挑战 Qwen 3.5：Gemma 4 私有部署与文本、图像、视频、音频处理及思考模式、工具调用全教程

GPUStack
0 阅读25分钟

随着多模态和智能体能力逐渐成为大模型的标配，如何在本地环境中高效部署正在成为越来越多开发者关注的重点。近期，Gemma 4 正式发布，模型能力对标 Qwen 3.5，在推理能力、多模态支持以及工具调用等方面都有明显提升。

以下为 Gemma 4 不同型号的简要介绍：

模型上下文能力多模态理解
Gemma 4 E2B128K思考模式，工具调用文本、图像、视频、音频
Gemma 4 E4B128K思考模式，工具调用文本、图像、视频、音频
Gemma 4 31B256K思考模式，工具调用文本、图像、视频
Gemma 4 26B A4B256K思考模式，工具调用文本、图像、视频

Gemma 4 在设计上更加全面：原生支持文本、图像、音频甚至视频输入，并具备思考模式、工具调用、结构化输出等能力，可以直接用于构建智能体工作流。这也意味着，单一模型即可完整实现一套接近生产形态的 AI 能力。

本文将基于 GPUStack，在本地环境中完整部署 Gemma 4 模型，并逐步测试其各项能力，包括文本生成、图像理解、音频与视频处理、思考模式（thinking）、以及工具调用（tool calling）等。同时，也会结合实际部署过程，介绍在推理后端、模型配置以及多模态支持方面需要注意的关键点。

如果你希望在本地搭建一套具备多模态与智能体能力的大模型环境，这篇文章可以作为一个完整的实践参考。

GPUStack 安装与集群初始化

GPUStack 是一个开源 GPU 集群管理与 AI 模型服务平台，旨在高效部署 AI 模型。它可以配置并编排多种推理引擎——如 vLLM、SGLang、TensorRT-LLM，甚至自定义引擎——以在 GPU 集群上实现最佳性能。核心功能包括多 GPU 集群管理、可插拔推理引擎架构、Day 0 模型支持、性能优化配置（低延迟/高吞吐）、以及企业级运维能力，如故障恢复、负载均衡、监控与权限管理。

在开始部署 Gemma 4 之前，首先需要完成 GPUStack 控制面的安装，并将 NVIDIA GPU 节点纳入管理。

准备容器环境

GPUStack 以容器方式运行，因此需要提前准备好容器运行环境（如 Docker、Podman 或 Kubernetes）。本文以 Docker 为例进行说明。

在各节点上安装 Docker，确保服务已正常启动：

docker info

启动 GPUStack Server

GPUStack Server 无需依赖 GPU，可运行在普通 CPU 节点上，也可运行在 GPU 节点。本文以一台双卡 NVIDIA 4090 48G 节点为实验环境，在该节点上启动 GPUStack Server 容器：

sudo docker run -d --name gpustack \
  --restart unless-stopped \
  -p 80:80 \
  --volume gpustack-data:/var/lib/gpustack \
  swr.cn-south-1.myhuaweicloud.com/gpustack/gpustack:v2.1.1 \
  --debug --bootstrap-password GPUStack@123

关键参数说明：

  • -p 80:80：用于对外暴露 Web 控制台端口；如需修改为其他端口（例如 9999），可调整为 -p 9999:80。
  • --volume：持久化平台数据（包括模型服务、计量数据、API Key 等）
  • --bootstrap-password：初始化 admin 用户密码
  • --debug：开启调试日志，便于排查问题

