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

"""
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])

// 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);

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();
  }
}

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
}