vLLM/vllm/entrypoints/launchers/api_server/entry.py

# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import asyncio
import multiprocessing
import multiprocessing.forkserver as forkserver
import os
import signal
import socket
import tempfile
from argparse import Namespace
from collections.abc import AsyncIterator
from contextlib import asynccontextmanager
from typing import Any

import vllm.envs as envs
from vllm.engine.arg_utils import AsyncEngineArgs
from vllm.engine.protocol import EngineClient
from vllm.logger import init_logger
from vllm.reasoning import ReasoningParserManager
from vllm.tool_parsers import ToolParserManager
from vllm.usage.usage_lib import UsageContext
from vllm.utils.system_utils import decorate_logs

from ..app import build_app
from ..launcher import serve_http, setup_server
from ..utils.server_utils import get_uvicorn_log_config
from .app_state import init_app_state

prometheus_multiproc_dir: tempfile.TemporaryDirectory

# Cannot use __name__ (https://github.com/vllm-project/vllm/pull/4765)
logger = init_logger("vllm.entrypoints.launchers.api_server.entry")


@asynccontextmanager
async def build_async_engine_client(
    args: Namespace,
    *,
    usage_context: UsageContext = UsageContext.OPENAI_API_SERVER,
    client_config: dict[str, Any] | None = None,
) -> AsyncIterator[EngineClient]:
    if os.getenv("VLLM_WORKER_MULTIPROC_METHOD") == "forkserver":
        # The executor is expected to be mp.
        # Pre-import heavy modules in the forkserver process
        logger.debug("Setup forkserver with pre-imports")
        multiprocessing.set_start_method("forkserver")
        multiprocessing.set_forkserver_preload(["vllm.v1.engine.async_llm"])
        forkserver.ensure_running()
        logger.debug("Forkserver setup complete!")

    # Context manager to handle engine_client lifecycle
    # Ensures everything is shutdown and cleaned up on error/exit
    engine_args = AsyncEngineArgs.from_cli_args(args)
    if client_config:
        engine_args._api_process_count = client_config.get("client_count", 1)
        engine_args._api_process_rank = client_config.get("client_index", 0)

    async with build_async_engine_client_from_engine_args(
        engine_args,
        usage_context=usage_context,
        client_config=client_config,
    ) as engine:
        yield engine


@asynccontextmanager
async def build_async_engine_client_from_engine_args(
    engine_args: AsyncEngineArgs,
    *,
    usage_context: UsageContext = UsageContext.OPENAI_API_SERVER,
    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)

    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(timeout=vllm_config.shutdown_timeout)


async def build_and_serve(
    engine_client: EngineClient,
    listen_address: str,
    sock: socket.socket,
    args: Namespace,
    **uvicorn_kwargs,
) -> asyncio.Task:
    """Build FastAPI app, initialize state, and start serving.

    Returns the shutdown task for the caller to await.
    """

    # 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

    supported_tasks = await engine_client.get_supported_tasks()
    model_config = engine_client.model_config

    logger.info("Supported tasks: %s", supported_tasks)
    app = build_app(args, supported_tasks, model_config)
    await init_app_state(engine_client, app.state, args, supported_tasks)

    logger.info("Starting vLLM server on %s", listen_address)

    return 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,
    )


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

    decorate_logs("APIServer", skip_if_decorated=True)

    # Interrupt initialization if SIGTERM arrives before uvicorn installs its
    # own signal handlers. Once uvicorn is running it replaces this.
    def _interrupt_init(*_) -> None:
        raise KeyboardInterrupt("terminated")

    signal.signal(signal.SIGTERM, _interrupt_init)

    listen_address, sock = setup_server(args, reuse_port=False)
    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)

    async with build_async_engine_client(
        args,
        client_config=client_config,
    ) as engine_client:
        shutdown_task = await build_and_serve(
            engine_client, listen_address, sock, args, **uvicorn_kwargs
        )
    # NB: Await server shutdown only after the backend context is exited
    try:
        await shutdown_task
    finally:
        sock.close()


def main():
    import uvloop

    from vllm.entrypoints.serve.utils.api_utils import cli_env_setup
    from vllm.utils.argparse_utils import FlexibleArgumentParser

    from ..cli_args import (
        make_arg_parser,
        validate_parsed_serve_args,
    )

    # NOTE(simon):
    # This section should be in sync with vllm/entrypoints/cli/main.py for CLI
    # entrypoints.
    cli_env_setup()
    parser = FlexibleArgumentParser(
        description="vLLM OpenAI-Compatible RESTful API server."
    )
    parser = make_arg_parser(parser)
    args = parser.parse_args()
    validate_parsed_serve_args(args)

    uvloop.run(run_server(args))


if __name__ == "__main__":
    main()