Deploy and inference Gemma using Model Garden and Vertex AI TPU-backed endpoints

---

In this tutorial, you use Model Garden to deploy the Gemma 2B open model to a TPU-backed Vertex AI endpoint. You must deploy a model to an endpoint before that model can be used to serve online predictions. Deploying a model associates physical resources with the model so it can serve online predictions with low latency.

After you deploy the Gemma 2B model, you inference the trained model by using the PredictionServiceClient to get online predictions. Online predictions are synchronous requests made to a model that is deployed to an endpoint.

Objectives

This tutorial shows you how to perform the following tasks:

• Deploy the Gemma 2B open model to a TPU backed endpoint by using Model Garden
• Use the PredictionServiceClient to get online predictions

Costs

In this document, you use the following billable components of Google Cloud:

• A ct5lp-hightpu-1t machine type with one TPU_V5 accelerator
• Vertex AI prediction and explanation

To generate a cost estimate based on your projected usage, use the pricing calculator. New Google Cloud users might be eligible for a free trial.

When you finish the tasks that are described in this document, you can avoid continued billing by deleting the resources that you created. For more information, see Clean up.

Before you begin

This tutorial requires you to:

• Set up a Google Cloud project and enable the Vertex AI API
• On your local machine:
  • Install, initialize, and authenticate with the Google Cloud CLI
  • Install the SDK for your language

Set up a Google Cloud project

Set up your Google Cloud project and enable the Vertex AI API.

1. Sign in to your Google Cloud account. If you're new to Google Cloud, create an account to evaluate how our products perform in real-world scenarios. New customers also get $300 in free credits to run, test, and deploy workloads.

2. In the Google Cloud console, on the project selector page, select or create a Google Cloud project.

   Note: If you don't plan to keep the resources that you create in this procedure, create a project instead of selecting an existing project. After you finish these steps, you can delete the project, removing all resources associated with the project.

   Go to project selector

3. Make sure that billing is enabled for your Google Cloud project.

4. Enable the Vertex AI API.

   Enable the API

5. In the Google Cloud console, on the project selector page, select or create a Google Cloud project.

   Note: If you don't plan to keep the resources that you create in this procedure, create a project instead of selecting an existing project. After you finish these steps, you can delete the project, removing all resources associated with the project.

   Go to project selector

6. Make sure that billing is enabled for your Google Cloud project.

7. Enable the Vertex AI API.

   Enable the API

Set up the Google Cloud CLI

On your local machine, set up the Google Cloud CLI.

1. Install and initialize the Google Cloud CLI.

2. If you previously installed the gcloud CLI, ensure your gcloud components are updated by running this command.

   gcloud components update

3. To authenticate with the gcloud CLI, generate a local Application Default Credentials (ADC) file by running this command. The web flow launched by the command is used to provide your user credentials.

   gcloud auth application-default login

   For more information, see gcloud CLI authentication configuration and ADC configuration.

Note: To avoid providing your project ID and region to the Google Cloud CLI, you can use the gcloud config set command to set a default project and region.

Set up the SDK for your programming language

To set up the environment used in this tutorial, you install the Vertex AI SDK for your language and the Protocol Buffers library. The code samples use functions from the Protocol Buffers library to convert the input dictionary to the JSON format that is expected by the API.

On your local machine, click one of the following tabs to install the SDK for your programming language.

Python

On your local machine, click one of the following tabs to install the SDK for your programming language.

• Install and update the Vertex AI SDK for Python by running this command.

  pip3 install --upgrade "google-cloud-aiplatform>=1.64"

• Install the Protocol Buffers library for Python by running this command.

  pip3 install --upgrade "protobuf>=5.28"

Node.js

Install or update the aiplatform SDK for Node.js by running the following command.

npm install @google-cloud/aiplatform

Java

To add google-cloud-aiplatform as a dependency, add the appropriate code for your environment.

