Serve an LLM with GKE Inference Gateway

Stay organized with collections Save and categorize content based on your preferences.

Get access to the model

To deploy the Llama3.1 model to GKE, sign the license consent agreement and generate a Hugging Face access token.

Sign the license consent agreement

You must sign the consent agreement to use the Llama3.1 model. Follow these instructions:

1. Access consent page and verify consent for using your Hugging Face account.
2. Accept the model terms.

Generate an access token

To access the model through Hugging Face, you need a Hugging Face token.

Follow these steps to generate a new token if you don't have one already:

1. Click Your Profile > Settings > Access Tokens.
2. Select New Token.
3. Specify a name of your choice and a role of at least Read.
4. Select Generate a token.
5. Copy the generated token to your clipboard.

Prepare your environment

In this tutorial, you use Cloud Shell to manage resources hosted on Google Cloud. Cloud Shell comes preinstalled with the software you need for this tutorial, including kubectl and gcloud CLI.

To set up your environment with Cloud Shell, perform the following steps:

1. In the Google Cloud console, launch a Cloud Shell session by clicking Activate Cloud Shell in the Google Cloud console. This launches a session in the bottom pane of Google Cloud console.

2. Set the default environment variables:

   gcloud config set project PROJECT_ID
   export PROJECT_ID=$(gcloud config get project)
   export REGION=REGION
   export CLUSTER_NAME=CLUSTER_NAME
   export HF_TOKEN=HF_TOKEN
   

   Replace the following values:

   • PROJECT_ID: your Google Cloud project ID.
   • REGION: a region that supports the accelerator type you want to use, for example, us-central1 for H100 GPU.
   • CLUSTER_NAME: the name of your cluster.
   • HF_TOKEN: the Hugging Face token you generated earlier.

Create and configure Google Cloud resources

To create the required resources, use these instructions.

Create a GKE cluster and node pool

Serve LLMs on GPUs in a GKE Autopilot or Standard cluster. We recommend that you use a Autopilot cluster for a fully managed Kubernetes experience. To choose the GKE mode of operation that's the best fit for your workloads, see Choose a GKE mode of operation.

gcloud container clusters create-auto CLUSTER_NAME \
    --project=PROJECT_ID \
    --region=REGION \
    --release-channel=rapid \
    --cluster-version=1.32.3-gke.1170000

1. To set up authorization to scrape metrics, create the inference-gateway-sa-metrics-reader-secret secret:

   kubectl apply -f - <<EOF
   ---
   apiVersion: rbac.authorization.k8s.io/v1
   kind: ClusterRole
   metadata:
     name: inference-gateway-metrics-reader
   rules:
   - nonResourceURLs:
     - /metrics
     verbs:
     - get
   ---
   apiVersion: v1
   kind: ServiceAccount
   metadata:
     name: inference-gateway-sa-metrics-reader
     namespace: default
   ---
   apiVersion: rbac.authorization.k8s.io/v1
   kind: ClusterRoleBinding
   metadata:
     name: inference-gateway-sa-metrics-reader-role-binding
     namespace: default
   subjects:
   - kind: ServiceAccount
     name: inference-gateway-sa-metrics-reader
     namespace: default
   roleRef:
     kind: ClusterRole
     name: inference-gateway-metrics-reader
     apiGroup: rbac.authorization.k8s.io
   ---
   apiVersion: v1
   kind: Secret
   metadata:
     name: inference-gateway-sa-metrics-reader-secret
     namespace: default
     annotations:
       kubernetes.io/service-account.name: inference-gateway-sa-metrics-reader
   type: kubernetes.io/service-account-token
   ---
   apiVersion: rbac.authorization.k8s.io/v1
   kind: ClusterRole
   metadata:
     name: inference-gateway-sa-metrics-reader-secret-read
   rules:
   - resources:
     - secrets
     apiGroups: [""]
     verbs: ["get", "list", "watch"]
     resourceNames: ["inference-gateway-sa-metrics-reader-secret"]
   ---
   apiVersion: rbac.authorization.k8s.io/v1
   kind: ClusterRoleBinding
   metadata:
     name: gmp-system:collector:inference-gateway-sa-metrics-reader-secret-read
     namespace: default
   roleRef:
     name: inference-gateway-sa-metrics-reader-secret-read
     kind: ClusterRole
     apiGroup: rbac.authorization.k8s.io
   subjects:
   - name: collector
     namespace: gmp-system
     kind: ServiceAccount
   EOF
   

Create a Kubernetes secret for Hugging Face credentials

In Cloud Shell, do the following:

1. To communicate with your cluster, configure kubectl:

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

   Replace the following values:

   • REGION: a region that supports the accelerator type you want to use, for example, us-central1 for H100 GPU.
   • CLUSTER_NAME: the name of your cluster.

