init

entrypoints/cli/benchmark/base.py

@@ -0,0 +1,39 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import argparse
+
+from vllm.entrypoints.cli.types import CLISubcommand
+from vllm.utils import FlexibleArgumentParser
+
+
+class BenchmarkSubcommandBase(CLISubcommand):
+    """ The base class of subcommands for vllm bench. """
+
+    @property
+    def help(self) -> str:
+        """The help message of the subcommand."""
+        raise NotImplementedError
+
+    def add_cli_args(self, parser: argparse.ArgumentParser) -> None:
+        """Add the CLI arguments to the parser."""
+        raise NotImplementedError
+
+    @staticmethod
+    def cmd(args: argparse.Namespace) -> None:
+        """Run the benchmark.
+
+        Args:
+            args: The arguments to the command.
+        """
+        raise NotImplementedError
+
+    def subparser_init(
+            self,
+            subparsers: argparse._SubParsersAction) -> FlexibleArgumentParser:
+        parser = subparsers.add_parser(
+            self.name,
+            help=self.help,
+            description=self.help,
+            usage=f"vllm bench {self.name} [options]")
+        self.add_cli_args(parser)
+        return parser

entrypoints/cli/benchmark/latency.py

@@ -0,0 +1,30 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import argparse
+
+from vllm.benchmarks.latency import add_cli_args, main
+from vllm.entrypoints.cli.benchmark.base import BenchmarkSubcommandBase
+from vllm.entrypoints.cli.types import CLISubcommand
+
+
+class BenchmarkLatencySubcommand(BenchmarkSubcommandBase):
+    """ The `latency` subcommand for vllm bench. """
+
+    def __init__(self):
+        self.name = "latency"
+        super().__init__()
+
+    @property
+    def help(self) -> str:
+        return "Benchmark the latency of a single batch of requests."
+
+    def add_cli_args(self, parser: argparse.ArgumentParser) -> None:
+        add_cli_args(parser)
+
+    @staticmethod
+    def cmd(args: argparse.Namespace) -> None:
+        main(args)
+
+
+def cmd_init() -> list[CLISubcommand]:
+    return [BenchmarkLatencySubcommand()]

entrypoints/cli/benchmark/main.py

@@ -0,0 +1,54 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import argparse
+
+import vllm.entrypoints.cli.benchmark.latency
+import vllm.entrypoints.cli.benchmark.serve
+import vllm.entrypoints.cli.benchmark.throughput
+from vllm.entrypoints.cli.types import CLISubcommand
+from vllm.utils import FlexibleArgumentParser
+
+BENCHMARK_CMD_MODULES = [
+    vllm.entrypoints.cli.benchmark.latency,
+    vllm.entrypoints.cli.benchmark.serve,
+    vllm.entrypoints.cli.benchmark.throughput,
+]
+
+
+class BenchmarkSubcommand(CLISubcommand):
+    """ The `bench` subcommand for the vLLM CLI. """
+
+    def __init__(self):
+        self.name = "bench"
+        super().__init__()
+
+    @staticmethod
+    def cmd(args: argparse.Namespace) -> None:
+        args.dispatch_function(args)
+
+    def validate(self, args: argparse.Namespace) -> None:
+        if args.bench_type in self.cmds:
+            self.cmds[args.bench_type].validate(args)
+
+    def subparser_init(
+            self,
+            subparsers: argparse._SubParsersAction) -> FlexibleArgumentParser:
+        bench_parser = subparsers.add_parser(
+            "bench",
+            help="vLLM bench subcommand.",
+            description="vLLM bench subcommand.",
+            usage="vllm bench <bench_type> [options]")
+        bench_subparsers = bench_parser.add_subparsers(required=True,
+                                                       dest="bench_type")
+        self.cmds = {}
+        for cmd_module in BENCHMARK_CMD_MODULES:
+            new_cmds = cmd_module.cmd_init()
+            for cmd in new_cmds:
+                cmd.subparser_init(bench_subparsers).set_defaults(
+                    dispatch_function=cmd.cmd)
+                self.cmds[cmd.name] = cmd
+        return bench_parser
+
+
+def cmd_init() -> list[CLISubcommand]:
+    return [BenchmarkSubcommand()]