容器启动后，可以通过日志确认服务是否正常运行：

docker logs -f gpustack

访问控制台并初始化

打开浏览器访问：http://<Server 主机 IP>:80

使用默认账号登录：

  • 用户名：admin
  • 密码：GPUStack@123

登录后，首先创建一个 Docker 类型的集群，用于统一管理后续接入的 GPU 节点。

image-20260404155239420

添加 NVIDIA GPU Worker 节点

在集群创建完成后，可以接入 NVIDIA GPU 节点。

在添加节点之前，先完成基础环境检查。

（1）驱动版本检查

在目标节点上执行以下命令：

nvidia-smi

该命令会显示当前安装的 NVIDIA 驱动版本。请确认驱动版本 ≥ 575，以保证对 Gemma 4 模型的兼容性和稳定性。

（2）Nvidia Container Toolkit 检查

执行以下命令检查 Docker 是否正确配置了 Nvidia Container Toolkit：

sudo docker info 2>/dev/null | grep -q "Runtime.*nvidia" && echo "Nvidia Container Toolkit OK" || (echo "Nvidia Container Toolkit not configured"; exit 1)
  • 该命令会从 docker info 输出中查找是否存在 nvidia 运行时配置。
  • 如果输出 "Nvidia Container Toolkit OK"，说明 Docker 已正确配置，可在容器中访问 GPU。
  • 如果输出 "Nvidia Container Toolkit not configured"，则说明未正确配置，需要安装并启用 Nvidia Container Toolkit，否则推理容器无法使用 GPU 资源。
（3）接入 Worker 节点

在 GPUStack 控制台中，选择添加节点（Worker），并复制系统生成的接入命令，在目标节点执行。

image-20260404154936516

该命令本质上会启动一个 Worker 容器，并自动注册到 Server。

（4）验证 Worker 状态

节点接入后，可以在节点上查看容器日志：

docker logs -f gpustack

同时，在 GPUStack 控制台中也可以看到节点状态是否为 Ready。

至此，GPUStack 的控制面已成功部署，NVIDIA GPU 节点也顺利接入集群，并能够正常采集 GPU 名称、索引、厂商信息、温度、利用率及显存使用等指标。接下来即可在该环境中部署具体的推理服务。

image-20260404155349407

添加自定义 vLLM 和 SGLang 版本

GPUStack 支持可插拔的推理引擎架构，允许自定义推理后端及其版本，用于引入 GPUStack 未内置的 vLLM / SGLang / MindIE 版本，或接入其他自定义推理引擎镜像。

以 vLLM 最新发布的 v0.19.0 和 SGLang 最新发布的 v0.5.10 为例。

vLLM

在推理后端菜单，编辑 vLLM，在版本配置中选择添加版本，添加一个新的 vLLM 版本，指向 vLLM 官方镜像：

配置值
版本0.19.0
镜像名称vllm/vllm-openai:v0.19.0
框架CUDA
覆盖镜像入口命令（ENTRYPOINT）vllm serve
执行命令{{model_path}} --host {{worker_ip}} --port {{port}} --served-model-name {{model_name}}

自定义添加 vLLM 0.19.0 配置如图所示：

image-20260404091212999

SGLang

在推理后端菜单，编辑 SGLang，在版本配置中选择添加版本，添加一个新的 SGLang 版本，指向 SGLang 官方镜像：

配置值
版本0.5.10
镜像名称lmsysorg/sglang:v0.5.10
框架CUDA
覆盖镜像入口命令（ENTRYPOINT）sglang serve
执行命令--model-path {{model_path}} --host {{worker_ip}} --port {{port}}

自定义添加 SGLang 0.5.10 配置如图所示：

image-20260406224602742

注意：

  1. GPUStack 会自动调用主机容器运行时拉取容器镜像，需要确保 Worker 节点可访问 Docker Hub，或者提前拉取好并重新 tag，并按需修改 UI 配置中的镜像地址；
  2. 保持执行命令中的 {{}} 不变，此为模板化配置。

也可以切换到 YAML 模式，直接使用以下的 YAML 导入：

vLLM

backend_name: vLLM
version_configs:
  0.19.0-custom:
    image_name: vllm/vllm-openai:v0.19.0
    entrypoint: vllm serve
    run_command: >-
      {{model_path}} --host {{worker_ip}} --port {{port}} --served-model-name
      {{model_name}}
    env: {}
    custom_framework: cuda

