fastembed-gpu 0.6.0__tar.gz → 0.6.1__tar.gz

{fastembed_gpu-0.6.0 → fastembed_gpu-0.6.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: fastembed-gpu
-Version: 0.6.0
+Version: 0.6.1
 Summary: Fast, light, accurate library built for retrieval embedding generation
 License: Apache License
 Keywords: vector,embedding,neural,search,qdrant,sentence-transformers
@@ -225,6 +225,23 @@ scores = list(encoder.rerank(query, documents))
 # [-11.48061752319336, 5.472434997558594]
 ```
 
+Text cross encoders can also be extended with models which are not in the list of supported models.
+
+```python
+from fastembed.rerank.cross_encoder import TextCrossEncoder 
+from fastembed.common.model_description import ModelSource
+
+TextCrossEncoder.add_custom_model(
+    model="Xenova/ms-marco-MiniLM-L-4-v2",
+    model_file="onnx/model.onnx",
+    sources=ModelSource(hf="Xenova/ms-marco-MiniLM-L-4-v2"),
+)
+model = TextCrossEncoder(model_name="Xenova/ms-marco-MiniLM-L-4-v2")
+scores = list(model.rerank_pairs(
+    [("What is AI?", "Artificial intelligence is ..."), ("What is ML?", "Machine learning is ..."),]
+))
+```
+
 ## ⚡️ FastEmbed on a GPU
 
 FastEmbed supports running on GPU devices.

{fastembed_gpu-0.6.0 → fastembed_gpu-0.6.1}/README.md

@@ -190,6 +190,23 @@ scores = list(encoder.rerank(query, documents))
 # [-11.48061752319336, 5.472434997558594]
 ```
 
+Text cross encoders can also be extended with models which are not in the list of supported models.
+
+```python
+from fastembed.rerank.cross_encoder import TextCrossEncoder 
+from fastembed.common.model_description import ModelSource
+
+TextCrossEncoder.add_custom_model(
+    model="Xenova/ms-marco-MiniLM-L-4-v2",
+    model_file="onnx/model.onnx",
+    sources=ModelSource(hf="Xenova/ms-marco-MiniLM-L-4-v2"),
+)
+model = TextCrossEncoder(model_name="Xenova/ms-marco-MiniLM-L-4-v2")
+scores = list(model.rerank_pairs(
+    [("What is AI?", "Artificial intelligence is ..."), ("What is ML?", "Machine learning is ..."),]
+))
+```
+
 ## ⚡️ FastEmbed on a GPU
 
 FastEmbed supports running on GPU devices.

{fastembed_gpu-0.6.0 → fastembed_gpu-0.6.1}/fastembed/common/model_description.py

@@ -7,6 +7,11 @@ from typing import Optional, Any
 class ModelSource:
     hf: Optional[str] = None
     url: Optional[str] = None
+    _deprecated_tar_struct: bool = False
+
+    @property
+    def deprecated_tar_struct(self) -> bool:
+        return self._deprecated_tar_struct
 
     def __post_init__(self) -> None:
         if self.hf is None and self.url is None:

{fastembed_gpu-0.6.0 → fastembed_gpu-0.6.1}/fastembed/common/model_management.py

@@ -330,9 +330,10 @@ class ModelManagement(Generic[T]):
         model_name: str,
         source_url: str,
         cache_dir: str,
+        deprecated_tar_struct: bool = False,
         local_files_only: bool = False,
     ) -> Path:
-        fast_model_name = f"fast-{model_name.split('/')[-1]}"
+        fast_model_name = f"{'fast-' if deprecated_tar_struct else ''}{model_name.split('/')[-1]}"
         cache_tmp_dir = Path(cache_dir) / "tmp"
         model_tmp_dir = cache_tmp_dir / fast_model_name
         model_dir = Path(cache_dir) / fast_model_name
@@ -438,6 +439,7 @@ class ModelManagement(Generic[T]):
                         model.model,
                         str(url_source),
                         str(cache_dir),
+                        deprecated_tar_struct=model.sources.deprecated_tar_struct,
                         local_files_only=local_files_only,
                     )
                 except Exception:

{fastembed_gpu-0.6.0 → fastembed_gpu-0.6.1}/fastembed/image/onnx_embedding.py

@@ -1,6 +1,5 @@
 from typing import Any, Iterable, Optional, Sequence, Type, Union
 