Maven with BOM

Add the following HTML to your pom.xml:

<dependencyManagement>
<dependencies>
  <dependency>
    <artifactId>libraries-bom</artifactId>
    <groupId>com.google.cloud</groupId>
    <scope>import</scope>
    <type>pom</type>
    <version>26.34.0</version>
  </dependency>
</dependencies>
</dependencyManagement>
<dependencies>
<dependency>
  <groupId>com.google.cloud</groupId>
  <artifactId>google-cloud-aiplatform</artifactId>
</dependency>
<dependency>
  <groupId>com.google.protobuf</groupId>
  <artifactId>protobuf-java-util</artifactId>
</dependency>
<dependency>
  <groupId>com.google.code.gson</groupId>
  <artifactId>gson</artifactId>
</dependency>
</dependencies>

Maven without BOM

Add the following to your pom.xml:

<dependency>
  <groupId>com.google.cloud</groupId>
  <artifactId>google-cloud-aiplatform</artifactId>
  <version>1.1.0</version>
</dependency>
<dependency>
  <groupId>com.google.protobuf</groupId>
  <artifactId>protobuf-java-util</artifactId>
  <version>5.28</version>
</dependency>
<dependency>
  <groupId>com.google.code.gson</groupId>
  <artifactId>gson</artifactId>
  <version>2.11.0</version>
</dependency>

Gradle without BOM

Add the following to your build.gradle:

implementation 'com.google.cloud:google-cloud-aiplatform:1.1.0'

Go

Install these Go packages by running the following commands.

• The aiplatform Go client library
• Go support for Protocol Buffers
• Google API Extensions for Go (gax-go)

go get cloud.google.com/go/aiplatform
go get google.golang.org/protobuf
go get github.com/googleapis/gax-go/v2

Deploy Gemma using Model Garden

You deploy the Gemma 2B model to a ct5lp-hightpu-1t Compute Engine machine type that is optimized for small to medium scale training. This machine has one TPU v5e accelerator. For more information on training models using TPUs, see Cloud TPU v5e training.

In this tutorial, you deploy the instruction-tuned Gemma 2B open model by using the model card in Model Garden. The specific model version is gemma2-2b-it — -it stands for instruction-tuned.

The Gemma 2B model has a lower parameter size which means lower resource requirements and more deployment flexibility.

1. In the Google Cloud console, go to the Model Garden page.

   Go to Model Garden

2. Click the Gemma 2 model card.

   Go to Gemma 2

3. Click Deploy to open the Deploy model pane.

4. In the Deploy model pane, specify these details.

   1. For Deployment environment click Vertex AI.

   2. In the Deploy model section:

      1. For Resource ID, choose gemma-2b-it.

      2. For Model name and Endpoint name, accept the default values. For example:

         • Model name: gemma2-2b-it-1234567891234
         • Endpoint name: gemma2-2b-it-mg-one-click-deploy

         Make a note of the endpoint name. You'll need it to find the endpoint ID used in the code samples.

   3. In the Deployment settings section:

      1. Accept the default option for Basic settings.

      2. For Region, accept the default value or choose a region from the list. Make a note of the region. You'll need it for the code samples.

      3. For Machine spec, choose the TPU backed instance: ct5lp-hightpu-1t (1 TPU_V5_LITEPOD; ct5lp-hightpu-1t).

5. Click Deploy. When the deployment is finished, receive an email that contains details about your new endpoint. You can also view the endpoint details by clicking Online prediction > Endpoints and selecting your region.

   Go to Endpoints

Inference Gemma 2B with the PredictionServiceClient

After you deploy Gemma 2B, you use the PredictionServiceClient to get online predictions for the prompt: "Why is the sky blue?"

Code parameters

The PredictionServiceClient code samples require you to update the following.

• PROJECT_ID: To find your project ID follow these steps.

  1. Go to the Welcome page in the Google Cloud console.

     Go to Welcome

  2. From the project picker at the top of the page, select your project.

     The project name, project number, and project ID appear after the Welcome heading.

• ENDPOINT_REGION: This is the region where you deployed the endpoint.

