Fix cache miss for gemini models with response_format (#10635)

* Fix unwanted mutation of kwargs in apply_response_schema_transformation

* add assertion the original dict is not mutated

litellm/llms/vertex_ai/gemini/vertex_and_google_ai_studio_gemini.py

@@ -337,21 +337,22 @@ class VertexGeminiConfig(VertexAIBaseConfig, BaseConfig):
         return old_schema
 
     def apply_response_schema_transformation(self, value: dict, optional_params: dict):
+        new_value = deepcopy(value)
         # remove 'additionalProperties' from json schema
-        value = _remove_additional_properties(value)
+        new_value = _remove_additional_properties(new_value)
         # remove 'strict' from json schema
-        value = _remove_strict_from_schema(value)
-        if value["type"] == "json_object":
+        new_value = _remove_strict_from_schema(new_value)
+        if new_value["type"] == "json_object":
             optional_params["response_mime_type"] = "application/json"
-        elif value["type"] == "text":
+        elif new_value["type"] == "text":
             optional_params["response_mime_type"] = "text/plain"
-        if "response_schema" in value:
+        if "response_schema" in new_value:
             optional_params["response_mime_type"] = "application/json"
-            optional_params["response_schema"] = value["response_schema"]
-        elif value["type"] == "json_schema":  # type: ignore
-            if "json_schema" in value and "schema" in value["json_schema"]:  # type: ignore
+            optional_params["response_schema"] = new_value["response_schema"]
+        elif new_value["type"] == "json_schema":  # type: ignore
+            if "json_schema" in new_value and "schema" in new_value["json_schema"]:  # type: ignore
                 optional_params["response_mime_type"] = "application/json"
-                optional_params["response_schema"] = value["json_schema"]["schema"]  # type: ignore
+                optional_params["response_schema"] = new_value["json_schema"]["schema"]  # type: ignore
 
         if "response_schema" in optional_params and isinstance(
             optional_params["response_schema"], dict

tests/litellm/llms/vertex_ai/gemini/test_vertex_and_google_ai_studio_gemini.py

@@ -1,5 +1,6 @@
 import asyncio
 from typing import List, cast
+from copy import deepcopy
 from unittest.mock import MagicMock
 
 import pytest
@@ -69,37 +70,39 @@ def test_get_model_name_from_gemini_spec_model():
 
 def test_vertex_ai_response_schema_dict():
     v = VertexGeminiConfig()
-    transformed_request = v.map_openai_params(
-        non_default_params={
-            "messages": [{"role": "user", "content": "Hello, world!"}],
-            "response_format": {
-                "type": "json_schema",
-                "json_schema": {
-                    "name": "math_reasoning",
-                    "schema": {
-                        "type": "object",
-                        "properties": {
-                            "steps": {
-                                "type": "array",
-                                "items": {
-                                    "type": "object",
-                                    "properties": {
-                                        "thought": {"type": "string"},
-                                        "output": {"type": "string"},
-                                    },
-                                    "required": ["thought", "output"],
-                                    "additionalProperties": False,
+    non_default_params = {
+        "messages": [{"role": "user", "content": "Hello, world!"}],
+        "response_format": {
+            "type": "json_schema",
+            "json_schema": {
+                "name": "math_reasoning",
+                "schema": {
+                    "type": "object",
+                    "properties": {
+                        "steps": {
+                            "type": "array",
+                            "items": {
+                                "type": "object",
+                                "properties": {
+                                    "thought": {"type": "string"},
+                                    "output": {"type": "string"},
                                 },
+                                "required": ["thought", "output"],
+                                "additionalProperties": False,
                             },
-                            "final_answer": {"type": "string"},
                         },
-                        "required": ["steps", "final_answer"],
-                        "additionalProperties": False,
+                        "final_answer": {"type": "string"},
                     },
-                    "strict": False,
+                    "required": ["steps", "final_answer"],
+                    "additionalProperties": False,
                 },
+                "strict": False,
             },
         },
+    }
+    original_non_default_params = deepcopy(non_default_params)
+    transformed_request = v.map_openai_params(
+        non_default_params=non_default_params,
         optional_params={},
         model="gemini-2.0-flash-lite",
         drop_params=False,
@@ -137,6 +140,8 @@ def test_vertex_ai_response_schema_dict():
         "required": ["steps", "final_answer"],
         "propertyOrdering": ["steps", "final_answer"],
     }
+    # should not mutate the original non_default_params
+    assert non_default_params == original_non_default_params
 
 
 class MathReasoning(BaseModel):