-import numpy as np
 
 from fastembed.common.types import NumpyArray
 from fastembed.common import ImageInput, OnnxProvider
@@ -195,7 +194,7 @@ class OnnxImageEmbedding(ImageEmbeddingBase, OnnxImageModel[NumpyArray]):
         return onnx_input
 
     def _post_process_onnx_output(self, output: OnnxOutputContext) -> Iterable[NumpyArray]:
-        return normalize(output.model_output).astype(np.float32)
+        return normalize(output.model_output)
 
 
 class OnnxImageEmbeddingWorker(ImageEmbeddingWorker[NumpyArray]):

{fastembed_gpu-0.6.0 → fastembed_gpu-0.6.1}/fastembed/late_interaction/colbert.py

@@ -46,7 +46,7 @@ class Colbert(LateInteractionTextEmbeddingBase, OnnxTextModel[NumpyArray]):
         self, output: OnnxOutputContext, is_doc: bool = True
     ) -> Iterable[NumpyArray]:
         if not is_doc:
-            return output.model_output.astype(np.float32)
+            return output.model_output
 
         if output.input_ids is None or output.attention_mask is None:
             raise ValueError(
@@ -58,11 +58,11 @@ class Colbert(LateInteractionTextEmbeddingBase, OnnxTextModel[NumpyArray]):
                 if token_id in self.skip_list or token_id == self.pad_token_id:
                     output.attention_mask[i, j] = 0
 
-        output.model_output *= np.expand_dims(output.attention_mask, 2).astype(np.float32)
+        output.model_output *= np.expand_dims(output.attention_mask, 2)
         norm = np.linalg.norm(output.model_output, ord=2, axis=2, keepdims=True)
         norm_clamped = np.maximum(norm, 1e-12)
         output.model_output /= norm_clamped
-        return output.model_output.astype(np.float32)
+        return output.model_output
 
     def _preprocess_onnx_input(
         self, onnx_input: dict[str, NumpyArray], is_doc: bool = True, **kwargs: Any

{fastembed_gpu-0.6.0 → fastembed_gpu-0.6.1}/fastembed/late_interaction_multimodal/colpali.py

@@ -142,7 +142,7 @@ class ColPali(LateInteractionMultimodalEmbeddingBase, OnnxMultimodalModel[NumpyA
         assert self.model_description.dim is not None, "Model dim is not defined"
         return output.model_output.reshape(
             output.model_output.shape[0], -1, self.model_description.dim
-        ).astype(np.float32)
+        )
 
     def _post_process_onnx_text_output(
         self,
@@ -157,7 +157,7 @@ class ColPali(LateInteractionMultimodalEmbeddingBase, OnnxMultimodalModel[NumpyA
         Returns:
             Iterable[NumpyArray]: Post-processed output as NumPy arrays.
         """
-        return output.model_output.astype(np.float32)
+        return output.model_output
 
     def tokenize(self, documents: list[str], **kwargs: Any) -> list[Encoding]:
         texts_query: list[str] = []

fastembed_gpu-0.6.1/fastembed/rerank/cross_encoder/custom_text_cross_encoder.py

@@ -0,0 +1,46 @@
+from typing import Optional, Sequence, Any
+
+from fastembed.common import OnnxProvider
+from fastembed.common.model_description import BaseModelDescription
+from fastembed.rerank.cross_encoder.onnx_text_cross_encoder import OnnxTextCrossEncoder
+
+
+class CustomTextCrossEncoder(OnnxTextCrossEncoder):
+    SUPPORTED_MODELS: list[BaseModelDescription] = []
+
+    def __init__(
+        self,
+        model_name: str,
+        cache_dir: Optional[str] = None,
+        threads: Optional[int] = None,
+        providers: Optional[Sequence[OnnxProvider]] = None,
+        cuda: bool = False,
+        device_ids: Optional[list[int]] = None,
+        lazy_load: bool = False,
+        device_id: Optional[int] = None,
+        specific_model_path: Optional[str] = None,
+        **kwargs: Any,
+    ):
+        super().__init__(
+            model_name=model_name,
+            cache_dir=cache_dir,
+            threads=threads,
+            providers=providers,
+            cuda=cuda,
+            device_ids=device_ids,
+            lazy_load=lazy_load,
+            device_id=device_id,
+            specific_model_path=specific_model_path,
+            **kwargs,
+        )
+
+    @classmethod
+    def _list_supported_models(cls) -> list[BaseModelDescription]:
+        return cls.SUPPORTED_MODELS
+
+    @classmethod
+    def add_model(
+        cls,
+        model_description: BaseModelDescription,
+    ) -> None:
+        cls.SUPPORTED_MODELS.append(model_description)