2. Create a Kubernetes Secret that contains the Hugging Face token:

     kubectl create secret generic HF_SECRET \
         --from-literal=hf_api_token=HF_TOKEN \
         --dry-run=client -o yaml | kubectl apply -f -
   

   Replace the following:

   • HF_TOKEN: the Hugging Face token you generated earlier.
   • HF_SECRET: the name for your Kubernetes secret. For example, hf-secret.

Install the InferenceModel and InferencePool CRDs

In this section, you install the necessary Custom Resource Definitions (CRDs) for GKE Inference Gateway.

CRDs extend the Kubernetes API. This lets you define new resource types. To use GKE Inference Gateway, install the InferencePool and InferenceModel CRDs in your GKE cluster by running the following command:

kubectl apply -f https://github.com/kubernetes-sigs/gateway-api-inference-extension/releases/download/v0.3.0/manifests.yaml

Deploy the model server

This example deploys a Llama3.1 model using a vLLM model server. The deployment is labeled as app:vllm-llama3-8b-instruct. This deployment also uses two LoRA adapters named food-review and cad-fabricator from Hugging Face. You can update this deployment with your own model server and model container, serving port, and deployment name. You can optionally configure LoRA adapters in the deployment, or deploy the base model.

1. To deploy on a nvidia-h100-80gb accelerator type, save the following manifest as vllm-llama3-8b-instruct.yaml. This manifest defines a Kubernetes Deployment with your model and model server:

   apiVersion: apps/v1
   kind: Deployment
   metadata:
     name: vllm-llama3-8b-instruct
   spec:
     replicas: 3
     selector:
       matchLabels:
         app: vllm-llama3-8b-instruct
     template:
       metadata:
         labels:
           app: vllm-llama3-8b-instruct
       spec:
         containers:
           - name: vllm
             image: "vllm/vllm-openai:latest"
             imagePullPolicy: Always
             command: ["python3", "-m", "vllm.entrypoints.openai.api_server"]
             args:
             - "--model"
             - "meta-llama/Llama-3.1-8B-Instruct"
             - "--tensor-parallel-size"
             - "1"
             - "--port"
             - "8000"
             - "--enable-lora"
             - "--max-loras"
             - "2"
             - "--max-cpu-loras"
             - "12"
             env:
               # Enabling LoRA support temporarily disables automatic v1, we want to force it on
               # until 0.8.3 vLLM is released.
               - name: VLLM_USE_V1
                 value: "1"
               - name: PORT
                 value: "8000"
               - name: HUGGING_FACE_HUB_TOKEN
                 valueFrom:
                   secretKeyRef:
                     name: hf-token
                     key: token
               - name: VLLM_ALLOW_RUNTIME_LORA_UPDATING
                 value: "true"
             ports:
               - containerPort: 8000
                 name: http
                 protocol: TCP
             lifecycle:
               preStop:
                 # vLLM stops accepting connections when it receives SIGTERM, so we need to sleep
                 # to give upstream gateways a chance to take us out of rotation. The time we wait
                 # is dependent on the time it takes for all upstreams to completely remove us from
                 # rotation. Older or simpler load balancers might take upwards of 30s, but we expect
                 # our deployment to run behind a modern gateway like Envoy which is designed to
                 # probe for readiness aggressively.
                 sleep:
                   # Upstream gateway probers for health should be set on a low period, such as 5s,
                   # and the shorter we can tighten that bound the faster that we release
                   # accelerators during controlled shutdowns. However, we should expect variance,
                   # as load balancers may have internal delays, and we don't want to drop requests
                   # normally, so we're often aiming to set this value to a p99 propagation latency
                   # of readiness -> load balancer taking backend out of rotation, not the average.
                   #
                   # This value is generally stable and must often be experimentally determined on
                   # for a given load balancer and health check period. We set the value here to
                   # the highest value we observe on a supported load balancer, and we recommend
                   # tuning this value down and verifying no requests are dropped.
                   #
                   # If this value is updated, be sure to update terminationGracePeriodSeconds.
                   #
                   seconds: 30
                 #
                 # IMPORTANT: preStop.sleep is beta as of Kubernetes 1.30 - for older versions
                 # replace with this exec action.
                 #exec:
                 #  command:
                 #  - /usr/bin/sleep
                 #  - 30
             livenessProbe:
               httpGet:
                 path: /health
                 port: http
                 scheme: HTTP
               # vLLM's health check is simple, so we can more aggressively probe it.  Liveness
               # check endpoints should always be suitable for aggressive probing.
               periodSeconds: 1
               successThreshold: 1
               # vLLM has a very simple health implementation, which means that any failure is
               # likely significant. However, any liveness triggered restart requires the very
               # large core model to be reloaded, and so we should bias towards ensuring the
               # server is definitely unhealthy vs immediately restarting. Use 5 attempts as
               # evidence of a serious problem.
               failureThreshold: 5
               timeoutSeconds: 1
             readinessProbe:
               httpGet:
                 path: /health
                 port: http
                 scheme: HTTP
               # vLLM's health check is simple, so we can more aggressively probe it.  Readiness
               # check endpoints should always be suitable for aggressive probing, but may be
               # slightly more expensive than readiness probes.
               periodSeconds: 1
               successThreshold: 1
               # vLLM has a very simple health implementation, which means that any failure is
               # likely significant,
               failureThreshold: 1
               timeoutSeconds: 1
             # We set a startup probe so that we don't begin directing traffic or checking
             # liveness to this instance until the model is loaded.
             startupProbe:
               # Failure threshold is when we believe startup will not happen at all, and is set
               # to the maximum possible time we believe loading a model will take. In our
               # default configuration we are downloading a model from HuggingFace, which may
               # take a long time, then the model must load into the accelerator. We choose
               # 10 minutes as a reasonable maximum startup time before giving up and attempting
               # to restart the pod.
               #
               # IMPORTANT: If the core model takes more than 10 minutes to load, pods will crash
               # loop forever. Be sure to set this appropriately.
               failureThreshold: 120
               # Set delay to start low so that if the base model changes to something smaller
               # or an optimization is deployed, we don't wait unnecessarily.
               initialDelaySeconds: 2
               # As a startup probe, this stops running and so we can more aggressively probe
               # even a moderately complex startup - this is a very important workload.
               periodSeconds: 1
               httpGet:
                 # vLLM does not start the OpenAI server (and hence make /health available)
                 # until models are loaded. This may not be true for all model servers.
                 path: /health
                 port: http
                 scheme: HTTP
   
             resources:
               limits:
                 nvidia.com/gpu: 1
               requests:
                 nvidia.com/gpu: 1
             volumeMounts:
               - mountPath: /data
                 name: data
               - mountPath: /dev/shm
                 name: shm
               - name: adapters
                 mountPath: "/adapters"
         initContainers:
           - name: lora-adapter-syncer
             tty: true
             stdin: true
             image: us-central1-docker.pkg.dev/k8s-staging-images/gateway-api-inference-extension/lora-syncer:main
             restartPolicy: Always
             imagePullPolicy: Always
             env:
               - name: DYNAMIC_LORA_ROLLOUT_CONFIG
                 value: "/config/configmap.yaml"
             volumeMounts: # DO NOT USE subPath, dynamic configmap updates don't work on subPaths
             - name: config-volume
               mountPath:  /config
         restartPolicy: Always
   
         # vLLM allows VLLM_PORT to be specified as an environment variable, but a user might
         # create a 'vllm' service in their namespace. That auto-injects VLLM_PORT in docker
         # compatible form as `tcp://<IP>:<PORT>` instead of the numeric value vLLM accepts
         # causing CrashLoopBackoff. Set service environment injection off by default.
         enableServiceLinks: false
   
