GitHub - InternLM/lagent: A lightweight framework for building LLM-based agents (original) (raw)

👋 join us on 𝕏 (Twitter), Discord and WeChat

Installation

Install from source:

git clone https://github.com/InternLM/lagent.git cd lagent pip install -e .

Usage

Lagent is inspired by the design philosophy of PyTorch. We expect that the analogy of neural network layers will make the workflow clearer and more intuitive, so users only need to focus on creating layers and defining message passing between them in a Pythonic way. This is a simple tutorial to get you quickly started with building multi-agent applications.

Models as Agents

Agents use AgentMessage for communication.

from typing import Dict, List from lagent.agents import Agent from lagent.schema import AgentMessage from lagent.llms import VllmModel, INTERNLM2_META

llm = VllmModel( path='Qwen/Qwen2-7B-Instruct', meta_template=INTERNLM2_META, tp=1, top_k=1, temperature=1.0, stop_words=['<|im_end|>'], max_new_tokens=1024, ) system_prompt = '你的回答只能从“典”、“孝”、“急”三个字中选一个。' agent = Agent(llm, system_prompt)

user_msg = AgentMessage(sender='user', content='今天天气情况') bot_msg = agent(user_msg) print(bot_msg)

content='急' sender='Agent' formatted=None extra_info=None type=None receiver=None stream_state=<AgentStatusCode.END: 0>

Memory as State

Both input and output messages will be added to the memory of Agent in each forward pass. This is performed in __call__ rather than forward. See the following pseudo code

def __call__(self, *message):
    message = pre_hooks(message)
    add_memory(message)
    message = self.forward(*message)
    add_memory(message)
    message = post_hooks(message)
    return message

Inspect the memory in two ways

memory: List[AgentMessage] = agent.memory.get_memory() print(memory) print('-' * 120) dumped_memory: Dict[str, List[dict]] = agent.state_dict() print(dumped_memory['memory'])

[AgentMessage(content='今天天气情况', sender='user', formatted=None, extra_info=None, type=None, receiver=None, stream_state=<AgentStatusCode.END: 0>), AgentMessage(content='急', sender='Agent', formatted=None, extra_info=None, type=None, receiver=None, stream_state=<AgentStatusCode.END: 0>)]
------------------------------------------------------------------------------------------------------------------------
[{'content': '今天天气情况', 'sender': 'user', 'formatted': None, 'extra_info': None, 'type': None, 'receiver': None, 'stream_state': <AgentStatusCode.END: 0>}, {'content': '急', 'sender': 'Agent', 'formatted': None, 'extra_info': None, 'type': None, 'receiver': None, 'stream_state': <AgentStatusCode.END: 0>}]

Clear the memory of this session(session_id=0 by default):

Custom Message Aggregation

DefaultAggregator is called under the hood to assemble and convert AgentMessage to OpenAI message format.

def forward(self, *message: AgentMessage, session_id=0, **kwargs) -> Union[AgentMessage, str]:
    formatted_messages = self.aggregator.aggregate(
        self.memory.get(session_id),
        self.name,
        self.output_format,
        self.template,
    )
    llm_response = self.llm.chat(formatted_messages, **kwargs)
    ...

Implement a simple aggregator that can receive few-shots

from typing import List, Union from lagent.memory import Memory from lagent.prompts import StrParser from lagent.agents.aggregator import DefaultAggregator

class FewshotAggregator(DefaultAggregator): def init(self, few_shot: List[dict] = None): self.few_shot = few_shot or []

def aggregate(self,
              messages: Memory,
              name: str,
              parser: StrParser = None,
              system_instruction: Union[str, dict, List[dict]] = None) -> List[dict]:
    _message = []
    if system_instruction:
        _message.extend(
            self.aggregate_system_intruction(system_instruction))
    _message.extend(self.few_shot)
    messages = messages.get_memory()
    for message in messages:
        if message.sender == name:
            _message.append(
                dict(role='assistant', content=str(message.content)))
        else:
            user_message = message.content
            if len(_message) > 0 and _message[-1]['role'] == 'user':
                _message[-1]['content'] += user_message
            else:
                _message.append(dict(role='user', content=user_message))
    return _message

agent = Agent( llm, aggregator=FewshotAggregator( [ {"role": "user", "content": "今天天气"}, {"role": "assistant", "content": "【晴】"}, ] ) ) user_msg = AgentMessage(sender='user', content='昨天天气') bot_msg = agent(user_msg) print(bot_msg)

content='【多云转晴，夜间有轻微降温】' sender='Agent' formatted=None extra_info=None type=None receiver=None stream_state=<AgentStatusCode.END: 0>

