fix(speech): honor request cache controls on Gemini and Vertex AI TTS

Backport of #50803 to lts/1.108.x.
Cherry-picked from fdc7648f43 (litellm_speech_bridge_api_base).

litellm/endpoints/speech/speech_to_completion_bridge/handler.py

@@ -2,6 +2,7 @@
 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
@@ -89,6 +90,7 @@ class SpeechToCompletionBridgeHandler:
         custom_llm_provider: str,
         api_base: str | None,
         timeout: "float | httpx.Timeout | None",
+        cache: Mapping[str, object] | None,
     ) -> "HttpxBinaryResponseContent":
         received_args: Final = locals()
         from litellm import completion
@@ -108,7 +110,7 @@ class SpeechToCompletionBridgeHandler:
             model=model,
             input=input,
             optional_params=optional_params,
-            litellm_params={**litellm_params, "api_base": api_base, "timeout": timeout},
+            litellm_params={**litellm_params, "api_base": api_base, "timeout": timeout, "cache": cache},
             headers=headers,
             litellm_logging_obj=logging_obj,
             custom_llm_provider=custom_llm_provider,

litellm/main.py

@@ -147,6 +147,7 @@ from litellm.types.llms.openai import AllMessageValues
 from litellm.types.router import GenericLiteLLMParams
 from litellm.types.utils import (
     BaseUrlRoutingHint,
+    CacheControlsHint,
     CustomPricingLiteLLMParams,
     ModelResponseStream,
     ProviderRoutingHints,
@@ -4137,6 +4138,7 @@ _REQUEST_KWARGS_ADAPTER: Final = TypeAdapter(Mapping[str, object])
 _MODEL_INFO_ADAPTER: Final = TypeAdapter(Mapping[str, object])
 _PROVIDER_ROUTING_HINTS: Final = TypeAdapter(ProviderRoutingHints)
 _BASE_URL_HINT: Final = TypeAdapter(BaseUrlRoutingHint)
+_CACHE_CONTROLS_HINT: Final = TypeAdapter(CacheControlsHint)
 
 
 _AWS_BEDROCK_CLIENT_CREDENTIAL_KEYS: Final = (
@@ -8753,6 +8755,7 @@ def speech(
                 custom_llm_provider=custom_llm_provider,
                 api_base=api_base,
                 timeout=timeout,
+                cache=_CACHE_CONTROLS_HINT.validate_python(kwargs).get("cache"),
             )
 
         # Vertex AI Text-to-Speech (Google Cloud TTS)
@@ -8806,6 +8809,7 @@ def speech(
             custom_llm_provider=custom_llm_provider,
             api_base=api_base,
             timeout=timeout,
+            cache=_CACHE_CONTROLS_HINT.validate_python(kwargs).get("cache"),
         )
     elif custom_llm_provider == "runwayml":
         from litellm.llms.runwayml.text_to_speech.transformation import (

litellm/types/utils.py

@@ -4868,3 +4868,7 @@ class ProviderRoutingHints(TypedDict, total=False):
 
 class BaseUrlRoutingHint(TypedDict, total=False):
     base_url: ReadOnly[str | None]
+
+
+class CacheControlsHint(TypedDict, total=False):
+    cache: ReadOnly[Mapping[str, object] | None]

tests/integration/providers/test_gemini_tts_voice_mapping_owned_proxy.py

@@ -1,7 +1,6 @@
 """The speech route, the audio_speech health check and deferred Vertex Live setups map OpenAI voices.
 
-The speech bridge calls the Gemini chat path without the deployment's ``api_base`` (tracked separately),
-so these rows boot their own proxy with ``GEMINI_API_BASE`` pointed at the scripted peer, and with
+These rows boot their own proxy with ``GEMINI_API_BASE`` pointed at the scripted peer, and with
 ``LITELLM_GEMINI_LIVE_DEFER_SETUP`` so a client ``session.update`` voice builds the Vertex Live setup.
 """
 

tests/unit/llms/vertex_ai/text_to_speech/test_transformation.py

@@ -11,6 +11,7 @@ import responses
 from pydantic import ValidationError
 
 import litellm
+from litellm.caching.caching import Cache
 from litellm.litellm_core_utils.exception_mapping_utils import exception_type
 from litellm.llms.vertex_ai.text_to_speech.transformation import (
     VertexAILyriaTextToSpeechConfig,
@@ -1144,3 +1145,25 @@ async def test_aspeech_vertex_ai_gemini_tts_routes_to_configured_api_base(
     assert gateway_route.call_count == 1
     assert gateway_route.calls.last.request.headers["authorization"] == "Bearer minted-google-token"
     assert _PCM_AUDIO in response.content
+
+
+@pytest.mark.asyncio
+async def test_aspeech_vertex_ai_gemini_tts_no_cache_request_reaches_vertex(
+    respx_mock: respx.MockRouter, google_token_endpoint: responses.RequestsMock, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    monkeypatch.setattr(litellm, "cache", Cache(type="local"))
+    vertex_route: Final = respx_mock.post(
+        url__regex=r"https://aiplatform\.googleapis\.com/.*/models/gemini-3\.8-flash-tts:generateContent$"
+    ).mock(return_value=httpx.Response(200, json=_GENERATE_CONTENT_RESPONSE))
+    speech_kwargs: Final = {
+        "model": "vertex_ai/gemini-3.8-flash-tts",
+        "input": "Say hello",
+        "voice": "Kore",
+        "vertex_credentials": _AUTHORIZED_USER,
+        "vertex_location": "global",
+    }
+
+    await litellm.aspeech(**speech_kwargs)
+    await litellm.aspeech(**speech_kwargs, cache={"no-cache": True})
+
+    assert vertex_route.call_count == 2

tests/unit/test_main.py

@@ -4838,6 +4838,38 @@ def test_speech_gemini_tts_without_a_timeout_waits_as_long_as_the_same_chat_comp
     assert speech_call.request.extensions["timeout"] == chat_call.request.extensions["timeout"]
 
 
+@pytest.mark.parametrize(
+    ("first_controls", "second_controls", "expected_upstream_calls"),
+    [
+        ({}, {}, 1),
+        ({}, {"cache": {"no-cache": True}}, 2),
+        ({"cache": {"no-store": True}}, {}, 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]],
+    expected_upstream_calls: int,
+):
+    monkeypatch.setattr(litellm, "cache", Cache(type="local"))
+    route: Final = respx_mock.post(url__regex=GEMINI_DEFAULT_GENERATE_CONTENT_URL).mock(
+        return_value=httpx.Response(200, json=_gemini_tts_generate_content_response())
+    )
+    speech_kwargs: Final = {
+        "model": f"gemini/{GEMINI_TTS_MODEL}",
+        "input": "hello",
+        "voice": "Kore",
+        "api_key": "fake-gemini-key",
+    }
+
+    litellm.speech(**speech_kwargs, **first_controls)
+    litellm.speech(**speech_kwargs, **second_controls)
+
+    assert route.call_count == expected_upstream_calls
+
+
 FOUNDRY_HOST: Final = "https://my-project.services.ai.azure.com"