entrypoints/cli/benchmark/serve.py

@@ -0,0 +1,30 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import argparse
+
+from vllm.benchmarks.serve import add_cli_args, main
+from vllm.entrypoints.cli.benchmark.base import BenchmarkSubcommandBase
+from vllm.entrypoints.cli.types import CLISubcommand
+
+
+class BenchmarkServingSubcommand(BenchmarkSubcommandBase):
+    """ The `serve` subcommand for vllm bench. """
+
+    def __init__(self):
+        self.name = "serve"
+        super().__init__()
+
+    @property
+    def help(self) -> str:
+        return "Benchmark the online serving throughput."
+
+    def add_cli_args(self, parser: argparse.ArgumentParser) -> None:
+        add_cli_args(parser)
+
+    @staticmethod
+    def cmd(args: argparse.Namespace) -> None:
+        main(args)
+
+
+def cmd_init() -> list[CLISubcommand]:
+    return [BenchmarkServingSubcommand()]

entrypoints/cli/benchmark/throughput.py

@@ -0,0 +1,30 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import argparse
+
+from vllm.benchmarks.throughput import add_cli_args, main
+from vllm.entrypoints.cli.benchmark.base import BenchmarkSubcommandBase
+from vllm.entrypoints.cli.types import CLISubcommand
+
+
+class BenchmarkThroughputSubcommand(BenchmarkSubcommandBase):
+    """ The `throughput` subcommand for vllm bench. """
+
+    def __init__(self):
+        self.name = "throughput"
+        super().__init__()
+
+    @property
+    def help(self) -> str:
+        return "Benchmark offline inference throughput."
+
+    def add_cli_args(self, parser: argparse.ArgumentParser) -> None:
+        add_cli_args(parser)
+
+    @staticmethod
+    def cmd(args: argparse.Namespace) -> None:
+        main(args)
+
+
+def cmd_init() -> list[CLISubcommand]:
+    return [BenchmarkThroughputSubcommand()]

entrypoints/cli/collect_env.py

@@ -0,0 +1,35 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import argparse
+
+from vllm.collect_env import main as collect_env_main
+from vllm.entrypoints.cli.types import CLISubcommand
+from vllm.utils import FlexibleArgumentParser
+
+
+class CollectEnvSubcommand(CLISubcommand):
+    """The `collect-env` subcommand for the vLLM CLI. """
+
+    def __init__(self):
+        self.name = "collect-env"
+        super().__init__()
+
+    @staticmethod
+    def cmd(args: argparse.Namespace) -> None:
+        """Collect information about the environment."""
+        collect_env_main()
+
+    def subparser_init(
+            self,
+            subparsers: argparse._SubParsersAction) -> FlexibleArgumentParser:
+        collect_env_parser = subparsers.add_parser(
+            "collect-env",
+            help="Start collecting environment information.",
+            description="Start collecting environment information.",
+            usage="vllm collect-env")
+        return collect_env_parser
+
+
+def cmd_init() -> list[CLISubcommand]:
+    return [CollectEnvSubcommand()]

entrypoints/cli/main.py

@@ -0,0 +1,65 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+# The CLI entrypoint to vLLM.
+import signal
+import sys
+
+import vllm.entrypoints.cli.benchmark.main
+import vllm.entrypoints.cli.collect_env
+import vllm.entrypoints.cli.openai
+import vllm.entrypoints.cli.run_batch
+import vllm.entrypoints.cli.serve
+import vllm.version
+from vllm.entrypoints.utils import VLLM_SUBCMD_PARSER_EPILOG, cli_env_setup
+from vllm.utils import FlexibleArgumentParser
+
+CMD_MODULES = [
+    vllm.entrypoints.cli.openai,
+    vllm.entrypoints.cli.serve,
+    vllm.entrypoints.cli.benchmark.main,
+    vllm.entrypoints.cli.collect_env,
+    vllm.entrypoints.cli.run_batch,
+]
+
+
+def register_signal_handlers():
+
+    def signal_handler(sig, frame):
+        sys.exit(0)
+
+    signal.signal(signal.SIGINT, signal_handler)
+    signal.signal(signal.SIGTSTP, signal_handler)
+
+
+def main():
+    cli_env_setup()
+
+    parser = FlexibleArgumentParser(
+        description="vLLM CLI",
+        epilog=VLLM_SUBCMD_PARSER_EPILOG,
+    )
+    parser.add_argument('-v',
+                        '--version',
+                        action='version',
+                        version=vllm.version.__version__)
+    subparsers = parser.add_subparsers(required=False, dest="subparser")
+    cmds = {}
+    for cmd_module in CMD_MODULES:
+        new_cmds = cmd_module.cmd_init()
+        for cmd in new_cmds:
+            cmd.subparser_init(subparsers).set_defaults(
+                dispatch_function=cmd.cmd)
+            cmds[cmd.name] = cmd
+    args = parser.parse_args()
+    if args.subparser in cmds:
+        cmds[args.subparser].validate(args)
+
+    if hasattr(args, "dispatch_function"):
+        args.dispatch_function(args)
+    else:
+        parser.print_help()
+
+
+if __name__ == "__main__":
+    main()