{fastembed_gpu-0.6.0 → fastembed_gpu-0.6.1}/fastembed/rerank/cross_encoder/text_cross_encoder.py

@@ -3,13 +3,19 @@ from dataclasses import asdict
 
 from fastembed.common import OnnxProvider
 from fastembed.rerank.cross_encoder.onnx_text_cross_encoder import OnnxTextCrossEncoder
+from fastembed.rerank.cross_encoder.custom_text_cross_encoder import CustomTextCrossEncoder
+
 from fastembed.rerank.cross_encoder.text_cross_encoder_base import TextCrossEncoderBase
-from fastembed.common.model_description import BaseModelDescription
+from fastembed.common.model_description import (
+    ModelSource,
+    BaseModelDescription,
+)
 
 
 class TextCrossEncoder(TextCrossEncoderBase):
     CROSS_ENCODER_REGISTRY: list[Type[TextCrossEncoderBase]] = [
         OnnxTextCrossEncoder,
+        CustomTextCrossEncoder,
     ]
 
     @classmethod
@@ -124,3 +130,34 @@ class TextCrossEncoder(TextCrossEncoderBase):
         yield from self.model.rerank_pairs(
             pairs, batch_size=batch_size, parallel=parallel, **kwargs
         )
+
+    @classmethod
+    def add_custom_model(
+        cls,
+        model: str,
+        sources: ModelSource,
+        model_file: str = "onnx/model.onnx",
+        description: str = "",
+        license: str = "",
+        size_in_gb: float = 0.0,
+        additional_files: Optional[list[str]] = None,
+    ) -> None:
+        registered_models = cls._list_supported_models()
+        for registered_model in registered_models:
+            if model == registered_model.model:
+                raise ValueError(
+                    f"Model {model} is already registered in CrossEncoderModel, if you still want to add this model, "
+                    f"please use another model name"
+                )
+
+        CustomTextCrossEncoder.add_model(
+            BaseModelDescription(
+                model=model,
+                sources=sources,
+                model_file=model_file,
+                description=description,
+                license=license,
+                size_in_GB=size_in_gb,
+                additional_files=additional_files or [],
+            )
+        )

{fastembed_gpu-0.6.0 → fastembed_gpu-0.6.1}/fastembed/sparse/splade_pp.py

@@ -18,7 +18,7 @@ supported_splade_models: list[SparseModelDescription] = [
         description="Independent Implementation of SPLADE++ Model for English.",
         license="apache-2.0",
         size_in_GB=0.532,
-        sources=ModelSource(hf="Qdrant/SPLADE_PP_en_v1"),
+        sources=ModelSource(hf="Qdrant/Splade_PP_en_v1"),
         model_file="model.onnx",
     ),
     SparseModelDescription(
@@ -27,7 +27,7 @@ supported_splade_models: list[SparseModelDescription] = [
         description="Independent Implementation of SPLADE++ Model for English.",
         license="apache-2.0",
         size_in_GB=0.532,
-        sources=ModelSource(hf="Qdrant/SPLADE_PP_en_v1"),
+        sources=ModelSource(hf="Qdrant/Splade_PP_en_v1"),
         model_file="model.onnx",
     ),
 ]

{fastembed_gpu-0.6.0 → fastembed_gpu-0.6.1}/fastembed/text/multitask_embedding.py

@@ -3,6 +3,7 @@ from typing import Any, Type, Iterable, Union, Optional
 
 import numpy as np
 
+from fastembed.common.onnx_model import OnnxOutputContext
 from fastembed.common.types import NumpyArray
 from fastembed.text.pooled_normalized_embedding import PooledNormalizedEmbedding
 from fastembed.text.onnx_embedding import OnnxTextEmbeddingWorker
@@ -44,9 +45,11 @@ class JinaEmbeddingV3(PooledNormalizedEmbedding):
     PASSAGE_TASK = Task.RETRIEVAL_PASSAGE
     QUERY_TASK = Task.RETRIEVAL_QUERY
 
-    def __init__(self, *args: Any, **kwargs: Any):
+    def __init__(self, *args: Any, task_id: Optional[int] = None, **kwargs: Any):
         super().__init__(*args, **kwargs)
-        self.current_task_id: Union[Task, int] = self.PASSAGE_TASK
+        self.default_task_id: Union[Task, int] = (
+            task_id if task_id is not None else self.PASSAGE_TASK
+        )
 