         # Generally, the termination grace period needs to last longer than the slowest request
         # we expect to serve plus any extra time spent waiting for load balancers to take the
         # model server out of rotation.
         #
         # An easy starting point is the p99 or max request latency measured for your workload,
         # although LLM request latencies vary significantly if clients send longer inputs or
         # trigger longer outputs. Since steady state p99 will be higher than the latency
         # to drain a server, you may wish to slightly this value either experimentally or
         # via the calculation below.
         #
         # For most models you can derive an upper bound for the maximum drain latency as
         # follows:
         #
         #   1. Identify the maximum context length the model was trained on, or the maximum
         #      allowed length of output tokens configured on vLLM (llama2-7b was trained to
         #      4k context length, while llama3-8b was trained to 128k).
         #   2. Output tokens are the more compute intensive to calculate and the accelerator
         #      will have a maximum concurrency (batch size) - the time per output token at
         #      maximum batch with no prompt tokens being processed is the slowest an output
         #      token can be generated (for this model it would be about 100ms TPOT at a max
         #      batch size around 50)
         #   3. Calculate the worst case request duration if a request starts immediately
         #      before the server stops accepting new connections - generally when it receives
         #      SIGTERM (for this model that is about 4096 / 10 ~ 40s)
         #   4. If there are any requests generating prompt tokens that will delay when those
         #      output tokens start, and prompt token generation is roughly 6x faster than
         #      compute-bound output token generation, so add 20% to the time from above (40s +
         #      16s ~ 55s)
         #
         # Thus we think it will take us at worst about 55s to complete the longest possible
         # request the model is likely to receive at maximum concurrency (highest latency)
         # once requests stop being sent.
         #
         # NOTE: This number will be lower than steady state p99 latency since we stop       receiving
         #       new requests which require continuous prompt token computation.
             # NOTE: The max timeout for backend connections from gateway to model servers should
         #       be configured based on steady state p99 latency, not drain p99 latency
         #
         #   5. Add the time the pod takes in its preStop hook to allow the load balancers have
         #      stopped sending us new requests (55s + 30s ~ 85s)
         #
         # Because the termination grace period controls when the Kubelet forcibly terminates a
         # stuck or hung process (a possibility due to a GPU crash), there is operational safety
         # in keeping the value roughly proportional to the time to finish serving. There is also
         # value in adding a bit of extra time to deal with unexpectedly long workloads.
         #
         #   6. Add a 50% safety buffer to this time since the operational impact should be low
         #      (85s * 1.5 ~ 130s)
         #
         # One additional source of drain latency is that some workloads may run close to
         # saturation and have queued requests on each server. Since traffic in excess of the
         # max sustainable QPS will result in timeouts as the queues grow, we assume that failure
         # to drain in time due to excess queues at the time of shutdown is an expected failure
         # mode of server overload. If your workload occasionally experiences high queue depths
         # due to periodic traffic, consider increasing the safety margin above to account for
         # time to drain queued requests.
         terminationGracePeriodSeconds: 130
         nodeSelector:
           cloud.google.com/gke-accelerator: "nvidia-h100-80gb"
         volumes:
           - name: data
             emptyDir: {}
           - name: shm
             emptyDir:
               medium: Memory
           - name: adapters
             emptyDir: {}
           - name: config-volume
             configMap:
               name: vllm-llama3-8b-adapters
   ---
   apiVersion: v1
   kind: ConfigMap
   metadata:
     name: vllm-llama3-8b-adapters
   data:
     configmap.yaml: |
         vLLMLoRAConfig:
           name: vllm-llama3.1-8b-instruct
           port: 8000
           defaultBaseModel: meta-llama/Llama-3.1-8B-Instruct
           ensureExist:
             models:
             - id: food-review
               source: Kawon/llama3.1-food-finetune_v14_r8
             - id: cad-fabricator
               source: redcathode/fabricator
   ---
   kind: HealthCheckPolicy
   apiVersion: networking.gke.io/v1
   metadata:
     name: health-check-policy
     namespace: default
   spec:
     targetRef:
       group: "inference.networking.x-k8s.io"
       kind: InferencePool
       name: vllm-llama3-8b-instruct
     default:
       config:
         type: HTTP
         httpHealthCheck:
             requestPath: /health
             port: 8000
   

2. Apply the manifest to your cluster:

   kubectl apply -f vllm-llama3-8b-instruct.yaml
   

Create an InferencePool resource

The InferencePool Kubernetes custom resource defines a group of Pods with a common base LLM and compute configuration.

