fix(proxy): honor explicit null to clear cache_control_injection_points on model update

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

litellm/proxy/management_endpoints/model_management_endpoints.py

@@ -129,6 +129,7 @@ if TYPE_CHECKING:
     from prisma import models as prisma_models
 
 router: Final = APIRouter()
+CLEARABLE_LITELLM_PARAMS: Final = frozenset({"cache_control_injection_points"})
 
 
 async def update_team(*args, **kwargs):
@@ -758,11 +759,16 @@ def update_db_model(db_model: Deployment, updated_patch: updateDeployment) -> Pr
     # model_info fields like team_id or access groups. SPECIAL_MODEL_INFO_PARAMS are
     # mirrored between litellm_params and model_info by Deployment.__init__, so the
     # clear propagates to both blobs.
+    # CLEARABLE_LITELLM_PARAMS lists non-mirrored litellm_params that explicit null may clear.
     if updated_patch.litellm_params:
         for field in updated_patch.litellm_params.model_fields_set:
-            if field in SPECIAL_MODEL_INFO_PARAMS and getattr(updated_patch.litellm_params, field) is None:
+            if getattr(updated_patch.litellm_params, field) is not None:
+                continue
+            if field in SPECIAL_MODEL_INFO_PARAMS:
                 merged_litellm_params.pop(field, None)
                 merged_model_info.pop(field, None)
+            elif field in CLEARABLE_LITELLM_PARAMS:
+                merged_litellm_params.pop(field, None)
     if updated_patch.model_info:
         for field in updated_patch.model_info.model_fields_set:
             if field in SPECIAL_MODEL_INFO_PARAMS and getattr(updated_patch.model_info, field) is None:

tests/test_litellm/proxy/management_endpoints/test_model_management_endpoints.py

@@ -3573,6 +3573,54 @@ class TestUpdateDBModelClearPricing:
         assert info["cache_creation_input_token_cost"] == 0.000003
 
 
+class TestUpdateDBModelClearCacheControlInjectionPoints:
+    def test_explicit_null_removes_stored_injection_points(self):
+        from litellm.proxy.management_endpoints.model_management_endpoints import (
+            update_db_model,
+        )
+        from litellm.types.router import LiteLLM_Params, ModelInfo, updateLiteLLMParams
+
+        db_model = Deployment(
+            model_name="haiku-cached",
+            litellm_params=LiteLLM_Params(
+                model="anthropic/claude-haiku-4-5",
+                cache_control_injection_points=[{"location": "message", "role": "system"}],
+            ),
+            model_info=ModelInfo(id="dep-cache-0"),
+        )
+        patch = updateDeployment(
+            litellm_params=updateLiteLLMParams(cache_control_injection_points=None)
+        )
+
+        result = update_db_model(db_model=db_model, updated_patch=patch)
+
+        params = json.loads(result["litellm_params"])
+        assert "cache_control_injection_points" not in params
+        assert params["model"] == "anthropic/claude-haiku-4-5"
+
+    def test_omitted_key_keeps_stored_injection_points(self):
+        from litellm.proxy.management_endpoints.model_management_endpoints import (
+            update_db_model,
+        )
+        from litellm.types.router import LiteLLM_Params, ModelInfo, updateLiteLLMParams
+
+        db_model = Deployment(
+            model_name="haiku-cached",
+            litellm_params=LiteLLM_Params(
+                model="anthropic/claude-haiku-4-5",
+                cache_control_injection_points=[{"location": "message", "role": "system"}],
+            ),
+            model_info=ModelInfo(id="dep-cache-0"),
+        )
+        patch = updateDeployment(litellm_params=updateLiteLLMParams(tpm=10))
+
+        result = update_db_model(db_model=db_model, updated_patch=patch)
+
+        params = json.loads(result["litellm_params"])
+        assert params["cache_control_injection_points"] == [{"location": "message", "role": "system"}]
+        assert params["tpm"] == 10
+
+
 class TestGetModelInfoWithIdBlocked:
     """`ProxyConfig.get_model_info_with_id` must propagate the DB-level `blocked`
     column into the in-memory `model_info` dict so the router filter can read it."""