fix(speech): forward the caching flag and drop malformed cache controls on Gemini and Vertex AI TTS

A non-bool caching or a cache dict with non-string keys is dropped instead of failing the speech call, so it behaves as it does today

litellm/endpoints/speech/speech_to_completion_bridge/handler.py

@@ -2,7 +2,6 @@
 Handler for transforming /chat/completions api requests to litellm.responses requests
 """
 
-from collections.abc import Mapping
 from typing import TYPE_CHECKING, Final
 
 from typing_extensions import TypedDict
@@ -14,6 +13,7 @@ if TYPE_CHECKING:
 
     from litellm import LiteLLMLoggingObj
     from litellm.types.llms.openai import HttpxBinaryResponseContent
+    from litellm.types.utils import CacheControlsHint
 
 
 class SpeechToCompletionBridgeHandlerInputKwargs(TypedDict):
@@ -92,7 +92,7 @@ class SpeechToCompletionBridgeHandler:
         custom_llm_provider: str,
         api_base: str | None,
         timeout: "float | httpx.Timeout | None",
-        cache: Mapping[str, object] | None,
+        cache_controls: "CacheControlsHint",
     ) -> "HttpxBinaryResponseContent":
         received_args: Final = locals()
         from litellm import completion
@@ -112,7 +112,7 @@ class SpeechToCompletionBridgeHandler:
             model=model,
             input=input,
             optional_params=optional_params,
-            litellm_params={**litellm_params, "api_base": api_base, "timeout": timeout, "cache": cache},
+            litellm_params={**litellm_params, "api_base": api_base, "timeout": timeout, **cache_controls},
             headers=headers,
             litellm_logging_obj=logging_obj,
             custom_llm_provider=custom_llm_provider,

litellm/main.py

@@ -8941,7 +8941,7 @@ def speech(
                 custom_llm_provider=custom_llm_provider,
                 api_base=api_base,
                 timeout=timeout,
-                cache=_CACHE_CONTROLS_HINT.validate_python(kwargs).get("cache"),
+                cache_controls=_CACHE_CONTROLS_HINT.validate_python(kwargs),
             )
 
         # Vertex AI Text-to-Speech (Google Cloud TTS)
@@ -8995,7 +8995,7 @@ def speech(
             custom_llm_provider=custom_llm_provider,
             api_base=api_base,
             timeout=timeout,
-            cache=_CACHE_CONTROLS_HINT.validate_python(kwargs).get("cache"),
+            cache_controls=_CACHE_CONTROLS_HINT.validate_python(kwargs),
         )
     elif custom_llm_provider == "runwayml":
         from litellm.llms.runwayml.text_to_speech.transformation import (

litellm/types/utils.py

@@ -42,8 +42,10 @@ from pydantic import (
     FieldSerializationInfo,
     GetCoreSchemaHandler,
     JsonValue,
+    OnErrorOmit,
     PrivateAttr,
     SkipValidation,
+    StrictBool,
     field_serializer,
     field_validator,
 )
@@ -5020,4 +5022,5 @@ class BaseUrlRoutingHint(TypedDict, total=False):
 
 
 class CacheControlsHint(TypedDict, total=False):
-    cache: ReadOnly[Mapping[str, object] | None]
+    cache: ReadOnly[OnErrorOmit[Mapping[str, object] | None]]
+    caching: ReadOnly[OnErrorOmit[StrictBool | None]]

tests/unit/test_main.py

@@ -5521,13 +5521,14 @@ def test_speech_gemini_tts_without_a_timeout_waits_as_long_as_the_same_chat_comp
         ({}, {}, 1),
         ({}, {"cache": {"no-cache": True}}, 2),
         ({"cache": {"no-store": True}}, {}, 2),
+        ({}, {"caching": False}, 2),
     ],
 )
 def test_speech_gemini_tts_honors_the_request_cache_controls(
     respx_mock: respx.MockRouter,
     monkeypatch: pytest.MonkeyPatch,
-    first_controls: Mapping[str, Mapping[str, bool]],
-    second_controls: Mapping[str, Mapping[str, bool]],
+    first_controls: Mapping[str, object],
+    second_controls: Mapping[str, object],
     expected_upstream_calls: int,
 ):
     monkeypatch.setattr(litellm, "cache", Cache(type="local"))
@@ -5547,6 +5548,29 @@ def test_speech_gemini_tts_honors_the_request_cache_controls(
     assert route.call_count == expected_upstream_calls
 
 
+@pytest.mark.parametrize("malformed_controls", [{"caching": "yes"}, {"caching": 1}, {"cache": {1: True}}])
+def test_speech_gemini_tts_without_a_cache_ignores_malformed_cache_controls(
+    respx_mock: respx.MockRouter,
+    monkeypatch: pytest.MonkeyPatch,
+    malformed_controls: Mapping[str, object],
+):
+    monkeypatch.setattr(litellm, "cache", None)
+    route: Final = respx_mock.post(url__regex=GEMINI_DEFAULT_GENERATE_CONTENT_URL).mock(
+        return_value=httpx.Response(200, json=_gemini_tts_generate_content_response())
+    )
+
+    response: Final = litellm.speech(
+        model=f"gemini/{GEMINI_TTS_MODEL}",
+        input="hello",
+        voice="Kore",
+        api_key="fake-gemini-key",
+        **malformed_controls,
+    )
+
+    assert route.call_count == 1
+    assert response.content.startswith(b"RIFF")
+
+
 FOUNDRY_HOST: Final = "https://my-project.services.ai.azure.com"