Microsoft Agent Lightning\u662f\u4e00\u79cd\u57fa\u4e8e\u5f3a\u5316\u5b66\u4e60\u7684\u5f00\u6e90AI Agent\u8bad\u7ec3\u6846\u67b6\uff0c\u5176\u6838\u5fc3\u8bbe\u8ba1\u7406\u5ff5\u5728\u4e8e\u5c06\u8fd0\u884c\u903b\u8f91\u4e0e\u5b66\u4e60\u903b\u8f91\u89e3\u8026\uff0c\u4ece\u800c\u80fd\u591f\u7075\u6d3b\u4e14\u9ad8\u6548\u5730\u96c6\u6210\u81f3\u73b0\u6709\u7684\u667a\u80fd\u4f53\u4e2d\u3002\u672c\u6587\u4ecb\u7ecd\u4e86\u5176\u5de5\u4f5c\u539f\u7406\u3001\u4f7f\u7528\u6b65\u9aa4\uff0c\u5305\u62ec\u67b6\u6784\u3001\u8fde\u63a5\u3001\u8c03\u8bd5\u3001\u8bad\u7ec3\u7b49\u5185\u5bb9\uff0c\u5e76\u8fdb\u4e00\u6b65\u5206\u6790\u4e86\u5176\u5178\u578b\u5e94\u7528\u573a\u666f\u3002<\/p>\n
<\/p>\n
AI Agent\u6b63\u5728\u6539\u53d8\u4eba\u4eec\u4f7f\u7528\u4eba\u5de5\u667a\u80fd\u7684\u65b9\u5f0f\u3002\u4f9d\u6258\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\uff0c\u5b83\u4eec\u80fd\u591f\u56de\u7b54\u95ee\u9898\u3001\u6267\u884c\u6307\u4ee4\uff0c\u5e76\u4e0e\u5404\u7c7b\u6570\u636e\u6e90\u6216API\u8fdb\u884c\u4ea4\u4e92\u3002\u7136\u800c\uff0c\u5b83\u4eec\u5728\u5904\u7406\u590d\u6742\u7684\u591a\u6b65\u9aa4\u4efb\u52a1\u65f6\u4ecd\u4f1a\u51fa\u9519\uff0c\u800c\u4eba\u5de5\u4fee\u590d\u8fd9\u4e9b\u95ee\u9898\u901a\u5e38\u65e2\u8017\u65f6\u53c8\u8d39\u529b\u3002<\/p>\n
\u5fae\u8f6f\u516c\u53f8\u6700\u8fd1\u63a8\u51fa\u7684Agent Lightning\u6846\u67b6\u4f7f\u8fd9\u4e00\u8fc7\u7a0b\u53d8\u5f97\u66f4\u52a0\u7b80\u5355\u3002\u8be5\u6846\u67b6\u5c06\u667a\u80fd\u4f53\uff08Agent\uff09\u7684\u8fd0\u884c\u673a\u5236\u4e0e\u5b66\u4e60\u8fc7\u7a0b\u5206\u79bb\u5f00\u6765\uff0c\u4f7f\u5f97\u667a\u80fd\u4f53\u80fd\u591f\u5728\u5b9e\u9645\u4ea4\u4e92\u4e2d\u6301\u7eed\u81ea\u6211\u4f18\u5316\u3002\u7528\u6237\u53ef\u57fa\u4e8e\u73b0\u6709\u7684\u804a\u5929\u5de5\u5177\u6216\u81ea\u52a8\u5316\u7cfb\u7edf\u8fdb\u884c\u90e8\u7f72\uff0c\u5e76\u501f\u52a9\u5f3a\u5316\u5b66\u4e60\u6280\u672f\uff0c\u8ba9\u667a\u80fd\u4f53\u5728\u6267\u884c\u4efb\u52a1\u8fc7\u7a0b\u4e2d\u9010\u6b65\u63d0\u5347\u667a\u80fd\u5316\u6c34\u5e73\u3002<\/p>\n
Agent Lightning\u662f\u5fae\u8f6f\u516c\u53f8\u63a8\u51fa\u7684\u4e00\u79cd\u5f00\u6e90\u6846\u67b6\uff0c\u65e8\u5728\u901a\u8fc7\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u6765\u8bad\u7ec3\u548c\u6539\u8fdbAI Agent\u3002\u5176\u4f18\u52bf\u5728\u4e8e\uff0c\u5b83\u53ef\u4ee5\u65e0\u7f1d\u5c01\u88c5\u4efb\u4f55\u4f7f\u7528\u73b0\u6709\u6846\u67b6\uff08\u5982LangChain\u3001OpenAI Agents SDK\u3001AutoGen\u3001CrewAI\u3001LangGraph\u6216\u81ea\u5b9a\u4e49Python\uff09\u5f00\u53d1\u7684\u667a\u80fd\u4f53\uff0c\u5e76\u4e14\u51e0\u4e4e\u65e0\u9700\u66f4\u6539\u4ee3\u7801\u3002<\/p>\n
