Use vLLM on GKE to run inference with Llama 4

This tutorial shows you how to deploy and serve a Llama 4 Scout (17Bx16E), a 17B large language model (LLM), and serve it by using the vLLM framework. You deploy this model on a single A4 virtual machine (VM) instance on Google Kubernetes Engine (GKE).

This tutorial is intended for machine learning (ML) engineers, platform administrators and operators, and for data and AI specialists who are interested in using Kubernetes container orchestration capabilities to handle inference workloads.

Access Llama 4 by using Hugging Face

To use Hugging Face to access Llama 4, do the following:

  1. Sign the consent agreement to use Llama 4.
  2. Create a Hugging Face read access token.
  3. Copy and save the read access token value. You use it later in this tutorial.

Prepare your environment

To prepare your environment, set the following variables:

gcloud config set project PROJECT_ID
gcloud config set billing/quota_project PROJECT_ID
export PROJECT_ID=$(gcloud config get project)
export RESERVATION_URL=RESERVATION_URL
export REGION=REGION
export CLUSTER_NAME=CLUSTER_NAME
export HUGGING_FACE_TOKEN=HUGGING_FACE_TOKEN
export NETWORK=NETWORK_NAME
export SUBNETWORK=SUBNETWORK_NAME

Replace the following:

  • PROJECT_ID: the ID of the Google Cloud project where you want to create the GKE cluster.

  • RESERVATION_URL: the URL of the reservation that you want to use to create your GKE cluster. Based on the project in which the reservation exists, specify one of the following values:

    • The reservation exists in your project: RESERVATION_NAME

    • The reservation exists in a different project, and your project can use the reservation: projects/RESERVATION_PROJECT_ID/reservations/RESERVATION_NAME

  • REGION: the region where you want to create your GKE cluster. You can only create the cluster in the region where you reservation exists.

  • CLUSTER_NAME: the name of the GKE cluster to create.

  • HUGGING_FACE_TOKEN: the Hugging Face access token that you created in the previous section.

  • NETWORK_NAME: the network that the GKE cluster uses. Specify one of the following values:

    • If you created a custom network, then specify the name of your network.

    • Otherwise, specify default.

  • SUBNETWORK_NAME: the subnetwork that the GKE cluster uses. Specify one of the following values:

    • If you created a custom subnetwork, then specify the name of your subnetwork. You can only specify a subnetwork that exists in the same region as the reservation.

    • Otherwise, specify default.

Create and configure Google Cloud resources

Follow these instructions in this section to create the required resources.

Create a GKE cluster in Autopilot mode

To create a GKE cluster in Autopilot mode, run the following command:

gcloud container clusters create-auto $CLUSTER_NAME \
    --project=$PROJECT_ID \
    --region=$REGION \
    --release-channel=rapid \
    --network=$NETWORK \
    --subnetwork=$SUBNETWORK

The creation of the GKE cluster might take some time to complete. To verify that Google Cloud has finished creating your cluster, go to Kubernetes clusters on the Google Cloud console.

Create a Kubernetes secret to store your Hugging Face credentials

To create a Kubernetes secret to store your Hugging Face credentials, do the following:

  1. Configure kubectl to communicate with your GKE cluster:

    gcloud container clusters get-credentials $CLUSTER_NAME \
    --location=$REGION

  2. Create a Kubernetes secret that contains the Hugging Face read access token that you created in an earlier step:

    kubectl create secret generic hf-secret \
    --from-literal=hf_api_token=${HUGGING_FACE_TOKEN} \
    --dry-run=client -o yaml | kubectl apply -f -

Deploy a vLLM container to your GKE cluster

To deploy the vLLM container to serve the Llama-4-Scout-17B-16E-Instruct model, do the following:

  1. Create a vllm-l4-17b.yaml file with your chosen vLLM deployment:

    apiVersion: apps/v1
kind: Deployment
metadata:
  name: vllm-llama4-deployment
spec:
  replicas: 1
  selector:
    matchLabels:
      app: llama4-server
  template:
    metadata:
      labels:
        app: llama4-server
        ai.gke.io/model: llama-4-scout-17b
        ai.gke.io/inference-server: vllm
        examples.ai.gke.io/source: user-guide
    spec:
      containers:
      - name: inference-server
        image: us-docker.pkg.dev/vertex-ai/vertex-vision-model-garden-dockers/pytorch-vllm-serve:20250722_0916_RC01
        resources:
          requests:
            cpu: "10"
            memory: "128Gi"
            ephemeral-storage: "240Gi"
            nvidia.com/gpu: "8"
          limits:
            cpu: "10"
            memory: "128Gi"
            ephemeral-storage: "240Gi"
            nvidia.com/gpu: "8"
        command: ["python3", "-m", "vllm.entrypoints.openai.api_server"]
        args:
        - --model=$(MODEL_ID)
        - --tensor-parallel-size=8
        - --host=0.0.0.0
        - --port=8000
        - --max-model-len=4096
        - --max-num-seqs=4
        env:
        - name: MODEL_ID
          value: meta-llama/Llama-4-Scout-17B-16E-Instruct
        - name: HUGGING_FACE_HUB_TOKEN
          valueFrom:
            secretKeyRef:
              name: hf-secret
              key: hf_api_token
        volumeMounts:
        - mountPath: /dev/shm
          name: dshm
        livenessProbe:
          httpGet:
            path: /health
            port: 8000
          initialDelaySeconds: 1800
          periodSeconds: 10
        readinessProbe:
          httpGet:
            path: /health
            port: 8000
          initialDelaySeconds: 1800
          periodSeconds: 5
      volumes:
      - name: dshm
        emptyDir:
            medium: Memory
      nodeSelector:
        cloud.google.com/gke-accelerator: nvidia-b200
        cloud.google.com/reservation-name: RESERVATION_URL
        cloud.google.com/reservation-affinity: "specific"
        cloud.google.com/gke-gpu-driver-version: latest
---
apiVersion: v1
kind: Service
metadata:
  name: llm-service
spec:
  selector:
    app: llama4-server
  type: ClusterIP
  ports:
    - protocol: TCP
      port: 8000
      targetPort: 8000
---
apiVersion: monitoring.googleapis.com/v1
kind: PodMonitoring
metadata:
  name: vllm-llama4-monitoring
spec:
  selector:
    matchLabels:
      app: llama4-server
  endpoints:
  - port: 8000
    path: /metrics
    interval: 30s

  2. Apply the vllm-l4-17b.yaml file to your GKE cluster:

      kubectl apply -f vllm-l4-17b.yaml

    During the deployment process, the container must download the Llama-4-Scout-17B-16E-Instruct model from Hugging Face. For this reason, deployment of the container might take up to 30 minutes to complete.

  3. To see the completion status, run the following command:

      kubectl wait \
          --for=condition=Available \
          --timeout=1800s deployment/vllm-llama4-deployment

    The--timeout=1800s flag allows the command to monitor the deployment for up to 30 minutes.

Interact with Llama 4 by using curl

To verify the Llama 4 Scout model that you deployed, do the following:

  1. Set up port forwarding to Llama 4 Scout:

    kubectl port-forward service/llm-service 8000:8000

  2. Open a new terminal window. You can then chat with your model by using curl:

    curl http://127.0.0.1:8000/v1/chat/completions \
     -X POST \
     -H "Content-Type: application/json" \
     -d '{
       "model": "meta-llama/Llama-4-Scout-17B-16E-Instruct",
       "messages": [
         {
           "role": "user",
           "content": "Describe a sailboat in one short sentence?"
         }
       ]
     }'

  3. The output that you see is similar to the following:

    {
      "id": "chatcmpl-ec0ad6310c494a889b17600881c06e3d",
      "object": "chat.completion",
      "created": 1754073279,
      "model": "meta-llama/Llama-4-Scout-17B-16E-Instruct",
      "choices": [
        {
          "index": 0,
          "message": {
            "role": "assistant",
            "content": "A sailboat is a type of watercraft that uses the wind for propulsion, typically featuring a hull, mast, and one or more sails.",
            "refusal": null,
            "annotations": null,
            "audio": null,
            "function_call": null,
            "tool_calls": [],
            "reasoning_content": null
          },
          "logprobs": null,
          "finish_reason": "stop",
          "stop_reason": null
        }
      ],
      "service_tier": null,
      "system_fingerprint": null,
      "usage": {
        "prompt_tokens": 19,
        "total_tokens": 49,
        "completion_tokens": 30,
        "prompt_tokens_details": null
      },
      "prompt_logprobs": null,
      "kv_transfer_params": null
    }

Observe model performance

To observe the performance of your model, you can use the vLLM dashboard integration in Cloud Monitoring. You can use this dashboard to view critical performance metrics like token throughput, request latency, and error rates.

For information about using Google Cloud Managed Service for Prometheus to collect metrics from your model, see the vLLM observability guidance in the Cloud Monitoring documentation.