entrypoints/cli/openai.py

@@ -0,0 +1,205 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# Commands that act as an interactive OpenAI API client
+
+import argparse
+import os
+import signal
+import sys
+from typing import Optional
+
+from openai import OpenAI
+from openai.types.chat import ChatCompletionMessageParam
+
+from vllm.entrypoints.cli.types import CLISubcommand
+from vllm.utils import FlexibleArgumentParser
+
+
+def _register_signal_handlers():
+
+    def signal_handler(sig, frame):
+        sys.exit(0)
+
+    signal.signal(signal.SIGINT, signal_handler)
+    signal.signal(signal.SIGTSTP, signal_handler)
+
+
+def _interactive_cli(args: argparse.Namespace) -> tuple[str, OpenAI]:
+    _register_signal_handlers()
+
+    base_url = args.url
+    api_key = args.api_key or os.environ.get("OPENAI_API_KEY", "EMPTY")
+    openai_client = OpenAI(api_key=api_key, base_url=base_url)
+
+    if args.model_name:
+        model_name = args.model_name
+    else:
+        available_models = openai_client.models.list()
+        model_name = available_models.data[0].id
+
+    print(f"Using model: {model_name}")
+
+    return model_name, openai_client
+
+
+def chat(system_prompt: Optional[str], model_name: str,
+         client: OpenAI) -> None:
+    conversation: list[ChatCompletionMessageParam] = []
+    if system_prompt is not None:
+        conversation.append({"role": "system", "content": system_prompt})
+
+    print("Please enter a message for the chat model:")
+    while True:
+        try:
+            input_message = input("> ")
+        except EOFError:
+            return
+        conversation.append({"role": "user", "content": input_message})
+
+        chat_completion = client.chat.completions.create(model=model_name,
+                                                         messages=conversation)
+
+        response_message = chat_completion.choices[0].message
+        output = response_message.content
+
+        conversation.append(response_message)  # type: ignore
+        print(output)
+
+
+def _add_query_options(
+        parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
+    parser.add_argument(
+        "--url",
+        type=str,
+        default="http://localhost:8000/v1",
+        help="url of the running OpenAI-Compatible RESTful API server")
+    parser.add_argument(
+        "--model-name",
+        type=str,
+        default=None,
+        help=("The model name used in prompt completion, default to "
+              "the first model in list models API call."))
+    parser.add_argument(
+        "--api-key",
+        type=str,
+        default=None,
+        help=(
+            "API key for OpenAI services. If provided, this api key "
+            "will overwrite the api key obtained through environment variables."
+        ))
+    return parser
+
+
+class ChatCommand(CLISubcommand):
+    """The `chat` subcommand for the vLLM CLI. """
+
+    def __init__(self):
+        self.name = "chat"
+        super().__init__()
+
+    @staticmethod
+    def cmd(args: argparse.Namespace) -> None:
+        model_name, client = _interactive_cli(args)
+        system_prompt = args.system_prompt
+        conversation: list[ChatCompletionMessageParam] = []
+
+        if system_prompt is not None:
+            conversation.append({"role": "system", "content": system_prompt})
+
+        if args.quick:
+            conversation.append({"role": "user", "content": args.quick})
+
+            chat_completion = client.chat.completions.create(
+                model=model_name, messages=conversation)
+            print(chat_completion.choices[0].message.content)
+            return
+
+        print("Please enter a message for the chat model:")
+        while True:
+            try:
+                input_message = input("> ")
+            except EOFError:
+                return
+            conversation.append({"role": "user", "content": input_message})
+
+            chat_completion = client.chat.completions.create(
+                model=model_name, messages=conversation)
+
+            response_message = chat_completion.choices[0].message
+            output = response_message.content
+
+            conversation.append(response_message)  # type: ignore
+            print(output)
+
+    def subparser_init(
+            self,
+            subparsers: argparse._SubParsersAction) -> FlexibleArgumentParser:
+        chat_parser = subparsers.add_parser(
+            "chat",
+            help="Generate chat completions via the running API server.",
+            description="Generate chat completions via the running API server.",
+            usage="vllm chat [options]")
+        _add_query_options(chat_parser)
+        chat_parser.add_argument(
+            "--system-prompt",
+            type=str,
+            default=None,
+            help=("The system prompt to be added to the chat template, "
+                  "used for models that support system prompts."))
+        chat_parser.add_argument("-q",
+                                 "--quick",
+                                 type=str,
+                                 metavar="MESSAGE",
+                                 help=("Send a single prompt as MESSAGE "
+                                       "and print the response, then exit."))
+        return chat_parser
+
+
+class CompleteCommand(CLISubcommand):
+    """The `complete` subcommand for the vLLM CLI. """
+
+    def __init__(self):
+        self.name = "complete"
+        super().__init__()
+
+    @staticmethod
+    def cmd(args: argparse.Namespace) -> None:
+        model_name, client = _interactive_cli(args)
+
+        if args.quick:
+            completion = client.completions.create(model=model_name,
+                                                   prompt=args.quick)
+            print(completion.choices[0].text)
+            return
+
+        print("Please enter prompt to complete:")
+        while True:
+            input_prompt = input("> ")
+            completion = client.completions.create(model=model_name,
+                                                   prompt=input_prompt)
+            output = completion.choices[0].text
+            print(output)
+
+    def subparser_init(
+            self,
+            subparsers: argparse._SubParsersAction) -> FlexibleArgumentParser:
+        complete_parser = subparsers.add_parser(
+            "complete",
+            help=("Generate text completions based on the given prompt "
+                  "via the running API server."),
+            description=("Generate text completions based on the given prompt "
+                         "via the running API server."),
+            usage="vllm complete [options]")
+        _add_query_options(complete_parser)
+        complete_parser.add_argument(
+            "-q",
+            "--quick",
+            type=str,
+            metavar="PROMPT",
+            help=
+            "Send a single prompt and print the completion output, then exit.")
+        return complete_parser
+
+
+def cmd_init() -> list[CLISubcommand]:
+    return [ChatCommand(), CompleteCommand()]

