fix(transcription): honor base_url alias for Groq Whisper and report it as the api base (#43917)

* fix(transcription): honor base_url alias for Groq Whisper and report it as the api base

transcription() and speech() only accepted api_base, so a deployment configured with base_url leaked the alias into the provider params, which Groq rejected as an unknown param, and the request never reached the internal gateway. get_api_base() now reads the same alias so response headers and logs show the configured endpoint instead of the provider default

Resolves LIT-9071

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

* fix(speech): keep base_url after existing audio params, route Vertex speech to it, skip empty alias

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

---------

Co-authored-by: yassin <yassin@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

litellm/litellm_core_utils/llm_response_utils/get_api_base.py

@@ -59,6 +59,10 @@ def get_api_base(model: str, optional_params: dict | LiteLLM_Params) -> str | No
 
     if _optional_params.api_base is not None:
         return _optional_params.api_base
+    extra_params: Final = _optional_params.model_extra
+    base_url_alias: Final = extra_params.get("base_url") if extra_params is not None else None
+    if isinstance(base_url_alias, str) and base_url_alias:
+        return base_url_alias
 
     if litellm.model_alias_map and model in litellm.model_alias_map:
         model = litellm.model_alias_map[model]

litellm/main.py

@@ -7931,6 +7931,7 @@ def transcription(
     api_version: str | None = None,
     max_retries: int | None = None,
     custom_llm_provider=None,
+    base_url: str | None = None,
     **kwargs,
 ) -> TranscriptionResponse | Coroutine[object, object, TranscriptionResponse]:
     """
@@ -7964,7 +7965,7 @@ def transcription(
     model, custom_llm_provider, dynamic_api_key, api_base = get_llm_provider(
         model=model,
         custom_llm_provider=custom_llm_provider,
-        api_base=api_base,
+        api_base=api_base or base_url,
         api_key=api_key,
     )
 
@@ -8237,6 +8238,7 @@ def speech(
     headers: dict | None = None,
     custom_llm_provider: str | None = None,
     aspeech: bool | None = None,
+    base_url: str | None = None,
     **kwargs,
 ) -> HttpxBinaryResponseContent | Coroutine[object, object, HttpxBinaryResponseContent]:
     user: Final = kwargs.get("user", None)
@@ -8246,7 +8248,7 @@ def speech(
     model_info: Final = kwargs.get("model_info", None)
     shared_session: Final = kwargs.get("shared_session", None)
     model, custom_llm_provider, dynamic_api_key, api_base = get_llm_provider(
-        model=model, custom_llm_provider=custom_llm_provider, api_base=api_base
+        model=model, custom_llm_provider=custom_llm_provider, api_base=api_base or base_url
     )
     kwargs.pop("tags", [])
 
@@ -8550,7 +8552,7 @@ def speech(
             extra_headers=headers,
             base_llm_http_handler=base_llm_http_handler,
             aspeech=aspeech or False,
-            api_base=generic_optional_params.api_base,
+            api_base=api_base,
             api_key=None,  # Vertex AI uses OAuth, not API key
             **kwargs,
         )

tests/unit/litellm_core_utils/llm_response_utils/test_get_api_base.py

@@ -91,3 +91,22 @@ def test_providers_with_a_fixed_base_still_get_it(model, expected, monkeypatch):
         monkeypatch.delenv(env, raising=False)
 
     assert litellm.get_api_base(model=model, optional_params={}) == expected
+
+
+def test_base_url_alias_is_reported_as_the_api_base():
+    api_base = litellm.get_api_base(
+        model="groq/whisper-large-v3", optional_params={"base_url": "https://groq.gateway.internal/openai/v1"}
+    )
+
+    assert api_base == "https://groq.gateway.internal/openai/v1"
+    assert (
+        litellm.get_api_base(
+            model="groq/whisper-large-v3",
+            optional_params={"api_base": "https://explicit.internal/v1", "base_url": "https://alias.internal/v1"},
+        )
+        == "https://explicit.internal/v1"
+    )
+    assert (
+        litellm.get_api_base(model="groq/whisper-large-v3", optional_params={"base_url": ""})
+        == "https://api.groq.com/openai/v1"
+    )

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

@@ -261,9 +261,7 @@ class TestVertexAILyriaTextToSpeechConfig:
         )
 
     def test_get_complete_url_encodes_injected_predict_path_segments(self, monkeypatch: pytest.MonkeyPatch) -> None:
-        injected: Final = (
-            "victim-project/locations/us-central1/publishers/google/models/other-model:predict?ignored="
-        )
+        injected: Final = "victim-project/locations/us-central1/publishers/google/models/other-model:predict?ignored="
         encoded: Final = (
             "victim-project%2Flocations%2Fus-central1%2Fpublishers%2Fgoogle"
             "%2Fmodels%2Fother-model%3Apredict%3Fignored%3D"
@@ -554,6 +552,33 @@ class TestVertexAILyriaTextToSpeechConfig:
         assert mock_post.call_args.kwargs["json"] == expected_body
 
