Merge remote-tracking branch 'origin/main' into litellm_team_member_temp_budget_increase

litellm/proxy/common_request_processing.py

@@ -1937,6 +1937,14 @@ class ProxyBaseLLMRequestProcessing:
     ) -> tuple[dict, LiteLLMLoggingObj]:
         start_time: Final = datetime.now()  # start before calling guardrail hooks
 
+        requested_model: Final = self.data.get("model")
+        if requested_model is not None and not isinstance(requested_model, str):
+            raise ProxyException(
+                message="'model' must be a string.",
+                type=ProxyErrorTypes.bad_request_error,
+                param="model",
+                code=status.HTTP_400_BAD_REQUEST,
+            )
         self.data = await add_litellm_data_to_request(
             data=self.data,
             request=request,

litellm/proxy/spend_tracking/spend_tracking_utils.py

@@ -485,10 +485,13 @@ def get_logging_payload(
         or None
     )
     custom_llm_provider: Final = logged_provider or _model_group_provider(_model_group, llm_router)
-    raw_model: Final = cast(str, kwargs.get("model") or "")
-    resolved_model: Final = (
-        standard_logging_payload.get("model") if standard_logging_payload is not None else None
-    ) or reconstruct_model_name(raw_model, logged_provider, metadata or {})
+    requested_model: Final = cast(object, kwargs.get("model"))
+    raw_model: Final = requested_model if isinstance(requested_model, str) else ""
+    model_is_malformed: Final = requested_model is not None and not isinstance(requested_model, str)
+    logged_model: Final = standard_logging_payload.get("model") if standard_logging_payload is not None else None
+    resolved_model: Final = (logged_model if isinstance(logged_model, str) else None) or reconstruct_model_name(
+        raw_model, logged_provider, metadata or {}
+    )
     failed_with_prompt_shaped_model: Final = (
         _get_status_for_spend_log(metadata=metadata) == "failure"
         and not _model_group
@@ -496,7 +499,7 @@ def get_logging_payload(
     )
     model_name: Final = (
         UNKNOWN_MODEL_SPEND_LOG_MODEL
-        if rejected_as_unknown_model or failed_with_prompt_shaped_model
+        if rejected_as_unknown_model or failed_with_prompt_shaped_model or model_is_malformed
         else resolved_model
     )
     litellm_call_id: Final = cast(

tests/test_litellm/proxy/spend_tracking/test_spend_tracking_utils.py

@@ -1049,6 +1049,27 @@ def test_get_logging_payload_replaces_rejected_or_prompt_shaped_models_with_the_
     assert payload["model"] == expected_model
 
 
+@pytest.mark.parametrize("requested_model", [{"bad": "value"}, ["gpt-5.2"], 1])
+def test_get_logging_payload_replaces_a_non_string_model_with_the_placeholder(
+    requested_model: dict[str, str] | list[str] | int,
+):
+    kwargs: Final = {
+        "model": requested_model,
+        "messages": [{"role": "user", "content": "hi"}],
+        "call_type": "acompletion",
+        "litellm_params": {"metadata": {"user_api_key": "sk-test", "status": "failure"}},
+    }
+
+    payload: Final = get_logging_payload(
+        kwargs=kwargs,
+        response_obj=ValueError("model must be a string"),
+        start_time=datetime.datetime.now(timezone.utc),
+        end_time=datetime.datetime.now(timezone.utc),
+    )
+
+    assert payload["model"] == UNKNOWN_MODEL_SPEND_LOG_MODEL
+
+
 @pytest.mark.parametrize(
     ("metadata", "response_obj"),
     [

tests/test_litellm/proxy/test_common_request_processing.py

@@ -327,6 +327,35 @@ class TestProxyBaseLLMRequestProcessing:
             pytest.fail("litellm_call_id is not a valid UUID")
         assert data_passed["litellm_call_id"] == returned_data["litellm_call_id"]
 
+    @pytest.mark.asyncio
+    @pytest.mark.parametrize("requested_model", [{"bad": "value"}, ["gpt-5.2"], 1])
+    async def test_common_processing_pre_call_logic_rejects_a_non_string_model_with_400(
+        self, monkeypatch, requested_model: dict[str, str] | list[str] | int
+    ):
+        processing_obj = ProxyBaseLLMRequestProcessing(
+            data={"model": requested_model, "messages": [{"role": "user", "content": "hi"}]}
+        )
+        mock_request = MagicMock(spec=Request)
+        mock_request.headers = {}
+        add_litellm_data_to_request = AsyncMock()
+        monkeypatch.setattr(
+            litellm.proxy.common_request_processing, "add_litellm_data_to_request", add_litellm_data_to_request
+        )
+
+        with pytest.raises(ProxyException) as exc_info:
+            await processing_obj.common_processing_pre_call_logic(
+                request=mock_request,
+                general_settings={},
+                user_api_key_dict=MagicMock(spec=UserAPIKeyAuth),
+                proxy_logging_obj=MagicMock(spec=ProxyLogging),
+                proxy_config=MagicMock(spec=ProxyConfig),
+                route_type="acompletion",
+            )
+
+        assert exc_info.value.code == str(status.HTTP_400_BAD_REQUEST)
+        assert exc_info.value.param == "model"
+        add_litellm_data_to_request.assert_not_awaited()
+
     @pytest.mark.asyncio
     async def test_common_processing_pre_call_logic_refreshes_proxy_server_request_body_after_guardrails(
         self, monkeypatch

tests/test_litellm/test_together_ai_model_metadata.py

@@ -95,7 +95,7 @@ def _successor(info: dict[str, object]) -> str | None:
     return successor if isinstance(successor, str) else None
 
 
-def test_together_successor_metadata_points_at_live_models(cost_map: CostMap):
+def test_together_successor_metadata_points_at_known_models(cost_map: CostMap):
     successors = {
         model: successor
         for model, info in cost_map.items()
@@ -103,9 +103,7 @@ def test_together_successor_metadata_points_at_live_models(cost_map: CostMap):
     }
     assert len(successors) >= 10
     for model, successor in successors.items():
-        target = cost_map.get(successor)
-        assert target is not None, f"{model} names successor {successor} that is not in the map"
-        assert "deprecation_date" not in target, f"{model} names deprecated successor {successor}"
+        assert successor in cost_map, f"{model} names successor {successor} that is not in the map"
 
 
 def test_together_backup_cost_map_in_sync(cost_map: CostMap):