GPULlama3.java

GPULlama3.java 基于 TornadoVM 构建，可利用 GPU 与异构计算，直接从 Java 加速 LLM 推理。目前，GPULlama3.java 通过 PTX 和 OPENCL 后端支持在 NVIDIA、AMD GPU 以及 Apple Silicon 上进行推理。

项目设置

要将 langchain4j 安装到你的项目中，请添加以下依赖：

对于 Maven 项目的 pom.xml

<dependency>
    <groupId>dev.langchain4j</groupId>
    <artifactId>langchain4j</artifactId>
    <version>1.18.1</version>
</dependency>

<dependency>
<groupId>dev.langchain4j</groupId>
<artifactId>langchain4j-gpu-llama3</artifactId>
<version>1.18.1-beta28</version>
</dependency>

对于 Gradle 项目的 build.gradle

implementation 'dev.langchain4j:langchain4j:1.18.1'
implementation 'dev.langchain4j:langchain4j-gpu-llama3:1.18.1-beta28'

模型兼容性

目前，GPULlama3.java 支持以下 GGUF 格式的模型，格式为 FP16、Q8 和 Q4：注意，对于 Q8 和 Q4 模型，加载时会反量化为 FP16。我们在 HuggingFace 仓库中维护了已测试模型的合集。

Llama3
Mistral
Qwen2.5
Qwen3.0
Phi-3
DeepSeek-R1-Distill-Qwen-1.5B-GGUF

聊天补全

聊天模型允许你使用在对话数据上微调的模型生成类人响应。

同步

创建一个类并添加以下代码。

prompt = "What is the capital of France?";
ChatRequest request = ChatRequest.builder().messages(
    UserMessage.from(prompt),
    SystemMessage.from("reply with extensive sarcasm"))
    .build();

Path modelPath = Paths.get("beehive-llama-3.2-1b-instruct-fp16.gguf");

GPULlama3ChatModel model = GPULlama3ChatModel.builder()
        .modelPath(modelPath)
        .onGPU(Boolean.TRUE) //if false, runs on CPU though a lightweight implementation of llama3.java
        .build();
ChatResponse response = model.chat(request);
System.out.println("\n" + response.aiMessage().text());

流式

创建一个类并添加以下代码。

public static void main(String[] args) {
    CompletableFuture<ChatResponse> futureResponse = new CompletableFuture<>();

    String prompt;

    if (args.length > 0) {
        prompt = args[0];
        System.out.println("User Prompt: " + prompt);
    } else {
        prompt = "What is the capital of France?";
        System.out.println("Example Prompt: " + prompt);
    }

    ChatRequest request = ChatRequest.builder().messages(
                    UserMessage.from(prompt),
                    SystemMessage.from("reply with extensive sarcasm"))
            .build();

    Path modelPath = Paths.get("beehive-llama-3.2-1b-instruct-fp16.gguf");


    GPULlama3StreamingChatModel model = GPULlama3StreamingChatModel.builder()
            .onGPU(Boolean.TRUE) // if false, runs on CPU though a lightweight implementation of llama3.java
            .modelPath(modelPath)
            .build();

    model.chat(request, new StreamingChatResponseHandler() {

        @Override
        public void onPartialResponse(String partialResponse) {
            System.out.print(partialResponse);
        }

        @Override
        public void onCompleteResponse(ChatResponse completeResponse) {
            futureResponse.complete(completeResponse);
            model.printLastMetrics();
        }

        @Override
        public void onError(Throwable error) {
            futureResponse.completeExceptionally(error);
        }
    });

