Merge cd4b188c85 into b781d157d7

litellm/proxy/management_endpoints/model_management_endpoints.py

@@ -340,6 +340,84 @@ def _raise_on_strategy_router_write_violation(
     )
 
 
+def _validate_routable_model_deployment(
+    incoming_params: GenericLiteLLMParams | None,
+    existing_params: GenericLiteLLMParams | None = None,
+) -> str | None:
+    """Validate that incoming litellm_params specifies a routable LLM provider before DB insertion.
+
+    When adding or patching a model deployment, the router must be able to resolve its provider.
+    Unroutable models get dropped by the router upon reload, causing a 500 degraded serving error
+    and leaving dead zombie rows in LiteLLM_ProxyModelTable. Pre-validating the provider returns
+    a clean 400 Bad Request instead.
+    """
+    if incoming_params is None:
+        return None
+
+    model = getattr(incoming_params, "model", None)
+    if model is None and existing_params is not None:
+        model = getattr(existing_params, "model", None)
+
+    if not model or not isinstance(model, str):
+        return None
+
+    # Auto-router pseudo-models are validated separately
+    if model.startswith("auto_router/"):
+        return None
+
+    # Prompt management models resolve provider dynamically at runtime
+    if "/" in model:
+        split_model = model.split("/")[0]
+        if split_model in litellm._known_custom_logger_compatible_callbacks:
+            return None
+
+    custom_llm_provider = getattr(incoming_params, "custom_llm_provider", None) or (
+        getattr(existing_params, "custom_llm_provider", None) if existing_params else None
+    )
+    api_base = getattr(incoming_params, "api_base", None) or (
+        getattr(existing_params, "api_base", None) if existing_params else None
+    )
+
+    try:
+        from litellm.litellm_core_utils.get_llm_provider_logic import (
+            get_llm_provider,
+            is_registered_custom_provider,
+        )
+        from litellm.llms.openai_like.json_loader import JSONProviderRegistry
+
+        _model, resolved_provider, _dynamic_api_key, _api_base = get_llm_provider(
+            model=model,
+            custom_llm_provider=custom_llm_provider,
+            api_base=api_base,
+        )
+        if (
+            resolved_provider not in litellm.provider_list
+            and not JSONProviderRegistry.exists(resolved_provider)
+            and not is_registered_custom_provider(resolved_provider)
+        ):
+            return (
+                f"Model '{model}' is not a supported or routable provider: unsupported provider '{resolved_provider}'"
+            )
+    except Exception as e:
+        return f"Model '{model}' is not a routable provider: {e}"
+    return None
+
+
+def _raise_on_unroutable_model_deployment(
+    incoming_params: GenericLiteLLMParams | None,
+    existing_params: GenericLiteLLMParams | None = None,
+) -> None:
+    violation = _validate_routable_model_deployment(incoming_params=incoming_params, existing_params=existing_params)
+    if violation is None:
+        return
+    raise ProxyException(
+        message=violation,
+        type=ProxyErrorTypes.bad_request_error.value,
+        code=status.HTTP_400_BAD_REQUEST,
+        param="litellm_params.model",
+    )
+
+
 def _stored_credential_name(existing_litellm_params: GenericLiteLLMParams | None) -> str | None:
     if existing_litellm_params is None or existing_litellm_params.litellm_credential_name is None:
         return None
@@ -1219,6 +1297,10 @@ async def patch_model(
             incoming_params=patch_data.litellm_params,
             existing_params=db_model.litellm_params,
         )
+        _raise_on_unroutable_model_deployment(
+            incoming_params=patch_data.litellm_params,
+            existing_params=db_model.litellm_params,
+        )
 
         effective_params: Final = _effective_complexity_router_params(
             patch_data.litellm_params, db_model.litellm_params
@@ -2436,6 +2518,10 @@ async def add_new_model(
             incoming_params=model_params.litellm_params,
             existing_params=None,
         )
+        _raise_on_unroutable_model_deployment(
+            incoming_params=model_params.litellm_params,
+            existing_params=None,
+        )
 
         _raise_if_rate_limits_required_but_missing(
             litellm_params=model_params.litellm_params,