使用带 GPU 加速器的 Gemma2 模型生成文本

此代码示例演示了如何对部署到具有 GPU 加速器的 Vertex AI 端点的 Gemma2 模型运行推理。

深入探索

如需查看包含此代码示例的详细文档，请参阅以下内容：

代码示例

Go

在尝试此示例之前，请按照《Vertex AI 快速入门：使用客户端库》中的 Go 设置说明执行操作。 如需了解详情，请参阅 Vertex AI Go API 参考文档。

如需向 Vertex AI 进行身份验证，请设置应用默认凭证。 如需了解详情，请参阅为本地开发环境设置身份验证。 

import (
	"context"
	"fmt"
	"io"

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

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

// predictGPU demonstrates how to run interference on a Gemma2 model deployed to a Vertex AI endpoint with GPU accelerators.
func predictGPU(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.
	// Pay attention that prompt should be set in "inputs" field.
	// Example format: [{'inputs': 'Why is the sky blue?', 'parameters': {'temperature': 0.9}}]
	promptValue, err := structpb.NewValue(map[string]interface{}{
		"inputs":     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
}

Java

在尝试此示例之前，请按照《Vertex AI 快速入门：使用客户端库》中的 Java 设置说明执行操作。 如需了解详情，请参阅 Vertex AI Java API 参考文档。

如需向 Vertex AI 进行身份验证，请设置应用默认凭证。 如需了解详情，请参阅为本地开发环境设置身份验证。 


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 Gemma2PredictGpu {

  private final PredictionServiceClient predictionServiceClient;

  // Constructor to inject the PredictionServiceClient
  public Gemma2PredictGpu(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-east4";
    String endpointId = "YOUR_ENDPOINT_ID";

    PredictionServiceSettings predictionServiceSettings =
        PredictionServiceSettings.newBuilder()
            .setEndpoint(String.format("%s-aiplatform.googleapis.com:443", endpointRegion))
            .build();
    PredictionServiceClient predictionServiceClient =
        PredictionServiceClient.create(predictionServiceSettings);
    Gemma2PredictGpu creator = new Gemma2PredictGpu(predictionServiceClient);

    creator.gemma2PredictGpu(projectId, endpointRegion, endpointId);
  }

  // Demonstrates how to run inference on a Gemma2 model
  // deployed to a Vertex AI endpoint with GPU accelerators.
  public String gemma2PredictGpu(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 = "{ \"inputs\": \"Why is the sky blue?\"}";
    Value.Builder instanceValue = Value.newBuilder();
    JsonFormat.parser().merge(instance, instanceValue);
    // Encapsulate the prompt in a correct format for GPUs
    // Example format: [{'inputs': 'Why is the sky blue?', 'parameters': {'temperature': 0.8}}]
    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();
  }
}

Node.js

在尝试此示例之前，请按照《Vertex AI 快速入门：使用客户端库》中的 Node.js 设置说明执行操作。 如需了解详情，请参阅 Vertex AI Node.js API 参考文档。

如需向 Vertex AI 进行身份验证，请设置应用默认凭证。 如需了解详情，请参阅为本地开发环境设置身份验证。 

async function gemma2PredictGpu(predictionServiceClient) {
  // 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';

  // Default configuration
  const config = {maxOutputTokens: 1024, temperature: 0.9, topP: 1.0, topK: 1};
  // Prompt used in the prediction
  const prompt = 'Why is the sky blue?';

  // Encapsulate the prompt in a correct format for GPUs
  // Example format: [{inputs: 'Why is the sky blue?', parameters: {temperature: 0.9}}]
  const input = {
    inputs: prompt,
    parameters: config,
  };

  // 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);
  return text;
}

module.exports = gemma2PredictGpu;

// TODO(developer): Uncomment below lines before running the sample.
// gemma2PredictGpu(...process.argv.slice(2)).catch(err => {
//   console.error(err.message);
//   process.exitCode = 1;
// });

Python

在尝试此示例之前，请按照《Vertex AI 快速入门：使用客户端库》中的 Python 设置说明执行操作。 如需了解详情，请参阅 Vertex AI Python API 参考文档。

如需向 Vertex AI 进行身份验证，请设置应用默认凭证。 如需了解详情，请参阅为本地开发环境设置身份验证。 

"""
Sample to run inference on a Gemma2 model deployed to a Vertex AI endpoint with GPU 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 GPUs
# Example format: [{'inputs': 'Why is the sky blue?', 'parameters': {'temperature': 0.9}}]
input = {"inputs": prompt, "parameters": 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])

后续步骤

如需搜索和过滤其他 Google Cloud 产品的代码示例，请参阅Google Cloud 示例浏览器。