new-release?

libs/client_infinity/infinity_client/infinity_client/models/open_ai_embedding_input_audio.py

@@ -18,13 +18,15 @@ class OpenAIEmbeddingInputAudio:
         model (Union[Unset, str]):  Default: 'default/not-specified'.
         encoding_format (Union[Unset, EmbeddingEncodingFormat]):
         user (Union[None, Unset, str]):
+        dimensions (Union[Unset, int]):  Default: 0.
         modality (Union[Unset, OpenAIEmbeddingInputAudioModality]):  Default: OpenAIEmbeddingInputAudioModality.AUDIO.
     """
 
     input_: Union[List[str], str]
     model: Union[Unset, str] = "default/not-specified"
     encoding_format: Union[Unset, EmbeddingEncodingFormat] = UNSET
     user: Union[None, Unset, str] = UNSET
+    dimensions: Union[Unset, int] = 0
     modality: Union[Unset, OpenAIEmbeddingInputAudioModality] = OpenAIEmbeddingInputAudioModality.AUDIO
     additional_properties: Dict[str, Any] = _attrs_field(init=False, factory=dict)
 
@@ -52,6 +54,8 @@ def to_dict(self) -> Dict[str, Any]:
         else:
             user = self.user
 
+        dimensions = self.dimensions
+
         modality: Union[Unset, str] = UNSET
         if not isinstance(self.modality, Unset):
             modality = self.modality.value
@@ -69,6 +73,8 @@ def to_dict(self) -> Dict[str, Any]:
             field_dict["encoding_format"] = encoding_format
         if user is not UNSET:
             field_dict["user"] = user
+        if dimensions is not UNSET:
+            field_dict["dimensions"] = dimensions
         if modality is not UNSET:
             field_dict["modality"] = modality
 
@@ -118,6 +124,8 @@ def _parse_user(data: object) -> Union[None, Unset, str]:
 
         user = _parse_user(d.pop("user", UNSET))
 
+        dimensions = d.pop("dimensions", UNSET)
+
         _modality = d.pop("modality", UNSET)
         modality: Union[Unset, OpenAIEmbeddingInputAudioModality]
         if isinstance(_modality, Unset):
@@ -130,6 +138,7 @@ def _parse_user(data: object) -> Union[None, Unset, str]:
             model=model,
             encoding_format=encoding_format,
             user=user,
+            dimensions=dimensions,
             modality=modality,
         )
 

libs/client_infinity/infinity_client/infinity_client/models/open_ai_embedding_input_image.py

@@ -18,13 +18,15 @@ class OpenAIEmbeddingInputImage:
         model (Union[Unset, str]):  Default: 'default/not-specified'.
         encoding_format (Union[Unset, EmbeddingEncodingFormat]):
         user (Union[None, Unset, str]):
+        dimensions (Union[Unset, int]):  Default: 0.
         modality (Union[Unset, OpenAIEmbeddingInputImageModality]):  Default: OpenAIEmbeddingInputImageModality.IMAGE.
     """
 
     input_: Union[List[str], str]
     model: Union[Unset, str] = "default/not-specified"
     encoding_format: Union[Unset, EmbeddingEncodingFormat] = UNSET
     user: Union[None, Unset, str] = UNSET
+    dimensions: Union[Unset, int] = 0
     modality: Union[Unset, OpenAIEmbeddingInputImageModality] = OpenAIEmbeddingInputImageModality.IMAGE
     additional_properties: Dict[str, Any] = _attrs_field(init=False, factory=dict)
 
@@ -52,6 +54,8 @@ def to_dict(self) -> Dict[str, Any]:
         else:
             user = self.user
 
+        dimensions = self.dimensions
+
         modality: Union[Unset, str] = UNSET
         if not isinstance(self.modality, Unset):
             modality = self.modality.value
@@ -69,6 +73,8 @@ def to_dict(self) -> Dict[str, Any]:
             field_dict["encoding_format"] = encoding_format
         if user is not UNSET:
             field_dict["user"] = user
+        if dimensions is not UNSET:
+            field_dict["dimensions"] = dimensions
         if modality is not UNSET:
             field_dict["modality"] = modality
 
@@ -118,6 +124,8 @@ def _parse_user(data: object) -> Union[None, Unset, str]:
 
         user = _parse_user(d.pop("user", UNSET))
 
