【再聊agent】Function call 详解
输入“/”快速插入内容
💥【再聊agent】Function call 详解
【再聊agent】Function call 详解
用户3092024年10月9日修改
👇
面向人员：能使用基本的bash，python，能够使用python处理数据即可。Agent理论部分需要部分的大模型和机器学习基础
机器要求：最少8G以上内存的机器，其余显存随上下文长度上升
VLLM server 方式（可适用gptq模型）
1.
下载模型
代码块
git clone https://huggingface.co/openbmb/MiniCPM3-4B
2.
按照MiniCPM 3.0 推理安装vllm
3.
获取官方代码
代码块
git clone https://github.com/OpenBMB/MiniCPM
4.
启动vllm服务
代码块
cd MiniCPM/demo/function_call
python openai_api_server.py --model /root/ld/ld_model_pretrained/minicpm3 --served-model-name MiniCPM3-4B --chat-template chatml.jinja --dtype auto --api-key token-abc123 --tensor-parallel-size 1 --max_model_len 4096 --gpu_memory_utilization 1 --trust-remote-code --enforce_eager
参数解析：
代码块
--model /root/ld/ld_model_pretrained/minicpm3 #模型地址
--served-model-name MiniCPM3-4B #模型的名字，可以自定义，在调用脚本中会用到
--chat-template chatml.jinja # 模型的template
--dtype auto # 设置的精度，比如bf16，fp16
--api-key token-abc123 # 可以理解为你的密码，调用脚本中会用到
--tensor-parallel-size 1 # 使用多少张卡，建议1，2，4，8
--max_model_len 4096 # 模型的最长上下文，越长占用显存越多，必须设置
--gpu_memory_utilization 0.9 # 使用多少显存来运行这个模型，1代表100%
--trust-remote-code # 不要改，这里是是否使用远程代码
--enforce_eager # 使用eager模式，这个模式下显存占用较少
5.
编写调用脚本
代码块
import requests
from transformers import AutoTokenizer
import json
tokenizer_path = "/root/ld/ld_model_pretrained/minicpm3"
url = "http://localhost:8000/v1/chat/completions" # 假设服务运行在本地机器上
headers = {
"Content-Type": "application/json",
"Authorization": "Bearer token-abc123" # 使用上面命令中的API密钥
}
tools = [
{
"type": "function",
"function": {
"name": "get_delivery_date",
"description": "Get the delivery date for a customer's order. Call this whenever you need to know the delivery date, for example when a customer asks 'Where is my package'",
"parameters": {
"type": "object",
"properties": {
"order_id": {
"type": "string",
"description": "The customer's order ID.",
},
},
"required": ["order_id"],
"additionalProperties": False,
},
},
},
{
"type": "function",
"function": {
"name": "image_gen",
"description": "image_gen is an AI painting (image generation) service that inputs a text description and returns the URL of the image drawn based on the text.",
"parameters": {
"type": "object",
"properties": {
"description_image": {
"type": "string",
"description": "Text description of the image the user wants to generate",
},
},
"required": ["description_image"],
"additionalProperties": False,
},
},
}
]
messages = [
{
"role": "system",
"content": "You are a helpful customer support assistant. Use the supplied tools to assist the user.",
},
{
"role": "user",
"content": "Hi, can you tell me the delivery date for my order? The order id is 1234 and 4321.",
},
]
tokenizer = AutoTokenizer.from_pretrained(tokenizer_path, trust_remote_code=True)
prompt = tokenizer.apply_chat_template(
messages, tools=tools, tokenize=False, add_generation_prompt=True
)
data = {
"model": "MiniCPM3-4B",
"messages": [
{"role": "system", "content": "You are a helpful assistant."},
{"role": "user", "content": prompt}
]
}
def fake_tool_execute(toolcall):
tool_name=toolcall['name']
arguments=eval(toolcall['arguments'])
if tool_name == "get_delivery_date":
order_id=arguments['order_id']
get_data="2024-09-05" if int(order_id)%2==0 else "2024-09-06"
return get_data
elif tool_name == "image_gen":
prompt=arguments['description_image']
import urllib.parse
prompt = urllib.parse.quote(prompt)
return json.dumps({'image_url': f'https://image.pollinations.ai/prompt/{prompt}'}, ensure_ascii=False)
return "没有找到该插件"
response = requests.post(url, headers=headers, json=data)
if response.status_code == 200:
for msg in response.json()["choices"][0]["message"]['tool_calls']:
if (
"function" in msg
and msg["function"] is not None
and len(msg["function"]) > 0
):
#print(msg)
toolcall=msg["function"]
tool_response = fake_tool_execute(toolcall)
tool_msg = {
"role": "tool",
"content": tool_response,
"plugin": toolcall["name"],
"input_param":toolcall['arguments']
}
print(tool_msg)
print(tool_response)
else:
print(f"Error: {response.status_code}, {response.text}")