vllm.entrypoints.openai.api_server ¶

@asynccontextmanager
async def build_async_engine_client_from_engine_args(
    engine_args: AsyncEngineArgs,
    *,
    usage_context: UsageContext = UsageContext.OPENAI_API_SERVER,
    disable_frontend_multiprocessing: bool = False,
    client_config: dict[str, Any] | None = None,
) -> AsyncIterator[EngineClient]:
    """
    Create EngineClient, either:
        - in-process using the AsyncLLMEngine Directly
        - multiprocess using AsyncLLMEngine RPC

    Returns the Client or None if the creation failed.
    """

    # Create the EngineConfig (determines if we can use V1).
    vllm_config = engine_args.create_engine_config(usage_context=usage_context)

    if disable_frontend_multiprocessing:
        logger.warning("V1 is enabled, but got --disable-frontend-multiprocessing.")

    from vllm.v1.engine.async_llm import AsyncLLM

    async_llm: AsyncLLM | None = None

    # Don't mutate the input client_config
    client_config = dict(client_config) if client_config else {}
    client_count = client_config.pop("client_count", 1)
    client_index = client_config.pop("client_index", 0)

    try:
        async_llm = AsyncLLM.from_vllm_config(
            vllm_config=vllm_config,
            usage_context=usage_context,
            enable_log_requests=engine_args.enable_log_requests,
            aggregate_engine_logging=engine_args.aggregate_engine_logging,
            disable_log_stats=engine_args.disable_log_stats,
            client_addresses=client_config,
            client_count=client_count,
            client_index=client_index,
        )

        # Don't keep the dummy data in memory
        assert async_llm is not None
        await async_llm.reset_mm_cache()

        yield async_llm
    finally:
        if async_llm:
            async_llm.shutdown()

async def run_server(args, **uvicorn_kwargs) -> None:
    """Run a single-worker API server."""

    # Add process-specific prefix to stdout and stderr.
    decorate_logs("APIServer")

    listen_address, sock = setup_server(args)
    await run_server_worker(listen_address, sock, args, **uvicorn_kwargs)

async def run_server_worker(
    listen_address, sock, args, client_config=None, **uvicorn_kwargs
) -> None:
    """Run a single API server worker."""

    if args.tool_parser_plugin and len(args.tool_parser_plugin) > 3:
        ToolParserManager.import_tool_parser(args.tool_parser_plugin)

    if args.reasoning_parser_plugin and len(args.reasoning_parser_plugin) > 3:
        ReasoningParserManager.import_reasoning_parser(args.reasoning_parser_plugin)

    # Get uvicorn log config (from file or with endpoint filter)
    log_config = get_uvicorn_log_config(args)
    if log_config is not None:
        uvicorn_kwargs["log_config"] = log_config

    async with build_async_engine_client(
        args,
        client_config=client_config,
    ) as engine_client:
        supported_tasks = await engine_client.get_supported_tasks()
        logger.info("Supported tasks: %s", supported_tasks)

        app = build_app(args, supported_tasks)
        await init_app_state(engine_client, app.state, args, supported_tasks)

        logger.info(
            "Starting vLLM API server %d on %s",
            engine_client.vllm_config.parallel_config._api_process_rank,
            listen_address,
        )
        shutdown_task = await serve_http(
            app,
            sock=sock,
            enable_ssl_refresh=args.enable_ssl_refresh,
            host=args.host,
            port=args.port,
            log_level=args.uvicorn_log_level,
            # NOTE: When the 'disable_uvicorn_access_log' value is True,
            # no access log will be output.
            access_log=not args.disable_uvicorn_access_log,
            timeout_keep_alive=envs.VLLM_HTTP_TIMEOUT_KEEP_ALIVE,
            ssl_keyfile=args.ssl_keyfile,
            ssl_certfile=args.ssl_certfile,
            ssl_ca_certs=args.ssl_ca_certs,
            ssl_cert_reqs=args.ssl_cert_reqs,
            ssl_ciphers=args.ssl_ciphers,
            h11_max_incomplete_event_size=args.h11_max_incomplete_event_size,
            h11_max_header_count=args.h11_max_header_count,
            **uvicorn_kwargs,
        )

    # NB: Await server shutdown only after the backend context is exited
    try:
        await shutdown_task
    finally:
        sock.close()

@instrument(span_name="API server setup")
def setup_server(args):
    """Validate API server args, set up signal handler, create socket
    ready to serve."""

    log_version_and_model(logger, VLLM_VERSION, args.model)
    log_non_default_args(args)

    if args.tool_parser_plugin and len(args.tool_parser_plugin) > 3:
        ToolParserManager.import_tool_parser(args.tool_parser_plugin)

    if args.reasoning_parser_plugin and len(args.reasoning_parser_plugin) > 3:
        ReasoningParserManager.import_reasoning_parser(args.reasoning_parser_plugin)

    validate_api_server_args(args)

    # workaround to make sure that we bind the port before the engine is set up.
    # This avoids race conditions with ray.
    # see https://github.com/vllm-project/vllm/issues/8204
    if args.uds:
        sock = create_server_unix_socket(args.uds)
    else:
        sock_addr = (args.host or "", args.port)
        sock = create_server_socket(sock_addr)

    # workaround to avoid footguns where uvicorn drops requests with too
    # many concurrent requests active
    set_ulimit()

    def signal_handler(*_) -> None:
        # Interrupt server on sigterm while initializing
        raise KeyboardInterrupt("terminated")

    signal.signal(signal.SIGTERM, signal_handler)

    if args.uds:
        listen_address = f"unix:{args.uds}"
    else:
        addr, port = sock_addr
        is_ssl = args.ssl_keyfile and args.ssl_certfile
        host_part = f"[{addr}]" if is_valid_ipv6_address(addr) else addr or "0.0.0.0"
        listen_address = f"http{'s' if is_ssl else ''}://{host_part}:{port}"
    return listen_address, sock