+        dimensions = d.pop("dimensions", UNSET)
+
         _modality = d.pop("modality", UNSET)
         modality: Union[Unset, OpenAIEmbeddingInputImageModality]
         if isinstance(_modality, Unset):
@@ -130,6 +138,7 @@ def _parse_user(data: object) -> Union[None, Unset, str]:
             model=model,
             encoding_format=encoding_format,
             user=user,
+            dimensions=dimensions,
             modality=modality,
         )
 

libs/client_infinity/infinity_client/infinity_client/models/open_ai_embedding_input_text.py

@@ -19,13 +19,15 @@ class OpenAIEmbeddingInputText:
         model (Union[Unset, str]):  Default: 'default/not-specified'.
         encoding_format (Union[Unset, EmbeddingEncodingFormat]):
         user (Union[None, Unset, str]):
+        dimensions (Union[Unset, int]):  Default: 0.
         modality (Union[Unset, OpenAIEmbeddingInputTextModality]):  Default: OpenAIEmbeddingInputTextModality.TEXT.
     """
 
     input_: Union[List[str], str]
     model: Union[Unset, str] = "default/not-specified"
     encoding_format: Union[Unset, EmbeddingEncodingFormat] = UNSET
     user: Union[None, Unset, str] = UNSET
+    dimensions: Union[Unset, int] = 0
     modality: Union[Unset, OpenAIEmbeddingInputTextModality] = OpenAIEmbeddingInputTextModality.TEXT
     additional_properties: Dict[str, Any] = _attrs_field(init=False, factory=dict)
 
@@ -49,6 +51,8 @@ def to_dict(self) -> Dict[str, Any]:
         else:
             user = self.user
 
+        dimensions = self.dimensions
+
         modality: Union[Unset, str] = UNSET
         if not isinstance(self.modality, Unset):
             modality = self.modality.value
@@ -66,6 +70,8 @@ def to_dict(self) -> Dict[str, Any]:
             field_dict["encoding_format"] = encoding_format
         if user is not UNSET:
             field_dict["user"] = user
+        if dimensions is not UNSET:
+            field_dict["dimensions"] = dimensions
         if modality is not UNSET:
             field_dict["modality"] = modality
 
@@ -106,6 +112,8 @@ def _parse_user(data: object) -> Union[None, Unset, str]:
 
         user = _parse_user(d.pop("user", UNSET))
 
+        dimensions = d.pop("dimensions", UNSET)
+
         _modality = d.pop("modality", UNSET)
         modality: Union[Unset, OpenAIEmbeddingInputTextModality]
         if isinstance(_modality, Unset):
@@ -118,6 +126,7 @@ def _parse_user(data: object) -> Union[None, Unset, str]:
             model=model,
             encoding_format=encoding_format,
             user=user,
+            dimensions=dimensions,
             modality=modality,
         )
 

libs/client_infinity/infinity_client/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "infinity_client"
-version = "0.0.72"
+version = "0.0.73"
 description = "A client library for accessing ♾️ Infinity - Embedding Inference Server"
 authors = []
 readme = "README.md"

libs/infinity_emb/infinity_emb/fastapi_schemas/pymodels.py

@@ -54,7 +54,7 @@ class _OpenAIEmbeddingInput(BaseModel):
     model: str = "default/not-specified"
     encoding_format: EmbeddingEncodingFormat = EmbeddingEncodingFormat.float
     user: Optional[str] = None
-    dimensions: Optional[Annotated[int, Field(strict=True, gt=0, lt=8193)]] = None
+    dimensions: int = 0
 
 
 class _OpenAIEmbeddingInput_Text(_OpenAIEmbeddingInput):

libs/infinity_emb/infinity_emb/inference/batch_handler.py

@@ -26,6 +26,7 @@
     ImageClassType,
     ModelCapabilites,
     ModelNotDeployedError,
+    MatryoshkaDimError,
     OverloadStatus,
     PredictSingle,
     PrioritizedQueueItem,
@@ -61,6 +62,18 @@ def submit(self, *args, **kwargs):
         return self._tp.submit(*args, **kwargs)
 
