Merge pull request #26077 from BerriAI/litellm_vertex_cached_content_mutual_exclusion

fix(vertex_ai): omit system_instruction/tools/toolConfig when cachedContent set

litellm/llms/vertex_ai/gemini/transformation.py

@@ -743,16 +743,22 @@ def _transform_request_body(  # noqa: PLR0915
                     ]
 
         data = RequestBody(contents=content)
-        if system_instructions is not None:
-            data["system_instruction"] = system_instructions
-        if tools is not None:
-            data["tools"] = tools
-        if tool_choice is not None:
-            data["toolConfig"] = tool_choice
-        if include_server_side_tool_invocations:
-            if "toolConfig" not in data:
-                data["toolConfig"] = {}
-            data["toolConfig"]["includeServerSideToolInvocations"] = True
+        # Vertex rejects system_instruction/tools/toolConfig alongside cachedContent.
+        # Treat dropping these fields as a request mutation guarded by modify_params.
+        can_send_cache_incompatible_fields = (
+            cached_content is None or litellm.modify_params is False
+        )
+        if can_send_cache_incompatible_fields:
+            if system_instructions is not None:
+                data["system_instruction"] = system_instructions
+            if tools is not None:
+                data["tools"] = tools
+            if tool_choice is not None:
+                data["toolConfig"] = tool_choice
+            if include_server_side_tool_invocations:
+                if "toolConfig" not in data:
+                    data["toolConfig"] = {}
+                data["toolConfig"]["includeServerSideToolInvocations"] = True
         if safety_settings is not None:
             data["safetySettings"] = safety_settings
         if generation_config is not None and len(generation_config) > 0:

tests/test_litellm/llms/vertex_ai/gemini/test_vertex_ai_gemini_transformation.py

@@ -86,6 +86,76 @@ def test_check_if_part_exists_in_parts_camel_case_snake_case():
     assert check_if_part_exists_in_parts(parts_mixed, part_mixed_casing)
 
 
+def test_cached_content_respects_modify_params_for_cache_incompatible_fields():
+    """Regression: cachedContent drops system/tools/toolConfig only when modify_params=True."""
+    import litellm
+
+    cache_name = "projects/p/locations/us-central1/cachedContents/abc123"
+    messages = [
+        {"role": "system", "content": "You are helpful"},
+        {"role": "user", "content": "hi"},
+    ]
+    optional_params = {
+        "tools": [
+            {
+                "functionDeclarations": [
+                    {"name": "get_weather", "description": "Get weather"},
+                ]
+            }
+        ],
+        "tool_choice": {"functionCallingConfig": {"mode": "AUTO"}},
+    }
+
+    original_modify_params = litellm.modify_params
+    try:
+        # With modify_params=False (default), keep fields even with cachedContent.
+        litellm.modify_params = False
+        result = _transform_request_body(
+            messages=list(messages),
+            model="gemini-2.5-pro",
+            optional_params=dict(optional_params),
+            custom_llm_provider="vertex_ai",
+            litellm_params={},
+            cached_content=cache_name,
+        )
+        assert result.get("cachedContent") == cache_name
+        assert "system_instruction" in result
+        assert "tools" in result
+        assert "toolConfig" in result
+        assert "contents" in result
+
+        # With modify_params=True, drop cache-incompatible fields.
+        litellm.modify_params = True
+        result_modify_true = _transform_request_body(
+            messages=list(messages),
+            model="gemini-2.5-pro",
+            optional_params=dict(optional_params),
+            custom_llm_provider="vertex_ai",
+            litellm_params={},
+            cached_content=cache_name,
+        )
+        assert result_modify_true.get("cachedContent") == cache_name
+        assert "system_instruction" not in result_modify_true
+        assert "tools" not in result_modify_true
+        assert "toolConfig" not in result_modify_true
+        assert "contents" in result_modify_true
+
+        # Without cache, fields are always included.
+        result_no_cache = _transform_request_body(
+            messages=list(messages),
+            model="gemini-2.5-pro",
+            optional_params=dict(optional_params),
+            custom_llm_provider="vertex_ai",
+            litellm_params={},
+            cached_content=None,
+        )
+        assert "system_instruction" in result_no_cache
+        assert "tools" in result_no_cache
+        assert "toolConfig" in result_no_cache
+    finally:
+        litellm.modify_params = original_modify_params
+
+
 # Tests for issue #14556: Labels field provider-aware filtering
 def test_google_genai_excludes_labels():
     """Test that Google GenAI/AI Studio endpoints exclude labels when custom_llm_provider='gemini'"""