\u4ece\u6280\u672f\u89d2\u5ea6\u6765\u770b\uff0c\u8be5\u6846\u67b6\u652f\u6301\u5bf9\u667a\u80fd\u4f53\u4e2d\u5d4c\u5165\u7684LLM\u8fdb\u884c\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\uff0c\u800c\u65e0\u9700\u6539\u53d8\u667a\u80fd\u4f53\u7684\u6838\u5fc3\u903b\u8f91\u3002\u5176\u57fa\u672c\u601d\u60f3\u662f\u5c06\u667a\u80fd\u4f53\u7684\u6267\u884c\u8fc7\u7a0b\u89c6\u4e3a\u4e00\u4e2a\u9a6c\u5c14\u53ef\u592b\u51b3\u7b56\u8fc7\u7a0b\u2014\u2014\u201c\u5728\u6bcf\u4e00\u6b65\uff0c\u667a\u80fd\u4f53\u90fd\u5904\u4e8e\u7279\u5b9a\u72b6\u6001\uff0c\u6267\u884c\u67d0\u4e2a\u884c\u52a8\uff08LLM\u8f93\u51fa\uff09\uff0c\u5e76\u5728\u8fd9\u4e9b\u884c\u52a8\u5728\u6210\u529f\u5b8c\u6210\u4efb\u52a1\u65f6\u83b7\u5f97\u76f8\u5e94\u7684\u5956\u52b1\u3002\u201d<\/p>\n
Agent Lightning\u6846\u67b6\u7531Python SDK\u4e0e\u8bad\u7ec3\u670d\u52a1\u5668\u4e24\u90e8\u5206\u7ec4\u6210\u3002\u7528\u6237\u53ea\u9700\u5c06\u667a\u80fd\u4f53\u903b\u8f91\u5c01\u88c5\u8fdbLitAgent\u7c7b\u6216\u517c\u5bb9\u63a5\u53e3\uff0c\u5e76\u5b9a\u4e49\u8f93\u51fa\u7ed3\u679c\uff08\u5956\u52b1\uff09\u8bc4\u4f30\u673a\u5236\uff0c\u5373\u53ef\u542f\u52a8\u8bad\u7ec3\u3002\u8be5\u6846\u67b6\u8d1f\u8d23\u91c7\u96c6\u4ea4\u4e92\u7ecf\u9a8c\uff0c\u5c06\u5176\u8f93\u5165\u81f3\u5206\u5c42\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5\uff08LightningRL\uff09\u4e2d\u4ee5\u5b8c\u6210\u4fe1\u7528\u5206\u914d\uff0c\u8fdb\u800c\u66f4\u65b0\u667a\u80fd\u4f53\u6a21\u578b\u6216\u63d0\u793a\u6a21\u677f\u3002\u5728\u8bad\u7ec3\u7ed3\u675f\u540e\uff0c\u7528\u6237\u5373\u53ef\u83b7\u5f97\u4e00\u4e2a\u7ecf\u8fc7\u5f3a\u5316\u3001\u6027\u80fd\u663e\u8457\u63d0\u5347\u7684\u667a\u80fd\u4f53\u3002<\/p>\n
\u4f20\u7edf\u7684\u667a\u80fd\u4f53\u6846\u67b6\uff08\u5982LangChain\u3001LangGraph\u3001CrewAI\u6216AutoGen\uff09\u80fd\u591f\u521b\u5efa\u53ef\u4ee5\u9010\u6b65\u63a8\u7406\u6216\u4f7f\u7528\u5de5\u5177\u7684AI Agent\uff0c\u4f46\u5b83\u4eec\u6ca1\u6709\u8bad\u7ec3\u6a21\u5757\u3002\u8fd9\u4e9b\u667a\u80fd\u4f53\u53ea\u662f\u57fa\u4e8e\u9759\u6001\u7684\u6a21\u578b\u53c2\u6570\u6216\u63d0\u793a\u8bcd\u6765\u8fd0\u884c\u6a21\u578b\uff0c\u8fd9\u610f\u5473\u7740\u5b83\u4eec\u65e0\u6cd5\u4ece\u6240\u7ecf\u5386\u7684\u60c5\u51b5\u4e2d\u5b66\u4e60\u3002\u73b0\u5b9e\u4e16\u754c\u4e2d\u7684\u6311\u6218\u5177\u6709\u4e00\u5b9a\u7a0b\u5ea6\u7684\u590d\u6742\u6027\uff0c\u9700\u8981\u4e00\u5b9a\u7a0b\u5ea6\u7684\u9002\u5e94\u6027\u3002Agent Lightning\u89e3\u51b3\u4e86\u8fd9\u4e00\u95ee\u9898\uff0c\u5c06\u5b66\u4e60\u5f15\u5165\u4e86\u667a\u80fd\u4f53\u6d41\u7a0b\u4e2d\u3002<\/p>\n
Agent Lightning\u901a\u8fc7\u5b9e\u73b0\u4e00\u4e2a\u81ea\u52a8\u4f18\u5316\u7ba1\u9053\u6765\u89e3\u51b3\u8fd9\u4e00\u9884\u671f\u5dee\u8ddd\uff0c\u5b83\u5229\u7528\u5f3a\u5316\u5b66\u4e60\u7684\u80fd\u529b\uff0c\u6839\u636e\u53cd\u9988\u4fe1\u53f7\u66f4\u65b0\u667a\u80fd\u4f53\u7684\u7b56\u7565\u3002\u7b80\u800c\u8a00\u4e4b\uff0c\u667a\u80fd\u4f53\u53ef\u4ee5\u4ece\u81ea\u5df1\u7684\u6210\u529f\u548c\u5931\u8d25\u4e2d\u5b66\u4e60\uff0c\u4ece\u800c\u53ef\u80fd\u4ea7\u751f\u66f4\u53ef\u9760\u3001\u66f4\u503c\u5f97\u4fe1\u8d56\u7684\u7ed3\u679c\u3002<\/p>\n
