diff --git a/litellm/llms/bedrock/embed/embedding.py b/litellm/llms/bedrock/embed/embedding.py index 69eb9b693b1..987c7cbf981 100644 --- a/litellm/llms/bedrock/embed/embedding.py +++ b/litellm/llms/bedrock/embed/embedding.py @@ -35,7 +35,7 @@ from .amazon_titan_multimodal_transformation import ( ) from .amazon_titan_v2_transformation import AmazonTitanV2Config from .cohere_transformation import BedrockCohereEmbeddingConfig -from .twelvelabs_marengo_transformation import TwelveLabsMarengoEmbeddingConfig +from .twelvelabs_marengo_transformation import TwelveLabsMarengoEmbeddingConfig, drop_params_enabled if TYPE_CHECKING: from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj @@ -480,6 +480,7 @@ class BedrockEmbedding(BaseAWSLLM): async_invoke_route=has_async_invoke, model_id=modelId, output_s3_uri=inference_params.get("output_s3_uri"), + drop_params=drop_params_enabled(litellm_params), ) batch_data.append(twelvelabs_request) elif provider == "nova": diff --git a/litellm/llms/bedrock/embed/twelvelabs_marengo_3_transformation.py b/litellm/llms/bedrock/embed/twelvelabs_marengo_3_transformation.py index 0f61d37258f..f4f9cb03dab 100644 --- a/litellm/llms/bedrock/embed/twelvelabs_marengo_3_transformation.py +++ b/litellm/llms/bedrock/embed/twelvelabs_marengo_3_transformation.py @@ -35,7 +35,7 @@ from litellm.types.llms.bedrock import ( ) from litellm.utils import get_base64_str -MARENGO_3_MODEL_MARKER: Final = "marengo-embed-3" +MARENGO_3_MODEL_MARKER: Final = "marengo-embed-3-" S3_URI_PREFIX: Final = "s3://" TIMED_MEDIA_OPTION_FIELDS: Final = MappingProxyType( { @@ -48,6 +48,9 @@ TIMED_MEDIA_OPTION_FIELDS: Final = MappingProxyType( } ) TIMED_MEDIA_OPTIONS: Final = TypeAdapter(TwelveLabsMarengo3TimedMediaOptions) +TIMED_INPUT_TYPES: Final = frozenset({"video", "audio"}) +MARENGO_2_7_ONLY_PARAMS: Final = ("textTruncate", "lengthSec", "useFixedLengthSec", "minClipSec") +MARENGO_2_7_ONLY_FIELDS: Final = MappingProxyType({name: True for name in MARENGO_2_7_ONLY_PARAMS}) def is_marengo_3_model(model: str | None) -> bool: @@ -69,15 +72,23 @@ class Marengo3Params(BaseModel): embeddingType: tuple[TWELVELABS_MARENGO_3_EMBEDDING_TYPES, ...] | None = None embeddingScope: tuple[TWELVELABS_MARENGO_3_EMBEDDING_SCOPES, ...] | None = None inferenceId: str | None = None + textTruncate: object = None + lengthSec: object = None + useFixedLengthSec: object = None + minClipSec: object = None @property def resolved_input_type(self) -> TWELVELABS_MARENGO_3_INPUT_TYPES: return self.inputType or self.input_type or "text" def timed_media_options(self) -> TwelveLabsMarengo3TimedMediaOptions: - return TIMED_MEDIA_OPTIONS.validate_python( - self.model_dump(include=TIMED_MEDIA_OPTION_FIELDS, exclude_none=True) - ) + return TIMED_MEDIA_OPTIONS.validate_python(self.given_timed_media_options()) + + def given_timed_media_options(self) -> dict[str, object]: + return self.model_dump(include=TIMED_MEDIA_OPTION_FIELDS, exclude_none=True) + + def given_2_7_only_params(self) -> dict[str, object]: + return self.model_dump(include=MARENGO_2_7_ONLY_FIELDS, exclude_none=True) def _s3_location(uri: str, bucket_owner: str | None) -> TwelveLabsS3Location: @@ -113,11 +124,23 @@ def _timed_media_input(media: str, params: Marengo3Params) -> TwelveLabsMarengo3 return timed +def _describe(error: ValidationError) -> str: + return "; ".join( + f"{'.'.join(str(part) for part in problem['loc'])}: {problem['msg']}" for problem in error.errors() + ) + + def _validated_params(inference_params: Mapping[str, object]) -> Marengo3Params: try: return Marengo3Params.model_validate(inference_params) except ValidationError as error: - raise BedrockError(status_code=400, message=f"Invalid Marengo 3.0 parameters: {error}") from error + raise BedrockError(status_code=400, message=f"Invalid Marengo 3.0 parameters: {_describe(error)}") from error + + +def _reject_unless_dropped(given: Mapping[str, object], drop_params: bool, reason: str) -> None: + if not given or drop_params: + return + raise BedrockError(status_code=400, message=f"{reason} {', '.join(given)}; set drop_params to drop them") def _require(value: str | None, input_type: str, param_name: str) -> str: @@ -143,10 +166,19 @@ def _request_base(inference_id: str | None) -> TwelveLabsMarengo3RequestBase: return identified -def build_marengo_3_request(input: str, inference_params: Mapping[str, object]) -> TwelveLabsMarengo3EmbeddingRequest: +def build_marengo_3_request( + input: str, inference_params: Mapping[str, object], drop_params: bool = False +) -> TwelveLabsMarengo3EmbeddingRequest: params: Final = _validated_params(inference_params) base: Final = _request_base(params.inferenceId) input_type: Final = params.resolved_input_type + _reject_unless_dropped( + params.given_2_7_only_params(), drop_params, "Marengo 3.0 does not accept the Marengo 2.7 parameters" + ) + if input_type not in TIMED_INPUT_TYPES: + _reject_unless_dropped( + params.given_timed_media_options(), drop_params, f"Input type '{input_type}' does not accept" + ) match input_type: case "text": text_request: Final[TwelveLabsMarengo3TextRequest] = { diff --git a/litellm/llms/bedrock/embed/twelvelabs_marengo_transformation.py b/litellm/llms/bedrock/embed/twelvelabs_marengo_transformation.py index ddf6dfbcc4d..35163ecf848 100644 --- a/litellm/llms/bedrock/embed/twelvelabs_marengo_transformation.py +++ b/litellm/llms/bedrock/embed/twelvelabs_marengo_transformation.py @@ -13,7 +13,9 @@ from typing import Final, cast from pydantic import BaseModel, ConfigDict, TypeAdapter from typing_extensions import assert_never +import litellm from litellm.llms.bedrock.embed.twelvelabs_marengo_3_transformation import ( + MARENGO_2_7_ONLY_PARAMS, build_marengo_3_request, is_marengo_3_model, ) @@ -99,6 +101,25 @@ def _billed_usage(batch_data: list[dict] | None) -> Usage: return Usage(prompt_tokens=0, completion_tokens=0, total_tokens=0, prompt_tokens_details=details) +MARENGO_SHARED_PARAMS: Final = ( + "encoding_format", + "embeddingOption", + "startSec", + "input_type", + "endSec", + "segmentation", + "embeddingType", + "embeddingScope", + "inferenceId", + "media_source", + "media_sources", +) + + +def drop_params_enabled(litellm_params: Mapping[str, object]) -> bool: + return litellm.drop_params is True or litellm_params.get("drop_params") is True + + class TwelveLabsMarengoEmbeddingConfig: """ Reference - https://docs.aws.amazon.com/bedrock/latest/userguide/model-parameters-marengo.html @@ -115,23 +136,9 @@ class TwelveLabsMarengoEmbeddingConfig: self.is_marengo_3: Final = is_marengo_3_model(model) def get_supported_openai_params(self) -> list[str]: - return [ - "encoding_format", - "textTruncate", - "embeddingOption", - "startSec", - "lengthSec", - "useFixedLengthSec", - "minClipSec", - "input_type", - "endSec", - "segmentation", - "embeddingType", - "embeddingScope", - "inferenceId", - "media_source", - "media_sources", - ] + if self.is_marengo_3: + return list(MARENGO_SHARED_PARAMS) + return [*MARENGO_SHARED_PARAMS, *MARENGO_2_7_ONLY_PARAMS] def map_openai_params(self, non_default_params: dict, optional_params: dict) -> dict: for k, v in non_default_params.items(): @@ -179,6 +186,7 @@ class TwelveLabsMarengoEmbeddingConfig: async_invoke_route: bool = False, model_id: str | None = None, output_s3_uri: str | None = None, + drop_params: bool = False, ) -> TwelveLabsMarengoEmbeddingRequest | TwelveLabsMarengo3EmbeddingRequest | TwelveLabsAsyncInvokeRequest: """ Transform OpenAI-style input to TwelveLabs Marengo format/async-invoke format. @@ -203,7 +211,9 @@ class TwelveLabsMarengoEmbeddingConfig: ) if self.is_marengo_3: - marengo_3_request: Final = build_marengo_3_request(input=input, inference_params=inference_params) + marengo_3_request: Final = build_marengo_3_request( + input=input, inference_params=inference_params, drop_params=drop_params + ) if async_invoke_route and model_id: return self._wrap_async_invoke_request( model_input=marengo_3_request, model_id=model_id, output_s3_uri=output_s3_uri diff --git a/tests/test_litellm/llms/bedrock/embed/test_twelvelabs_marengo_3_transformation.py b/tests/test_litellm/llms/bedrock/embed/test_twelvelabs_marengo_3_transformation.py index 0bf86352a5e..5033256d089 100644 --- a/tests/test_litellm/llms/bedrock/embed/test_twelvelabs_marengo_3_transformation.py +++ b/tests/test_litellm/llms/bedrock/embed/test_twelvelabs_marengo_3_transformation.py @@ -2,13 +2,16 @@ import json import pytest +import litellm from