new: gpu package (#224)

* fix: stage GCS downloads instead of deleting the caller's cache dir

* fix: verify archive integrity and give each download its own staging dir

.github/ISSUE_TEMPLATE/bug-report.yml

@@ -39,7 +39,7 @@ body:
     attributes:
       label: FastEmbed version
       description: What version of FastEmbed are you running? python -c "import fastembed; print(fastembed.__version__)". If you're not on the latest, please upgrade and see if the problem persists.
-      placeholder: v0.5.1
+      placeholder: v0.7.4
     validations:
       required: true
   - type: dropdown

.github/dependabot.yml

@@ -0,0 +1,40 @@
+version: 2
+updates:
+  - package-ecosystem: "pip"
+    directory: "/"
+    schedule:
+      interval: "weekly"
+    open-pull-requests-limit: 5
+    groups:
+      security-updates:
+        applies-to: security-updates
+        patterns:
+          - "*"
+      version-updates:
+        applies-to: version-updates
+        update-types:
+          - "minor"
+          - "patch"
+        patterns:
+          - "*"
+    cooldown:
+      default-days: 7
+  - package-ecosystem: "github-actions"
+    directory: "/"
+    schedule:
+      interval: "weekly"
+    open-pull-requests-limit: 5
+    groups:
+      security-updates:
+        applies-to: security-updates
+        patterns:
+          - "*"
+      version-updates:
+        applies-to: version-updates
+        update-types:
+          - "minor"
+          - "patch"
+        patterns:
+          - "*"
+    cooldown:
+      default-days: 7

.github/workflows/ci.yml

@@ -10,12 +10,12 @@ jobs:
   deploy:
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v3
-      - uses: actions/setup-python@v4
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+      - uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7.0.0
         with:
           python-version: 3.x
       - run: echo "cache_id=$(date --utc '+%V')" >> $GITHUB_ENV
-      - uses: actions/cache@v3
+      - uses: actions/cache@55cc8345863c7cc4c66a329aec7e433d2d1c52a9 # v6.1.0
         with:
           key: mkdocs-material-${{ env.cache_id }}
           path: .cache

.github/workflows/python-publish.yml

@@ -21,11 +21,11 @@ jobs:
     runs-on: ubuntu-latest
 
     steps:
-    - uses: actions/checkout@v2
+    - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
     - name: Set up Python
-      uses: actions/setup-python@v2
+      uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7.0.0
       with:
-        python-version: '3.9.x'
+        python-version: '3.10.x'
     - name: Install dependencies
       run: |
         python -m pip install poetry
@@ -33,7 +33,7 @@ jobs:
     - name: Build package
       run: poetry build
     - name: Publish package
-      uses: pypa/gh-action-pypi-publish@27b31702a0e7fc50959f5ad993c78deac1bdfc29
+      uses: pypa/gh-action-pypi-publish@dc37677b2e1c63e2034f94d8a5b11f265b73ba33 # v1.14.2
       with:
         user: __token__
         password: ${{ secrets.PYPI_API_TOKEN }}

.github/workflows/python-tests.yml

@@ -1,9 +1,11 @@
 name: Tests
+run-name: Tests (gpu)
 
 on:
-  push:
-    branches: [ master, main, gpu ]
   pull_request:
+    branches: [ master, main, gpu ]
+  workflow_dispatch:
+
 
 env:
   CARGO_TERM_COLOR: always
@@ -14,24 +16,21 @@ jobs:
     strategy:
       matrix:
         python-version:
-          - '3.9.x'
           - '3.10.x'
           - '3.11.x'
           - '3.12.x'
           - '3.13.x'
         os:
           - ubuntu-latest
-          - macos-latest
-          - windows-latest
 
     runs-on: ${{ matrix.os }}
 
     name: Python ${{ matrix.python-version }} on ${{ matrix.os }} test
 
     steps:
-      - uses: actions/checkout@v3
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
       - name: Set up Python
-        uses: actions/setup-python@v5
+        uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7.0.0
         with:
           python-version: ${{ matrix.python-version }}
       - name: Install dependencies
@@ -41,5 +40,7 @@ jobs:
           poetry install --no-interaction --no-ansi --without dev,docs
 
       - name: Run pytest
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
         run: |
-          poetry run pytest 
+          poetry run pytest