SGLang

backend_name: SGLang
version_configs:
  0.5.10-custom:
    image_name: lmsysorg/sglang:v0.5.10
    entrypoint: sglang serve
    run_command: '--model-path {{model_path}} --host {{worker_ip}} --port {{port}}'
    env: {}
    custom_framework: cuda

注意：如果当前已经有其它自定义版本，需要将其它自定义版本一同添加在 version_configs 中一起导入。

image-20260404091554773

截止目前（2026.4.6）， SGLang 对 Gemm4 的支持代码还未合并，以下使用 vLLM 来运行 Gemma 4。

由于运行 Gemma 4 需要 transformers 5.5.0 以及 vllm[audio] 依赖，建议自行构建支持 Gemma 4 的镜像。参考 Dockerfile 如下：

FROM vllm/vllm-openai:v0.19.0
RUN uv pip install --system vllm[audio] \
  && uv pip install --system transformers==5.5.0

构建 vllm/vllm-openai:v0.19.0-gemma4 镜像：

docker build -t vllm/vllm-openai:v0.19.0-gemma4 .

然后，在推理后端菜单中编辑 vLLM，切换到 YAML 模式，直接导入以下 YAML。该配置中声明了两个版本：官方的 0.19.0 版本，以及上方构建的 0.19.0-gemma4 版本。请确保用于部署 Gemma 4 的节点上已提前存在对应镜像：

backend_name: vLLM
version_configs:
  0.19.0-custom:
    image_name: vllm/vllm-openai:v0.19.0
    entrypoint: vllm serve
    run_command: >-
      {{model_path}} --host {{worker_ip}} --port {{port}} --served-model-name
      {{model_name}}
    env: {}
    custom_framework: cuda
  0.19.0-gemma4-custom:
    image_name: vllm/vllm-openai:v0.19.0-gemma4
    entrypoint: vllm serve
    run_command: >-
      {{model_path}} --host {{worker_ip}} --port {{port}} --served-model-name
      {{model_name}}
    env: {}
    custom_framework: cuda

成功导入后，点击 vLLM 卡片或编辑 vLLM ，查看两个自定义版本配置如下：

image-20260404161202006

image-20260404150129051

部署 Gemma4 模型

vLLM 已提供关于 Gemma4 模型的部署与使用教程，详情可参考：

docs.vllm.ai/projects/re…

以下将介绍在 GPUStack 上部署 Gemma4 模型的配置流程。

  1. 在在线环境下，可直接通过 HuggingFace 或 ModelScope 搜索 Gemma4 模型并进行部署，具体步骤参考下方。

  2. 在离线环境中，需要提前下载好模型权重，并将其分发到所有 Worker 节点，同时挂载到对应的 Worker 容器中。随后，在 GPUStack 控制台 - 模型文件菜单中，选择添加模型文件 - 本地路径，填写对应的模型权重路径。需要注意，这里填写的应为容器内路径，例如：

image-20260404092651530

以下测试在 2 块 4090 48GB GPU 环境下进行。

联网环境：在 GPUStack 控制台 - 部署菜单下，选择 部署模型 → ModelScope，直接搜索 Gemma4 模型进行部署。

离线环境：可从 GPUStack 控制台 - 模型文件菜单中，选择已添加的 Gemma4 模型进行部署。

以 google/gemma-4-31B-it 模型为例（it 为指令微调版本）：

  • 后端：选择 vLLM
  • 版本：选择前面自定义添加的 0.19.0-gemma4-custom

使用以下后端参数启动，后端参数支持单行或多行形式（注意 --tensor-parallel-size 2 已设置双卡张量并行，请确保有两块 GPU 可分配；其它环境请根据实际情况调整并行策略）：

--tensor-parallel-size 2 --max-model-len 32768 --gpu-memory-utilization 0.9 --enable-auto-tool-choice --reasoning-parser gemma4 --tool-call-parser gemma4  --async-scheduling --limit-mm-per-prompt '{"image": 4, "video": 1, "audio": 1}'

