[router] allow one router to support different model families and serving mode (#10244)

sgl-router/py_test/e2e/conftest.py

@@ -128,6 +128,7 @@ def _popen_launch_router_only(
     timeout: float = 120.0,
     *,
     dp_aware: bool = False,
+    enable_igw: bool = False,
     api_key: str | None = None,
 ) -> subprocess.Popen:
     host, port = _parse_url(base_url)
@@ -146,6 +147,8 @@ def _popen_launch_router_only(
     ]
     if dp_aware:
         cmd += ["--dp-aware"]
+    if enable_igw:
+        cmd += ["--enable-igw"]
     if api_key is not None:
         cmd += ["--api-key", api_key]
     cmd += [

sgl-router/py_test/integration/test_retries.py

@@ -35,7 +35,7 @@ def test_retry_reroutes_to_healthy_worker(router_manager, mock_workers):
     )
     assert r.status_code == 200
     wid = r.headers.get("X-Worker-Id") or r.json().get("worker_id")
-    assert wid == id_b  # should have retried onto healthy worker
+    assert wid in [id_b, id_c]  # should have retried onto a healthy worker (B or C)
     # mock_workers fixture handles cleanup