.github/workflows/type-checkers.yml

@@ -8,16 +8,16 @@ jobs:
     strategy:
       fail-fast: true
       matrix:
-        python-version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
+        python-version: ["3.10", "3.11", "3.12", "3.13"]
         os: [ubuntu-latest]
 
     name: Python ${{ matrix.python-version }} test
 
     steps:
-      - uses: actions/checkout@v1
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
 
       - name: Set up Python ${{ matrix.python-version }}
-        uses: actions/setup-python@v2
+        uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7.0.0
         with:
           python-version: ${{ matrix.python-version }}
 
@@ -26,8 +26,6 @@ jobs:
           python -m pip install --upgrade pip poetry
           poetry install --no-interaction --no-ansi --without dev,docs,test
 
-          poetry run pip install "numpy<2.0.0"  # https://github.com/python/mypy/issues/17396
-
       - name: mypy
         run: |
             poetry run mypy fastembed \

LICENSE

@@ -186,7 +186,7 @@
       same "printed page" as the copyright notice for easier
       identification within third-party archives.
 
-   Copyright [yyyy] [name of copyright owner]
+   Copyright 2026 Qdrant Solutions GmbH
 
    Licensed under the Apache License, Version 2.0 (the "License");
    you may not use this file except in compliance with the License.

NOTICE

@@ -15,6 +15,8 @@ These models are developed by Jina (https://jina.ai/) and are subject to Jina AI
 This distribution includes the following Google models, each with its respective license:
 - vidore/colpali-v1.3
     - License: gemma
+- google/embeddinggemma-300m
+    - License: gemma
 
 Gemma is provided under and subject to the Gemma Terms of Use found at ai.google.dev/gemma/terms
 

README.md

@@ -63,6 +63,23 @@ embeddings = list(model.embed(documents))
 
 ```
 
+Dense text embedding can also be extended with models which are not in the list of supported models.
+
+```python
+from fastembed import TextEmbedding
+from fastembed.common.model_description import PoolingType, ModelSource
+
+TextEmbedding.add_custom_model(
+    model="intfloat/multilingual-e5-small",
+    pooling=PoolingType.MEAN,
+    normalization=True,
+    sources=ModelSource(hf="intfloat/multilingual-e5-small"),  # can be used with an `url` to load files from a private storage
+    dim=384,
+    model_file="onnx/model.onnx",  # can be used to load an already supported model with another optimization or quantization, e.g. onnx/model_O4.onnx
+)
+model = TextEmbedding(model_name="intfloat/multilingual-e5-small")
+embeddings = list(model.embed(documents))
+```
 
 
 ### 🔱 Sparse text embeddings
@@ -137,6 +154,27 @@ embeddings = list(model.embed(images))
 # ]
 ```
 
+### Late interaction multimodal models (ColPali)
+
+```python
+from fastembed import LateInteractionMultimodalEmbedding
+
+doc_images = [
+    "./path/to/qdrant_pdf_doc_1_screenshot.jpg",
+    "./path/to/colpali_pdf_doc_2_screenshot.jpg",
+]
+
+query = "What is Qdrant?"
+
+model = LateInteractionMultimodalEmbedding(model_name="Qdrant/colpali-v1.3-fp16")
+doc_images_embeddings = list(model.embed_image(doc_images))
+# shape (2, 1030, 128)
+# [array([[-0.03353882, -0.02090454, ..., -0.15576172, -0.07678223]], dtype=float32)]
+query_embedding = model.embed_text(query)
+# shape (1, 20, 128)
+# [array([[-0.00218201,  0.14758301, ...,  -0.02207947,  0.16833496]], dtype=float32)]
+```
+
 ### 🔄 Rerankers
 ```python
 from fastembed.rerank.cross_encoder import TextCrossEncoder
@@ -152,6 +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.
@@ -191,36 +246,36 @@ pip install qdrant-client[fastembed-gpu]
 You might have to use quotes ```pip install 'qdrant-client[fastembed]'``` on zsh.
 
 ```python
-from qdrant_client import QdrantClient
+from qdrant_client import QdrantClient, models
 
 # Initialize the client
 client = QdrantClient("localhost", port=6333) # For production
-# client = QdrantClient(":memory:") # For small experiments
+# client = QdrantClient(":memory:") # For experimentation
 
-# Prepare your documents, metadata, and IDs
-docs = ["Qdrant has Langchain integrations", "Qdrant also has Llama Index integrations"]
-metadata = [
-    {"source": "Langchain-docs"},
-    {"source": "Llama-index-docs"},
+model_name = "sentence-transformers/all-MiniLM-L6-v2"
+payload = [
+    {"document": "Qdrant has Langchain integrations", "source": "Langchain-docs", },
+    {"document": "Qdrant also has Llama Index integrations", "source": "LlamaIndex-docs"},
 ]
+docs = [models.Document(text=data["document"], model=model_name) for data in payload]
 ids = [42, 2]
 