对于不同型号的显存建议如下：

  • 31B / 26B 模型：建议至少 80GB 显存
  • E2B / E4B 模型：建议至少 24GB 显存

image-20260404151825539

部署其它 Gemma 4 型号的流程与前面介绍的 31B 模型类似，包括选择 vLLM 后端、指定自定义版本、配置启动参数和并行策略等，此处不再赘述。

如果 NVIDIA 驱动版本 > 575，由于镜像中内置的 CUDA 版本为 12.9，可能会出现驱动版本与 CUDA 版本不匹配的提示。

说明：

  • 这是因为检查到当前驱动版本高于容器镜像内 CUDA 12.9 的预期驱动版本，但 CUDA 向上兼容更高版本的驱动。
  • 因此，这种提示可以忽略，直接选择仍然提交进行部署即可，模型仍能正常运行。

等待模型启动时，可以在操作中点击查看日志，实时观察 vLLM 的启动过程：

image-20260404160907974

当模型实例状态显示为 Running 时，说明模型已经成功启动，可以进行后续的文本、图像、音频、视频以及工具调用等功能测试。

模型能力测试

试验场

文本生成

Gemma 4 31B

image-20260404110901660

Gemma 4 26B A4B

image-20260404161506506

图像理解

Gemma 4 31B

image-20260404111021734

Gemma 4 26B A4B

image-20260404161654250

API 调用

在 GPUStack 控制台 - 路由菜单中，选择对应模型，点击操作 → API 接入信息，即可查看模型 API 的调用说明：

image-20260404114607960

  • 接入地址：统一通过 GPUStack 的认证 API 入口访问。
  • 模型名称：调用时在请求中填写对应的 model。
  • API 密钥：用于身份认证和权限校验。可按指引跳转到API 密钥管理页面创建 API Key，并可设置访问控制，仅允许访问指定模型。

image-20260404114721299

在试验场右上角点击查看代码，即可根据当前输入和参数设置，动态生成调用该模型的示例代码：

image-20260404114823610

音频识别

仅 Gemma 4 E4B 和 Gemma 4 E2B 模型支持音频识别（注意根据实际环境修改 URL、Bearer token 和 model）：

curl http://192.168.50.16/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer gpustack_0dd6f29690c7da2d_40d4a9ede6b5592209b5288b10b325f2" \
-d '{
  "seed": null,
  "stop": null,
  "temperature": 1,
  "top_p": 1,
  "max_tokens": 4096,
  "frequency_penalty": 0,
  "presence_penalty": 0,
  "model": "gemma-4-e4b-it",
  "messages": [
    {
      "role": "user",
      "content": [
        {
          "type": "input_audio",
          "input_audio": {
            "data": "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",
            "format": "mp3"
          }
        }
      ]
    },
    {
      "role": "user",
      "content": "Provide a verbatim, word-for-word transcription of the audio."
    }
  ]
}' | jq
视频理解

安装 OpenAI 依赖：

pip install openai

创建测试 Python 脚本：

vim video_understanding.py

填写以下内容（注意根据实际环境修改 base_url、api_key 和 model）。示例中使用的视频来源于 Hugging Face，请确保运行环境可以正常访问 Hugging Face：

from openai import OpenAI

client = OpenAI(
    base_url="http://192.168.50.16/v1",
    api_key="gpustack_02d1d03b2f1e73d1_9488d69fce3119d86bab20396e34984c"
)

response = client.chat.completions.create(
    model="gemma-4-31b-it",
    messages=[
        {
            "role": "user",
            "content": [
                {
                    "type": "video_url",
                    "video_url": {"url": "https://huggingface.co/datasets/merve/vlm_test_images/resolve/main/concert.mp4"}
                },
                {
                    "type": "text",
                    "text": "Summarize what happens in this video."
                }
            ]
        }
    ],
    max_tokens=1024
)

print(response.choices[0].message.content)