\u5728\u670d\u52a1\u5668-\u5ba2\u6237\u7aef\u67b6\u6784\u4e2d\uff0cAgent Lightning\u5229\u7528\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5\u6765\u751f\u6210\u4efb\u52a1\u5e76\u4f18\u5316\u5efa\u8bae\uff0c\u4f8b\u5982\u8c03\u6574\u63d0\u793a\u6216\u66f4\u65b0\u6a21\u578b\u6743\u91cd\u3002\u4efb\u52a1\u7531Runner\u6267\u884c\uff0c\u5b83\u6536\u96c6\u667a\u80fd\u4f53\u7684\u6bcf\u4e2a\u884c\u52a8\u548c\u6700\u7ec8\u5956\u52b1\uff0c\u5e76\u5c06\u8fd9\u4e9b\u6570\u636e\u8fd4\u56de\u7ed9\u7b97\u6cd5\u3002\u8fd9\u79cd\u53cd\u9988\u673a\u5236\u4f7f\u5f97\u667a\u80fd\u4f53\u80fd\u591f\u968f\u7740\u65f6\u95f4\u7684\u63a8\u79fb\uff0c\u5229\u7528\u201c\u81ea\u52a8\u4e2d\u95f4\u5956\u52b1\u201d\u529f\u80fd\u6765\u8fdb\u4e00\u6b65\u5fae\u8c03\u5176\u63d0\u793a\u6216\u6743\u91cd\uff0c\u8be5\u529f\u80fd\u53ef\u4e3a\u6210\u529f\u7684\u4e2d\u95f4\u884c\u52a8\u63d0\u4f9b\u5373\u65f6\u7684\u5c0f\u5e45\u5956\u52b1\u4ee5\u52a0\u901f\u5b66\u4e60\u8fc7\u7a0b\u3002<\/p>\n
Agent Lightning\u672c\u8d28\u4e0a\u5c06Agent\u64cd\u4f5c\u89c6\u4e3a\u4e00\u4e2a\u5faa\u73af\uff1a\u72b6\u6001\u662f\u5176\u5f53\u524d\u4e0a\u4e0b\u6587\uff1b\u884c\u52a8\u662f\u5b83\u7684\u4e0b\u4e00\u6b65\uff0c\u5956\u52b1\u662f\u4efb\u52a1\u6210\u529f\u7684\u6307\u6807\u3002\u901a\u8fc7\u8bbe\u8ba1\u201c\u72b6\u6001-\u884c\u52a8-\u5956\u52b1\u201d\u7684\u8f6c\u6362\uff0cAgent Lightning\u6700\u7ec8\u53ef\u4ee5\u4e3a\u4efb\u4f55\u7c7b\u578b\u7684\u667a\u80fd\u4f53\u63d0\u4f9b\u8bad\u7ec3\u652f\u6301\u3002<\/p>\n
Agent Lightning\u91c7\u7528\u4e86\u89e3\u8026\u5316\u8bbe\u8ba1\uff0c\u5c06\u667a\u80fd\u4f53\u7684\u5b66\u4e60\u4e0e\u6267\u884c\u76f8\u5206\u79bb\u3002\u5728\u8fd9\u4e00\u67b6\u6784\u4e2d\uff0c\u670d\u52a1\u5668\u8d1f\u8d23\u6a21\u578b\u66f4\u65b0\u4e0e\u4f18\u5316\uff0c\u5ba2\u6237\u7aef\u5219\u8d1f\u8d23\u6267\u884c\u5b9e\u9645\u4efb\u52a1\u5e76\u62a5\u544a\u7ed3\u679c\u3002\u8fd9\u79cd\u4efb\u52a1\u5206\u5de5\u4f7f\u667a\u80fd\u4f53\u5728\u9ad8\u6548\u5b8c\u6210\u5177\u4f53\u4efb\u52a1\u7684\u540c\u65f6\uff0c\u80fd\u6301\u7eed\u901a\u8fc7\u5f3a\u5316\u5b66\u4e60\u4f18\u5316\u5176\u6027\u80fd\u3002<\/p>\n
\u6ce8\uff1a<\/strong>Agent Lightning\u4f7f\u7528\u5206\u5c42\u5f3a\u5316\u5b66\u4e60\u7cfb\u7edfLightningRL\uff0c\u80fd\u591f\u5206\u89e3\u590d\u6742\u7684\u591a\u6b65\u9aa4\u667a\u80fd\u4f53\u884c\u4e3a\u8fdb\u884c\u8bad\u7ec3\u3002LightningRL\u8fd8\u652f\u6301\u591a\u667a\u80fd\u4f53\u3001\u590d\u6742\u5de5\u5177\u4f7f\u7528\u548c\u5ef6\u8fdf\u53cd\u9988\u3002<\/p>\n \u672c\u8282\u5c06\u4ecb\u7ecd\u5982\u4f55\u4f7f\u7528Agent Lightning\u8bad\u7ec3\u4e00\u4e2aSQL\u667a\u80fd\u4f53\uff0c\u5e76\u5c55\u793a\u5982\u4f55\u96c6\u6210\u5176\u6838\u5fc3\u7ec4\u4ef6\uff1a\u57fa\u4e8eLangGraph\u6784\u5efa\u7684SQL\u667a\u80fd\u4f53\u3001VERL\u5f3a\u5316\u5b66\u4e60\u6846\u67b6\uff0c\u4ee5\u53ca\u7528\u4e8e\u63a7\u5236\u8bad\u7ec3\u4e0e\u8c03\u8bd5\u6d41\u7a0b\u7684Trainer\u3002<\/p>\n \u672c\u6587\u63d0\u4f9b\u4e86\u4e00\u4e2a\u53ef\u76f4\u63a5\u8fd0\u884c\u7684\u547d\u4ee4\u884c\u793a\u4f8b\uff08examples\/spider\/train_sql_agent.py\uff09\uff0c\u4f46\u91cd\u70b9\u5728\u4e8e\u5e2e\u52a9\u5f00\u53d1\u8005\u6df1\u5165\u7406\u89e3\u7cfb\u7edf\u67b6\u6784\u4e0e\u5de5\u4f5c\u6d41\u7a0b\uff0c\u4ece\u800c\u80fd\u591f\u81ea\u4fe1\u5730\u5c06Agent Lightning\u5e94\u7528\u4e8e\u5b9e\u9645\u4e1a\u52a1\u573a\u666f\u3002<\/p>\n Agent-Lightning\u53ef\u4ee5\u4e0eAutoGen\u3001CrewAI\u3001LangGraph\u3001OpenAI Agents SDK\u4ee5\u53ca\u5176\u4ed6\u81ea\u5b9a\u4e49Python\u903b\u8f91\u7b49\u6846\u67b6\u65e0\u7f1d\u534f\u4f5c\u3002\u5728\u8fd9\u4e2a\u4f8b\u5b50\u4e2d\uff0c\u4f7f\u7528LangGraph\u6784\u5efa\u4e86\u4e00\u4e2a\u5faa\u73af\u5de5\u4f5c\u6d41\uff0c\u6a21\u62df\u6570\u636e\u5206\u6790\u5e08\u8fed\u4ee3\u7f16\u5199\u548c\u4fee\u590dSQL\u67e5\u8be2\u7684\u8fc7\u7a0b\uff1a<\/p>\n \u8fd9\u4e2a\u5de5\u4f5c\u6d41\u5305\u542b\u56db\u4e2a\u529f\u80fd\u9636\u6bb5\uff1a<\/p>\n \u5faa\u73af\u7ee7\u7eed\uff0c\u76f4\u5230\u67e5\u8be2\u9a8c\u8bc1\u6216\u8fbe\u5230\u6700\u5927\u8fed\u4ee3\u8ba1\u6570\uff08max_turns\uff09\u3002\u5f3a\u5316\u5b66\u4e60\u4f18\u5316\u4e86write_query\u548crewrite_query\u9636\u6bb5\u3002<\/p>\n \u4e3a\u4fdd\u8bc1\u4ee3\u7801\u7684\u6a21\u5757\u5316\u4e0e\u53ef\u7ef4\u62a4\u6027\uff0c\u5efa\u8bae\u4f7f\u7528\u72ec\u7acb\u7684\u6784\u5efa\u5668\u51fd\u6570\u6765\u5b9a\u4e49 LangGraph \u5de5\u4f5c\u6d41\uff0c\u5177\u4f53\u5b9e\u73b0\u5982\u4e0b\u6240\u793a\uff1a<\/p>\n python<\/p>\n \u590d\u5236<\/p>\n from langgraph import StateGraph def build_langgraph_sql_agent( database_path: str, openai_base_url: str, model: str, sampling_parameters: dict, max_turns: int, truncate_length: int ): # Step 1: Define the LangGraph workflow builder = StateGraph() # Step 2: Add agent nodes for each step builder.add_node(“write_query”) builder.add_node(“execute_query”) builder.add_node(“check_query”) builder.add_node(“rewrite_query”) # Step 3: Connect the workflow edges builder.add_edge(“__start__”, “write_query”) builder.add_edge(“write_query”, “execute_query”) builder.add_edge(“execute_query”, “check_query”) builder.add_edge(“check_query”, “rewrite_query”) builder.add_edge(“rewrite_query”, “__end__”) # Step 4: Compile the graph return builder.compile().graph()<\/p>\n \u8fd9\u6837\u505a\u53ef\u4ee5\u5c06LangGraph\u903b\u8f91\u4e0eAgent Lightning\u7684\u672a\u6765\u66f4\u65b0\u76f8\u5206\u79bb\uff0c\u4ece\u800c\u63d0\u9ad8\u4ee3\u7801\u7684\u53ef\u8bfb\u6027\u548c\u53ef\u7ef4\u62a4\u6027\u3002<\/p>\n LitSQLAgent\u7c7b\u4f5c\u4e3aLangGraph\u4e0eAgent Lightning\u4e4b\u95f4\u7684\u6865\u6881\u3002\u5b83\u7ee7\u627f\u81eaagl.LitAgent\uff0c\u56e0\u6b64Runner\u53ef\u4ee5\u4e3a\u6bcf\u6b21\u8fed\u4ee3\u7ba1\u7406\u5171\u4eab\u8d44\u6e90\uff08\u5982LLM\uff09\u3002<\/p>\n python<\/p>\n \u590d\u5236<\/p>\n import agentlightning as agl class LitSQLAgent(agl.LitAgent[dict]): def __init__(self, max_turns: int, truncate_length: int): super().