The InferencePool custom resource includes the following key fields:

• selector: specifies which Pods belong to this pool. The labels in this selector must exactly match the labels applied to your model server Pods.
• targetPort: defines the ports used by the model server within the Pods.

The InferencePool resource enables GKE Inference Gateway to route traffic to your model server Pods.

To create an InferencePool using Helm, perform the following steps:

helm install vllm-llama3-8b-instruct \
  --set inferencePool.modelServers.matchLabels.app=vllm-llama3-8b-instruct \
  --set provider.name=gke \
  --version v0.3.0 \
  oci://registry.k8s.io/gateway-api-inference-extension/charts/inferencepool

Change the following field to match your Deployment:

• inferencePool.modelServers.matchLabels.app: the key of the label used to select your model server Pods.

This command creates an InferencePool object that logically represents your model server deployment and references the model endpoint services within the Pods that the Selector selects.

Create an InferenceModel resource with a serving criticality

The InferenceModel Kubernetes custom resource defines a specific model, including LoRA-tuned models, and its serving criticality.

The InferenceModel custom resource includes the following key fields:

• modelName: specifies the name of the base model or LoRA adapter.
• Criticality: specifies the serving criticality of the model.
• poolRef: references the InferencePool that the model is served on.

The InferenceModel enables the GKE Inference Gateway to route traffic to your model server Pods based on the model name and criticality.

To create an InferenceModel, perform the following steps:

1. Save the following sample manifest as inferencemodel.yaml:

   apiVersion: inference.networking.x-k8s.io/v1alpha2
   kind: InferenceModel
   metadata:
     name: inferencemodel-sample
   spec:
     modelName: MODEL_NAME
     criticality: CRITICALITY
     poolRef:
       name: INFERENCE_POOL_NAME
   

   Replace the following:

   • MODEL_NAME: the name of your base model or LoRA adapter. For example, food-review.
   • CRITICALITY: the chosen serving criticality. Choose from Critical, Standard, or Sheddable. For example, Standard.
   • INFERENCE_POOL_NAME: the name of the InferencePool you created in the previous step. For example, vllm-llama3-8b-instruct.

2. Apply the sample manifest to your cluster:

   kubectl apply -f inferencemodel.yaml
   

The following example creates an InferenceModel object that configures the food-review LoRA model on the vllm-llama3-8b-instruct InferencePool with a Standard serving criticality. The InferenceModel object also configures the base model to be served with a Critical priority level.

apiVersion: inference.networking.x-k8s.io/v1alpha2
kind: InferenceModel
metadata:
  name: food-review
spec:
  modelName: food-review
  criticality: Standard
  poolRef:
    name: vllm-llama3-8b-instruct
  targetModels:
  - name: food-review
    weight: 100

---
apiVersion: inference.networking.x-k8s.io/v1alpha2
kind: InferenceModel
metadata:
  name: llama3-base-model
spec:
  modelName: meta-llama/Llama-3.1-8B-Instruct
  criticality: Critical
  poolRef:
    name: vllm-llama3-8b-instruct

Create the Gateway

The Gateway resource acts as the entry point for external traffic into your Kubernetes cluster. It defines the listeners that accept incoming connections.

GKE Inference Gateway supports the gke-l7-rilb and gke-l7-regional-external-managed Gateway Class. For more information, see the GKE documentation on Gateway Classes.

To create a Gateway, perform the following steps:

1. Save the following sample manifest as gateway.yaml:

   apiVersion: gateway.networking.k8s.io/v1
   kind: Gateway
   metadata:
     name: GATEWAY_NAME
   spec:
     gatewayClassName: gke-l7-regional-external-managed
     listeners:
       - protocol: HTTP # Or HTTPS for production
         port: 80 # Or 443 for HTTPS
         name: http
   

   Replace GATEWAY_NAME with a unique name for your Gateway resource. For example, inference-gateway.

