源码精读：sglang serve → run_server → HTTP → 子进程拉起

def serve(args, extra_argv):
    backend_name, argv = _extract_model_type_override(extra_argv)
    argv, positional = _normalize_positional_model_path(argv)
    request = ServeRequest(argv=tuple(argv), ...)
    registry = _create_backend_registry()
    load_plugins()
    try:
        registered = registry.auto_detect(request)         # LLM / diffusion / 插件
        registered.backend.run(request)
    finally:
        kill_process_tree(os.getpid(), include_parent=False)

def run_server(server_args):
    server_args.resolve_once()
    cfg = resolving_view(server_args)
    if cfg.encoder_only:
        if cfg.smg_grpc_mode or cfg.grpc_mode:
            from sglang.srt.disaggregation.encoder.grpc_server import serve_grpc_encoder
            asyncio.run(serve_grpc_encoder(server_args))
        else:
            from sglang.srt.disaggregation.encoder.http_server import launch_server
            launch_server(server_args)
    elif cfg.smg_grpc_mode:
        # 旧 SMG gRPC-only 路径
        from sglang.srt.entrypoints.grpc_server import serve_grpc
        asyncio.run(serve_grpc(server_args))
    elif cfg.use_ray:
        # Ray 模式
        from sglang.srt.ray.http_server import launch_server
        launch_server(server_args)
    else:
        # 默认 HTTP 模式（90% 用户走这条路）
        from sglang.srt.entrypoints.http_server import launch_server
        launch_server(server_args)

def launch_server(
    server_args: ServerArgs,
    init_tokenizer_manager_func: Callable = init_tokenizer_manager,
    run_scheduler_process_func: Callable = run_scheduler_process,
    run_detokenizer_process_func: Callable = run_detokenizer_process,
    execute_warmup_func: Callable = _execute_server_warmup,
    launch_callback: Optional[Callable[[], None]] = None,
):
    """Launch SRT Server.
    ...
    The engine consists of three components:
        1. TokenizerManager: Tokenizes requests, sends to scheduler.
        2. Scheduler (subprocess): Receives, schedules, forwards.
        3. DetokenizerManager (subprocess): Detokenizes outputs.
    """
    # Launch subprocesses
    (tokenizer_manager, template_manager, port_args,
     scheduler_init_result, subprocess_watchdog) = Engine._launch_subprocesses(...)

    _setup_and_run_http_server(
        server_args, tokenizer_manager, template_manager, port_args,
        scheduler_init_result.scheduler_infos, subprocess_watchdog, ...
    )

def _launch_subprocesses(server_args, ...):
    port_args = PortArgs.init_new(server_args)              # 算所有端口
    tokenizer_manager = init_tokenizer_manager_func(server_args, port_args)

    # 起 Scheduler 子进程（每个 TP rank 一个）
    scheduler_procs = []
    for tp_rank in range(server_args.tp):
        proc = multiprocessing.Process(
            target=run_scheduler_process_func,
            args=(server_args, port_args, tp_rank, ...)
        )
        proc.start()
        scheduler_procs.append(proc)

    # 起 Detokenizer 子进程
    detokenizer_proc = multiprocessing.Process(
        target=run_detokenizer_process_func,
        args=(server_args, port_args)
    )
    detokenizer_proc.start()

    # 等子进程 ready
    scheduler_init_result = wait_for_scheduler_init(...)

    # Watchdog 监控子进程存活
    watchdog = SubprocessWatchdog(scheduler_procs + [detokenizer_proc])

    return tokenizer_manager, template_manager, port_args, scheduler_init_result, watchdog

def _setup_and_run_http_server(server_args, tokenizer_manager, ...):
    # 1. 建 FastAPI app
    app = build_fastapi_app(server_args, tokenizer_manager, ...)

    # 2. 注册路由：/v1/chat/completions, /v1/completions, /v1/embeddings, ...
    register_openai_routes(app, tokenizer_manager, ...)
    register_admin_routes(app, ...)
    register_metrics_routes(app, ...)

    # 3. 启动 (Granian / Uvicorn)
    if server_args.http_server == "granian":
        _run_granian_server(server_args)
    else:
        uvicorn.run(app, host=server_args.host, port=server_args.port, ...)

[INFO] server_args=ServerArgs(...)                        ← server_args.py 解析后的 dataclass repr
[INFO] Init torch distributed begin.                      ← TP 集合通信初始化
[INFO] Init torch distributed ends. mem usage=...
[INFO] Load weight begin. ...                             ← model_loader 加载权重
[INFO] Load weight end. Type: float16, ...
[INFO] Memory pool end. avail mem=XX.X GB                 ← KV pool 预分配后剩余
[INFO] max_total_num_tokens=...                           ← KV pool 总 slot 数
[INFO] Use cuda graph for decoding mode.
[INFO] Capture cuda graph begin. ...                      ← 在录 graph，可能花 10-60s
[INFO] Capture cuda graph end. Time elapsed: X.X s.
[INFO] max_running_requests=..., max_total_num_tokens=...
[INFO] Application startup complete.                      ← FastAPI ready
[INFO] The server is fired up and ready to roll!          ← 可以接请求了

from sglang.srt.entrypoints.http_server import launch_server
from sglang.srt.managers.scheduler import run_scheduler_process

def my_scheduler(server_args, port_args, tp_rank, ...):
    # 自定义调度逻辑，最后还是要 run_scheduler_process
    print("Custom scheduler wrapper")
    return run_scheduler_process(server_args, port_args, tp_rank, ...)

launch_server(server_args, run_scheduler_process_func=my_scheduler)