__init__() self.max_turns = max_turns self.truncate_length = truncate_length def rollout(self, task: dict, resources: agl.NamedResources, rollout: agl.Rollout) -> float: # Step 1: Load shared LLM resource llm: agl.LLM = resources[“main_llm”] # Step 2: Build LangGraph agent dynamically agent = build_langgraph_sql_agent( database_path=”sqlite:\/\/\/” + task[“db_id”], openai_base_url=llm.get_base_url(rollout.rollout_id, rollout.attempt.attempt_id), model=llm.model, sampling_parameters=llm.sampling_parameters, max_turns=self.max_turns, truncate_length=self.truncate_length, ) # Step 3: Invoke agent result = agent.invoke({“question”: task[“question”]}, { “callbacks”: [self.tracer.get_langchain_handler()], “recursion_limit”: 100, }) # Step 4: Evaluate query to generate reward reward = evaluate_query( result[“query”], task[“ground_truth”], task[“db_path”], raise_on_error=False ) return reward<\/p>\n \u6ce8\uff1a<\/strong>\u201cmain_llm\u201d\u8d44\u6e90\u952e\u662f\u667a\u80fd\u4f53\u4e0eVERL\u4e4b\u95f4\u7684\u4e00\u79cd\u534f\u4f5c\u7ea6\u5b9a\uff0c\u7528\u4e8e\u5728\u670d\u52a1\u4e0a\u4e0b\u6587\u4e2d\u4e3a\u6bcf\u6b21\u4efb\u52a1\u63d0\u4f9b\u6b63\u786e\u7684\u7aef\u70b9\u8bbf\u95ee\u3002<\/p>\n evaluate_query\u51fd\u6570\u5b9a\u4e49\u4e86\u7528\u4e8e\u5f3a\u5316\u8bad\u7ec3\u7684\u5956\u52b1\u673a\u5236\u3002Spider\u6570\u636e\u96c6\u4e0a\u7684\u6bcf\u4e2a\u4efb\u52a1\u90fd\u5305\u542b\u4e00\u4e2a\u81ea\u7136\u8bed\u8a00\u95ee\u9898\u3001\u4e00\u4e2a\u6570\u636e\u5e93\u6a21\u5f0f\u548c\u4e00\u4e2a\u57fa\u51c6SQL\u67e5\u8be2\u3002\u5956\u52b1\u673a\u5236\u5c06\u6a21\u578b\u751f\u6210\u7684SQL\u67e5\u8be2\u4e0e\u53c2\u8003SQL\u67e5\u8be2\u8fdb\u884c\u6bd4\u8f83\uff1a<\/p>\n python<\/p>\n \u590d\u5236<\/p>\n def evaluate_query(predicted_query, ground_truth_query, db_path, raise_on_error=False): result_pred = run_sql(predicted_query, db_path) result_true = run_sql(ground_truth_query, db_path) return 1.0 if result_pred == result_true else 0.0<\/p>\n \u6ce8\uff1a<\/strong>\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0c\u667a\u80fd\u4f53\u7edd\u4e0d\u80fd\u770b\u5230\u771f\u5b9e\u67e5\u8be2\uff0c\u5426\u5219\u4f1a\u5bfc\u81f4\u4fe1\u606f\u6cc4\u9732\u3002<\/p>\n VERL\u662f\u667a\u80fd\u4f53\u7684\u5f3a\u5316\u5b66\u4e60\u540e\u7aef\u3002\u5176\u914d\u7f6e\u5982\u540c\u5b9a\u4e49Python\u5b57\u5178\u4e00\u6837\u7b80\u5355\uff0c\u7528\u6237\u9700\u8981\u8f93\u5165\u7b97\u6cd5\u3001\u6a21\u578b\u3001rollout\u53c2\u6570\u548c\u8bad\u7ec3\u9009\u9879\u3002\u4ee5\u4e0b\u662f\u4e00\u4e2a\u57fa\u7840\u914d\u7f6e\u793a\u4f8b\uff1a<\/p>\n python<\/p>\n \u590d\u5236<\/p>\n verl_config = { “algorithm”: {“adv_estimator”: “grpo”, “use_kl_in_reward”: False}, “data”: { “train_batch_size”: 32, “max_prompt_length”: 4096, “max_response_length”: 2048, }, “actor_rollout_ref”: { “rollout”: {“name”: “vllm”, “n”: 4, “multi_turn”: {“format”: “hermes”}}, “actor”: {“ppo_mini_batch_size”: 32, “optim”: {“lr”: 1e-6}}, “model”: {“path”: “Qwen\/Qwen2.5-Coder-1.5B-Instruct”}, }, “trainer”: { “n_gpus_per_node”: 1, “val_before_train”: True, “test_freq”: 32, “save_freq”: 64, “total_epochs”: 2, }, }<\/p>\n \u8fd9\u7c7b\u4f3c\u4e8e\u53ef\u4ee5\u5728CLI\u4e2d\u8fd0\u884c\u7684\u547d\u4ee4\uff1a<\/p>\n python<\/p>\n \u590d\u5236<\/p>\n python3 -m verl.trainer.main_ppo \\ algorithm.adv_estimator=grpo \\ data.train_batch_size=32 \\ actor_rollout_ref.model.path=Qwen\/Qwen2.5-Coder-1.5B-Instruct<\/p>\n Trainer\u662f\u8fde\u63a5\u667a\u80fd\u4f53\u3001\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5\u3001\u6570\u636e\u96c6\u548c\u5206\u5e03\u5f0fRunner\u7684\u9ad8\u7ea7\u534f\u8c03\u6a21\u5757\u3002<\/p>\n python<\/p>\n \u590d\u5236<\/p>\n import pandas as pd import agentlightning as agl # Step 1: Initialize agent and algorithm agent = LitSQLAgent(max_turns=3, truncate_length=1024) algorithm = agl.VERL(verl_config) # Step 2: Initialize Trainer trainer = agl.Trainer( n_runners=10, algorithm=algorithm, adapter={“agent_match”: “write|rewrite”} # Optimize both query stages ) # Step 3: Load dataset train_data = pd.read_parquet(“data\/train_spider.parquet”).to_dict(“records”) val_data = pd.read_parquet(“data\/test_dev_500.parquet”).to_dict(“records”) # Step 4: Train trainer.fit(agent, train_dataset=train_data, val_dataset=val_data)<\/p>\n \u5e55\u540e\u53d1\u751f\u7684\u8fc7\u7a0b\u5982\u4e0b\uff1a<\/p>\n \u5728\u5f00\u59cb\u5b8c\u6574\u7684\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u8bad\u7ec3\u4e4b\u524d\uff0c\u5efa\u8bae\u5bf9\u6574\u4e2a\u6d41\u7a0b\u8fdb\u884c\u8bd5\u8fd0\u884c\uff0c\u4ee5\u68c0\u67e5\u8fde\u63a5\u60c5\u51b5\u548c\u8f68\u8ff9\u4fe1\u606f\u3002<\/p>\n python<\/p>\n \u590d\u5236<\/p>\n trainer = agl.Trainer( n_workers=1, initial_resources={ “main_llm”: agl.LLM( endpoint=os.environ[“OPENAI_API_BASE”], model=”gpt-4.1-nano”, sampling_parameters={“temperature”: 0.7}, ) }, ) # Load a small subset for dry-run import pandas as pd dev_data = pd.read_parquet(“data\/test_dev_500.parquet”).to_dict(“records”)[:10] # Run dry-run mode trainer.dev(agent, dev_dataset=dev_data) \u8be5\u6b65\u9aa4\u786e\u4fdd\u5728\u6295\u5165\u5927\u91cfGPU\u8d44\u6e90\u8fdb\u884c\u957f\u65f6\u95f4\u8bad\u7ec3\u4e4b\u524d\uff0c\u6574\u4e2aLangGraph\u63a7\u5236\u6d41\u3001\u6570\u636e\u5e93\u8fde\u63a5\u53ca\u5956\u52b1\u903b\u8f91\u5747\u5df2\u6b63\u786e\u914d\u7f6e\u5e76\u9a8c\u8bc1\u901a\u8fc7\u3002<\/p>\n \u8981\u8bbe\u7f6e\u73af\u5883\uff0c\u9700\u5b89\u88c5\u4f9d\u8d56\u9879\uff08\u5373\u4f7f\u7528pip install -r requirements.txt\u547d\u4ee4\uff09\uff0c\u7136\u540e\u8fd0\u884c\u5b8c\u6574\u7684\u8bad\u7ec3\u811a\u672c\uff1a<\/p>\n python<\/p>\n \u590d\u5236<\/p>\n # Step 1: Install dependencies pip install “agentlightning[verl]” langchain pandas gdown # Step 2: Download Spider dataset cd examples\/spider gdown –fuzzy https:\/\/drive.google.com\/file\/d\/1oi9J1jZP9TyM35L85CL3qeGWl2jqlnL6\/view unzip -q spider-data.zip -d data && rm spider-data.zip # Step 3: Launch training python train_sql_agent.py qwen # Qwen-2.5-Coder-1.5B # or python train_sql_agent.py llama # LLaMA 3.2 1B<\/p>\n \u5982\u679c\u4f7f\u7528\u7684\u662f\u6258\u7ba1\u5728Hugging Face\u5e73\u53f0\u4e0a\u7684\u6a21\u578b\uff0c\u90a3\u4e48\u52a1\u5fc5\u6309\u7167\u4ee5\u4e0b\u6b65\u9aa4\u5bfc\u8bbf\u95ee\u4ee4\u724c\uff1a<\/p>\n python<\/p>\n \u590d\u5236<\/p>\n export HF_TOKEN=”your_huggingface_token”<\/p>\n \u5982\u679c\u5e0c\u671b\u5728\u6ca1\u6709\u5f3a\u5316\u5b66\u4e60\u7684\u60c5\u51b5\u4e0b\u9a8c\u8bc1\u4ee3\u7406\u903b\u8f91\uff0c\u53ef\u4ee5\u4f7f\u7528\u5185\u7f6e\u7684\u8c03\u8bd5\u52a9\u624b\uff1a<\/p>\n python<\/p>\n \u590d\u5236<\/p>\n export OPENAI_API_BASE=”https:\/\/api.openai.com\/v1″ export OPENAI_API_KEY=”your_api_key_here” cd examples\/spider python sql_agent.py<\/p>\n \u8fd9\u5c06\u5141\u8bb8\u7528\u6237\u4f7f\u7528\u5f53\u524dLLM\u7aef\u70b9\u8fd0\u884cSQL\u00a0Agent\uff0c\u4ee5\u786e\u8ba4\u67e5\u8be2\u5df2\u6267\u884c\u4e14\u63a7\u5236\u6d41\u6309\u9884\u671f\u5de5\u4f5c\u3002<\/p>\n \u6ce8\uff1a<\/strong>\u5728\u5355\u4e2a80GB GPU\u4e0a\u8fd0\u884cpython train_sql_agent.py qwen\u901a\u5e38\u4f1a\u5728\u5927\u7ea612\u5c0f\u65f6\u540e\u5b8c\u6210\u3002\u4f60\u5c06\u770b\u5230\u8bad\u7ec3\u83b7\u5f97\u7684\u5956\u52b1\u6301\u7eed\u589e\u52a0\uff0c\u8fd9\u8868\u660e\u667a\u80fd\u4f53\u7684SQL\u751f\u6210\u8fc7\u7a0b\u968f\u7740\u65f6\u95f4\u7684\u63a8\u79fb\u800c\u4e0d\u65ad\u6539\u8fdb\u3002\u56e0\u6b64\uff0c\u7531\u4e8e\u8d44\u6e90\u9650\u5236\uff0c\u5728\u8fd9\u91cc\u4f7f\u7528\u4e86\u5b98\u65b9\u6587\u6863\u4e2d\u5c55\u793a\u7684\u7ed3\u679c\u3002<\/p>\n \u5728\u5b9e\u9645\u5e94\u7528\u4e2d\uff0c\u5047\u8bbe\u7528\u6237\u6709\u4e00\u4e2a\u57fa\u4e8eLLM\u7684\u667a\u80fd\u4f53\uff0c\u5b83\u5728\u67d0\u4e2a\u5e94\u7528\u7a0b\u5e8f\uff08\u5982\u5ba2\u6237\u652f\u6301\u804a\u5929\u673a\u5668\u4eba\u3001\u81ea\u52a8\u5316\u7f16\u7801\u52a9\u624b\u7b49\uff09\u4e2d\u626e\u6f14\u7740\u91cd\u8981\u89d2\u8272\uff0c\u5e76\u4e14\u5e0c\u671b\u5bf9\u5176\u8fdb\u884c\u4f18\u5316\uff0c\u90a3\u4e48Agent Lightning\u662f\u4e00\u4e2a\u7406\u60f3\u7684\u5019\u9009\u65b9\u6848\u3002\u8be5\u6846\u67b6\u5df2\u5728SQL\u67e5\u8be2\u751f\u6210\u7b49\u4efb\u52a1\u4e2d\u5f97\u5230\u9a8c\u8bc1\u3002\u7ecf\u8fc7\u5f15\u5165\u5f3a\u5316\u5b66\u4e60\u6216\u63d0\u793a\u4f18\u5316\u673a\u5236\uff0cAgent Lightning\u80fd\u591f\u6301\u7eed\u8fed\u4ee3\u5e76\u4f18\u5316\u5df2\u6709\u667a\u80fd\u4f53\uff0c\u4ece\u800c\u751f\u6210\u66f4\u51c6\u786e\u3001\u66f4\u53ef\u9760\u7684\u8f93\u51fa\u7ed3\u679c\u3002<\/p>\n Microsoft Agent Lightning\u662f\u4e00\u79cd\u7528\u4e8e\u63d0\u5347AI