fix(router): track routers with empty model_list in _live_routers

litellm/router.py

@@ -1034,9 +1034,6 @@ class Router:
         if model_list is not None:
             # set_model_list will build indices automatically
             self.set_model_list(model_list)
-            # Track this router so a price data reload can rebuild its deployments'
-            # cost-map entries from the list it is serving at that moment.
-            _live_routers.add(self)
             self.healthy_deployments: list = self.model_list
             for m in model_list:
                 if "model" in m["litellm_params"]:
@@ -1044,6 +1041,10 @@ class Router:
         else:
             self.model_list: list = []  # initialize an empty list - to allow _add_deployment and delete_deployment to work
 
+        # Track every live router so price-data reload can rebuild cost-map entries
+        # even when the proxy started with an empty / omitted model_list.
+        _live_routers.add(self)
+
         if allowed_fails is not None:
             self.allowed_fails = allowed_fails
         else:

tests/unit/test_router_model_cost_isolation.py

@@ -1865,6 +1865,55 @@ def test_reload_keeps_custom_pricing_configured_on_litellm_params_for_a_db_model
         _invalidate_model_cost_lowercase_map()
 
 
+@pytest.mark.timeout(15)
+def test_reload_keeps_cost_map_when_router_starts_with_no_model_list():
+    """
+    An empty proxy config builds Router() with model_list omitted. Deployments
+    added afterwards (DB-stored models) must stay in the cost map across a
+    price data reload. The model is a local provider so add_deployment does
+    not start a remote login.
+    """
+    from litellm.router import _live_routers
+
+    saved_catalog = litellm.model_cost
+    fetched_catalog = copy.deepcopy(litellm.model_cost)
+    model_key = "hosted_vllm/empty-init-priced"
+    deployment_id = "empty-init-priced-id"
+    input_cost = 2e-6
+    output_cost = 1e-5
+    try:
+        router = Router()
+        assert router in _live_routers
+        router.add_deployment(
+            deployment=Deployment(
+                model_name=model_key,
+                litellm_params=LiteLLM_Params(
+                    model=model_key,
+                    api_key="sk-fake",
+                    input_cost_per_token=input_cost,
+                    output_cost_per_token=output_cost,
+                ),
+                model_info=ModelInfo(id=deployment_id),
+            )
+        )
+
+        assert model_key in litellm.model_cost
+        assert deployment_id in litellm.model_cost
+        assert litellm.model_cost[deployment_id]["input_cost_per_token"] == input_cost
+        assert litellm.model_cost[deployment_id]["output_cost_per_token"] == output_cost
+
+        _simulate_price_data_reload(copy.deepcopy(fetched_catalog))
+
+        assert model_key in litellm.model_cost
+        assert deployment_id in litellm.model_cost
+        assert litellm.model_cost[deployment_id]["input_cost_per_token"] == input_cost
+        assert litellm.model_cost[deployment_id]["output_cost_per_token"] == output_cost
+        assert router.model_list
+    finally:
+        litellm.model_cost = saved_catalog
+        _invalidate_model_cost_lowercase_map()
+
+
 def test_replay_live_router_model_cost_rebuilds_every_live_router():
     """
     A process can hold more than one Router, so the rebuild has to fan out across