[Frontend] Disallow passing model as both argument and option (vllm…

…-project#7347)

vllm/engine/arg_utils.py

@@ -32,7 +32,7 @@ def nullable_str(val: str):
 @dataclass
 class EngineArgs:
     """Arguments for vLLM engine."""
-    model: str
+    model: str = 'facebook/opt-125m'
     served_model_name: Optional[Union[List[str]]] = None
     tokenizer: Optional[str] = None
     skip_tokenizer_init: bool = False
@@ -133,7 +133,7 @@ def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
         parser.add_argument(
             '--model',
             type=str,
-            default='facebook/opt-125m',
+            default=EngineArgs.model,
             help='Name or path of the huggingface model to use.')
         parser.add_argument(
             '--tokenizer',

vllm/scripts.py

@@ -9,6 +9,7 @@
 from openai import OpenAI
 from openai.types.chat import ChatCompletionMessageParam
 
+from vllm.engine.arg_utils import EngineArgs
 from vllm.entrypoints.openai.api_server import run_server
 from vllm.entrypoints.openai.cli_args import make_arg_parser
 from vllm.utils import FlexibleArgumentParser
@@ -24,6 +25,12 @@ def signal_handler(sig, frame):
 
 
 def serve(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.")
+
     # EngineArgs expects the model name to be passed as --model.
     args.model = args.model_tag