 
+@pytest.mark.parametrize("endpoint_kwarg", ["api_base", "base_url"])
+def test_litellm_speech_vertex_ai_sends_request_to_the_configured_endpoint(endpoint_kwarg: str):
+    mock_response = Mock(spec=httpx.Response)
+    mock_response.status_code = 200
+    mock_response.headers = {"content-type": "application/json"}
+    mock_response.json.return_value = {"audioContent": "SGVsbG8gV29ybGQ="}
+    with (
+        patch.object(  # test-quality-ok: litellm.speech has no seam for Vertex token minting
+            VertexAITextToSpeechConfig, "_ensure_access_token", return_value=("mock-token", "test-project")
+        ),
+        patch(  # test-quality-ok: litellm.speech has no seam for the HTTP handler
+            "litellm.llms.custom_httpx.llm_http_handler.HTTPHandler.post", return_value=mock_response
+        ) as mock_post,
+    ):
+        response = litellm.speech(
+            model="vertex_ai/chirp",
+            input="Hello",
+            voice="en-US-Chirp3-HD-Charon",
+            vertex_project="test-project",
+            vertex_location="us-central1",
+            **{endpoint_kwarg: "https://tts.gateway.internal/v1/text:synthesize"},
+        )
+
+    assert mock_post.call_args.kwargs["url"] == "https://tts.gateway.internal/v1/text:synthesize"
+    assert response.content == b"Hello World"
+
+
 @patch("litellm.llms.custom_httpx.llm_http_handler.HTTPHandler.post")
 @patch.object(VertexAITextToSpeechConfig, "_ensure_access_token")
 @patch.object(VertexAITextToSpeechConfig, "_get_token_and_url")

tests/unit/test_main.py

@@ -4188,6 +4188,62 @@ def test_azure_ai_speech_on_a_foundry_host_uses_the_azure_openai_deployment_rout
     assert response.content == b"mp3-bytes"
 
 
+GROQ_INTERNAL_BASE: Final = "https://groq.gateway.internal/openai/v1"
+GROQ_WAV_FILE: Final = ("tone.wav", b"RIFF\x00\x00\x00\x00WAVE", "audio/wav")
+
+
+def test_groq_transcription_honors_base_url_alias(respx_mock: respx.MockRouter):
+    route: Final = respx_mock.post(f"{GROQ_INTERNAL_BASE}/audio/transcriptions").mock(
+        return_value=httpx.Response(200, json={"text": "hello"})
+    )
+
+    response: Final = litellm.transcription(
+        model="groq/whisper-large-v3",
+        file=GROQ_WAV_FILE,
+        base_url=GROQ_INTERNAL_BASE,
+        api_key="fake-key",
+    )
+
+    assert route.called
+    assert response.text == "hello"
+
+
+async def test_groq_atranscription_honors_base_url_alias(
+    respx_mock: respx.MockRouter, monkeypatch: pytest.MonkeyPatch
+):
+    monkeypatch.setattr(litellm, "disable_aiohttp_transport", True)
+    route: Final = respx_mock.post(f"{GROQ_INTERNAL_BASE}/audio/transcriptions").mock(
+        return_value=httpx.Response(200, json={"text": "hello"})
+    )
+
+    response: Final = await litellm.atranscription(
+        model="groq/whisper-large-v3",
+        file=GROQ_WAV_FILE,
+        base_url=GROQ_INTERNAL_BASE,
+        api_key="fake-key",
+    )
+
+    assert route.called
+    assert response.text == "hello"
+
+
+def test_groq_speech_honors_base_url_alias(respx_mock: respx.MockRouter):
+    route: Final = respx_mock.post(f"{GROQ_INTERNAL_BASE}/audio/speech").mock(
+        return_value=httpx.Response(200, content=b"mp3-bytes")
+    )
+
+    response: Final = litellm.speech(
+        model="groq/playai-tts",
+        input="hello",
+        voice="Fritz-PlayAI",
+        base_url=GROQ_INTERNAL_BASE,
+        api_key="fake-key",
+    )
+
+    assert route.called
+    assert response.content == b"mp3-bytes"
+
+
 FORWARDED_CLIENT_HEADERS: Final = {"x-forwarded-for": "10.0.0.1", "x-amzn-trace-id": "Root=1-lit7694"}