[TRTLLM-17267][test] Let the multi-frontend tests re-enact the composition root with the tokenize serving object

Both tests build a partial OpenAIServer and register its routes the way
the constructor does. The constructor now puts the tokenize serving object
on app.state and attaches the role's routers after the local table, so the
re-enactments do the same; the tokenize router refuses to attach without
its serving object.

Signed-off-by: JunyiXu-nv <219237550+JunyiXu-nv@users.noreply.github.com>

Author: JunyiXu-nv

tests/unittest/llmapi/test_serve_multi_frontend_helpers.py

@@ -37,6 +37,8 @@
 from tensorrt_llm.serve import _frontend_processes, openai_server
 from tensorrt_llm.serve._base.routing import ServerCapabilities, attach_routers
 from tensorrt_llm.serve._frontend_processes import _watch_attached_frontends
+from tensorrt_llm.serve.api.openai.tokenize import api_router as tokenize_router
+from tensorrt_llm.serve.api.openai.tokenize.serving import TokenizeServing
 from tensorrt_llm.serve.api.ops.system.serving import SystemServing
 from tensorrt_llm.serve.multi_frontend import (
     FORWARDED_ROUTES,
@@ -113,13 +115,19 @@ async def lifespan(_: FastAPI):
     server._enable_runtime_control_endpoints = False
     server.resource_governor = None
     server._launcher_forwarder = forwarder
-    # The order OpenAIServer.__init__ registers them in: the system router
-    # (/health, /version, /v1/models) first, then the launcher forwarding
-    # routes, then the local table.
+    # The order OpenAIServer.__init__ registers them in: the launcher forwarding
+    # routes, then the local table, then the routers of the role (the system
+    # router with /health, /version and /v1/models, and the tokenize router),
+    # which read their serving objects from app.state.
     server.app.state.ops_serving_system = SystemServing(model="m", health_probe=lambda: True)
-    attach_routers(server.app, ServerCapabilities.for_role(None))
+    setattr(
+        server.app.state,
+        tokenize_router.APP_STATE_KEY,
+        TokenizeServing(model="m", tokenizer=None),
+    )
     server._register_launcher_forwarding_routes()
     server.register_routes()
+    attach_routers(server.app, ServerCapabilities.for_role(None))
     return server
 
 

tests/unittest/llmapi/test_serve_multi_frontend_route_gate.py

@@ -28,6 +28,8 @@
 from test_common.serve.routes import effective_routes
 
 from tensorrt_llm.serve._base.routing import ServerCapabilities, attach_routers
+from tensorrt_llm.serve.api.openai.tokenize import api_router as tokenize_router
+from tensorrt_llm.serve.api.openai.tokenize.serving import TokenizeServing
 from tensorrt_llm.serve.api.ops.system.serving import SystemServing
 from tensorrt_llm.serve.multi_frontend import FORWARDED_ROUTES
 from tensorrt_llm.serve.openai_server import OpenAIServer
@@ -76,11 +78,18 @@ def _llm_server_routes() -> list[tuple[str, str]]:
     server._enable_rl_control_endpoints = True
     server._enable_runtime_control_endpoints = True
     server.resource_governor = None
-    # As OpenAIServer.__init__ does: the system router (/health, /version,
-    # /v1/models) is attached for every role before the role's own table.
+    # As OpenAIServer.__init__ does: the serving objects the routers read go on
+    # app.state, the role's own table is registered, then the routers of the role
+    # (the system router with /health, /version and /v1/models, and the tokenize
+    # router) are attached.
     server.app.state.ops_serving_system = SystemServing(model="m", health_probe=lambda: True)
-    attach_routers(server.app, ServerCapabilities.for_role(None))
+    setattr(
+        server.app.state,
+        tokenize_router.APP_STATE_KEY,
+        TokenizeServing(model="m", tokenizer=None),
+    )
     server.register_routes()
+    attach_routers(server.app, ServerCapabilities.for_role(None))
     # effective_routes walks included routers, which FastAPI 0.139 keeps as
     # objects without .path or .methods in app.routes.
     return sorted(