[Bugfix][Frontend] Run the serve arg checks for vllm launch too (#52825)
Signed-off-by: Vineeth Sai <vineethsai4444@gmail.com>
This commit is contained in:
@@ -214,6 +214,50 @@ def test_per_request_metrics_requires_log_stats(serve_parser):
|
||||
validate_parsed_serve_args(args)
|
||||
|
||||
|
||||
def _build_launch_render_parser():
|
||||
"""Mirror `vllm launch render`.
|
||||
|
||||
`vllm/entrypoints/cli/main.py` parses subcommands with ``dest="subparser"``,
|
||||
and `LaunchSubcommandBase.add_cli_args` builds the component parser with
|
||||
`make_arg_parser`, so a launch component carries serve args under
|
||||
``args.subparser == "launch"``.
|
||||
"""
|
||||
vllm_parser = FlexibleArgumentParser()
|
||||
subparsers = vllm_parser.add_subparsers(required=False, dest="subparser")
|
||||
launch_parser = subparsers.add_parser("launch")
|
||||
launch_subparsers = launch_parser.add_subparsers(
|
||||
required=True, dest="launch_component"
|
||||
)
|
||||
render_parser = launch_subparsers.add_parser("render")
|
||||
make_arg_parser(render_parser)
|
||||
return vllm_parser
|
||||
|
||||
|
||||
@pytest.fixture
|
||||
def launch_render_parser():
|
||||
return _build_launch_render_parser()
|
||||
|
||||
|
||||
def test_launch_render_validates_serve_args(launch_render_parser):
|
||||
"""`vllm launch render` reuses the serve parser, so it gets the serve checks"""
|
||||
args = launch_render_parser.parse_args(
|
||||
args=["launch", "render", "--enable-auto-tool-choice"]
|
||||
)
|
||||
assert args.subparser == "launch"
|
||||
with pytest.raises(TypeError):
|
||||
validate_parsed_serve_args(args)
|
||||
|
||||
|
||||
def test_subcommand_without_serve_args_skips_validation():
|
||||
"""A subcommand that does not use the serve parser is still skipped"""
|
||||
vllm_parser = FlexibleArgumentParser()
|
||||
subparsers = vllm_parser.add_subparsers(required=False, dest="subparser")
|
||||
subparsers.add_parser("chat")
|
||||
args = vllm_parser.parse_args(args=["chat"])
|
||||
|
||||
validate_parsed_serve_args(args)
|
||||
|
||||
|
||||
@pytest.mark.parametrize(
|
||||
"cli_args, expected_middleware",
|
||||
[
|
||||
|
||||
@@ -412,7 +412,9 @@ def make_arg_parser(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
|
||||
|
||||
def validate_parsed_serve_args(args: argparse.Namespace):
|
||||
"""Quick checks for model serve args that raise prior to loading."""
|
||||
if hasattr(args, "subparser") and args.subparser != "serve":
|
||||
# `vllm launch <component>` builds its parser with make_arg_parser too (see
|
||||
# LaunchSubcommandBase.add_cli_args), so its args are serve args as well.
|
||||
if hasattr(args, "subparser") and args.subparser not in ("serve", "launch"):
|
||||
return
|
||||
|
||||
# Ensure that the chat template is valid; raises if it likely isn't
|
||||
|
||||
Reference in New Issue
Block a user