diff --git a/litellm/__init__.py b/litellm/__init__.py index 1784deeb797..ea049ea60cf 100644 --- a/litellm/__init__.py +++ b/litellm/__init__.py @@ -1438,6 +1438,7 @@ from .exceptions import ( Timeout, APIConnectionError, UnsupportedParamsError, + UnsupportedEndpointError as UnsupportedEndpointError, APIResponseValidationError, UnprocessableEntityError, InternalServerError, diff --git a/litellm/exceptions.py b/litellm/exceptions.py index 733f1aa72d5..b900b2e6983 100644 --- a/litellm/exceptions.py +++ b/litellm/exceptions.py @@ -317,6 +317,10 @@ class MalformedRequestFieldError(BadRequestError): pass +class UnsupportedEndpointError(BadRequestError): + pass + + class UnprocessableEntityError(openai.UnprocessableEntityError): def __init__( self, @@ -1054,6 +1058,7 @@ LITELLM_EXCEPTION_TYPES: Final = [ BadRequestError, UnprocessableEntityError, UnsupportedParamsError, + UnsupportedEndpointError, Timeout, PermissionDeniedError, RateLimitError, diff --git a/litellm/main.py b/litellm/main.py index 327a7a1c4b5..e160daabb98 100644 --- a/litellm/main.py +++ b/litellm/main.py @@ -8335,7 +8335,7 @@ def transcription( custom_llm_provider=custom_llm_provider, ) - response: TranscriptionResponse | Coroutine[object, object, TranscriptionResponse] | None = None + response: TranscriptionResponse | Coroutine[object, object, TranscriptionResponse] provider_config: Final = ProviderConfigManager.get_provider_audio_transcription_config( model=model, @@ -8497,10 +8497,16 @@ def transcription( provider_config=provider_config, shared_session=shared_session, ) + else: + raise litellm.UnsupportedEndpointError( + message=f"{custom_llm_provider}/{model} does not support audio transcription.", + model=model, + llm_provider=custom_llm_provider, + ) # Store duration in _hidden_params for cost calculation without # exposing it in the response body (see sync path comment above). - if response is not None and not isinstance(response, Coroutine): + if not isinstance(response, Coroutine): existing_duration: Final = getattr(response, "duration", None) if existing_duration is None: calculated_duration: Final = calculate_request_duration(file) @@ -8510,8 +8516,6 @@ def transcription( ) response_hidden_params["audio_transcription_duration"] = calculated_duration - if response is None: - raise ValueError("Unmapped provider passed in. Unable to get the response.") return response diff --git a/litellm/router.py b/litellm/router.py index 705accbcf8d..bd85fa91756 100644 --- a/litellm/router.py +++ b/litellm/router.py @@ -612,6 +612,7 @@ _DEPLOYMENT_PARAMS_ADAPTER: Final[TypeAdapter[Mapping[str, object] | GenericLite _OPAQUE_VALUE_ADAPTER: Final = TypeAdapter(object) _DEPLOYMENT_SELECTED_EVENT: Final = "litellm.request.deployment_selected" _CONTENT_BLOCKS_ADAPTER: Final = TypeAdapter(list[dict[str, object]]) +_DEPLOYMENTS_ADAPTER: Final = TypeAdapter(Sequence[Mapping[str, object]]) def _extend_trailing_text_block(blocks: Sequence[Mapping[str, object]], text: str) -> Sequence[Mapping[str, object]]: @@ -673,6 +674,12 @@ def _configured_model_info(deployment: Mapping[str, object]) -> Mapping[str, obj return _MODEL_INFO_ADAPTER.validate_python(deployment.get("model_info") or {}) +def _every_deployment_refused( + deployments: Sequence[Mapping[str, object]], refused_deployment_ids: tuple[str, ...] +) -> bool: + return all(_configured_model_info(deployment).get("id") in refused_deployment_ids for deployment in deployments) + + def _as_retry_skipped_deployment_ids(value: object) -> tuple[str, ...]: return tuple(item for item in value if isinstance(item, str)) if isinstance(value, tuple) else () @@ -8972,6 +8979,10 @@ class Router: ) if first_skipped_ids: kwargs["_retry_skipped_deployment_ids"] = first_skipped_ids # rebind-ok: the next attempt reads it + if isinstance(original_exception, litellm.UnsupportedEndpointError) and _every_deployment_refused( + _DEPLOYMENTS_ADAPTER.validate_python(_healthy_deployments), first_skipped_ids + ): + raise else: raise @@ -9051,6 +9062,10 @@ class Router: ) if skipped_ids: kwargs["_retry_skipped_deployment_ids"] = skipped_ids # rebind-ok: the next attempt reads it + if isinstance(e, litellm.UnsupportedEndpointError) and _every_deployment_refused( + _DEPLOYMENTS_ADAPTER.validate_python(_healthy_deployments), skipped_ids + ): + raise _timeout = self._time_to_sleep_before_retry( e=e, remaining_retries=remaining_retries, @@ -9147,7 +9162,7 @@ class Router: status_code: Final = getattr(error, "status_code", None) if status_code is not None and not litellm.should_retry(status_code): # 401/403 are special cases - allow retry if multiple deployments exist (handled below) - if status_code not in (401, 403): + if status_code not in (401, 403) and not isinstance(error, litellm.UnsupportedEndpointError): raise error if isinstance(error, litellm.NotFoundError): diff --git a/litellm/router_utils/fallback_event_handlers.py b/litellm/router_utils/fallback_event_handlers.py index 6c93fe6d239..0a77ad5b9ea 100644 --- a/litellm/router_utils/fallback_event_handlers.py +++ b/litellm/router_utils/fallback_event_handlers.py @@ -261,7 +261,10 @@ def is_bad_request_without_fallback(exception: Exception, fallback_on_unsupporte return False if isinstance(exception, litellm.UnsupportedParamsError): return True - if isinstance(exception, (litellm.ContextWindowExceededError, litellm.ContentPolicyViolationError)): + if isinstance( + exception, + (litellm.ContextWindowExceededError, litellm.ContentPolicyViolationError, litellm.UnsupportedEndpointError), + ): return False return exception.status_code == 400 and bool(getattr(exception, "llm_provider", "")) diff --git a/tests/code_coverage_tests/router_code_coverage.py b/tests/code_coverage_tests/router_code_coverage.py index af2f386a362..ab264f11412 100644 --- a/tests/code_coverage_tests/router_code_coverage.py +++ b/tests/code_coverage_tests/router_code_coverage.py @@ -95,6 +95,7 @@ ignored_function_names = [ "arm_routing_read_prefetch", # Tested in tests/unit/caching/test_request_redis_batch_pre_call.py (file lacks "router" in name) "_deployment_params_base_model", # Tested through log_retry in test_router_helper_utils.py "_configured_model_info", # Tested through get_configured_service_tiers in test_router.py + "_every_deployment_refused", # Tested through the unsupported transcription retry tests in test_router.py "_routable_deployments", # Tested through get_configured_service_tiers and get_routable_upstream_model in test_router.py "_async_get_available_deployment", # Body of the `route {model}` phase wrapper, exercised through async_get_available_deployment in test_router.py "_async_get_available_deployment_for_pass_through", # Same, through async_get_available_deployment_for_pass_through in test_router.py diff --git a/tests/unit/test_main.py b/tests/unit/test_main.py index 70b268bf2a2..24e01eec6e4 100644 --- a/tests/unit/test_main.py +++ b/tests/unit/test_main.py @@ -9779,6 +9779,37 @@ async def test_atranscription_answers_a_non_string_provider_hint_with_a_bad_requ assert not respx_mock.calls +def test_transcription_rejects_an_unsupported_provider_without_an_http_call(respx_mock: respx.MockRouter) -> None: + model: Final = "test-transcription-model" + provider: Final = "anthropic" + + with pytest.raises(litellm.UnsupportedEndpointError) as raised: + litellm.transcription(model=f"{provider}/{model}", file=("speech.wav", b"RIFF-audio", "audio/wav")) + + assert raised.value.status_code == 400 + assert raised.value.message == f"litellm.BadRequestError: {provider}/{model} does not support audio transcription." + assert raised.value.model == model + assert raised.value.llm_provider == provider + assert not respx_mock.calls + + +@pytest.mark.asyncio +async def test_atranscription_rejects_an_unsupported_provider_without_an_http_call( + respx_mock: respx.MockRouter, +) -> None: + model: Final = "test-transcription-model" + provider: Final = "anthropic" + + with pytest.raises(litellm.UnsupportedEndpointError) as raised: + await