     @classmethod
     def _get_worker_class(cls) -> Type[OnnxTextEmbeddingWorker]:
@@ -57,9 +60,14 @@ class JinaEmbeddingV3(PooledNormalizedEmbedding):
         return supported_multitask_models
 
     def _preprocess_onnx_input(
-        self, onnx_input: dict[str, NumpyArray], **kwargs: Any
+        self,
+        onnx_input: dict[str, NumpyArray],
+        task_id: Optional[Union[int, Task]] = None,
+        **kwargs: Any,
     ) -> dict[str, NumpyArray]:
-        onnx_input["task_id"] = np.array(self.current_task_id, dtype=np.int64)
+        if task_id is None:
+            raise ValueError(f"task_id must be provided for JinaEmbeddingV3, got <{task_id}>")
+        onnx_input["task_id"] = np.array(task_id, dtype=np.int64)
         return onnx_input
 
     def embed(
@@ -67,20 +75,19 @@ class JinaEmbeddingV3(PooledNormalizedEmbedding):
         documents: Union[str, Iterable[str]],
         batch_size: int = 256,
         parallel: Optional[int] = None,
-        task_id: int = PASSAGE_TASK,
+        task_id: Optional[int] = None,
         **kwargs: Any,
     ) -> Iterable[NumpyArray]:
-        self.current_task_id = task_id
-        kwargs["task_id"] = task_id
-        yield from super().embed(documents, batch_size, parallel, **kwargs)
+        task_id = (
+            task_id if task_id is not None else self.default_task_id
+        )  # required for multiprocessing
+        yield from super().embed(documents, batch_size, parallel, task_id=task_id, **kwargs)
 
     def query_embed(self, query: Union[str, Iterable[str]], **kwargs: Any) -> Iterable[NumpyArray]:
-        self.current_task_id = self.QUERY_TASK
-        yield from super().embed(query, **kwargs)
+        yield from super().embed(query, task_id=self.QUERY_TASK, **kwargs)
 
     def passage_embed(self, texts: Iterable[str], **kwargs: Any) -> Iterable[NumpyArray]:
-        self.current_task_id = self.PASSAGE_TASK
-        yield from super().embed(texts, **kwargs)
+        yield from super().embed(texts, task_id=self.PASSAGE_TASK, **kwargs)
 
 
 class JinaEmbeddingV3Worker(OnnxTextEmbeddingWorker):
@@ -90,11 +97,15 @@ class JinaEmbeddingV3Worker(OnnxTextEmbeddingWorker):
         cache_dir: str,
         **kwargs: Any,
     ) -> JinaEmbeddingV3:
-        model = JinaEmbeddingV3(
+        return JinaEmbeddingV3(
             model_name=model_name,
             cache_dir=cache_dir,
             threads=1,
             **kwargs,
         )
-        model.current_task_id = kwargs["task_id"]
-        return model
+
+    def process(self, items: Iterable[tuple[int, Any]]) -> Iterable[tuple[int, OnnxOutputContext]]:
+        self.model: JinaEmbeddingV3  # mypy complaints `self.model` does not have `default_task_id`
+        for idx, batch in items:
+            onnx_output = self.model.onnx_embed(batch, task_id=self.model.default_task_id)
+            yield idx, onnx_output

{fastembed_gpu-0.6.0 → fastembed_gpu-0.6.1}/fastembed/text/onnx_embedding.py

@@ -1,6 +1,5 @@
 from typing import Any, Iterable, Optional, Sequence, Type, Union
 
