Docker下使用llama.cpp部署带Function calling和Json Mode功能的Mistral 7B模型

Docker下使用llama.cpp部署带Function calling和Json Mode功能的Mistral 7B模型

说明：

• 首次发表日期：2024-08-27
• 参考：
  • https://www.markhneedham.com/blog/2024/06/23/mistral-7b-function-calling-llama-cpp/
  • https://github.com/abetlen/llama-cpp-python?tab=readme-ov-file#function-calling
  • https://github.com/abetlen/llama-cpp-python/tree/main/docker#cuda_simple
  • https://docs.mistral.ai/capabilities/json_mode/
  • https://huggingface.co/MaziyarPanahi/Mistral-7B-Instruct-v0.3-GGUF
  • https://stackoverflow.com/questions/30905674/newer-versions-of-docker-have-cap-add-what-caps-can-be-added
  • https://man7.org/linux/man-pages/man7/capabilities.7.html
  • https://docs.docker.com/engine/containers/run/#runtime-privilege-and-linux-capabilities
  • https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/install-guide.html
  • https://www.cnblogs.com/davis12/p/14453690.html

下载GGUF模型

使用HuggingFace的镜像 https://hf-mirror.com/

方式一：

pip install -U huggingface_hub
export HF_ENDPOINT=https://hf-mirror.com

huggingface-cli download --resume-download MaziyarPanahi/Mistral-7B-Instruct-v0.3-GGUF --include *Q4_K_M.gguf

方式二（推荐）：

sudo apt update
sudo apt install aria2 git-lfs

wget https://hf-mirror.com/hfd/hfd.sh

chmod a+x hfd.sh

./hfd.sh MaziyarPanahi/Mistral-7B-Instruct-v0.3-GGUF --include *Q4_K_M.gguf --tool aria2c -x 16 --local-dir MaziyarPanahi--Mistral-7B-Instruct-v0.3-GGUF

使用Docker部署服务

构建之前需要先安装NVIDIA Container Toolkit

安装NVIDIA Container Toolkit

准备：

curl -fsSL https://nvidia.github.io/libnvidia-container/gpgkey | sudo gpg --dearmor -o /usr/share/keyrings/nvidia-container-toolkit-keyring.gpg \
  && curl -s -L https://nvidia.github.io/libnvidia-container/stable/deb/nvidia-container-toolkit.list | \
    sed 's#deb https://#deb [signed-by=/usr/share/keyrings/nvidia-container-toolkit-keyring.gpg] https://#g' | \
    sudo tee /etc/apt/sources.list.d/nvidia-container-toolkit.list

安装：

sudo apt-get update
sudo apt-get install -y nvidia-container-toolkit

配置docker

sudo nvidia-ctk runtime configure --runtime=docker

NVIDIA Container Toolkit 安装的更多信息请参考官方文档： https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/install-guide.html

构建镜像

使用官方的Dockerfile: https://github.com/abetlen/llama-cpp-python/blob/main/docker/cuda_simple/Dockerfile

ARG CUDA_IMAGE="12.2.0-devel-ubuntu22.04"
FROM nvidia/cuda:${CUDA_IMAGE}

# We need to set the host to 0.0.0.0 to allow outside access
ENV HOST 0.0.0.0

RUN apt-get update && apt-get upgrade -y \
    && apt-get install -y git build-essential \
    python3 python3-pip gcc wget \
    ocl-icd-opencl-dev opencl-headers clinfo \
    libclblast-dev libopenblas-dev \
    && mkdir -p /etc/OpenCL/vendors && echo "libnvidia-opencl.so.1" > /etc/OpenCL/vendors/nvidia.icd

COPY . .

# setting build related env vars
ENV CUDA_DOCKER_ARCH=all
ENV GGML_CUDA=1

# Install depencencies
RUN python3 -m pip install --upgrade pip pytest cmake scikit-build setuptools fastapi uvicorn sse-starlette pydantic-settings starlette-context

# Install llama-cpp-python (build with cuda)
RUN CMAKE_ARGS="-DGGML_CUDA=on" pip install llama-cpp-python

# Run the server
CMD python3 -m llama_cpp.server

因为我本地安装的CUDA版本为12.2，所以将base镜像改为nvidia/cuda:12.2.0-devel-ubuntu22.04

docker build -t llama_cpp_cuda_simple .

