| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import argparse |
| 5 | import os |
| 6 | import sys |
| 7 | import typing |
| 8 | |
| 9 | from vllm import envs |
| 10 | from vllm.entrypoints.cli.benchmark.base import BenchmarkSubcommandBase |
| 11 | from vllm.entrypoints.cli.types import CLISubcommand |
| 12 | from vllm.entrypoints.serve.utils.api_utils import VLLM_SUBCMD_PARSER_EPILOG |
| 13 | from vllm.logger import init_logger |
| 14 | |
| 15 | if typing.TYPE_CHECKING: |
| 16 | from vllm.utils.argparse_utils import FlexibleArgumentParser |
| 17 | else: |
| 18 | FlexibleArgumentParser = argparse.ArgumentParser |
| 19 | |
| 20 | logger = init_logger(__name__) |
| 21 | |
| 22 | |
| 23 | def maybe_exec_rust_bench() -> None: |
| 24 | if sys.argv[1:3] != ["bench", "serve"] or not envs.VLLM_USE_RUST_BENCH: |
| 25 | return |
| 26 | |
| 27 | rust_cli = envs.VLLM_RUST_FRONTEND_PATH |
| 28 | if rust_cli is None: |
| 29 | raise RuntimeError( |
| 30 | "VLLM_USE_RUST_BENCH=1 requires VLLM_RUST_FRONTEND_PATH " |
| 31 | "to resolve to the vllm-rs binary." |
| 32 | ) |
| 33 | |
| 34 | logger.info("Delegating `vllm bench serve` to Rust binary at %s.", rust_cli) |
| 35 | os.execv(rust_cli, [rust_cli, "bench", "serve", *sys.argv[3:]]) |
| 36 | |
| 37 | |
| 38 | def _import_bench_subcommand_modules() -> None: |
| 39 | # Imported lazily so `BenchmarkSubcommandBase` subclasses register only |
| 40 | # when `vllm bench` is actually invoked. |
| 41 | import vllm.entrypoints.cli.benchmark.latency # noqa: F401 |
| 42 | import vllm.entrypoints.cli.benchmark.mm_processor # noqa: F401 |
| 43 | import vllm.entrypoints.cli.benchmark.serve # noqa: F401 |
| 44 | import vllm.entrypoints.cli.benchmark.startup # noqa: F401 |
| 45 | import vllm.entrypoints.cli.benchmark.sweep # noqa: F401 |
| 46 | import vllm.entrypoints.cli.benchmark.throughput # noqa: F401 |
| 47 | |
| 48 | |
| 49 | class BenchmarkSubcommand(CLISubcommand): |
| 50 | """The `bench` subcommand for the vLLM CLI.""" |
| 51 | |
| 52 | name = "bench" |
| 53 | help = "vLLM bench subcommand." |
| 54 | |
| 55 | @staticmethod |
| 56 | def cmd(args: argparse.Namespace) -> None: |
| 57 | args.dispatch_function(args) |
| 58 | |
| 59 | def validate(self, args: argparse.Namespace) -> None: |
| 60 | pass |
| 61 | |
| 62 | def subparser_init( |
| 63 | self, subparsers: argparse._SubParsersAction |
| 64 | ) -> FlexibleArgumentParser: |
| 65 | bench_parser = subparsers.add_parser( |
| 66 | self.name, |
| 67 | help=self.help, |
| 68 | description=self.help, |
| 69 | usage=f"vllm {self.name} <bench_type> [options]", |
| 70 | ) |
| 71 | bench_subparsers = bench_parser.add_subparsers(required=True, dest="bench_type") |
| 72 | |
| 73 | # Only build the nested bench subparsers when the user is actually |
| 74 | # invoking `bench`; otherwise we'd drag in imports |
| 75 | # unnecessarily on every `vllm --help` and `vllm serve`. |
| 76 | # Scan for the first positional arg so global flags (e.g. `-v`) |
| 77 | # before the subcommand don't break detection. |
| 78 | first_positional = next( |
| 79 | (arg for arg in sys.argv[1:] if not arg.startswith("-")), None |
| 80 | ) |
| 81 | if first_positional == self.name: |
| 82 | _import_bench_subcommand_modules() |
| 83 | for cmd_cls in BenchmarkSubcommandBase.__subclasses__(): |
| 84 | cmd_subparser = bench_subparsers.add_parser( |
| 85 | cmd_cls.name, |
| 86 | help=cmd_cls.help, |
| 87 | description=cmd_cls.help, |
| 88 | usage=f"vllm {self.name} {cmd_cls.name} [options]", |
| 89 | ) |
| 90 | cmd_subparser.set_defaults(dispatch_function=cmd_cls.cmd) |
| 91 | cmd_cls.add_cli_args(cmd_subparser) |
| 92 | cmd_subparser.epilog = VLLM_SUBCMD_PARSER_EPILOG.format( |
| 93 | subcmd=f"{self.name} {cmd_cls.name}" |
| 94 | ) |
| 95 | return bench_parser |
| 96 | |
| 97 | |
| 98 | def cmd_init() -> list[CLISubcommand]: |
| 99 | return [BenchmarkSubcommand()] |