entrypoints/cli/run_batch.py

@@ -0,0 +1,62 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import argparse
+import asyncio
+
+from prometheus_client import start_http_server
+
+from vllm.entrypoints.cli.types import CLISubcommand
+from vllm.entrypoints.logger import logger
+from vllm.entrypoints.openai.run_batch import main as run_batch_main
+from vllm.entrypoints.openai.run_batch import make_arg_parser
+from vllm.entrypoints.utils import (VLLM_SUBCMD_PARSER_EPILOG,
+                                    show_filtered_argument_or_group_from_help)
+from vllm.utils import FlexibleArgumentParser
+from vllm.version import __version__ as VLLM_VERSION
+
+
+class RunBatchSubcommand(CLISubcommand):
+    """The `run-batch` subcommand for vLLM CLI."""
+
+    def __init__(self):
+        self.name = "run-batch"
+        super().__init__()
+
+    @staticmethod
+    def cmd(args: argparse.Namespace) -> None:
+        logger.info("vLLM batch processing API version %s", VLLM_VERSION)
+        logger.info("args: %s", args)
+
+        # Start the Prometheus metrics server.
+        # LLMEngine uses the Prometheus client
+        # to publish metrics at the /metrics endpoint.
+        if args.enable_metrics:
+            logger.info("Prometheus metrics enabled")
+            start_http_server(port=args.port, addr=args.url)
+        else:
+            logger.info("Prometheus metrics disabled")
+
+        asyncio.run(run_batch_main(args))
+
+    def subparser_init(
+            self,
+            subparsers: argparse._SubParsersAction) -> FlexibleArgumentParser:
+        run_batch_parser = subparsers.add_parser(
+            "run-batch",
+            help="Run batch prompts and write results to file.",
+            description=(
+                "Run batch prompts using vLLM's OpenAI-compatible API.\n"
+                "Supports local or HTTP input/output files."),
+            usage=
+            "vllm run-batch -i INPUT.jsonl -o OUTPUT.jsonl --model <model>",
+        )
+        run_batch_parser = make_arg_parser(run_batch_parser)
+        show_filtered_argument_or_group_from_help(run_batch_parser,
+                                                  "run-batch")
+        run_batch_parser.epilog = VLLM_SUBCMD_PARSER_EPILOG
+        return run_batch_parser
+
+
+def cmd_init() -> list[CLISubcommand]:
+    return [RunBatchSubcommand()]