 
+def matryososka_slice(
+    embeddings: list[np.ndarray], matryoshka_dim: Optional[int]
+) -> list[np.ndarray]:
+    if matryoshka_dim:
+        if 1 > matryoshka_dim or matryoshka_dim > len(embeddings[0]):
+            raise MatryoshkaDimError(
+                f"matryoshka_dim={matryoshka_dim} is not in a valid range. Select between 1 and {len(embeddings[0])}."
+            )
+        return [e[:matryoshka_dim] for e in embeddings]
+    return embeddings
+
+
 class BatchHandler:
     def __init__(
         self,
@@ -159,9 +172,7 @@ async def embed(
         input_sentences = [EmbeddingSingle(sentence=s) for s in sentences]
 
         embeddings, usage = await self._schedule(input_sentences)
-        if matryoshka_dim:
-            embeddings = [embedding[:matryoshka_dim] for embedding in embeddings]
-        return embeddings, usage
+        return matryososka_slice(embeddings, matryoshka_dim), usage
 
     async def rerank(
         self,
@@ -267,9 +278,7 @@ async def image_embed(
 
         items = await resolve_images(images)
         embeddings, usage = await self._schedule(items)
-        if matryoshka_dim:
-            embeddings = [embedding[:matryoshka_dim] for embedding in embeddings]
-        return embeddings, usage
+        return matryososka_slice(embeddings, matryoshka_dim), usage
 
     async def audio_embed(
         self, *, audios: list[Union[str, bytes]], matryoshka_dim: Optional[int] = None
@@ -299,9 +308,7 @@ async def audio_embed(
             getattr(self.model_worker[0]._model, "sampling_rate", -42),
         )
         embeddings, usage = await self._schedule(items)
-        if matryoshka_dim:
-            embeddings = [embedding[:matryoshka_dim] for embedding in embeddings]
-        return embeddings, usage
+        return matryososka_slice(embeddings, matryoshka_dim), usage
 
     async def _schedule(self, list_queueitem: Sequence[AbstractSingle]) -> tuple[list[Any], int]:
         """adds list of items to the queue and awaits until these are completed."""

libs/infinity_emb/infinity_emb/infinity_server.py

@@ -29,6 +29,7 @@
     InferenceEngine,
     Modality,
     ModelCapabilites,
+    MatryoshkaDimError,
     ModelNotDeployedError,
     PoolingMethod,
 )
@@ -390,14 +391,9 @@ def url_to_base64(url, modality = "image"):
                 f"ModelNotDeployedError: model=`{data_root.model}` does not support `embed` for modality `{modality.value}`. Reason: {ex}",
                 code=status.HTTP_400_BAD_REQUEST,
             )
-        except (ImageCorruption, AudioCorruption) as ex:
-            # get urls_or_bytes if not defined
-            try:
-                urls_or_bytes = urls_or_bytes
-            except NameError:
-                urls_or_bytes = []
+        except (ImageCorruption, AudioCorruption, MatryoshkaDimError) as ex:
             raise errors.OpenAIException(
-                f"{modality.value}Corruption, could not open {[b if isinstance(b, str) else 'bytes' for b in urls_or_bytes]} -> {ex}",
+                f"{ex.__class__} -> {ex}",
                 code=status.HTTP_400_BAD_REQUEST,
             )
         except Exception as ex:
@@ -545,9 +541,9 @@ async def _embeddings_image(data: ImageEmbeddingInput):
                 encoding_format=data.encoding_format,
                 usage=usage,
             )
-        except ImageCorruption as ex:
+        except (ImageCorruption, MatryoshkaDimError) as ex:
             raise errors.OpenAIException(
-                f"ImageCorruption, could not open {[b if isinstance(b, str) else 'bytes' for b in urls_or_bytes]} -> {ex}",
+                f"{ex.__class__} -> {ex}",
                 code=status.HTTP_400_BAD_REQUEST,
             )
         except ModelNotDeployedError as ex:
@@ -604,9 +600,9 @@ async def _embeddings_audio(data: AudioEmbeddingInput):
                 encoding_format=data.encoding_format,
                 usage=usage,
             )
-        except AudioCorruption as ex:
+        except (AudioCorruption, MatryoshkaDimError) as ex:
             raise errors.OpenAIException(
-                f"AudioCorruption, could not open {[b if isinstance(b, str) else 'bytes' for b in urls_or_bytes]} -> {ex}",
+                f"{ex.__class__} -> {ex}",
                 code=status.HTTP_400_BAD_REQUEST,
             )
         except ModelNotDeployedError as ex:

libs/infinity_emb/infinity_emb/primitives.py

@@ -436,6 +436,10 @@ class ModelNotDeployedError(Exception):
     pass
 
 
+class MatryoshkaDimError(Exception):
+    pass
+
+
 class ImageCorruption(Exception):
     pass
 

libs/infinity_emb/infinity_emb/transformer/audio/utils.py

@@ -51,13 +51,9 @@ async def resolve_audios(
     CHECK_AIOHTTP.mark_required()
     CHECK_SOUNDFILE.mark_required()
 
