[Feature] specify model in config.yaml (#14855)

Signed-off-by: weizeng <weizeng@roblox.com>
This commit is contained in:
Wei Zeng
2025-03-21 00:26:03 -07:00
committed by GitHub
parent da6ea29f7a
commit 0fa3970deb
7 changed files with 102 additions and 30 deletions
+13 -9
View File
@@ -21,14 +21,16 @@ class ServeSubcommand(CLISubcommand):
@staticmethod
def cmd(args: argparse.Namespace) -> None:
# The default value of `--model`
if args.model != EngineArgs.model:
raise ValueError(
"With `vllm serve`, you should provide the model as a "
"positional argument instead of via the `--model` option.")
# 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
# Otherwise use model from config (already in args.model)
# EngineArgs expects the model name to be passed as --model.
args.model = args.model_tag
# Check if we have a model specified somewhere
if args.model == EngineArgs.model: # Still has default value
raise ValueError(
"With `vllm serve`, you should provide the model either as a "
"positional argument or in config file.")
uvloop.run(run_server(args))
@@ -41,10 +43,12 @@ class ServeSubcommand(CLISubcommand):
serve_parser = subparsers.add_parser(
"serve",
help="Start the vLLM OpenAI Compatible API server",
usage="vllm serve <model_tag> [options]")
usage="vllm serve [model_tag] [options]")
serve_parser.add_argument("model_tag",
type=str,
help="The model tag to serve")
nargs='?',
help="The model tag to serve "
"(optional if specified in config)")
serve_parser.add_argument(
"--config",
type=str,
+25 -9
View File
@@ -1264,19 +1264,29 @@ class FlexibleArgumentParser(argparse.ArgumentParser):
config_args = self._load_config_file(file_path)
# 0th index is for {serve,chat,complete}
# followed by model_tag (only for serve)
# optionally followed by model_tag (only for serve)
# followed by config args
# followed by rest of cli args.
# maintaining this order will enforce the precedence
# of cli > config > defaults
if args[0] == "serve":
if index == 1:
model_in_cli = len(args) > 1 and not args[1].startswith('-')
model_in_config = any(arg == '--model' for arg in config_args)
if not model_in_cli and not model_in_config:
raise ValueError(
"No model_tag specified! Please check your command-line"
" arguments.")
args = [args[0]] + [
args[1]
] + config_args + args[2:index] + args[index + 2:]
"No model specified! Please specify model either in "
"command-line arguments or in config file.")
if model_in_cli:
# Model specified as positional arg, keep CLI version
args = [args[0]] + [
args[1]
] + config_args + args[2:index] + args[index + 2:]
else:
# No model in CLI, use config if available
args = [args[0]
] + config_args + args[1:index] + args[index + 2:]
else:
args = [args[0]] + config_args + args[1:index] + args[index + 2:]
@@ -1294,9 +1304,7 @@ class FlexibleArgumentParser(argparse.ArgumentParser):
'--port': '12323',
'--tensor-parallel-size': '4'
]
"""
extension: str = file_path.split('.')[-1]
if extension not in ('yaml', 'yml'):
raise ValueError(
@@ -1321,7 +1329,15 @@ class FlexibleArgumentParser(argparse.ArgumentParser):
if isinstance(action, StoreBoolean)
]
# Skip model from config if it's provided as positional argument
skip_model = (hasattr(self, '_parsed_args') and self._parsed_args
and len(self._parsed_args) > 1
and self._parsed_args[0] == 'serve'
and not self._parsed_args[1].startswith('-'))
for key, value in config.items():
if skip_model and key == 'model':
continue
if isinstance(value, bool) and key not in store_boolean_arguments:
if value:
processed_args.append('--' + key)