Agent\u667a\u80fd\u6c34\u5e73\u7684\u65b0\u673a\u5236\u3002\u5176\u6838\u5fc3\u7406\u5ff5\u5728\u4e8e\uff0c\u667a\u80fd\u4f53\u4e0d\u5e94\u662f\u56fa\u5b9a\u7684\u4ee3\u7801\u5bf9\u8c61\uff0c\u800c\u5e94\u8be5\u80fd\u591f\u901a\u8fc7\u6301\u7eed\u7684\u8bad\u7ec3\u5faa\u73af\u4ece\u7ecf\u9a8c\u4e2d\u5b66\u4e60\u3002\u901a\u8fc7\u5c06\u8bad\u7ec3\u8fc7\u7a0b\u4e0e\u6267\u884c\u8fc7\u7a0b\u89e3\u8026\uff0c\u5b83\u53ef\u4ee5\u5728\u65e0\u9700\u66f4\u6539\u4efb\u4f55\u4ee3\u7801\u7684\u60c5\u51b5\u4e0b\u4f18\u5316\u4efb\u4f55\u667a\u80fd\u4f53\u5de5\u4f5c\u6d41\u3002<\/p>\n \u8fd9\u610f\u5473\u7740\uff0c\u53ef\u4ee5\u901a\u8fc7\u4f7f\u7528Agent Lightning\u7684\u5f3a\u5316\u5b66\u4e60\u673a\u5236\u8f7b\u677e\u589e\u5f3a\u667a\u80fd\u4f53\u5de5\u4f5c\u6d41\u7a0b\uff0c\u65e0\u8bba\u5b83\u662f\u5b9a\u5236\u667a\u80fd\u4f53\u3001LangChain\u673a\u5668\u4eba\u3001CrewAI\u3001LangGraph\u3001AutoGen\u8fd8\u662f\u66f4\u5177\u4f53\u7684OpenAI SDK agent,\u3002\u5b9e\u9645\u4e0a\uff0c\u667a\u80fd\u4f53\u6b63\u4ece\u5176\u81ea\u8eab\u7684\u6570\u636e\u4e2d\u53d8\u5f97\u66f4\u52a0\u667a\u80fd\u3002<\/p>\n Q1.\u4ec0\u4e48\u662fMicrosoft Agent Lightning\uff1f<\/strong><\/p>\n A\uff1a<\/strong>\u5b83\u662f\u5fae\u8f6f\u7684\u4e00\u4e2a\u5f00\u6e90\u6846\u67b6\uff0c\u901a\u8fc7\u5f3a\u5316\u5b66\u4e60\u6765\u8bad\u7ec3AI\u00a0Agent\uff0c\u800c\u65e0\u9700\u6539\u53d8\u5176\u6838\u5fc3\u903b\u8f91\u6216\u5de5\u4f5c\u6d41\u3002<\/p>\n Q2. Agent Lightning\u5982\u4f55\u6539\u8fdbAI Agent\uff1f<\/strong><\/p>\n A\uff1a<\/strong>\u5b83\u5141\u8bb8\u667a\u80fd\u4f53\u901a\u8fc7\u5f3a\u5316\u5b66\u4e60\u4ece\u5b9e\u9645\u4efb\u52a1\u53cd\u9988\u4e2d\u5b66\u4e60\uff0c\u4ece\u800c\u4e0d\u65ad\u4f18\u5316\u63d0\u793a\u6216\u6a21\u578b\u6743\u91cd\u4ee5\u63d0\u9ad8\u6027\u80fd\u3002<\/p>\n Q3. Agent Lightning\u80fd\u5426\u4e0e\u73b0\u6709\u7684\u667a\u80fd\u4f53\u7684\u6846\u67b6\u534f\u540c\u5de5\u4f5c\uff1f<\/strong><\/p>\n A\uff1a<\/strong>\u662f\u7684\uff0c\u5b83\u53ef\u4ee5\u4e0eLangChain\u3001AutoGen\u3001CrewAI\u3001LangGraph\u4ee5\u53ca\u81ea\u5b9a\u4e49Python\u667a\u80fd\u4f53\u8f7b\u677e\u96c6\u6210\uff0c\u5e76\u4e14\u51e0\u4e4e\u65e0\u9700\u4fee\u6539\u4ee3\u7801\u3002<\/p>\n \u539f\u6587\u6807\u9898\uff1aTrain Your AI Agents Like a Pro with Microsoft Agent Lightning (Full Setup & Workflow)\u00a0<\/u>\uff0c\u4f5c\u8005\uff1aVipin Vashisth<\/p>\n \u6587\u7ae0\u6765\u81ea\uff1a51CTO<\/p>\n\u4f7f\u7528\u5fae\u8f6fAgent Lightning\u8bad\u7ec3\u667a\u80fd\u4f53\u7684\u5206\u6b65\u6307\u5357<\/h3>\n
\u667a\u80fd\u4f53\u67b6\u6784<\/h3>\n
\n
\u6784\u5efaLangGraph\u667a\u80fd\u4f53<\/h3>\n
\n
\u8fde\u63a5LangGraph\u4e0eAgent Lightning<\/h3>\n
\n
\u5956\u52b1\u4fe1\u53f7\u548c\u8bc4\u4f30\u673a\u5236<\/h3>\n
\n
\u914d\u7f6eVERL\u4ee5\u8fdb\u884c\u5f3a\u5316\u5b66\u4e60<\/h3>\n
\n
\n
\u4f7f\u7528Trainer\u7f16\u6392\u8bad\u7ec3\u8fc7\u7a0b<\/h3>\n
\n
\n
\u4f7f\u7528trainer.dev()\u8c03\u8bd5\u667a\u80fd\u4f53<\/h3>\n
\n
\u8fd0\u884c\u5b8c\u6574\u793a\u4f8b<\/h3>\n
\n
\n
\u5728\u6ca1\u6709VERL\u60c5\u51b5\u4e0b\u7684\u8c03\u8bd5<\/h3>\n
\n
\u8bc4\u4f30\u7ed3\u679c<\/h3>\n
Agent Lightning\u7684\u9002\u7528\u573a\u666f<\/h3>\n
\n
\u7ed3\u8bba<\/h3>\n
\u5e38\u89c1\u95ee\u9898\u89e3\u7b54<\/h3>\n