2. Apply the manifest to your cluster:

   kubectl apply -f gateway.yaml
   

Create the HTTPRoute resource

In this section, you create an HTTPRoute resource to define how the Gateway routes incoming HTTP requests to your InferencePool.

The HTTPRoute resource defines how the GKE Gateway routes incoming HTTP requests to backend services, which is your InferencePool. It specifies matching rules (for example, headers, or paths) and the backend to which traffic should be forwarded.

To create an HTTPRoute, perform the following steps:

1. Save the following sample manifest as httproute.yaml:

   apiVersion: gateway.networking.k8s.io/v1
   kind: HTTPRoute
   metadata:
     name: HTTPROUTE_NAME
   spec:
     parentRefs:
     - name: GATEWAY_NAME
     rules:
     - matches:
       - path:
           type: PathPrefix
           value: PATH_PREFIX
       backendRefs:
       - name: INFERENCE_POOL_NAME
         kind: InferencePool
   

   Replace the following:

   • HTTPROUTE_NAME: a unique name for your HTTPRoute resource. For example, my-route.
   • GATEWAY_NAME: the name of the Gateway resource that you created. For example, inference-gateway.
   • PATH_PREFIX: the path prefix that you use to match incoming requests. For example, / to match all.
   • INFERENCE_POOL_NAME: the name of the InferencePool resource that you want to route traffic to. For example, vllm-llama3-8b-instruct.

2. Apply the manifest to your cluster:

   kubectl apply -f httproute.yaml
   

Send an inference request

After you have configured GKE Inference Gateway, you can send inference requests to your deployed model.

To send inference requests, perform the following steps:

• Retrieve the Gateway endpoint.
• Construct a properly formatted JSON request.
• Use curl to send the request to the /v1/completions endpoint.

This lets you generate text based on your input prompt and specified parameters.

1. To get the Gateway endpoint, run the following command:

   IP=$(kubectl get gateway/GATEWAY_NAME -o jsonpath='{.status.addresses[0].address}')
   PORT=PORT_NUMBER # Use 443 for HTTPS, or 80 for HTTP
   

   Replace the following:

   • GATEWAY_NAME: the name of your Gateway resource.
   • PORT_NUMBER: the port number you configured in the Gateway.

2. To send a request to the /v1/completions endpoint using curl, run the following command:

   curl -i -X POST https://${IP}:${PORT}/v1/completions \
   -H 'Content-Type: application/json' \
   -H 'Authorization: Bearer $(gcloud auth print-access-token)' \
   -d '{
       "model": "MODEL_NAME",
       "prompt": "PROMPT_TEXT",
       "max_tokens": MAX_TOKENS,
       "temperature": "TEMPERATURE"
   }'
   

   Replace the following:

   • MODEL_NAME: the name of the model or LoRA adapter to use.
   • PROMPT_TEXT: the input prompt for the model.
   • MAX_TOKENS: the maximum number of tokens to generate in the response.
   • TEMPERATURE: controls the randomness of the output. Use the value 0 for deterministic output, or a higher number for more creative output.

Be aware of the following behaviors:

• Request body: the request body can include additional parameters like stop and top_p. Refer to the OpenAI API specification for a complete list of options.
• Error handling: implement proper error handling in your client code to handle potential errors in the response. For example, check the HTTP status code in the curl response. A non-200 status code generally indicates an error.
• Authentication and authorization: for production deployments, secure your API endpoint with authentication and authorization mechanisms. Include the appropriate headers (for example, Authorization) in your requests.

Configure observability for your Inference Gateway

GKE Inference Gateway provides observability into the health, performance, and behavior of your inference workloads. This helps you to identify and resolve issues, optimize resource utilization, and ensure the reliability of your applications. You can view these observability metrics in Cloud Monitoring through the Metrics Explorer.

To configure observability for GKE Inference Gateway, see Configure observability.

Delete the deployed resources

To avoid incurring charges to your Google Cloud account for the resources that you created from this guide, run the following command:

gcloud container clusters delete CLUSTER_NAME \
    --region=REGION

Replace the following values:

• REGION: a region that supports the accelerator type you want to use, for example, us-central1 for H100 GPU.
• CLUSTER_NAME: the name of your cluster.

What's next

• Read about GKE Inference Gateway.
• Read about Deploying GKE Inference Gateway.