• ENDPOINT_ID: To find your endpoint ID, view it in the console or run the gcloud ai endpoints list command. You'll need the endpoint name and region from the Deploy model pane.

  Console

  You can view the endpoint details by clicking Online prediction > Endpoints and selecting your region. Note the number that appears in the ID column.

  Go to Endpoints

  gcloud

  You can view the endpoint details by running the gcloud ai endpoints list command.

  gcloud ai endpoints list \
    --region=ENDPOINT_REGION \
    --filter=display_name=ENDPOINT_NAME
  

  The output looks like this.

  Using endpoint [https://us-central1-aiplatform.googleapis.com/]
  ENDPOINT_ID: 1234567891234567891
  DISPLAY_NAME: gemma2-2b-it-mg-one-click-deploy
  

Sample code

In the sample code for your language, update the PROJECT_ID, ENDPOINT_REGION, and ENDPOINT_ID. Then run your code.

Python

To learn how to install or update the Vertex AI SDK for Python, see Install the Vertex AI SDK for Python. For more information, see the Python API reference documentation.

"""
Sample to run inference on a Gemma2 model deployed to a Vertex AI endpoint with TPU accellerators.
"""

from google.cloud import aiplatform
from google.protobuf import json_format
from google.protobuf.struct_pb2 import Value

# TODO(developer): Update & uncomment lines below
# PROJECT_ID = "your-project-id"
# ENDPOINT_REGION = "your-vertex-endpoint-region"
# ENDPOINT_ID = "your-vertex-endpoint-id"

# Default configuration
config = {"max_tokens": 1024, "temperature": 0.9, "top_p": 1.0, "top_k": 1}

# Prompt used in the prediction
prompt = "Why is the sky blue?"

# Encapsulate the prompt in a correct format for TPUs
# Example format: [{'prompt': 'Why is the sky blue?', 'temperature': 0.9}]
input = {"prompt": prompt}
input.update(config)

# Convert input message to a list of GAPIC instances for model input
instances = [json_format.ParseDict(input, Value())]

# Create a client
api_endpoint = f"{ENDPOINT_REGION}-aiplatform.googleapis.com"
client = aiplatform.gapic.PredictionServiceClient(
    client_options={"api_endpoint": api_endpoint}
)

# Call the Gemma2 endpoint
gemma2_end_point = (
    f"projects/{PROJECT_ID}/locations/{ENDPOINT_REGION}/endpoints/{ENDPOINT_ID}"
)
response = client.predict(
    endpoint=gemma2_end_point,
    instances=instances,
)
text_responses = response.predictions
print(text_responses[0])

Node.js

Before trying this sample, follow the Node.js setup instructions in the Vertex AI quickstart using client libraries. For more information, see the Vertex AI Node.js API reference documentation.

To authenticate to Vertex AI, set up Application Default Credentials. For more information, see Set up authentication for a local development environment.

// Imports the Google Cloud Prediction Service Client library
const {
  // TODO(developer): Uncomment PredictionServiceClient before running the sample.
  // PredictionServiceClient,
  helpers,
} = require('@google-cloud/aiplatform');
/**
 * TODO(developer): Update these variables before running the sample.
 */
const projectId = 'your-project-id';
const endpointRegion = 'your-vertex-endpoint-region';
const endpointId = 'your-vertex-endpoint-id';

// Prompt used in the prediction
const prompt = 'Why is the sky blue?';

// Encapsulate the prompt in a correct format for TPUs
// Example format: [{prompt: 'Why is the sky blue?', temperature: 0.9}]
const input = {
  prompt,
  // Parameters for default configuration
  maxOutputTokens: 1024,
  temperature: 0.9,
  topP: 1.0,
  topK: 1,
};

// Convert input message to a list of GAPIC instances for model input
const instances = [helpers.toValue(input)];

// TODO(developer): Uncomment apiEndpoint and predictionServiceClient before running the sample.
// const apiEndpoint = `${endpointRegion}-aiplatform.googleapis.com`;

