fix(frontend): Fix smg serve reject vLLM OpenAI args

bindings/python/src/smg/serve.py

@@ -185,7 +185,9 @@ def build_command(
         ):
             cmd.append("--enable-prompt-tokens-details")
 
-        cmd.extend(self._filter_backend_args(backend_args, ["--model", "--host", "--port"]))
+        cmd.extend(
+            self._filter_backend_args(backend_args, ["--model", "--host", "--port", "--uds"])
+        )
 
         return cmd
 
@@ -384,15 +386,37 @@ def _add_sglang_args(parser: argparse.ArgumentParser) -> None:
 
 
 def _add_vllm_args(parser: argparse.ArgumentParser) -> None:
-    """Add vllm-specific arguments."""
+    """Add vLLM engine CLI arguments."""
     try:
-        from vllm.engine.arg_utils import EngineArgs
+        from vllm.engine import arg_utils as vllm_arg_utils
+
+        engine_args_cls = getattr(vllm_arg_utils, "AsyncEngineArgs", None) or getattr(
+            vllm_arg_utils, "EngineArgs", None
+        )
+        if engine_args_cls is None:
+            raise ImportError("vllm.engine.arg_utils is missing EngineArgs")
 
-        EngineArgs.add_cli_args(parser)
+        engine_args_cls.add_cli_args(parser)
     except ImportError:
         parser.error("vllm is not installed. Install it with: pip install vllm")
 
 
+def _add_vllm_frontend_args(parser: argparse.ArgumentParser) -> None:
+    """Add vLLM OpenAI frontend arguments."""
+    try:
+        from vllm.entrypoints.openai import cli_args as vllm_openai_cli_args
+    except ImportError:
+        vllm_openai_cli_args = None
+
+    frontend_args_cls = (
+        getattr(vllm_openai_cli_args, "FrontendArgs", None)
+        if vllm_openai_cli_args is not None
+        else None
+    )
+    if frontend_args_cls is not None:
+        frontend_args_cls.add_cli_args(parser)
+
+
 def _add_trtllm_stub_args(parser: argparse.ArgumentParser) -> None:
     """Add TensorRT-LLM specific arguments.
 
@@ -524,12 +548,16 @@ def parse_serve_args(
     )
     add_serve_args(parser)
     _import_backend_args(backend, parser)
+    if backend == "vllm" and serve_router_args.connection_mode == "http":
+        logger.debug("Adding vLLM OpenAI frontend CLI arguments for HTTP workers.")
+        _add_vllm_frontend_args(parser)
     RouterArgs.add_cli_args(parser, use_router_prefix=True, exclude_host_port=True)
 
     if backend == "trtllm":
         args, _ = parser.parse_known_args(argv)
     else:
         args = parser.parse_args(argv)
+
     return backend, args, backend_args
 
 

bindings/python/tests/test_serve.py

@@ -286,6 +286,73 @@ def test_vllm_explicit_exits(self):
             parse_serve_args(["--backend", "vllm"])
         assert exc_info.value.code == 2
 
+    def test_vllm_http_accepts_frontend_args(self):
+        def _mock_vllm_args(backend, parser):
+            if backend == "vllm":
+                parser.add_argument("--model", type=str)
+            else:
+                _import_backend_args(backend, parser)
+
+        def _mock_vllm_frontend_args(parser):
+            parser.add_argument("--enable-auto-tool-choice", action="store_true")
+            parser.add_argument("--tool-call-parser", type=str)
+
+        with (
+            patch("smg.serve._import_backend_args", side_effect=_mock_vllm_args),
+            patch("smg.serve._add_vllm_frontend_args", side_effect=_mock_vllm_frontend_args),
+        ):
+            backend, args, backend_args = parse_serve_args(
+                [
+                    "--backend",
+                    "vllm",
+                    "--connection-mode",
+                    "http",
+                    "--model",
+                    "/tmp/model",
+                    "--enable-auto-tool-choice",
+                    "--tool-call-parser",
+                    "minimax_m2",
+                ]
+            )
+
+        assert backend == "vllm"
+        assert args.connection_mode == "http"
+        assert args.model == "/tmp/model"
+        assert args.enable_auto_tool_choice is True
+        assert args.tool_call_parser == "minimax_m2"
+        assert "--enable-auto-tool-choice" in backend_args
+        assert "--tool-call-parser" in backend_args
+        assert "minimax_m2" in backend_args
+
+    def test_vllm_grpc_rejects_frontend_args(self):
+        def _mock_vllm_args(backend, parser):
+            if backend == "vllm":
+                parser.add_argument("--model", type=str)
+            else:
+                _import_backend_args(backend, parser)
+
+        with (
+            patch("smg.serve._import_backend_args", side_effect=_mock_vllm_args),
+            patch("smg.serve._add_vllm_frontend_args") as mock_frontend_args,
+        ):
+            with pytest.raises(SystemExit) as exc_info:
+                parse_serve_args(
+                    [
+                        "--backend",
+                        "vllm",
+                        "--connection-mode",
+                        "grpc",
+                        "--model",
+                        "/tmp/model",
+                        "--enable-auto-tool-choice",
+                        "--tool-call-parser",
+                        "minimax_m2",
+                    ]
+                )
+
+        assert exc_info.value.code == 2
+        mock_frontend_args.assert_not_called()
+
     def test_invalid_backend_exits(self):
         with pytest.raises(SystemExit):
             parse_serve_args(["--backend", "nonexistent"])