entrypoints/cli/serve.py

@@ -0,0 +1,328 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import argparse
+import os
+import signal
+import sys
+
+import uvloop
+import zmq
+
+import vllm.envs as envs
+from vllm import AsyncEngineArgs
+from vllm.entrypoints.cli.types import CLISubcommand
+from vllm.entrypoints.openai.api_server import (run_server, run_server_worker,
+                                                setup_server)
+from vllm.entrypoints.openai.cli_args import (make_arg_parser,
+                                              validate_parsed_serve_args)
+from vllm.entrypoints.utils import (VLLM_SUBCMD_PARSER_EPILOG,
+                                    show_filtered_argument_or_group_from_help)
+from vllm.executor.multiproc_worker_utils import _add_prefix
+from vllm.logger import init_logger
+from vllm.usage.usage_lib import UsageContext
+from vllm.utils import FlexibleArgumentParser, get_tcp_uri, zmq_socket_ctx
+from vllm.v1.engine.coordinator import DPCoordinator
+from vllm.v1.engine.core import EngineCoreProc
+from vllm.v1.engine.core_client import CoreEngineProcManager
+from vllm.v1.executor.abstract import Executor
+from vllm.v1.metrics.prometheus import setup_multiprocess_prometheus
+from vllm.v1.utils import (APIServerProcessManager, CoreEngine,
+                           CoreEngineActorManager, EngineZmqAddresses,
+                           get_engine_client_zmq_addr,
+                           wait_for_completion_or_failure,
+                           wait_for_engine_startup)
+
+logger = init_logger(__name__)
+
+
+class ServeSubcommand(CLISubcommand):
+    """The `serve` subcommand for the vLLM CLI. """
+
+    def __init__(self):
+        self.name = "serve"
+        super().__init__()
+
+    @staticmethod
+    def cmd(args: argparse.Namespace) -> None:
+        # If model is specified in CLI (as positional arg), it takes precedence
+        if hasattr(args, 'model_tag') and args.model_tag is not None:
+            args.model = args.model_tag
+
+        if args.headless or args.api_server_count < 1:
+            run_headless(args)
+        elif args.api_server_count > 1:
+            run_multi_api_server(args)
+        else:
+            # Single API server (this process).
+            uvloop.run(run_server(args))
+
+    def validate(self, args: argparse.Namespace) -> None:
+        validate_parsed_serve_args(args)
+
+    def subparser_init(
+            self,
+            subparsers: argparse._SubParsersAction) -> FlexibleArgumentParser:
+        serve_parser = subparsers.add_parser(
+            "serve",
+            help="Start the vLLM OpenAI Compatible API server.",
+            description="Start the vLLM OpenAI Compatible API server.",
+            usage="vllm serve [model_tag] [options]")
+        serve_parser.add_argument("model_tag",
+                                  type=str,
+                                  nargs='?',
+                                  help="The model tag to serve "
+                                  "(optional if specified in config)")
+        serve_parser.add_argument(
+            "--headless",
+            action='store_true',
+            default=False,
+            help="Run in headless mode. See multi-node data parallel "
+            "documentation for more details.")
+        serve_parser.add_argument(
+            '--data-parallel-start-rank',
+            '-dpr',
+            type=int,
+            default=0,
+            help='Starting data parallel rank for secondary nodes.')
+        serve_parser.add_argument('--api-server-count',
+                                  '-asc',
+                                  type=int,
+                                  default=1,
+                                  help='How many API server processes to run.')
+        serve_parser.add_argument(
+            "--config",
+            type=str,
+            default='',
+            required=False,
+            help="Read CLI options from a config file."
+            "Must be a YAML with the following options:"
+            "https://docs.vllm.ai/en/latest/serving/openai_compatible_server.html#cli-reference"
+        )
+
+        serve_parser = make_arg_parser(serve_parser)
+        show_filtered_argument_or_group_from_help(serve_parser, "serve")
+        serve_parser.epilog = VLLM_SUBCMD_PARSER_EPILOG
+        return serve_parser
+
+
+def cmd_init() -> list[CLISubcommand]:
+    return [ServeSubcommand()]
+
+
+def run_headless(args: argparse.Namespace):
+
+    if args.api_server_count > 1:
+        raise ValueError("api_server_count can't be set in headless mode")
+
+    # Create the EngineConfig.
+    engine_args = AsyncEngineArgs.from_cli_args(args)
+    usage_context = UsageContext.OPENAI_API_SERVER
+    vllm_config = engine_args.create_engine_config(usage_context=usage_context)
+
+    if not envs.VLLM_USE_V1:
+        raise ValueError("Headless mode is only supported for V1")
+
+    parallel_config = vllm_config.parallel_config
+    local_engine_count = parallel_config.data_parallel_size_local
+    host = parallel_config.data_parallel_master_ip
+    port = engine_args.data_parallel_rpc_port  # add to config too
+    handshake_address = get_tcp_uri(host, port)
+
+    if local_engine_count <= 0:
+        raise ValueError("data_parallel_size_local must be > 0 in "
+                         "headless mode")
+
+    # Catch SIGTERM and SIGINT to allow graceful shutdown.
+    def signal_handler(signum, frame):
+        logger.debug("Received %d signal.", signum)
+        raise SystemExit
+
+    signal.signal(signal.SIGTERM, signal_handler)
+    signal.signal(signal.SIGINT, signal_handler)
+
+    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(
+        target_fn=EngineCoreProc.run_engine_core,
+        local_engine_count=local_engine_count,
+        start_index=args.data_parallel_start_rank,
+        local_start_index=0,
+        vllm_config=vllm_config,
+        on_head_node=False,
+        handshake_address=handshake_address,
+        executor_class=Executor.get_class(vllm_config),
+        log_stats=not engine_args.disable_log_stats,
+    )
+
+    try:
+        engine_manager.join_first()
+    finally:
+        logger.info("Shutting down.")
+        engine_manager.close()
+
+
+def run_multi_api_server(args: argparse.Namespace):
+
+    assert not args.headless
+    num_api_servers = args.api_server_count
+    assert num_api_servers > 0
+
+    if num_api_servers > 1:
+        setup_multiprocess_prometheus()
+
+    listen_address, sock = setup_server(args)
+
+    engine_args = AsyncEngineArgs.from_cli_args(args)
+    usage_context = UsageContext.OPENAI_API_SERVER
+    vllm_config = engine_args.create_engine_config(usage_context=usage_context)
+    model_config = vllm_config.model_config
+
+    if num_api_servers > 1:
+        if not envs.VLLM_USE_V1:
+            raise ValueError("api_server_count > 1 is only supported for V1")
+
+        if envs.VLLM_ALLOW_RUNTIME_LORA_UPDATING:
+            raise ValueError("VLLM_ALLOW_RUNTIME_LORA_UPDATING cannot be used "
+                             "with api_server_count > 1")
+
+        if model_config.is_multimodal_model and not (
+                model_config.disable_mm_preprocessor_cache):
+            logger.warning(
+                "Multi-model preprocessor cache will be disabled for"
+                " api_server_count > 1")
+            model_config.disable_mm_preprocessor_cache = True
+
+    parallel_config = vllm_config.parallel_config
+
+    assert parallel_config.data_parallel_rank == 0
+
+    dp_size = parallel_config.data_parallel_size
+    local_engine_count = parallel_config.data_parallel_size_local
+    host = parallel_config.data_parallel_master_ip
+    local_only = local_engine_count == dp_size
+
+    # Set up input and output addresses.
+    input_addresses = [
+        get_engine_client_zmq_addr(local_only, host)
+        for _ in range(num_api_servers)
+    ]
+    output_addresses = [
+        get_engine_client_zmq_addr(local_only, host)
+        for _ in range(num_api_servers)
+    ]
+
+    addresses = EngineZmqAddresses(
+        inputs=input_addresses,
+        outputs=output_addresses,
+    )
+
+    # Set up coordinator for dp > 1.
+    coordinator = None
+    stats_update_address = None
+    if dp_size > 1:
+        coordinator = DPCoordinator(parallel_config)
+        addresses.coordinator_input, addresses.coordinator_output = (
+            coordinator.get_engine_socket_addresses())
+        stats_update_address = coordinator.get_stats_publish_address()
+        logger.info("Started DP Coordinator process (PID: %d)",
+                    coordinator.proc.pid)
+
+    if parallel_config.data_parallel_backend == "ray":
+        logger.info("Starting ray-based data parallel backend")
+