启动服务

docker run --gpus=all --cap-add SYS_RESOURCE -e USE_MLOCK=0 -e model=/models/downloaded/MaziyarPanahi--Mistral-7B-Instruct-v0.3-GGUF/Mistral-7B-Instruct-v0.3.Q4_K_M.gguf -e n_gpu_layers=-1 -e chat_format=chatml-function-calling -v /mnt/d/16-LLM-Cache/llama_cpp_gnuf:/models -p 8000:8000 -t llama_cpp_cuda_simple

其中：

• -v 将本地文件夹映射到容器内部文件夹/models
• --gpus=all 表示使用所有的GPU
• --cap-add SYS_RESOURCE 表示容器将有SYS_RESOURCE的权限
• 其中以-e开头的表示设置环境变量，实际上是设置llama_cpp.server的参数，相关代码详见 https://github.com/abetlen/llama-cpp-python/blob/259ee151da9a569f58f6d4979e97cfd5d5bc3ecd/llama_cpp/server/main.py#L79 和 https://github.com/abetlen/llama-cpp-python/blob/259ee151da9a569f58f6d4979e97cfd5d5bc3ecd/llama_cpp/server/settings.py#L17 这里设置的环境变量是大小写不敏感的，见 https://docs.pydantic.dev/latest/concepts/pydantic_settings/#case-sensitivity
  • -e model 指向模型文件
  • -e n_gpu_layers=-1 表示将所有神经网络层移到GPU
    • 假设模型一共有N层，其中n_gpu_layers层被放在GPU上，那么剩下的 N - n_gpu_layers 就会被放在CPU上
  • -e chat_format=chatml-function-calling 设置以支持Function Calling功能

启动完成后，在浏览器打开 http://localhost:8000/docs 查看API文档

调用测试

Function Calling

curl --location 'http://localhost:8000/v1/chat/completions' \
--header 'Content-Type: application/json' \
--header 'Authorization: Bearer sk-xxxxxxxxxxxxxxxxxxxxxx' \
--data '{
  "model": "gpt-3.5-turbo",
  "messages": [
    {
      "role": "system",
      "content": "You are a helpful assistant.\nYou can call functions with appropriate input when necessary"
    },
    {
      "role": "user",
      "content": "What'\''s the weather like in Mauritius?"
    }
  ],
  "tools": [
    {
      "type": "function",
      "function": {
        "name": "get_current_weather",
        "description": "Get the current weather in a given latitude and longitude",
        "parameters": {
          "type": "object",
          "properties": {
            "latitude": {
              "type": "number",
              "description": "The latitude of a place"
            },
            "longitude": {
              "type": "number",
              "description": "The longitude of a place"
            }
          },
          "required": ["latitude", "longitude"]
        }
      }
    }
  ],
  "tool_choice": "auto"
}'

输出：

{
    "id": "chatcmpl-50c8e261-2b1a-4285-a6ee-e18a07ce92d9",
    "object": "chat.completion",
    "created": 1724757544,
    "model": "gpt-3.5-turbo",
    "choices": [
        {
            "index": 0,
            "message": {
                "content": null,
                "tool_calls": [
                    {
                        "id": "call__0_get_current_weather_cmpl-97515c72-d214-4ed9-b183-7736199e5be1",
                        "type": "function",
                        "function": {
                            "name": "get_current_weather",
                            "arguments": "{\"latitude\": -20.375, \"longitude\": 57.568} "
                        }
                    }
                ],
                "role": "assistant",
                "function_call": {
                    "name": "",
                    "arguments": "{\"latitude\": -20.375, \"longitude\": 57.568} "
                }
            },
            "logprobs": null,
            "finish_reason": "tool_calls"
        }
    ],
    "usage": {
        "prompt_tokens": 299,
        "completion_tokens": 25,
        "total_tokens": 324
    }
}

JSON Mode

curl --location "http://localhost:8000/v1/chat/completions" \
     --header 'Content-Type: application/json' \
     --header 'Accept: application/json' \
     --header "Authorization: Bearer sk-xxxxxxxxxxxxxxxxxxxxxx" \
     --data '{
    "model": "gpt-3.5-turbo",
    "messages": [
     {
        "role": "user",
        "content": "What is the best French cheese? Return the product and produce location in JSON format"
      }
    ],
    "response_format": {"type": "json_object"}
  }'

输出：

{
    "id": "chatcmpl-bbfecfc5-2ea9-4052-93b2-08f1733e8219",
    "object": "chat.completion",
    "created": 1724757752,
    "model": "gpt-3.5-turbo",
    "choices": [
        {
            "index": 0,
            "message": {
                "content": "{\n  \"product\": \"Roquefort\",\n  \"produce_location\": \"France, South of France\"\n}\n  \t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t",
                "role": "assistant"
            },
            "logprobs": null,
            "finish_reason": "stop"
        }
    ],
    "usage": {
        "prompt_tokens": 44,
        "completion_tokens": 50,
        "total_tokens": 94
    }
}

使用以下代码将content部分写入到文本：

text = "{\n  \"product\": \"Roquefort\",\n  \"location\": \"France, South of France\"\n}\n \t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t"

with open('resp.txt', 'w') as f:
    f.write(text)

可以看到内容：

{
  "product": "Roquefort",
  "location": "France, South of France"
}