|
| 1 | +.. _function_calling: |
| 2 | + |
| 3 | +工具调用(Function Calling) |
| 4 | +============================ |
| 5 | + |
| 6 | +LightLLM 支持多种主流模型的工具调用功能,提供 OpenAI 兼容的 API。 |
| 7 | + |
| 8 | +支持的模型 |
| 9 | +---------- |
| 10 | + |
| 11 | +Qwen2.5/Qwen3 |
| 12 | +~~~~~~~~~~~~~ |
| 13 | + |
| 14 | +**解析器**: ``qwen25`` |
| 15 | + |
| 16 | +**格式**: |
| 17 | + |
| 18 | +.. code-block:: xml |
| 19 | +
|
| 20 | + <tool_call> |
| 21 | + {"name": "function_name", "arguments": {"param": "value"}} |
| 22 | + </tool_call> |
| 23 | +
|
| 24 | +**启动**: |
| 25 | + |
| 26 | +.. code-block:: bash |
| 27 | +
|
| 28 | + python -m lightllm.server.api_server \ |
| 29 | + --model_dir /path/to/qwen2.5 \ |
| 30 | + --tool_call_parser qwen25 \ |
| 31 | + --tp 1 |
| 32 | +
|
| 33 | +Llama 3.2 |
| 34 | +~~~~~~~~~ |
| 35 | + |
| 36 | +**解析器**: ``llama3`` |
| 37 | + |
| 38 | +**格式**: ``<|python_tag|>{"name": "func", "arguments": {...}}`` |
| 39 | + |
| 40 | +**启动**: |
| 41 | + |
| 42 | +.. code-block:: bash |
| 43 | +
|
| 44 | + python -m lightllm.server.api_server \ |
| 45 | + --model_dir /path/to/llama-3.2 \ |
| 46 | + --tool_call_parser llama3 \ |
| 47 | + --tp 1 |
| 48 | +
|
| 49 | +Mistral |
| 50 | +~~~~~~~ |
| 51 | + |
| 52 | +**解析器**: ``mistral`` |
| 53 | + |
| 54 | +**格式**: ``[TOOL_CALLS] [{"name": "func", "arguments": {...}}, ...]`` |
| 55 | + |
| 56 | +DeepSeek-V3 |
| 57 | +~~~~~~~~~~~ |
| 58 | + |
| 59 | +**解析器**: ``deepseekv3`` |
| 60 | + |
| 61 | +**格式**: |
| 62 | + |
| 63 | +.. code-block:: xml |
| 64 | +
|
| 65 | + <|tool▁calls▁begin|> |
| 66 | + <|tool▁call▁begin|>function<|tool▁sep|>func_name |
| 67 | + ```json |
| 68 | + {"param": "value"} |
| 69 | + ``` |
| 70 | + <|tool▁call▁end|> |
| 71 | + <|tool▁calls▁end|> |
| 72 | +
|
| 73 | +DeepSeek-V3.1 |
| 74 | +~~~~~~~~~~~~~ |
| 75 | + |
| 76 | +**解析器**: ``deepseekv31`` |
| 77 | + |
| 78 | +**格式**: 简化的 V3 格式,参数直接内联,无代码块包围 |
| 79 | + |
| 80 | +基本使用 |
| 81 | +-------- |
| 82 | + |
| 83 | +定义工具 |
| 84 | +~~~~~~~~ |
| 85 | + |
| 86 | +.. code-block:: python |
| 87 | +
|
| 88 | + tools = [ |
| 89 | + { |
| 90 | + "type": "function", |
| 91 | + "function": { |
| 92 | + "name": "get_weather", |
| 93 | + "description": "获取指定城市的天气信息", |
| 94 | + "parameters": { |
| 95 | + "type": "object", |
| 96 | + "properties": { |
| 97 | + "city": { |
| 98 | + "type": "string", |
| 99 | + "description": "城市名称" |
| 100 | + } |
| 101 | + }, |
| 102 | + "required": ["city"] |
| 103 | + } |
| 104 | + } |
| 105 | + } |
| 106 | + ] |
| 107 | +
|
| 108 | +非流式调用 |
| 109 | +~~~~~~~~~~ |
| 110 | + |
| 111 | +.. code-block:: python |
| 112 | +
|
| 113 | + import requests |
| 114 | + import json |
| 115 | +
|
| 116 | + url = "http://localhost:8088/v1/chat/completions" |
| 117 | + data = { |
| 118 | + "model": "model_name", |
| 119 | + "messages": [ |
| 120 | + {"role": "user", "content": "北京今天天气怎么样?"} |
| 121 | + ], |
| 122 | + "tools": tools, |
| 123 | + "tool_choice": "auto" # "auto" | "none" | "required" |
| 124 | + } |
| 125 | +
|
| 126 | + response = requests.post(url, json=data).json() |
| 127 | + message = response["choices"][0]["message"] |
| 128 | +
|
| 129 | + if message.get("tool_calls"): |
| 130 | + for tc in message["tool_calls"]: |
| 131 | + print(f"工具: {tc['function']['name']}") |
| 132 | + print(f"参数: {tc['function']['arguments']}") |
| 133 | +
|
| 134 | +流式调用 |
| 135 | +~~~~~~~~ |
| 136 | + |
| 137 | +.. code-block:: python |
| 138 | +
|
| 139 | + data = { |
| 140 | + "model": "model_name", |
| 141 | + "messages": [{"role": "user", "content": "查询北京和上海的天气"}], |
| 142 | + "tools": tools, |
| 143 | + "stream": True |
| 144 | + } |
| 145 | +
|
| 146 | + response = requests.post(url, json=data, stream=True) |
| 147 | + tool_calls = {} |
| 148 | +
|
| 149 | + for line in response.iter_lines(): |
| 150 | + if line and line.startswith(b"data: "): |
| 151 | + chunk = json.loads(line[6:]) |