Flexible Response Formatting

In AgentMessage, formatted is reserved to store information parsed by output_format from the model output.

def forward(self, *message: AgentMessage, session_id=0, **kwargs) -> Union[AgentMessage, str]:
    ...
    llm_response = self.llm.chat(formatted_messages, **kwargs)
    if self.output_format:
        formatted_messages = self.output_format.parse_response(llm_response)
        return AgentMessage(
            sender=self.name,
            content=llm_response,
            formatted=formatted_messages,
        )
    ...

Use a tool parser as follows

from lagent.prompts.parsers import ToolParser

system_prompt = "逐步分析并编写Python代码解决以下问题。" parser = ToolParser(tool_type='code interpreter', begin='python\n', end='\n\n') llm.gen_params['stop_words'].append('\n```\n') agent = Agent(llm, system_prompt, output_format=parser)

user_msg = AgentMessage( sender='user', content='Marie is thinking of a multiple of 63, while Jay is thinking of a ' 'factor of 63. They happen to be thinking of the same number. There are ' 'two possibilities for the number that each of them is thinking of, one ' 'positive and one negative. Find the product of these two numbers.') bot_msg = agent(user_msg) print(bot_msg.model_dump_json(indent=4))

{
    "content": "首先，我们需要找出63的所有正因数和负因数。63的正因数可以通过分解63的质因数来找出，即\\(63 = 3^2 \\times 7\\)。因此，63的正因数包括1, 3, 7, 9, 21, 和 63。对于负因数，我们只需将上述正因数乘以-1。\n\n接下来，我们需要找出与63的正因数相乘的结果为63的数，以及与63的负因数相乘的结果为63的数。这可以通过将63除以每个正因数和负因数来实现。\n\n最后，我们将找到的两个数相乘得到最终答案。\n\n下面是Python代码实现：\n\n```python\ndef find_numbers():\n    # 正因数\n    positive_factors = [1, 3, 7, 9, 21, 63]\n    # 负因数\n    negative_factors = [-1, -3, -7, -9, -21, -63]\n    \n    # 找到与正因数相乘的结果为63的数\n    positive_numbers = [63 / factor for factor in positive_factors]\n    # 找到与负因数相乘的结果为63的数\n    negative_numbers = [-63 / factor for factor in negative_factors]\n    \n    # 计算两个数的乘积\n    product = positive_numbers[0] * negative_numbers[0]\n    \n    return product\n\nresult = find_numbers()\nprint(result)",
    "sender": "Agent",
    "formatted": {
        "tool_type": "code interpreter",
        "thought": "首先，我们需要找出63的所有正因数和负因数。63的正因数可以通过分解63的质因数来找出，即\\(63 = 3^2 \\times 7\\)。因此，63的正因数包括1, 3, 7, 9, 21, 和 63。对于负因数，我们只需将上述正因数乘以-1。\n\n接下来，我们需要找出与63的正因数相乘的结果为63的数，以及与63的负因数相乘的结果为63的数。这可以通过将63除以每个正因数和负因数来实现。\n\n最后，我们将找到的两个数相乘得到最终答案。\n\n下面是Python代码实现：\n\n",
        "action": "def find_numbers():\n    # 正因数\n    positive_factors = [1, 3, 7, 9, 21, 63]\n    # 负因数\n    negative_factors = [-1, -3, -7, -9, -21, -63]\n    \n    # 找到与正因数相乘的结果为63的数\n    positive_numbers = [63 / factor for factor in positive_factors]\n    # 找到与负因数相乘的结果为63的数\n    negative_numbers = [-63 / factor for factor in negative_factors]\n    \n    # 计算两个数的乘积\n    product = positive_numbers[0] * negative_numbers[0]\n    \n    return product\n\nresult = find_numbers()\nprint(result)",
        "status": 1
    },
    "extra_info": null,
    "type": null,
    "receiver": null,
    "stream_state": 0
}