litellm.atranscription(model=f"{provider}/{model}", file=("speech.wav", b"RIFF-audio", "audio/wav")) + + assert raised.value.status_code == 400 + assert raised.value.message == f"litellm.BadRequestError: {provider}/{model} does not support audio transcription." + assert raised.value.model == model + assert raised.value.llm_provider == provider + assert not respx_mock.calls + + @pytest.mark.asyncio async def test_aspeech_infers_the_provider_from_base_url(respx_mock: respx.MockRouter) -> None: base_url: Final = "https://api.groq.com/openai/v1" diff --git a/tests/unit/test_router/test_router.py b/tests/unit/test_router/test_router.py index f06f9e08f7e..3e12f3e9bef 100644 --- a/tests/unit/test_router/test_router.py +++ b/tests/unit/test_router/test_router.py @@ -21429,6 +21429,93 @@ def test_router_retry_policy_400_moves_off_the_refusing_deployment_on_every_sync assert accepting.call_count == 1 +_TRANSCRIPTION_AUDIO: Final = ("speech.wav", b"RIFF", "audio/wav") +_TRANSCRIPTION_API_BASE: Final = "https://transcriber.local/v1" + + +def _transcription_deployment(model_name: str, deployment_id: str, litellm_model: str, weight: int) -> dict: + return { + "model_name": model_name, + "litellm_params": { + "model": litellm_model, + "api_key": "sk-fake", + "api_base": _TRANSCRIPTION_API_BASE, + "weight": weight, + }, + "model_info": {"id": deployment_id}, + } + + +@pytest.mark.asyncio +async def test_router_retries_a_transcription_off_a_deployment_whose_provider_cannot_transcribe( + monkeypatch: pytest.MonkeyPatch, +): + monkeypatch.setattr(litellm, "disable_aiohttp_transport", True) + router: Final = litellm.Router( + model_list=[ + _transcription_deployment("transcriber", "cannot-transcribe", "anthropic/claude-sonnet-5-5", weight=1), + _transcription_deployment("transcriber", "can-transcribe", "openai/gpt-transcribe", weight=0), + ], + num_retries=1, + disable_cooldowns=True, + ) + + with respx.mock as respx_mock: + upstream: Final = respx_mock.post(f"{_TRANSCRIPTION_API_BASE}/audio/transcriptions").mock( + return_value=httpx.Response(200, json={"text": "hi"}) + ) + response: Final = await router.atranscription(model="transcriber", file=_TRANSCRIPTION_AUDIO) + + assert response.text == "hi" + assert upstream.call_count == 1 + + +@pytest.mark.parametrize("deployment_count", [1, 2, 3]) +@pytest.mark.asyncio +async def test_router_tries_each_deployment_that_cannot_transcribe_once_then_answers_400(deployment_count: int): + router: Final = litellm.Router( + model_list=[ + _transcription_deployment("transcriber", f"cannot-transcribe-{index}", "anthropic/claude-sonnet-5-5", 1) + for index in range(deployment_count) + ], + num_retries=5, + disable_cooldowns=True, + ) + + with respx.mock as respx_mock, pytest.raises(litellm.UnsupportedEndpointError) as raised: + await router.atranscription(model="transcriber", file=_TRANSCRIPTION_AUDIO) + + assert raised.value.status_code == 400 + assert router.fail_calls["transcriber"] == deployment_count + assert not respx_mock.calls + + +@pytest.mark.asyncio +async def test_router_falls_back_from_a_group_that_cannot_transcribe_when_unsupported_params_do_not( + monkeypatch: pytest.MonkeyPatch, +): + monkeypatch.setattr(litellm, "disable_aiohttp_transport", True) + router: Final = litellm.Router( + model_list=[ + _transcription_deployment("chat-only", "cannot-transcribe", "anthropic/claude-sonnet-5-5", weight=1), + _transcription_deployment("transcriber", "can-transcribe", "openai/gpt-transcribe", weight=1), + ], + fallbacks=[{"chat-only": ["transcriber"]}], + fallback_on_unsupported_params=False, + num_retries=0, + disable_cooldowns=True, + ) + + with respx.mock as respx_mock: + upstream: Final = respx_mock.post(f"{_TRANSCRIPTION_API_BASE}/audio/transcriptions").mock( + return_value=httpx.Response(200, json={"text": "hi"}) + ) + response: Final = await router.atranscription(model="chat-only", file=_TRANSCRIPTION_AUDIO) + + assert response.text == "hi" + assert upstream.call_count == 1 + + def _make_failure_logging_obj(): return LiteLLMLogging( model="gpt-5.6",