downloads the model natively at startup. Defaults to false.

Configure KV caching and intelligent routing for improved performance

Enable KV caching by setting enableL1Cache and enableL2Cache to true.Then, set l2CacheSpec to redis and update l2CacheLocalUrl with the Redis cluster URL.
```
  kvCacheSpec:
    enableL1Cache: true
    enableL2Cache: true
    l2CacheSpec:
      l2CacheBackend: <redis | tieredstorage>
      l2CacheLocalUrl: <redis cluster URL if l2CacheBackend is redis >
```
Note
If the redis cluster is not within the same Amazon VPC as the HyperPod cluster, encryption for the data in transit is not guaranteed.

Note
Do not need l2CacheLocalUrl if tieredstorage is selected.
Enable intelligent routing by setting enabled to true under intelligentRoutingSpec. You can specify which routing strategy to use under routingStrategy. If no routing strategy is specified, it defaults to prefixaware.
```
intelligentRoutingSpec:
    enabled: true
    routingStrategy: <routing strategy to use>
```
Enable router metrics and caching metrics by setting enabled to true under metrics. The port value needs to be the same as the containerPort value under modelInvocationPort.
```
metrics:
    enabled: true
    modelMetrics:
      port: <port value>
    ...
    modelInvocationPort:
      containerPort: <port value>
```

Deploy your model from Amazon S3, Amazon FSx, or Hugging Face Hub

Get the Amazon EKS cluster name from the HyperPod cluster ARN for kubectl authentication.


export EKS_CLUSTER_NAME=$(aws --region $REGION sagemaker describe-cluster --cluster-name $HYPERPOD_CLUSTER_NAME \
  --query 'Orchestrator.Eks.ClusterArn' --output text | \
  cut -d'/' -f2)
aws eks update-kubeconfig --name $EKS_CLUSTER_NAME --region $REGION

Deploy your InferenceEndpointConfig model with one of the following options:
Deploy with Amazon FSx as a source
```
kubectl apply -f deploy_fsx_luster_inference.yaml
```
Deploy with Amazon S3 as a source
```
kubectl apply -f deploy_s3_inference.yaml
```
Deploy with Hugging Face Hub as a source
```
kubectl apply -f deploy_hf_inference.yaml
```
If the deployment fails, check the InferenceEndpointConfig events for diagnostic information. For common issues such as token errors, network connectivity, and model not found, see Hugging Face Hub model deployment failures.

Verify the status of your deployment

Check if the model successfully deployed.


kubectl describe InferenceEndpointConfig $SAGEMAKER_ENDPOINT_NAME -n $CLUSTER_NAMESPACE

Check that the endpoint is successfully created.


kubectl describe SageMakerEndpointRegistration $SAGEMAKER_ENDPOINT_NAME -n $CLUSTER_NAMESPACE

Test the deployed endpoint to verify it's working correctly. This step confirms that your model is successfully deployed and can process inference requests.


aws sagemaker-runtime invoke-endpoint \
  --endpoint-name $SAGEMAKER_ENDPOINT_NAME \
  --content-type "application/json" \
  --body '{"inputs": "What is AWS SageMaker?"}' \
  --region $REGION \
  --cli-binary-format raw-in-base64-out \
  /dev/stdout

Manage your deployment

When you're finished testing your deployment, use the following commands to clean up your resources.

Note

Verify that you no longer need the deployed model or stored data before proceeding.

Clean up your resources

Delete the inference deployment and associated Kubernetes resources. This stops the running model containers and removes the SageMaker endpoint.
```
kubectl delete inferenceendpointconfig $SAGEMAKER_ENDPOINT_NAME -n $CLUSTER_NAMESPACE
```

Verify the cleanup was done successfully.


# # Check that Kubernetes resources are removed
kubectl get pods,svc,deployment,InferenceEndpointConfig,sagemakerendpointregistration -n $CLUSTER_NAMESPACE


# Verify SageMaker endpoint is deleted (should return error or empty)
aws sagemaker describe-endpoint --endpoint-name $SAGEMAKER_ENDPOINT_NAME --region $REGION

Troubleshooting

Use these debugging commands if your deployment isn't working as expected.

Check the Kubernetes deployment status.


kubectl describe deployment $SAGEMAKER_ENDPOINT_NAME -n $CLUSTER_NAMESPACE

Check the InferenceEndpointConfig status to see the high-level deployment state and any configuration issues.
```
kubectl describe InferenceEndpointConfig $SAGEMAKER_ENDPOINT_NAME -n $CLUSTER_NAMESPACE
```
Check status of all Kubernetes objects. Get a comprehensive view of all related Kubernetes resources in your namespace. This gives you a quick overview of what's running and what might be missing.
```
kubectl get pods,svc,deployment,InferenceEndpointConfig,sagemakerendpointregistration -n $CLUSTER_NAMESPACE
```

Warning Javascript is disabled or is unavailable in your browser.

To use the Amazon Web Services Documentation, Javascript must be enabled. Please refer to your browser's Help pages for instructions.

Document Conventions

Deploy models from JumpStart using kubectl

Deploy models from local NVMe storage using kubectl

Deploy models from Amazon S3, Amazon FSx, or Hugging Face Hub using kubectl

Prerequisites

Setup and configuration

Note

Note

Note

Important

Note

Note

Configure KV caching and intelligent routing for improved performance

Note

Note

Deploy your model from Amazon S3, Amazon FSx, or Hugging Face Hub

Verify the status of your deployment

Manage your deployment

Note

Clean up your resources

Troubleshooting