Merge pull request #36518 from kunal2002/enforce-rpm-tpm-on-model-add

feat(proxy): enforce rpm/tpm on model add + fix validation error title in UI

litellm/proxy/management_endpoints/model_management_endpoints.py

@@ -252,6 +252,38 @@ def _raise_on_strategy_router_write_violation(
     )
 
 
+ENFORCE_RPM_TPM_ON_MODEL_ADD_SETTING: Final = "enforce_rpm_tpm_on_model_add"
+_REQUIRED_RATE_LIMIT_FIELDS: Final = ("rpm", "tpm")
+
+
+def _raise_if_rate_limits_required_but_missing(*, litellm_params: GenericLiteLLMParams, enforced: bool) -> None:
+    """Require both rpm and tpm (each a positive value) when the operator opts in via config.yaml.
+
+    Off by default, so deployments keep adding models without limits. When
+    ``enforce_rpm_tpm_on_model_add: true`` is set under general_settings, a model added
+    without both rpm and tpm set to a positive value is rejected rather than stored
+    unbounded (or effectively excluded from routing by a zero/negative limit).
+    """
+    if not enforced:
+        return
+    missing: Final = tuple(
+        field
+        for field in _REQUIRED_RATE_LIMIT_FIELDS
+        if (value := getattr(litellm_params, field)) is None or value <= 0
+    )
+    if not missing:
+        return
+    raise ProxyException(
+        message=(
+            f"{' and '.join(missing)} must be set to a positive value when "
+            f"'{ENFORCE_RPM_TPM_ON_MODEL_ADD_SETTING}' is enabled in general_settings"
+        ),
+        type=ProxyErrorTypes.validation_error.value,
+        code=status.HTTP_400_BAD_REQUEST,
+        param=f"litellm_params.{missing[0]}",
+    )
+
+
 _PTU_PRICED_PAIR: Final = frozenset({"ptu_count", "cost_per_ptu_per_hour"})
 
 
@@ -1748,6 +1780,11 @@ async def add_new_model(
             existing_params=None,
         )
 
+        _raise_if_rate_limits_required_but_missing(
+            litellm_params=model_params.litellm_params,
+            enforced=bool(general_settings.get(ENFORCE_RPM_TPM_ON_MODEL_ADD_SETTING, False)),
+        )
+
         model_response: prisma_models.LiteLLM_ProxyModelTable | LiteLLM_ProxyModelTable | None = None
         # update DB
         incoming_model_info: Final = model_params.model_info.model_dump(exclude_none=True)

tests/test_litellm/proxy/management_endpoints/test_model_management_endpoints.py

@@ -20,6 +20,7 @@ from litellm.proxy._types import (
 from litellm.proxy.management_endpoints.model_management_endpoints import (
     ModelManagementAuthChecks,
     _get_team_deployments,
+    _raise_if_rate_limits_required_but_missing,
     clear_cache,
     delete_team_models,
 )
@@ -4256,3 +4257,41 @@ class TestAutoRouterClassifierDefaultPrompt:
         for empty in (None, "", "{}"):
             response = await get_auto_router_classifier_default_prompt(context_window_size=5, tier_labels=empty)
             assert response.system_prompt == classification_system_prompt(5)
+
+
+class TestEnforceRpmTpmOnModelAdd:
+    def test_passes_when_disabled_even_without_limits(self):
+        assert (
+            _raise_if_rate_limits_required_but_missing(
+                litellm_params=LiteLLM_Params(model="azure/gpt-5.2"),
+                enforced=False,
+            )
+            is None
+        )
+
+    def test_passes_when_enabled_and_both_set(self):
+        assert (
+            _raise_if_rate_limits_required_but_missing(
+                litellm_params=LiteLLM_Params(model="azure/gpt-5.2", rpm=10, tpm=1000),
+                enforced=True,
+            )
+            is None
+        )
+
+    @pytest.mark.parametrize(
+        "params, expected_missing",
+        [
+            (LiteLLM_Params(model="azure/gpt-5.2"), "rpm and tpm"),
+            (LiteLLM_Params(model="azure/gpt-5.2", rpm=10), "tpm"),
+            (LiteLLM_Params(model="azure/gpt-5.2", tpm=1000), "rpm"),
+            (LiteLLM_Params(model="azure/gpt-5.2", rpm=0, tpm=1000), "rpm"),
+            (LiteLLM_Params(model="azure/gpt-5.2", rpm=10, tpm=-1), "tpm"),
+        ],
+    )
+    def test_raises_when_enabled_and_missing(self, params, expected_missing):
+        from litellm.proxy._types import ProxyException
+
+        with pytest.raises(ProxyException) as exc_info:
+            _raise_if_rate_limits_required_but_missing(litellm_params=params, enforced=True)
+        assert expected_missing in str(exc_info.value.message)
+        assert exc_info.value.code == "400"