Feature: support async qwen agent

[Zhuofeng-Li]

Overview

This commit introduces non-streaming asynchronous agents built on the Qwen-Agent framework, integrated with a vLLM server.

Usage

python examples/async_assistant_qwen_tool_call.py

Implementation Principles

1. Async Agent Architecture

The async agent implementation leverages asyncio and aiohttp for concurrent execution. All components are designed to align with the existing synchronous implementations closely:

• AsyncAssistant: The main asynchronous agent class that follows the Assistant logic while enabling fully async execution.
• TextChatAtOAIAsync: The asynchronous LLM interface that supports non-blocking OpenAI API calls.
• Async Tools: Custom tools implement an async def call() method to support non-blocking I/O.

2. Key Components

# Async tool definition
@register_tool('get_weather')
class GetWeatherTool(BaseTool):
    async def call(self, params: str, **kwargs) -> str:
        # Async implementation here
        pass

# Async agent setup
bot = AsyncAssistant(
    llm={'model_type': 'oai_async', ...},
    function_list=['get_weather']
)

# Async execution
response = await bot.run(messages=messages)

3. Benefits of Async

• Concurrent I/O: Non-blocking API calls and tool executions
• Better Performance: Handle multiple requests simultaneously
• Resource Efficiency: Lower memory footprint during I/O waits
• Scalability: Suitable for production environments with high concurrency

Changes

• Add AsyncAssitant and TextChatAtOAIAsync class
• Add async example to examples/ directory