// Create a client
// predictionServiceClient = new PredictionServiceClient({apiEndpoint});

// Call the Gemma2 endpoint
const gemma2Endpoint = `projects/${projectId}/locations/${endpointRegion}/endpoints/${endpointId}`;

const [response] = await predictionServiceClient.predict({
  endpoint: gemma2Endpoint,
  instances,
});

const predictions = response.predictions;
const text = predictions[0].stringValue;

console.log('Predictions:', text);

Java

Before trying this sample, follow the Java setup instructions in the Vertex AI quickstart using client libraries. For more information, see the Vertex AI Java API reference documentation.

To authenticate to Vertex AI, set up Application Default Credentials. For more information, see Set up authentication for a local development environment.

import com.google.cloud.aiplatform.v1.EndpointName;
import com.google.cloud.aiplatform.v1.PredictResponse;
import com.google.cloud.aiplatform.v1.PredictionServiceClient;
import com.google.cloud.aiplatform.v1.PredictionServiceSettings;
import com.google.gson.Gson;
import com.google.protobuf.InvalidProtocolBufferException;
import com.google.protobuf.Value;
import com.google.protobuf.util.JsonFormat;
import java.io.IOException;
import java.util.ArrayList;
import java.util.HashMap;
import java.util.List;
import java.util.Map;

public class Gemma2PredictTpu {
  private final PredictionServiceClient predictionServiceClient;

  // Constructor to inject the PredictionServiceClient
  public Gemma2PredictTpu(PredictionServiceClient predictionServiceClient) {
    this.predictionServiceClient = predictionServiceClient;
  }

  public static void main(String[] args) throws IOException {
    // TODO(developer): Replace these variables before running the sample.
    String projectId = "YOUR_PROJECT_ID";
    String endpointRegion = "us-west1";
    String endpointId = "YOUR_ENDPOINT_ID";

    PredictionServiceSettings predictionServiceSettings =
        PredictionServiceSettings.newBuilder()
            .setEndpoint(String.format("%s-aiplatform.googleapis.com:443", endpointRegion))
            .build();
    PredictionServiceClient predictionServiceClient =
        PredictionServiceClient.create(predictionServiceSettings);
    Gemma2PredictTpu creator = new Gemma2PredictTpu(predictionServiceClient);

    creator.gemma2PredictTpu(projectId, endpointRegion, endpointId);
  }

  // Demonstrates how to run inference on a Gemma2 model
  // deployed to a Vertex AI endpoint with TPU accelerators.
  public String gemma2PredictTpu(String projectId, String region,
           String endpointId) throws IOException {
    Map<String, Object> paramsMap = new HashMap<>();
    paramsMap.put("temperature", 0.9);
    paramsMap.put("maxOutputTokens", 1024);
    paramsMap.put("topP", 1.0);
    paramsMap.put("topK", 1);
    Value parameters = mapToValue(paramsMap);
    // Prompt used in the prediction
    String instance = "{ \"prompt\": \"Why is the sky blue?\"}";
    Value.Builder instanceValue = Value.newBuilder();
    JsonFormat.parser().merge(instance, instanceValue);
    // Encapsulate the prompt in a correct format for TPUs
    // Example format: [{'prompt': 'Why is the sky blue?', 'temperature': 0.9}]
    List<Value> instances = new ArrayList<>();
    instances.add(instanceValue.build());

    EndpointName endpointName = EndpointName.of(projectId, region, endpointId);

    PredictResponse predictResponse = this.predictionServiceClient
        .predict(endpointName, instances, parameters);
    String textResponse = predictResponse.getPredictions(0).getStringValue();
    System.out.println(textResponse);
    return textResponse;
  }

  private static Value mapToValue(Map<String, Object> map) throws InvalidProtocolBufferException {
    Gson gson = new Gson();
    String json = gson.toJson(map);
    Value.Builder builder = Value.newBuilder();
    JsonFormat.parser().merge(json, builder);
    return builder.build();
  }
}

Go

Before trying this sample, follow the Go setup instructions in the Vertex AI quickstart using client libraries. For more information, see the Vertex AI Go API reference documentation.