| 152 | + delta = chunk["choices"][0]["delta"] |
| 153 | +
|
| 154 | + if delta.get("tool_calls"): |
| 155 | + for tc in delta["tool_calls"]: |
| 156 | + idx = tc.get("index", 0) |
| 157 | + if idx not in tool_calls: |
| 158 | + tool_calls[idx] = {"function": {"name": "", "arguments": ""}} |
| 159 | +
|
| 160 | + if tc["function"].get("name"): |
| 161 | + tool_calls[idx]["function"]["name"] = tc["function"]["name"] |
| 162 | + if tc["function"].get("arguments"): |
| 163 | + tool_calls[idx]["function"]["arguments"] += tc["function"]["arguments"] |
| 164 | +
|
| 165 | +多轮对话 |
| 166 | +~~~~~~~~ |
| 167 | + |
| 168 | +.. code-block:: python |
| 169 | +
|
| 170 | + # 1. 用户提问 |
| 171 | + messages = [{"role": "user", "content": "北京天气如何?"}] |
| 172 | +
|
| 173 | + # 2. 模型调用工具 |
| 174 | + response1 = requests.post(url, json={ |
| 175 | + "messages": messages, |
| 176 | + "tools": tools |
| 177 | + }).json() |
| 178 | +
|
| 179 | + tool_call = response1["choices"][0]["message"]["tool_calls"][0] |
| 180 | + messages.append(response1["choices"][0]["message"]) |
| 181 | +
|
| 182 | + # 3. 返回工具结果 |
| 183 | + weather_result = {"temperature": 15, "condition": "晴朗"} |
| 184 | + messages.append({ |
| 185 | + "role": "tool", |
| 186 | + "tool_call_id": tool_call["id"], |
| 187 | + "name": tool_call["function"]["name"], |
| 188 | + "content": json.dumps(weather_result, ensure_ascii=False) |
| 189 | + }) |
| 190 | +
|
| 191 | + # 4. 生成最终回答 |
| 192 | + response2 = requests.post(url, json={"messages": messages}).json() |
| 193 | + print(response2["choices"][0]["message"]["content"]) |
| 194 | +
|
| 195 | +高级功能 |
| 196 | +-------- |
| 197 | + |
| 198 | +并行工具调用 |
| 199 | +~~~~~~~~~~~~ |
| 200 | + |
| 201 | +.. code-block:: python |
| 202 | +
|
| 203 | + data = { |
| 204 | + "messages": messages, |
| 205 | + "tools": tools, |
| 206 | + "parallel_tool_calls": True # 启用并行调用 |
| 207 | + } |
| 208 | +
|
| 209 | +强制调用特定工具 |
| 210 | +~~~~~~~~~~~~~~~~ |
| 211 | + |
| 212 | +.. code-block:: python |
| 213 | +
|
| 214 | + data = { |
| 215 | + "tools": tools, |
| 216 | + "tool_choice": { |
| 217 | + "type": "function", |
| 218 | + "function": {"name": "get_weather"} |
| 219 | + } |
| 220 | + } |
| 221 | +
|
| 222 | +与推理模型集成 |
| 223 | +~~~~~~~~~~~~~~ |
| 224 | + |
| 225 | +.. code-block:: python |
| 226 | +
|
| 227 | + data = { |
| 228 | + "model": "deepseek-r1", |
| 229 | + "tools": tools, |
| 230 | + "chat_template_kwargs": {"enable_thinking": True}, |
| 231 | + "separate_reasoning": True # 分离推理内容 |
| 232 | + } |
| 233 | +
|
| 234 | + response = requests.post(url, json=data).json() |
| 235 | + message = response["choices"][0]["message"] |
| 236 | +
|
| 237 | + print("推理:", message.get("reasoning_content")) |
| 238 | + print("工具调用:", message.get("tool_calls")) |
| 239 | +
|
| 240 | +常见问题 |
| 241 | +-------- |
| 242 | + |
| 243 | +**工具调用未触发** |
| 244 | + 检查 ``--tool_call_parser`` 参数和工具描述是否清晰 |
| 245 | + |
| 246 | +**参数解析错误** |
| 247 | + 确认使用了正确的解析器,检查模型输出格式 |
| 248 | + |
| 249 | +**流式模式不完整** |
| 250 | + 正确处理所有 chunks,使用 ``index`` 字段组装多个工具调用 |
| 251 | + |
| 252 | +**与推理模型集成失败** |
| 253 | + 确保使用最新版本,正确配置 ``separate_reasoning`` 和 ``chat_template_kwargs`` |
| 254 | + |
| 255 | +技术细节 |
| 256 | +-------- |
| 257 | + |
| 258 | +**核心文件**: |
| 259 | +- ``lightllm/server/function_call_parser.py`` - 解析器实现 |
| 260 | +- ``lightllm/server/api_openai.py`` - API 集成 |
| 261 | +- ``lightllm/server/build_prompt.py`` - 工具注入 |
| 262 | +- ``test/test_api/test_openai_api.py`` - 测试示例 |
| 263 | + |
| 264 | +**相关 PR**: |
| 265 | +- PR #1158: 支持推理内容中的函数调用 |
| 266 | + |
| 267 | +参考资料 |
| 268 | +-------- |
| 269 | + |
| 270 | +- OpenAI Function Calling: https://platform.openai.com/docs/guides/function-calling |
| 271 | +- JSON Schema: https://json-schema.org/ |
| 272 | +- LightLLM GitHub: https://github.com/ModelTC/lightllm |
0 commit comments