Merge pull request #40632 from BerriAI/litellm_fix_ui_disable_cache_control_injection_points

fix(ui): persist disabling cache control injection points on model update

litellm/proxy/management_endpoints/model_management_endpoints.py

@@ -144,6 +144,8 @@ if TYPE_CHECKING:
     from prisma import types as prisma_types
 
 router: Final = APIRouter()
+CLEARABLE_LITELLM_PARAMS: Final = frozenset({"cache_control_injection_points"})
+NULL_CLEARABLE_LITELLM_PARAMS: Final = frozenset((*SPECIAL_MODEL_INFO_PARAMS, *CLEARABLE_LITELLM_PARAMS))
 
 
 async def update_team(*args, **kwargs):
@@ -898,7 +900,7 @@ def update_db_model(db_model: Deployment, updated_patch: updateDeployment) -> Pr
     # clear propagates to both blobs.
     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 None and field in NULL_CLEARABLE_LITELLM_PARAMS:
                 merged_litellm_params.pop(field, None)
                 merged_model_info.pop(field, None)
             elif (

tests/test_litellm/proxy/management_endpoints/test_model_management_endpoints.py

@@ -3864,6 +3864,54 @@ class TestModelInfoServerDerivedPricingFilter:
         assert written["access_groups"] == ["prod"]
 
 
+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."""

ui/litellm-dashboard/src/components/model_info_view.test.tsx

@@ -1785,7 +1785,7 @@ describe("ModelInfoView", () => {
         expect(payload.litellm_params.cache_control_injection_points).toEqual([{ location: "message", role: "user" }]);
       });
 
-      it("drops the stored injection points when the operator turns the toggle off", async () => {
+      it("sends an explicit null when the operator turns the toggle off so the backend clears the stored points", async () => {
         withCachePoints([{ location: "message", role: "user" }]);
         const user = userEvent.setup();
         await enterEditMode(user);
@@ -1793,7 +1793,7 @@ describe("ModelInfoView", () => {
         await user.click(screen.getByRole("switch"));
         const payload = await save(user);
 
-        expect(payload.litellm_params).not.toHaveProperty("cache_control_injection_points");
+        expect(payload.litellm_params.cache_control_injection_points).toBeNull();
       });
 
       it("adds a typed index as a string, matching what the deployment already stores", async () => {

ui/litellm-dashboard/src/components/model_info_view.tsx

@@ -352,8 +352,11 @@ export default function ModelInfoView({
       }
 
       // Handle cache control settings
+      const hadInjectionPoints = Boolean(localModelData?.litellm_params?.cache_control_injection_points);
       if (values.cache_control && (values.cache_control_injection_points?.length ?? 0) > 0) {
         updatedLitellmParams.cache_control_injection_points = values.cache_control_injection_points;
+      } else if (hadInjectionPoints) {
+        updatedLitellmParams.cache_control_injection_points = null;
       } else {
         delete updatedLitellmParams.cache_control_injection_points;
       }