全Java方案：本地部署DeepSeek并集成Spring Boot项目

原创已于 2025-03-08 16:57:13 修改 · 972 阅读

10 ·

CC 4.0 BY-SA版权

文章标签：

#java #spring boot #ai

于 2025-03-08 16:52:58 首次发布

全Java方案：本地部署DeepSeek并集成Spring Boot项目

一、方案特点

纯Java生态：基于DJL（Deep Java Library）实现模型推理
零Python依赖：全程使用Java技术栈
生产就绪：内置线程池管理与性能优化

二、环境准备

1. 硬件要求

CPU：Intel i7+（支持AVX512指令集）
内存：16GB+（模型加载需要）
磁盘：10GB+可用空间

2. 开发环境

JDK 17+
Maven 3.8+
Spring Boot 3.1+

三、本地模型部署（Java版）

1. 添加Maven依赖

<!-- pom.xml -->
<dependency>
    <groupId>ai.djl</groupId>
    <artifactId>api</artifactId>
    <version>0.25.0</version>
</dependency>
<dependency>
    <groupId>ai.djl.pytorch</groupId>
    <artifactId>pytorch-engine</artifactId>
    <version>0.25.0</version>
    <scope>runtime</scope>
</dependency>
<dependency>
    <groupId>ai.djl.huggingface</groupId>
    <artifactId>tokenizers</artifactId>
    <version>0.25.0</version>
</dependency>

2. 实现推理服务

@Service
public class DeepSeekService {

    private static final String MODEL_NAME = "deepseek-ai/deepseek-llm-1.3b";
    private final Predictor<String, String> predictor;

    public DeepSeekService() throws ModelException, IOException {
        Criteria<String, String> criteria = Criteria.builder()
                .setTypes(String.class, String.class)
                .optModelUrls("djl://ai.djl.huggingface.pytorch/" + MODEL_NAME)
                .optEngine("PyTorch")
                .optTranslator(new TextTranslator())
                .build();

        this.predictor = criteria.loadModel().newPredictor();
    }

    @PreDestroy
    public void cleanup() {
        predictor.close();
    }

    public String generate(String prompt) throws TranslateException {
        return predictor.predict(prompt);
    }

    static class TextTranslator implements Translator<String, String> {
        @Override
        public Pipeline getPipeline() {
            return new Pipeline(new HuggingFaceTokenizer("deepseek-ai/deepseek-llm-1.3b"));
        }

        @Override
        public String processOutput(TranslatorContext ctx, NDList list) {
            return ctx.getTokenizer().decode(list.get(0).toLongArray());
        }
    }
}

四、Spring Boot集成

1. 配置线程池

@Configuration
public class ModelConfig {

    @Bean
    public ExecutorService modelExecutor() {
        return Executors.newFixedThreadPool(4);
    }
}

2. 实现REST接口

@RestController
@RequestMapping("/api/ai")
public class AIController {

    private final DeepSeekService deepSeekService;
    private final ExecutorService executor;

    public AIController(DeepSeekService deepSeekService, 
                       @Qualifier("modelExecutor") ExecutorService executor) {
        this.deepSeekService = deepSeekService;
        this.executor = executor;
    }

    @PostMapping("/generate")
    public CompletableFuture<ResponseEntity<String>> generateText(
            @RequestBody Map<String, String> request) {
        String prompt = request.get("prompt");
        return CompletableFuture.supplyAsync(
                () -> {
                    try {
                        return ResponseEntity.ok(deepSeekService.generate(prompt));
                    } catch (Exception e) {
                        return ResponseEntity.internalServerError().build();
                    }
                }, 
                executor
        );
    }
}

五、性能优化配置

1. 应用配置

# application.properties
# 模型缓存目录
djl.pytorch.model_dir=classpath:/models
# 开启原生加速
djl.pytorch.num_interop_threads=4
djl.pytorch.num_threads=8

2. 内存管理

// 在DeepSeekService中添加
.optArgument("mapLocation", "true") // GPU显存优化
.optArgument("inferenceMode", "true") // 减少内存占用

六、本地模型管理

1. 模型下载脚本

#!/bin/bash
wget https://deepseek-model.oss-cn-beijing.aliyuncs.com/deepseek-llm-1.3b.zip
unzip deepseek-llm-1.3b.zip -d src/main/resources/models/

2. 模型验证工具

@Component
public class ModelValidator implements CommandLineRunner {

    @Override
    public void run(String... args) throws Exception {
        try (Model model = Model.newInstance("deepseek")) {
            model.load(Paths.get("src/main/resources/models"));
            System.out.println("模型加载验证成功");
        }
    }
}

七、生产级增强

1. 健康检查端点

@Component
public class ModelHealthIndicator implements HealthIndicator {

    @Override
    public Health health() {
        try {
            Model.newInstance("healthcheck").load(Paths.get("models"));
            return Health.up().build();
        } catch (Exception e) {
            return Health.down().withDetail("error", e.getMessage()).build();
        }
    }
}

2. 限流保护

@Bean
public RateLimiter rateLimiter() {
    return RateLimiter.create(10); // 每秒10个请求
}

@Around("@annotation(rateLimited)")
public Object rateLimit(ProceedingJoinPoint joinPoint) throws Throwable {
    if (rateLimiter.tryAcquire()) {
        return joinPoint.proceed();
    }
    throw new ResponseStatusException(HttpStatus.TOO_MANY_REQUESTS);
}

八、典型问题解决方案

1. 中文乱码修复

// 在Translator实现中添加
@Override
protected void prepareInput(TranslatorContext ctx, String input) {
    ctx.setAttachment("charset", StandardCharsets.UTF_8.name());
}

2. 长文本处理

public String generate(String prompt) {
    // 分块处理长文本
    return splitText(prompt).stream()
            .parallel()
            .map(this::processChunk)
            .collect(Collectors.joining());
}

private List<String> splitText(String text) {
    // 实现基于token的分块逻辑
}