    futureResponse.join();
}

如何运行：

需要配置 TornadoVM 才能运行示例详细说明见 Setup & Configure

步骤 1 — 获取 Tornado JVM 标志

运行以下命令（需要已安装 Tornado）：

tornado --printJavaFlags

示例输出：

/home/mikepapadim/.sdkman/candidates/java/current/bin/java -server \
 -XX:+UnlockExperimentalVMOptions -XX:+EnableJVMCI \
-XX:-UseCompressedClassPointers --enable-preview \
-Djava.library.path=/home/mikepapadim/java-ai-demos/GPULlama3.java/external/tornadovm/bin/sdk/lib \
--module-path .:/home/mikepapadim/java-ai-demos/GPULlama3.java/external/tornadovm/bin/sdk/share/java/tornado \
-Dtornado.load.api.implementation=uk.ac.manchester.tornado.runtime.tasks.TornadoTaskGraph \
-Dtornado.load.runtime.implementation=uk.ac.manchester.tornado.runtime.TornadoCoreRuntime \
-Dtornado.load.tornado.implementation=uk.ac.manchester.tornado.runtime.common.Tornado \
-Dtornado.load.annotation.implementation=uk.ac.manchester.tornado.annotation.ASMClassVisitor \
-Dtornado.load.annotation.parallel=uk.ac.manchester.tornado.api.annotations.Parallel \
--upgrade-module-path /home/mikepapadim/java-ai-demos/GPULlama3.java/external/tornadovm/bin/sdk/share/java/graalJars \
-XX:+UseParallelGC \
@/home/mikepapadim/java-ai-demos/GPULlama3.java/external/tornadovm/bin/sdk/etc/exportLists/common-exports \
@/home/mikepapadim/java-ai-demos/GPULlama3.java/external/tornadovm/bin/sdk/etc/exportLists/opencl-exports \
--add-modules ALL-SYSTEM,tornado.runtime,tornado.annotation,tornado.drivers.common,tornado.drivers.opencl

步骤 2 — 构建 Maven classpath

在项目根目录运行：

mvn dependency:build-classpath -Dmdep.outputFile=cp.txt

步骤 3 — 构建 Maven classpath

mvn clean package

主 JAR 将位于：

target/gpullama3.java-example-1.18.1-beta28.jar

步骤 4 — 直接用 Java 运行程序

现在你可以用所有 JVM 和 Tornado 标志运行示例：

JAVA_BIN=/home/mikepapadim/.sdkman/candidates/java/current/bin/java
CP="target/gpullama3.java-example-1.18.1-beta28.jar:$(cat cp.txt)"

$JAVA_BIN \
  -server \
  -XX:+UnlockExperimentalVMOptions \
  -XX:+EnableJVMCI \
  --enable-preview \
  -Djava.library.path=/home/mikepapadim/java-ai-demos/GPULlama3.java/external/tornadovm/bin/sdk/lib \
  --module-path .:/home/mikepapadim/java-ai-demos/GPULlama3.java/external/tornadovm/bin/sdk/share/java/tornado \
  -Dtornado.load.api.implementation=uk.ac.manchester.tornado.runtime.tasks.TornadoTaskGraph \
  -Dtornado.load.runtime.implementation=uk.ac.manchester.tornado.runtime.TornadoCoreRuntime \
  -Dtornado.load.tornado.implementation=uk.ac.manchester.tornado.runtime.common.Tornado \
  -Dtornado.load.annotation.implementation=uk.ac.manchester.tornado.annotation.ASMClassVisitor \
  -Dtornado.load.annotation.parallel=uk.ac.manchester.tornado.api.annotations.Parallel \
  --upgrade-module-path /home/mikepapadim/java-ai-demos/GPULlama3.java/external/tornadovm/bin/sdk/share/java/graalJars \
  -XX:+UseParallelGC \
  @/home/mikepapadim/java-ai-demos/GPULlama3.java/external/tornadovm/bin/sdk/etc/exportLists/common-exports \
  @/home/mikepapadim/java-ai-demos/GPULlama3.java/external/tornadovm/bin/sdk/etc/exportLists/opencl-exports \
  --add-modules ALL-SYSTEM,tornado.runtime,tornado.annotation,tornado.drivers.common,tornado.drivers.opencl \
  -Xms6g -Xmx6g \
  -Dtornado.device.memory=6GB \
  -cp "$CP" \
  GPULlama3ChatModelExamples

预期输出：

WARNING: Using incubator modules: jdk.incubator.vector
Example Prompt: What is the capital of France?
SLF4J(W): No SLF4J providers were found.
SLF4J(W): Defaulting to no-operation (NOP) logger implementation
SLF4J(W): See https://www.slf4j.org/codes.html#noProviders for further details.
Wow, I'm so glad you asked. I've been waiting for someone to finally ask me this question. It's not like I have better things to do, like take a nap or something. So, yes, the capital of France is... (dramatic pause) ...Paris!

achieved tok/s: 48.86. Tokens: 87, seconds: 1.78

说明：

GPU 利用率可用 nvidia-smi（针对 NVIDIA GPU）或 'nvtop' 等适用于 AMD/Apple Silicon 的工具进行监控。

项目设置​

模型兼容性​

聊天补全​

同步​

流式​

如何运行：​

步骤 1 — 获取 Tornado JVM 标志​

步骤 2 — 构建 Maven classpath​

步骤 3 — 构建 Maven classpath​

步骤 4 — 直接用 Java 运行程序​

预期输出：​

说明：​