To authenticate to Vertex AI, set up Application Default Credentials. For more information, see Set up authentication for a local development environment.

import (
	"context"
	"fmt"
	"io"

	"cloud.google.com/go/aiplatform/apiv1/aiplatformpb"

	"google.golang.org/protobuf/types/known/structpb"
)

// predictTPU demonstrates how to run interference on a Gemma2 model deployed to a Vertex AI endpoint with TPU accelerators.
func predictTPU(w io.Writer, client PredictionsClient, projectID, location, endpointID string) error {
	ctx := context.Background()

	// Note: client can be initialized in the following way:
	// apiEndpoint := fmt.Sprintf("%s-aiplatform.googleapis.com:443", location)
	// client, err := aiplatform.NewPredictionClient(ctx, option.WithEndpoint(apiEndpoint))
	// if err != nil {
	// 	return fmt.Errorf("unable to create prediction client: %v", err)
	// }
	// defer client.Close()

	gemma2Endpoint := fmt.Sprintf("projects/%s/locations/%s/endpoints/%s", projectID, location, endpointID)
	prompt := "Why is the sky blue?"
	parameters := map[string]interface{}{
		"temperature":     0.9,
		"maxOutputTokens": 1024,
		"topP":            1.0,
		"topK":            1,
	}

	// Encapsulate the prompt in a correct format for TPUs.
	// Example format: [{'prompt': 'Why is the sky blue?', 'temperature': 0.9}]
	promptValue, err := structpb.NewValue(map[string]interface{}{
		"prompt":     prompt,
		"parameters": parameters,
	})
	if err != nil {
		fmt.Fprintf(w, "unable to convert prompt to Value: %v", err)
		return err
	}

	req := &aiplatformpb.PredictRequest{
		Endpoint:  gemma2Endpoint,
		Instances: []*structpb.Value{promptValue},
	}

	resp, err := client.Predict(ctx, req)
	if err != nil {
		return err
	}

	prediction := resp.GetPredictions()
	value := prediction[0].GetStringValue()
	fmt.Fprintf(w, "%v", value)

	return nil
}

Clean up

To avoid incurring charges to your Google Cloud account for the resources used in this tutorial, either delete the project that contains the resources, or keep the project and delete the individual resources.

Delete the project

Caution: Deleting a project has the following effects:

• Everything in the project is deleted. If you used an existing project for the tasks in this document, when you delete it, you also delete any other work you've done in the project.
• Custom project IDs are lost. When you created this project, you might have created a custom project ID that you want to use in the future. To preserve the URLs that use the project ID, such as an appspot.com URL, delete selected resources inside the project instead of deleting the whole project.

If you plan to explore multiple architectures, tutorials, or quickstarts, reusing projects can help you avoid exceeding project quota limits.

1. In the Google Cloud console, go to the Manage resources page.

   Go to Manage resources

2. In the project list, select the project that you want to delete, and then click Delete.
3. In the dialog, type the project ID, and then click Shut down to delete the project.

Delete individual resources

If you're keeping your project, delete the resources used in this tutorial:

• Undeploy the model and delete the endpoint
• Delete the model from Model Registry

Undeploy the model and delete the endpoint

Use one of the following methods to undeploy a model and delete the endpoint.

Note: You can only delete the endpoint after all models have been undeployed from it.

Console

1. In the Google Cloud console, click Online prediction and then click Endpoints.

   Go to the Endpoints page

2. In the Region drop-down list, choose the region where you deployed your endpoint.

3. Click the endpoint name to open the details page. For example: gemma2-2b-it-mg-one-click-deploy.

4. On the row for the Gemma 2 (Version 1) model, click Actions, and then click Undeploy model from endpoint.

5. In the Undeploy model from endpoint dialog, click Undeploy.

6. Click the Back button to return to the Endpoints page.

   Go to the Endpoints page

7. At the end of the gemma2-2b-it-mg-one-click-deploy row, click Actions, and then select Delete endpoint.

8. In the confirmation prompt, click Confirm.

gcloud

To undeploy the model and delete the endpoint using the Google Cloud CLI, follow these steps.

