fix(python): add router arg fallback disable flag

bindings/python/src/smg/router_args.py

@@ -241,6 +241,17 @@ def add_cli_args(
             "Control Plane Authentication", "API key and JWT/OIDC authentication"
         )
 
+        if use_router_prefix:
+            parser.add_argument(
+                "--router-disable-arg-fallback",
+                action="store_true",
+                default=False,
+                help=(
+                    "When set, only use explicitly provided --router-* arguments and do not"
+                    " fall back to backend arguments with the same name."
+                ),
+            )
+
         # Worker configuration
         if not exclude_host_port:
             worker_group.add_argument(
@@ -1107,6 +1118,7 @@ def from_cli_args(cls, args: argparse.Namespace, use_router_prefix: bool = False
         prefix = "router_" if use_router_prefix else ""
         cli_args_dict = vars(args)
         args_dict = {}
+        disable_arg_fallback = bool(cli_args_dict.get(f"{prefix}disable_arg_fallback", False))
 
         for attr in dataclasses.fields(cls):
             # Auto strip prefix from args.
@@ -1117,7 +1129,11 @@ def from_cli_args(cls, args: argparse.Namespace, use_router_prefix: bool = False
             prefixed_key = f"{prefix}{attr.name}"
             if prefixed_key in cli_args_dict and cli_args_dict[prefixed_key] is not None:
                 args_dict[attr.name] = cli_args_dict[prefixed_key]
-            elif attr.name in cli_args_dict and cli_args_dict[attr.name] not in (None, ""):
+            elif (
+                not disable_arg_fallback
+                and attr.name in cli_args_dict
+                and cli_args_dict[attr.name] not in (None, "")
+            ):
                 args_dict[attr.name] = cli_args_dict[attr.name]
 
             # Special handling for CLI args with dashes vs dataclass fields with underscores

bindings/python/tests/test_arg_parser.py

@@ -406,6 +406,52 @@ def test_from_cli_args_without_prefix(self):
         assert router_args.policy == "random"
         assert router_args.pd_disaggregation is False
 
+    def test_prefixed_args_fall_back_to_backend_args_by_default(self):
+        """Prefixed router args should still fall back to backend args unless disabled."""
+        args = SimpleNamespace(
+            router_model_path=None,
+            router_disable_arg_fallback=False,
+            model_path="backend/model",
+            router_tokenizer_path=None,
+            tokenizer_path="backend/tokenizer",
+            router_worker_urls=[],
+            worker_urls=[],
+            router_prefill=None,
+            router_decode=None,
+            router_selector=None,
+            router_prefill_selector=None,
+            router_decode_selector=None,
+            router_router_selector=None,
+        )
+
+        router_args = RouterArgs.from_cli_args(args, use_router_prefix=True)
+
+        assert router_args.model_path == "backend/model"
+        assert router_args.tokenizer_path == "backend/tokenizer"
+
+    def test_prefixed_args_can_disable_backend_fallback(self):
+        """When router fallback is disabled, backend args should not fill router args."""
+        args = SimpleNamespace(
+            router_model_path=None,
+            router_disable_arg_fallback=True,
+            model_path="backend/model",
+            router_tokenizer_path=None,
+            tokenizer_path="backend/tokenizer",
+            router_worker_urls=[],
+            worker_urls=[],
+            router_prefill=None,
+            router_decode=None,
+            router_selector=None,
+            router_prefill_selector=None,
+            router_decode_selector=None,
+            router_router_selector=None,
+        )
+
+        router_args = RouterArgs.from_cli_args(args, use_router_prefix=True)
+
+        assert router_args.model_path is None
+        assert router_args.tokenizer_path is None
+
 
 class TestPolicyFromStr:
     """Test policy string to enum conversion."""

bindings/python/tests/test_serve.py

@@ -263,6 +263,17 @@ def test_trtllm_includes_router_args(self):
         )
         assert args.router_policy == "round_robin"
 
+    def test_trtllm_accepts_router_disable_arg_fallback_flag(self):
+        """--router-disable-arg-fallback should parse and be available on the namespace."""
+        _, args, _ = parse_serve_args(
+            [
+                "--backend",
+                "trtllm",
+                "--router-disable-arg-fallback",
+            ]
+        )
+        assert args.router_disable_arg_fallback is True
+
     def test_trtllm_router_args_defaults(self):
         """Router args should have sensible defaults."""
         _, args, _ = parse_serve_args(["--backend", "trtllm"])