Consistency of Tool Calling

ActionExecutor uses the same communication data structure as Agent, but requires the content of input AgentMessage to be a dict containing:

name: tool name, e.g. 'IPythonInterpreter', 'WebBrowser.search'.
parameters: keyword arguments of the tool API, e.g. {'command': 'import math;math.sqrt(2)'}, {'query': ['recent progress in AI']}.

You can register custom hooks for message conversion.

from lagent.hooks import Hook from lagent.schema import ActionReturn, ActionStatusCode, AgentMessage from lagent.actions import ActionExecutor, IPythonInteractive

class CodeProcessor(Hook): def before_action(self, executor, message, session_id): message = message.copy(deep=True) message.content = dict( name='IPythonInteractive', parameters={'command': message.formatted['action']} ) return message

def after_action(self, executor, message, session_id):
    action_return = message.content
    if isinstance(action_return, ActionReturn):
        if action_return.state == ActionStatusCode.SUCCESS:
            response = action_return.format_result()
        else:
            response = action_return.errmsg
    else:
        response = action_return
    message.content = response
    return message

executor = ActionExecutor(actions=[IPythonInteractive()], hooks=[CodeProcessor()]) bot_msg = AgentMessage( sender='Agent', content='首先，我们需要...', formatted={ 'tool_type': 'code interpreter', 'thought': '首先，我们需要...', 'action': 'def find_numbers():\n # 正因数\n positive_factors = [1, 3, 7, 9, 21, 63]\n # 负因数\n negative_factors = [-1, -3, -7, -9, -21, -63]\n \n # 找到与正因数相乘的结果为63的数\n positive_numbers = [63 / factor for factor in positive_factors]\n # 找到与负因数相乘的结果为63的数\n negative_numbers = [-63 / factor for factor in negative_factors]\n \n # 计算两个数的乘积\n product = positive_numbers[0] * negative_numbers[0]\n \n return product\n\nresult = find_numbers()\nprint(result)', 'status': 1 }) executor_msg = executor(bot_msg) print(executor_msg)

content='3969.0' sender='ActionExecutor' formatted=None extra_info=None type=None receiver=None stream_state=<AgentStatusCode.END: 0>

For convenience, Lagent provides InternLMActionProcessor which is adapted to messages formatted by ToolParser as mentioned above.

Dual Interfaces

Lagent adopts dual interface design, where almost every component(LLMs, actions, action executors...) has the corresponding asynchronous variant by prefixing its identifier with 'Async'. It is recommended to use synchronous agents for debugging and asynchronous ones for large-scale inference to make the most of idle CPU and GPU resources.

However, make sure the internal consistency of agents, i.e. asynchronous agents should be equipped with asynchronous LLMs and asynchronous action executors that drive asynchronous tools.

from lagent.llms import VllmModel, AsyncVllmModel, LMDeployPipeline, AsyncLMDeployPipeline from lagent.actions import ActionExecutor, AsyncActionExecutor, WebBrowser, AsyncWebBrowser from lagent.agents import Agent, AsyncAgent, AgentForInternLM, AsyncAgentForInternLM

Practice

Try to implement forward instead of __call__ of subclasses unless necessary.
Always include the session_id argument explicitly, which is designed for isolation of memory, LLM requests and tool invocation(e.g. maintain multiple independent IPython environments) in concurrency.