+        engine_actor_manager = CoreEngineActorManager(
+            vllm_config=vllm_config,
+            addresses=addresses,
+            executor_class=Executor.get_class(vllm_config),
+            log_stats=not engine_args.disable_log_stats,
+        )
+        # Start API servers using the manager
+        api_server_manager = APIServerProcessManager(
+            target_server_fn=run_api_server_worker_proc,
+            listen_address=listen_address,
+            sock=sock,
+            args=args,
+            num_servers=num_api_servers,
+            input_addresses=input_addresses,
+            output_addresses=output_addresses,
+            stats_update_address=stats_update_address)
+
+        wait_for_completion_or_failure(api_server_manager=api_server_manager,
+                                       engine_manager=engine_actor_manager,
+                                       coordinator=coordinator)
+        return
+
+    handshake_address = get_engine_client_zmq_addr(
+        local_only, host, parallel_config.data_parallel_rpc_port)
+
+    with zmq_socket_ctx(handshake_address, zmq.ROUTER,
+                        bind=True) as handshake_socket:
+
+        # Start local engines.
+        if not local_engine_count:
+            local_engine_manager = None
+        else:
+            local_engine_manager = CoreEngineProcManager(
+                EngineCoreProc.run_engine_core,
+                vllm_config=vllm_config,
+                executor_class=Executor.get_class(vllm_config),
+                log_stats=not engine_args.disable_log_stats,
+                handshake_address=handshake_address,
+                on_head_node=True,
+                local_engine_count=local_engine_count,
+                start_index=0,
+                local_start_index=0)
+
+        # Start API servers using the manager
+        api_server_manager = APIServerProcessManager(
+            target_server_fn=run_api_server_worker_proc,
+            listen_address=listen_address,
+            sock=sock,
+            args=args,
+            num_servers=num_api_servers,
+            input_addresses=input_addresses,
+            output_addresses=output_addresses,
+            stats_update_address=stats_update_address)
+
+        # Wait for engine handshakes to complete.
+        core_engines = [
+            CoreEngine(index=i, local=(i < local_engine_count))
+            for i in range(dp_size)
+        ]
+        wait_for_engine_startup(
+            handshake_socket,
+            addresses,
+            core_engines,
+            parallel_config,
+            vllm_config.cache_config,
+            local_engine_manager,
+            coordinator.proc if coordinator else None,
+        )
+
+        # Wait for API servers
+        wait_for_completion_or_failure(api_server_manager=api_server_manager,
+                                       engine_manager=local_engine_manager,
+                                       coordinator=coordinator)
+
+
+def run_api_server_worker_proc(listen_address,
+                               sock,
+                               args,
+                               client_config=None,
+                               **uvicorn_kwargs) -> None:
+    """Entrypoint for individual API server worker processes."""
+
+    # Add process-specific prefix to stdout and stderr.
+    from multiprocessing import current_process
+    process_name = current_process().name
+    pid = os.getpid()
+    _add_prefix(sys.stdout, process_name, pid)
+    _add_prefix(sys.stderr, process_name, pid)
+
+    uvloop.run(
+        run_server_worker(listen_address, sock, args, client_config,
+                          **uvicorn_kwargs))

entrypoints/cli/types.py

@@ -0,0 +1,25 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import argparse
+
+from vllm.utils import FlexibleArgumentParser
+
+
+class CLISubcommand:
+    """Base class for CLI argument handlers."""
+
+    name: str
+
+    @staticmethod
+    def cmd(args: argparse.Namespace) -> None:
+        raise NotImplementedError("Subclasses should implement this method")
+
+    def validate(self, args: argparse.Namespace) -> None:
+        # No validation by default
+        pass
+
+    def subparser_init(
+            self,
+            subparsers: argparse._SubParsersAction) -> FlexibleArgumentParser:
+        raise NotImplementedError("Subclasses should implement this method")