-import numpy as np
 from fastembed.common.types import NumpyArray, OnnxProvider
 from fastembed.common.onnx_model import OnnxOutputContext
 from fastembed.common.utils import define_cache_dir, normalize
@@ -21,6 +20,7 @@ supported_onnx_models: list[DenseModelDescription] = [
         sources=ModelSource(
             hf="Qdrant/fast-bge-base-en",
             url="https://storage.googleapis.com/qdrant-fastembed/fast-bge-base-en.tar.gz",
+            _deprecated_tar_struct=True,
         ),
         model_file="model_optimized.onnx",
     ),
@@ -36,6 +36,7 @@ supported_onnx_models: list[DenseModelDescription] = [
         sources=ModelSource(
             hf="qdrant/bge-base-en-v1.5-onnx-q",
             url="https://storage.googleapis.com/qdrant-fastembed/fast-bge-base-en-v1.5.tar.gz",
+            _deprecated_tar_struct=True,
         ),
         model_file="model_optimized.onnx",
     ),
@@ -63,6 +64,7 @@ supported_onnx_models: list[DenseModelDescription] = [
         sources=ModelSource(
             hf="Qdrant/bge-small-en",
             url="https://storage.googleapis.com/qdrant-fastembed/BAAI-bge-small-en.tar.gz",
+            _deprecated_tar_struct=True,
         ),
         model_file="model_optimized.onnx",
     ),
@@ -90,6 +92,7 @@ supported_onnx_models: list[DenseModelDescription] = [
         sources=ModelSource(
             hf="Qdrant/bge-small-zh-v1.5",
             url="https://storage.googleapis.com/qdrant-fastembed/fast-bge-small-zh-v1.5.tar.gz",
+            _deprecated_tar_struct=True,
         ),
         model_file="model_optimized.onnx",
     ),
@@ -309,7 +312,7 @@ class OnnxTextEmbedding(TextEmbeddingBase, OnnxTextModel[NumpyArray]):
             processed_embeddings = embeddings
         else:
             raise ValueError(f"Unsupported embedding shape: {embeddings.shape}")
-        return normalize(processed_embeddings).astype(np.float32)
+        return normalize(processed_embeddings)
 
     def load_onnx_model(self) -> None:
         self._load_onnx_model(

{fastembed_gpu-0.6.0 → fastembed_gpu-0.6.1}/fastembed/text/onnx_text_model.py

@@ -115,7 +115,7 @@ class OnnxTextModel(OnnxModel[T]):
             if not hasattr(self, "model") or self.model is None:
                 self.load_onnx_model()
             for batch in iter_batch(documents, batch_size):
-                yield from self._post_process_onnx_output(self.onnx_embed(batch))
+                yield from self._post_process_onnx_output(self.onnx_embed(batch, **kwargs))
         else:
             if parallel == 0:
                 parallel = os.cpu_count()

{fastembed_gpu-0.6.0 → fastembed_gpu-0.6.1}/fastembed/text/pooled_embedding.py

@@ -82,6 +82,7 @@ supported_pooled_models: list[DenseModelDescription] = [
         sources=ModelSource(
             hf="qdrant/multilingual-e5-large-onnx",
             url="https://storage.googleapis.com/qdrant-fastembed/fast-multilingual-e5-large.tar.gz",
+            _deprecated_tar_struct=True,
         ),
         model_file="model.onnx",
         additional_files=["model.onnx_data"],
@@ -115,7 +116,7 @@ class PooledEmbedding(OnnxTextEmbedding):
 
         embeddings = output.model_output
         attn_mask = output.attention_mask
-        return self.mean_pooling(embeddings, attn_mask).astype(np.float32)
+        return self.mean_pooling(embeddings, attn_mask)
 
 
 class PooledEmbeddingWorker(OnnxTextEmbeddingWorker):

{fastembed_gpu-0.6.0 → fastembed_gpu-0.6.1}/fastembed/text/pooled_normalized_embedding.py

@@ -1,6 +1,5 @@
 from typing import Any, Iterable, Type
 
-import numpy as np
 
 from fastembed.common.types import NumpyArray
 from fastembed.common.onnx_model import OnnxOutputContext
@@ -22,6 +21,7 @@ supported_pooled_normalized_models: list[DenseModelDescription] = [
         sources=ModelSource(
             url="https://storage.googleapis.com/qdrant-fastembed/sentence-transformers-all-MiniLM-L6-v2.tar.gz",
             hf="qdrant/all-MiniLM-L6-v2-onnx",
+            _deprecated_tar_struct=True,
         ),
         model_file="model.onnx",
     ),
@@ -144,7 +144,7 @@ class PooledNormalizedEmbedding(PooledEmbedding):
 
         embeddings = output.model_output
         attn_mask = output.attention_mask
-        return normalize(self.mean_pooling(embeddings, attn_mask)).astype(np.float32)
+        return normalize(self.mean_pooling(embeddings, attn_mask))
 
 
 class PooledNormalizedEmbeddingWorker(OnnxTextEmbeddingWorker):

{fastembed_gpu-0.6.0 → fastembed_gpu-0.6.1}/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "fastembed-gpu"
-version = "0.6.0"
+version = "0.6.1"
 description = "Fast, light, accurate library built for retrieval embedding generation"
 authors = ["Qdrant Team <info@qdrant.tech>", "NirantK <nirant.bits@gmail.com>"]
 license = "Apache License"