测试：

python video_understanding.py

image-20260404142750148

思考模式

cURL 可以通过 "chat_template_kwargs": {"enable_thinking": true} 开启思考，示例如下（注意修改 URL，Bearer token 和 model）：

curl http://192.168.50.16/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer gpustack_3d545327e9b81d66_9ed17ad891348c4e1d6a7f62e75aa34b" \
-d '{
  "max_tokens": 16384,
  "model": "gemma-4-31b-it",
  "messages": [
    {
      "role": "user",
      "content": "鸡兔同笼问题，40个头，130只脚，分别有几只鸡和兔子？"
    }
  ],
  "chat_template_kwargs": {"enable_thinking": true}
}' | jq

注意：Gemma4 思考模式的思维链采用额外的标记 thought\n 表示，而不是放在 vLLM 返回的 "reasoning" 字段中。

image-20260404122125684

工具调用

安装 OpenAI 依赖：

pip install openai

创建测试 Python 脚本：

vim toolcalling.py

填写以下内容（注意根据实际环境修改 base_url、api_key 以及两处 model 参数）：

from openai import OpenAI
import json

client = OpenAI(
    base_url="http://192.168.50.16/v1",
    api_key="gpustack_3d545327e9b81d66_9ed17ad891348c4e1d6a7f62e75aa34b"
)

# Define tools
tools = [
    {
        "type": "function",
        "function": {
            "name": "get_weather",
            "description": "Get the current weather for a location",
            "parameters": {
                "type": "object",
                "properties": {
                    "location": {
                        "type": "string",
                        "description": "City name, e.g. 'San Francisco, CA'"
                    },
                    "unit": {
                        "type": "string",
                        "enum": ["celsius", "fahrenheit"],
                        "description": "Temperature unit"
                    }
                },
                "required": ["location"]
            }
        }
    }
]

# Step 1: Send user message with tools
response = client.chat.completions.create(
    model="gemma-4-31b-it",
    messages=[
        {"role": "user", "content": "What is the weather in Tokyo today?"}
    ],
    tools=tools,
    max_tokens=1024
)

message = response.choices[0].message

# Step 2: Process tool calls
if message.tool_calls:
    tool_call = message.tool_calls[0]
    print(f"Tool: {tool_call.function.name}")
    print(f"Args: {tool_call.function.arguments}")
    
    # Step 3: Feed back tool result and get final answer
    response = client.chat.completions.create(
        model="gemma-4-31b-it",
        messages=[
            {"role": "user", "content": "What is the weather in Tokyo today?"},
            message,  # assistant's tool call message
            {
                "role": "tool",
                "tool_call_id": tool_call.id,
                "content": json.dumps({"temperature": 22, "condition": "Partly cloudy", "unit": "celsius"})
            }
        ],
        tools=tools,
        max_tokens=1024
    )

    print(f"\nFinal answer: {response.choices[0].message.content}")

测试：

python toolcalling.py

image-20260404123511958

更多关于 Gemma 4 模型的用法，请参考：docs.vllm.ai/projects/re…

总结

GPUStack 采用可插拔的推理引擎架构，将平台与具体推理引擎解耦。当有新模型发布且依赖特定版本或定制推理环境时，可以直接通过接入对应的引擎镜像或自定义后端来完成部署，而不需要等待平台侧适配或升级。

这样做的好处是，在实际使用中可以更灵活地应对模型和引擎版本的变化。一方面可以更快验证和上线新模型，另一方面也避免了不同模型之间因运行环境不一致带来的冲突问题，使整体部署和运维更加可控。

如果希望进一步了解 GPUStack 的功能和使用方式，可以参考以下资源：

image-20260404112334023

加入 GPUStack 社区

GPUStack 社区聚焦 AI 基础设施与大模型实践。

社区中持续分享真实环境下的部署经验、问题排查案例，以及推理引擎、算力管理和系统架构相关讨论。

欢迎扫码加入 GPUStack 社区，与更多关注 AI Infra 与大模型推理实践的伙伴一起交流、学习与分享。

9f0c1539a0477e224fb65e7eeecbb4b6

若群聊已满或二维码失效，请访问以下页面查看最新群二维码： github.com/gpustack/gp…

avatar
文章
阅读
粉丝