Single Agent

Math agents that solve problems by programming

from lagent.agents.aggregator import InternLMToolAggregator

class Coder(Agent): def init(self, model_path, system_prompt, max_turn=3): super().init() llm = VllmModel( path=model_path, meta_template=INTERNLM2_META, tp=1, top_k=1, temperature=1.0, stop_words=['\n\n', '<|im_end|>'], max_new_tokens=1024, ) self.agent = Agent( llm, system_prompt, output_format=ToolParser( tool_type='code interpreter', begin='python\n', end='\n```\n' ), # InternLMToolAggregator is adapted to ToolParser for aggregating # messages with tool invocations and execution results aggregator=InternLMToolAggregator(), ) self.executor = ActionExecutor([IPythonInteractive()], hooks=[CodeProcessor()]) self.max_turn = max_turn

def forward(self, message: AgentMessage, session_id=0) -> AgentMessage:
    for _ in range(self.max_turn):
        message = self.agent(message, session_id=session_id)
        if message.formatted['tool_type'] is None:
            return message
        message = self.executor(message, session_id=session_id)
    return message

coder = Coder('Qwen/Qwen2-7B-Instruct', 'Solve the problem step by step with assistance of Python code') query = AgentMessage( sender='user', content='Find the projection of mathbfa\mathbf{a}mathbfa onto $\mathbf{b} = ' '\begin{pmatrix} 1 \\ -3 \end{pmatrix}$ if mathbfacdotmathbfb=2.\mathbf{a} \cdot \mathbf{b} = 2.mathbfacdotmathbfb=2.' ) answer = coder(query) print(answer.content) print('-' * 120) for msg in coder.state_dict()['agent.memory']: print('*' * 80) print(f'{msg["sender"]}:\n\n{msg["content"]}')

Multiple Agents

Asynchronous blogging agents that improve writing quality by self-refinement (original AutoGen example)

import asyncio import os from lagent.llms import AsyncGPTAPI from lagent.agents import AsyncAgent os.environ['OPENAI_API_KEY'] = 'YOUR_API_KEY'

class PrefixedMessageHook(Hook): def init(self, prefix: str, senders: list = None): self.prefix = prefix self.senders = senders or []

def before_agent(self, agent, messages, session_id):
    for message in messages:
        if message.sender in self.senders:
            message.content = self.prefix + message.content

class AsyncBlogger(AsyncAgent): def init(self, model_path, writer_prompt, critic_prompt, critic_prefix='', max_turn=3): super().init() llm = AsyncGPTAPI(model_type=model_path, retry=5, max_new_tokens=2048) self.writer = AsyncAgent(llm, writer_prompt, name='writer') self.critic = AsyncAgent( llm, critic_prompt, name='critic', hooks=[PrefixedMessageHook(critic_prefix, ['writer'])] ) self.max_turn = max_turn

async def forward(self, message: AgentMessage, session_id=0) -> AgentMessage:
    for _ in range(self.max_turn):
        message = await self.writer(message, session_id=session_id)
        message = await self.critic(message, session_id=session_id)
    return await self.writer(message, session_id=session_id)

blogger = AsyncBlogger( 'gpt-4o-2024-05-13', writer_prompt="You are an writing assistant tasked to write engaging blogpost. You try to generate the best blogpost possible for the user's request. " "If the user provides critique, then respond with a revised version of your previous attempts", critic_prompt="Generate critique and recommendations on the writing. Provide detailed recommendations, including requests for length, depth, style, etc..", critic_prefix='Reflect and provide critique on the following writing. \n\n', ) user_prompt = ( "Write an engaging blogpost on the recent updates in {topic}. " "The blogpost should be engaging and understandable for general audience. " "Should have more than 3 paragraphes but no longer than 1000 words.") bot_msgs = asyncio.get_event_loop().run_until_complete( asyncio.gather( [ blogger(AgentMessage(sender='user', content=user_prompt.format(topic=topic)), session_id=i) for i, topic in enumerate(['AI', 'Biotechnology', 'New Energy', 'Video Games', 'Pop Music']) ] ) ) print(bot_msgs[0].content) print('-' * 120) for msg in blogger.state_dict(session_id=0)['writer.memory']: print('' * 80) print(f'{msg["sender"]}:\n\n{msg["content"]}') print('-' * 120) for msg in blogger.state_dict(session_id=0)['critic.memory']: print('*' * 80) print(f'{msg["sender"]}:\n\n{msg["content"]}')

