Merge ed682aba7b into 22a349ee70

litellm/router.py

@@ -4856,6 +4856,16 @@ class Router:
             if custom_llm_provider is not None:
                 response_kwargs["custom_llm_provider"] = custom_llm_provider
 
+            # kwargs["session"] (e.g. realtime client_secrets) still has whatever
+            # model string the caller/proxy originally sent (a model_name/alias),
+            # not model_name resolved above. acreate_realtime_client_secret
+            # prioritizes session.model over the model kwarg, so leaving it
+            # unresolved would silently route with the alias instead of the
+            # deployment litellm picked. See BerriAI/litellm#36742.
+            session = response_kwargs.get("session")
+            if isinstance(session, dict) and session.get("model"):
+                response_kwargs["session"] = {**session, "model": model_name}
+
             response = original_generic_function(**response_kwargs)
 
             rpm_semaphore: Final = self._get_client(

tests/router_unit_tests/test_router_endpoints.py

@@ -520,6 +520,42 @@ def test_generic_api_call_with_fallbacks_basic(sync_mode):
     assert response == mock_response
 
 
+@pytest.mark.asyncio
+async def test_ageneric_api_call_with_fallbacks_resolves_session_model():
+    """
+    Regression test for https://github.com/BerriAI/litellm/issues/36742: when a
+    call has a nested session.model (realtime client_secrets), it must be
+    rewritten to the deployment litellm actually picked for the model_group,
+    not left as the caller's original alias.
+    """
+    mock_function = AsyncMock()
+    mock_function.__name__ = "acreate_realtime_client_secret"
+    mock_function.return_value = {"value": "ek_test"}
+
+    router = Router(
+        model_list=[
+            {
+                "model_name": "gpt-realtime-2-1-mini",
+                "litellm_params": {
+                    "model": "azure/gpt-realtime-2-1-mini-deployment",
+                    "api_key": "fake-api-key",
+                    "api_base": "https://fake.openai.azure.com",
+                },
+            }
+        ]
+    )
+
+    await router._ageneric_api_call_with_fallbacks(
+        model="gpt-realtime-2-1-mini",
+        original_function=mock_function,
+        session={"type": "realtime", "model": "gpt-realtime-2-1-mini"},
+    )
+
+    call_kwargs = mock_function.call_args.kwargs
+    assert call_kwargs["model"] == "azure/gpt-realtime-2-1-mini-deployment"
+    assert call_kwargs["session"]["model"] == "azure/gpt-realtime-2-1-mini-deployment"
+
+
 @pytest.mark.asyncio
 async def test_aadapter_completion():
     """