litellm.llms.bedrock.common_utils import BedrockError from litellm.llms.bedrock.embed.twelvelabs_marengo_3_transformation import ( + MARENGO_2_7_ONLY_PARAMS, build_marengo_3_request, is_marengo_3_model, ) from litellm.llms.bedrock.embed.twelvelabs_marengo_transformation import ( TwelveLabsMarengoEmbeddingConfig, + drop_params_enabled, ) MARENGO_3_BASE = "twelvelabs.marengo-embed-3-0-v1:0" @@ -27,6 +30,7 @@ OUTPUT_S3_URI = "s3://out-bucket/marengo/" ("async_invoke/twelvelabs.marengo-embed-3-0-v1:0", True), (MARENGO_27_US, False), ("twelvelabs.marengo-embed-2-7-v1:0", False), + ("twelvelabs.marengo-embed-30-v1:0", False), (None, False), ], ) @@ -266,3 +270,89 @@ def test_marengo_3_only_params_are_forwarded_by_map_openai_params(): "embeddingScope": ["clip"], "inferenceId": "req-1", } + + +@pytest.mark.parametrize( + "params,problem", + [ + ( + {"input_type": "clip"}, + "input_type: Input should be 'text', 'image', 'video', 'audio', 'text_image' or 'multi_input'", + ), + ({"input_type": "video", "embeddingOption": "visual"}, "embeddingOption: Input should be a valid tuple"), + ( + {"input_type": "multi_input", "media_sources": ["not", "a", "mapping"]}, + "media_sources: Input should be a valid dictionary", + ), + ], +) +def test_invalid_marengo_3_params_name_the_field_and_the_reason(params, problem): + with pytest.raises(BedrockError) as excinfo: + build_marengo_3_request("s3://media/clip.mp4", params) + assert excinfo.value.message == f"Invalid Marengo 3.0 parameters: {problem}" + + +MARENGO_2_7_ONLY_VALUES = {"textTruncate": "end", "lengthSec": 5, "useFixedLengthSec": True, "minClipSec": 2} + + +@pytest.mark.parametrize("name", MARENGO_2_7_ONLY_PARAMS) +def test_marengo_2_7_only_params_are_rejected_on_3_0_unless_dropped(name): + params = {"input_type": "text", name: MARENGO_2_7_ONLY_VALUES[name]} + with pytest.raises(BedrockError) as excinfo: + build_marengo_3_request("hello", params) + assert excinfo.value.status_code == 400 + assert excinfo.value.message == ( + f"Marengo 3.0 does not accept the Marengo 2.7 parameters {name}; set drop_params to drop them" + ) + assert build_marengo_3_request("hello", params, drop_params=True) == { + "inputType": "text", + "text": {"inputText": "hello"}, + } + + +def test_marengo_2_7_only_params_are_advertised_only_for_2_7(): + marengo_3 = TwelveLabsMarengoEmbeddingConfig(model=MARENGO_3_US).get_supported_openai_params() + marengo_27 = TwelveLabsMarengoEmbeddingConfig(model=MARENGO_27_US).get_supported_openai_params() + assert set(MARENGO_2_7_ONLY_PARAMS).isdisjoint(marengo_3) + assert set(MARENGO_2_7_ONLY_PARAMS) <= set(marengo_27) + assert set(marengo_3) <= set(marengo_27) + + +def test_drop_params_comes_from_the_call_or_the_global(monkeypatch): + monkeypatch.setattr(litellm, "drop_params", False) + assert drop_params_enabled({}) is False + assert drop_params_enabled({"drop_params": True}) is True + monkeypatch.setattr(litellm, "drop_params", True) + assert drop_params_enabled({}) is True + + +def test_config_drops_marengo_2_7_only_params_only_when_asked(): + config = TwelveLabsMarengoEmbeddingConfig(model=MARENGO_3_US) + with pytest.raises(BedrockError, match=r"Marengo 2\.7 parameters textTruncate"): + config._transform_request("hello", {"textTruncate": "end"}) + assert config._transform_request("hello", {"textTruncate": "end"}, drop_params=True) == { + "inputType": "text", + "text": {"inputText": "hello"}, + } + + +@pytest.mark.parametrize( + "params", + [ + {"input_type": "text"}, + {"input_type": "image"}, + {"input_type": "text_image", "media_source": DUCK_DATA_URL}, + {"input_type": "multi_input", "media_sources": {"bird": DUCK_DATA_URL}}, + ], +) +def test_timed_media_options_are_rejected_on_untimed_input_types_unless_dropped(params): + timed = {**params, "startSec": 0, "embeddingOption": ["visual"]} + with pytest.raises(BedrockError) as excinfo: + build_marengo_3_request(DUCK_DATA_URL, timed) + assert excinfo.value.status_code == 400 + assert excinfo.value.message == ( + f"Input type '{params['input_type']}' does not accept startSec, embeddingOption; set drop_params to drop them" + ) + assert build_marengo_3_request(DUCK_DATA_URL, timed, drop_params=True) == build_marengo_3_request( + DUCK_DATA_URL, params + )