A multi-agent workflow that performs information retrieval, data collection and chart plotting (original LangGraph example)

import json from lagent.actions import IPythonInterpreter, WebBrowser, ActionExecutor from lagent.agents.stream import get_plugin_prompt from lagent.llms import GPTAPI from lagent.hooks import InternLMActionProcessor

TOOL_TEMPLATE = ( "You are a helpful AI assistant, collaborating with other assistants. Use the provided tools to progress" " towards answering the question. If you are unable to fully answer, that's OK, another assistant with" " different tools will help where you left off. Execute what you can to make progress. If you or any of" " the other assistants have the final answer or deliverable, prefix your response with {finish_pattern}" " so the team knows to stop. You have access to the following tools:\n{tool_description}\nPlease provide" " your thought process when you need to use a tool, followed by the call statement in this format:" "\n{invocation_format}\\n**{system_prompt}**" )

class DataVisualizer(Agent): def init(self, model_path, research_prompt, chart_prompt, finish_pattern="Final Answer", max_turn=10): super().init() llm = GPTAPI(model_path, key='YOUR_OPENAI_API_KEY', retry=5, max_new_tokens=1024, stop_words=["\n"]) interpreter, browser = IPythonInterpreter(), WebBrowser("BingSearch", api_key="YOUR_BING_API_KEY") self.researcher = Agent( llm, TOOL_TEMPLATE.format( finish_pattern=finish_pattern, tool_description=get_plugin_prompt(browser), invocation_format='json\n{"name": {{tool name}}, "parameters": {{keyword arguments}}}\n\n', system_prompt=research_prompt, ), output_format=ToolParser( "browser", begin="json\n", end="\n\n", validate=lambda x: json.loads(x.rstrip('`')), ), aggregator=InternLMToolAggregator(), name="researcher", ) self.charter = Agent( llm, TOOL_TEMPLATE.format( finish_pattern=finish_pattern, tool_description=interpreter.name, invocation_format='python\n{{code}}\n\n', system_prompt=chart_prompt, ), output_format=ToolParser( "interpreter", begin="python\n", end="\n```\n", validate=lambda x: x.rstrip('`'), ), aggregator=InternLMToolAggregator(), name="charter", ) self.executor = ActionExecutor([interpreter, browser], hooks=[InternLMActionProcessor()]) self.finish_pattern = finish_pattern self.max_turn = max_turn

def forward(self, message, session_id=0):
    for _ in range(self.max_turn):
        message = self.researcher(message, session_id=session_id, stop_words=["```\n", "```python"]) # override llm stop words
        while message.formatted["tool_type"]:
            message = self.executor(message, session_id=session_id)
            message = self.researcher(message, session_id=session_id, stop_words=["```\n", "```python"])
        if self.finish_pattern in message.content:
            return message
        message = self.charter(message)
        while message.formatted["tool_type"]:
            message = self.executor(message, session_id=session_id)
            message = self.charter(message, session_id=session_id)
        if self.finish_pattern in message.content:
            return message
    return message

visualizer = DataVisualizer( "gpt-4o-2024-05-13", research_prompt="You should provide accurate data for the chart generator to use.", chart_prompt="Any charts you display will be visible by the user.", ) user_msg = AgentMessage( sender='user', content="Fetch the China's GDP over the past 5 years, then draw a line graph of it. Once you code it up, finish.") bot_msg = visualizer(user_msg) print(bot_msg.content) json.dump(visualizer.state_dict(), open('visualizer.json', 'w'), ensure_ascii=False, indent=4)

Citation

If you find this project useful in your research, please consider cite:

@misc{lagent2023, title={{Lagent: InternLM} a lightweight open-source framework that allows users to efficiently build large language model(LLM)-based agents}, author={Lagent Developer Team}, howpublished = {\url{https://github.com/InternLM/lagent}}, year={2023} }

License

This project is released under the Apache 2.0 license.

🔼 Back to top