-    resolved_audios: list[AudioSingle] = []
     async with aiohttp.ClientSession(trust_env=True) as session:
-        try:
-            resolved_audios = await asyncio.gather(
-                *[resolve_audio(audio, allowed_sampling_rate, session) for audio in audio_urls]
-            )
-        except Exception as e:
-            raise AudioCorruption(f"Failed to resolve audio: {e}")
+        resolved_audios = await asyncio.gather(
+            *[resolve_audio(audio, allowed_sampling_rate, session) for audio in audio_urls]
+        )
 
     return resolved_audios

libs/infinity_emb/infinity_emb/transformer/classifier/optimum.py

@@ -39,7 +39,7 @@ def __init__(self, *, engine_args: EngineArgs):
             prefer_quantized=("cpu" in provider.lower() or "openvino" in provider.lower()),
         )
 
-        self.model = optimize_model(
+        model = optimize_model(
             model_name_or_path=engine_args.model_name_or_path,
             model_class=ORTModelForSequenceClassification,
             revision=engine_args.revision,
@@ -48,7 +48,7 @@ def __init__(self, *, engine_args: EngineArgs):
             file_name=onnx_file.as_posix(),
             optimize_model=not os.environ.get("INFINITY_ONNX_DISABLE_OPTIMIZE", False),
         )
-        self.model.use_io_binding = False
+        model.use_io_binding = False
 
         self.tokenizer = AutoTokenizer.from_pretrained(
             engine_args.model_name_or_path,
@@ -60,12 +60,11 @@ def __init__(self, *, engine_args: EngineArgs):
 
         self._pipe = pipeline(
             task="text-classification",
-            model=self.model,
+            model=model,
             trust_remote_code=engine_args.trust_remote_code,
             top_k=None,
             revision=engine_args.revision,
             tokenizer=self.tokenizer,
-            device=engine_args.device,
         )
 
     def encode_pre(self, sentences: list[str]):

libs/infinity_emb/infinity_emb/transformer/vision/utils.py

@@ -78,11 +78,7 @@ async def resolve_images(
     CHECK_PIL.mark_required()
 
     resolved_imgs = []
-
-    try:
-        async with aiohttp.ClientSession(trust_env=True) as session:
-            resolved_imgs = await asyncio.gather(*[resolve_image(img, session) for img in images])
-    except Exception as e:
-        raise ImageCorruption(f"Failed to resolve image: {images}.\nError msg: {str(e)}")
+    async with aiohttp.ClientSession(trust_env=True) as session:
+        resolved_imgs = await asyncio.gather(*[resolve_image(img, session) for img in images])
 
     return resolved_imgs

libs/infinity_emb/pyproject.toml

@@ -1,7 +1,7 @@
 
 [tool.poetry]
 name = "infinity_emb"
-version = "0.0.72"
+version = "0.0.73"
 description = "Infinity is a high-throughput, low-latency REST API for serving text-embeddings, reranking models and clip."
 authors = ["michaelfeil <noreply@michaelfeil.eu>"]
 license = "MIT"

libs/infinity_emb/tests/unit_test/transformer/classifier/test_optimum_classifier.py

@@ -1,23 +1,19 @@
-import torch
-from optimum.pipelines import pipeline  # type: ignore
-from optimum.onnxruntime import ORTModelForSequenceClassification
+from transformers.pipelines import pipeline  # type: ignore
 from infinity_emb.args import EngineArgs
+
 from infinity_emb.transformer.classifier.optimum import OptimumClassifier
 
 
 def test_classifier(model_name: str = "SamLowe/roberta-base-go_emotions-onnx"):
     model = OptimumClassifier(
         engine_args=EngineArgs(
             model_name_or_path=model_name,
-            device="cuda" if torch.cuda.is_available() else "cpu",
         )  # type: ignore
     )
 
     pipe = pipeline(
         task="text-classification",
-        model=ORTModelForSequenceClassification.from_pretrained(
-            model_name, file_name="onnx/model_quantized.onnx"
-        ),
+        model="SamLowe/roberta-base-go_emotions",  # hoping that this is the same model as model_name
         top_k=None,
     )