夜雨聆风声明: 纯兴趣爱好，如有疏漏敬请谅解。
源码版本: v0.21.0
学习相关源码路径:
vllm/vllm/entrypoints/cli/serve.py at v0.21.0 · vllm-project/vllm · GitHub
概述:
今天来详细剖析，serve三种启动方式之一的headless：
1. run_headless:
无头模式入口函数(api_server_count小于1时)
启动命令: vllm serve --headless
但是会出现报错RuntimeError: Did not receive response from front-end process within 5 minutes，目前还没找错误原因。哪位大佬知道可以指导我一下。
2. run_multi_api_server:
多API服务进程启动入口(api_server_count大于1时)
启动命令: vllm serve --api-server-count=2
3. uvloop.run(run_server(args))
单实例HTTP API服务顶层启动入口(api_server_count等于1时)
启动命令: vllm serve
先回顾下上篇文章(如下图)，启动方式由以下参数决定
data_parallel_exteran_lb、data_parallel_rank参数决定了is_exteranl_lb
data_parallel_hybrid_lb、data_parallel_start_rank参数决定了is_hybrid_lb
(注: is_exteranl_lb、is_hybrid_lb互斥， 如果is_exteranl_lb、is_hybrid_lb同时为true则程序中断)
enable_elastic_ep: 对api_server_count进行修正
headless函数源码剖析
1. api_server_count验证
if args.api_server_count > 1:raise ValueError("api_server_count can't be set in headless mode")
2. 创建运行配置vllm_config
# Create the EngineConfig.engine_args = vllm.AsyncEngineArgs.from_cli_args(args)usage_context = UsageContext.OPENAI_API_SERVERvllm_config = engine_args.create_engine_config(usage_context=usage_context, headless=True)
3. data_parallel_hybrid_lb验证
if engine_args.data_parallel_hybrid_lb:raise ValueError("data_parallel_hybrid_lb is not applicable in headless mode")
4.获取并行配置 local_engine_count参数验证
parallel_config = vllm_config.parallel_configlocal_engine_count = parallel_config.data_parallel_size_localif local_engine_count <= 0:raise ValueError("data_parallel_size_local must be > 0 in headless mode")
5. 注册终止信号回调函数，实现优雅关闭
shutdown_requested = Falsedef signal_handler(signum, frame):nonlocal shutdown_requestedlogger.debug("Received %d signal.", signum)if not shutdown_requested:shutdown_requested = Trueraise SystemExitsignal.signal(signal.SIGTERM, signal_handler)signal.signal(signal.SIGINT, signal_handler)
6. 分布式从节点逻辑
单机部署永远不会触发
if parallel_config.node_rank_within_dp > 0:from vllm.version import __version__ as VLLM_VERSION# Run headless workers (for multi-node PP/TP).host = parallel_config.master_addrhead_node_address = f"{host}:{parallel_config.master_port}"logger.info("Launching vLLM (v%s) headless multiproc executor, ""with head node address %s for torch.distributed process group.",VLLM_VERSION,head_node_address,)executor = MultiprocExecutor(vllm_config, monitor_workers=False)executor.start_worker_monitor(inline=True)return
7.实例化 CoreEngineProcManager 引擎进程管理器
专门负责创建、管理、监控所有本地 DP 推理 Worker 子进程。
host = parallel_config.data_parallel_master_ipport = parallel_config.data_parallel_rpc_porthandshake_address = get_tcp_uri(host, port)logger.info("Launching %d data parallel engine(s) in headless mode, ""with head node address %s.",local_engine_count,handshake_address,)# Create the engines.engine_manager = CoreEngineProcManager(local_engine_count=local_engine_count,start_index=vllm_config.parallel_config.data_parallel_rank,local_start_index=0,vllm_config=vllm_config,local_client=False,handshake_address=handshake_address,executor_class=Executor.get_class(vllm_config),log_stats=not engine_args.disable_log_stats,)
8.启动监听
try:engine_manager.monitor_engine_liveness()finally:timeout = Noneif shutdown_requested:timeout = vllm_config.shutdown_timeoutlogger.info("Waiting up to %d seconds for processes to exit", timeout)engine_manager.shutdown(timeout=timeout)logger.info("Shutting down.")
无限阻塞循环，主线程全程卡在这一行，是程序常驻运行的核心；持续维护父子进程 TCP 心跳（127.0.0.1:29550），实时监控所有 Worker 子进程存活状态；
本来想一次性写完 headless、multi_api_server、uvloop.run三种启动方式，奈何最近项目太忙，其他两种方式下期整理。
知识点
Python嵌套函数
def outer():# 外层函数x = 10# 内层函数，定义在 outer 里面def inner():print(x) # 可以访问外层变量# 调用内层函数inner()outer() # 输出 10
优点
封装：工具逻辑只在当前函数可见，命名不污染全局；
闭包可捕获上层变量，不用反复传参；
代码内聚，逻辑就近存放，可读性高。
缺点
多层嵌套可读性变差（不建议超过 2 层）；
调试栈会多一层，复杂嵌套排错稍麻烦。
nonlocal关键字
nonlocal 专门用在内层嵌套函数，代表： 变量不在当前函数局部作用域，也不是全局global，而是取自外层上层函数的局部变量。