In these commands, replace:

• PROJECT_ID with your project name
• LOCATION_ID with the region where you deployed the model and endpoint
• ENDPOINT_ID with the endpoint ID
• DEPLOYED_MODEL_NAME with the model's display name
• DEPLOYED_MODEL_ID with the model ID

1. Get the endpoint ID by running the gcloud ai endpoints list command. This command lists the endpoint IDs for all endpoints in your project. Make a note of the ID of the endpoint used in this tutorial.

   gcloud ai endpoints list \
       --project=PROJECT_ID \
       --region=LOCATION_ID
   

   The output looks like this. In the output, the ID is called ENDPOINT_ID.

   Using endpoint [https://us-central1-aiplatform.googleapis.com/]
   ENDPOINT_ID: 1234567891234567891
   DISPLAY_NAME: gemma2-2b-it-mg-one-click-deploy
   

2. Get the model ID by running the gcloud ai models describe command. Make a note of the ID of the model you deployed in this tutorial.

   gcloud ai models describe DEPLOYED_MODEL_NAME \
       --project=PROJECT_ID \
       --region=LOCATION_ID
   

   The abbreviated output looks like this. In the output, the ID is called deployedModelId.

   Using endpoint [https://us-central1-aiplatform.googleapis.com/]
   artifactUri: [URI removed]
   baseModelSource:
     modelGardenSource:
       publicModelName: publishers/google/models/gemma2
   ...
   deployedModels:
   - deployedModelId: '1234567891234567891'
     endpoint: projects/12345678912/locations/us-central1/endpoints/12345678912345
   displayName: gemma2-2b-it-12345678912345
   etag: [ETag removed]
   modelSourceInfo:
     sourceType: MODEL_GARDEN
   name: projects/123456789123/locations/us-central1/models/gemma2-2b-it-12345678912345
   ...
   

3. Undeploy the model from the endpoint. You'll need the endpoint ID and model ID from the previous commands.

   gcloud ai endpoints undeploy-model ENDPOINT_ID \
       --project=PROJECT_ID \
       --region=LOCATION_ID \
       --deployed-model-id=DEPLOYED_MODEL_ID
   

   This command produces no output.

4. Run the gcloud ai endpoints delete command to delete the endpoint.

   gcloud ai endpoints delete ENDPOINT_ID \
       --project=PROJECT_ID \
       --region=LOCATION_ID
   

   When promted, type y to confirm. This command produces no output.

Delete the model

Console

1. Go to the Model Registry page from the Vertex AI section in the Google Cloud console.

   Go to the Model Registry page

2. In the Region drop-down list, choose the region where you deployed your model.

3. At the end of the gemma2-2b-it-1234567891234 row, click Actions.

4. Select Delete model.

   When you delete the model, all associated model versions and evaluations are deleted from your Google Cloud project.

5. In the confirmation prompt, click Delete.

gcloud

To delete the model using the Google Cloud CLI, provide the model's display name and region to the gcloud ai models delete command.

gcloud ai models delete DEPLOYED_MODEL_NAME \
    --project=PROJECT_ID \
    --region=LOCATION_ID

Replace DEPLOYED_MODEL_NAME with the model's display name. Replace PROJECT_ID with your project name. Replace LOCATION_ID with the region where you deployed the model.

What's next

• Learn more about Gemma open models.
• Read the Gemma Terms of Use.
• Learn more about open models.
• Learn how to deploy a tuned model.
• Learn how to deploy Gemma 2 to Google Kubernetes Engine using HuggingFace Textgen Inference (TGI).
• Learn more about the PredictionServiceClient in your preferred language: Python, Node.js, Java, or Go.