-# If you want to change the model:
-# client.set_model("sentence-transformers/all-MiniLM-L6-v2")
-# List of supported models: https://qdrant.github.io/fastembed/examples/Supported_Models
-
-# Use the new add() instead of upsert()
-# This internally calls embed() of the configured embedding model
-client.add(
-    collection_name="demo_collection",
-    documents=docs,
-    metadata=metadata,
-    ids=ids
+client.create_collection(
+    "demo_collection",
+    vectors_config=models.VectorParams(
+        size=client.get_embedding_size(model_name), distance=models.Distance.COSINE)
 )
 
-search_result = client.query(
+client.upload_collection(
     collection_name="demo_collection",
-    query_text="This is a query document"
+    vectors=docs,
+    ids=ids,
+    payload=payload,
 )
+
+search_result = client.query_points(
+    collection_name="demo_collection",
+    query=models.Document(text="This is a query document", model=model_name)
+).points
 print(search_result)
 ```

experiments/if_splade_to_onnx.py

@@ -0,0 +1,134 @@
+"""Export an inference-free SPLADE document encoder to ONNX.
+
+Converts `opensearch-project/opensearch-neural-sparse-encoding-doc-v3-gte` (an MLM head
+over a GTE backbone) into an onnx model producing token logits, and assembles a model dir
+with everything fastembed's `IfSplade` needs: model.onnx, tokenizer files and idf.json.
+
+Usage:
+    python experiments/if_splade_to_onnx.py --output-dir models/opensearch-neural-sparse-encoding-doc-v3-gte
+"""
+
+import argparse
+import shutil
+from pathlib import Path
+
+import torch
+from huggingface_hub import hf_hub_download
+from transformers import AutoModelForMaskedLM, AutoTokenizer
+
+MODEL_ID = "opensearch-project/opensearch-neural-sparse-encoding-doc-v3-gte"
+# revision of the remote modeling code (Alibaba-NLP/new-impl), pinned in the model card
+CODE_REVISION = "40ced75c3017eb27626c9d4ea981bde21a2662f4"
+
+TOKENIZER_FILES = [
+    "config.json",
+    "tokenizer.json",
+    "tokenizer_config.json",
+    "special_tokens_map.json",
+    "vocab.txt",
+    "idf.json",
+]
+
+
+class LogitsOnly(torch.nn.Module):
+    def __init__(self, model: torch.nn.Module):
+        super().__init__()
+        self.model = model
+
+    def forward(self, input_ids: torch.Tensor, attention_mask: torch.Tensor) -> torch.Tensor:
+        return self.model(input_ids=input_ids, attention_mask=attention_mask).logits
+
+
+def export(model_id: str, output_dir: Path, opset: int = 14) -> Path:
+    output_dir.mkdir(parents=True, exist_ok=True)
+
+    model = AutoModelForMaskedLM.from_pretrained(
+        model_id, trust_remote_code=True, code_revision=CODE_REVISION
+    )
+    model.eval()
+    wrapped = LogitsOnly(model)
+
+    tokenizer = AutoTokenizer.from_pretrained(model_id)
+    dummy = tokenizer(
+        ["fastembed is a library", "onnx export"],
+        padding=True,
+        truncation=True,
+        return_tensors="pt",
+        return_token_type_ids=False,
+    )
+
+    onnx_path = output_dir / "model.onnx"
+    with torch.inference_mode():
+        torch.onnx.export(
+            wrapped,
+            (dummy["input_ids"], dummy["attention_mask"]),
+            f=onnx_path.as_posix(),
+            input_names=["input_ids", "attention_mask"],
+            output_names=["logits"],
+            dynamic_axes={
+                "input_ids": {0: "batch_size", 1: "sequence_length"},
+                "attention_mask": {0: "batch_size", 1: "sequence_length"},
+                "logits": {0: "batch_size", 1: "sequence_length"},
+            },
+            do_constant_folding=True,
+            opset_version=opset,
+            dynamo=False,
+        )
+
+    for file_name in TOKENIZER_FILES:
+        local_path = hf_hub_download(repo_id=model_id, filename=file_name)
+        shutil.copy(local_path, output_dir / file_name)
+
+    return onnx_path
+
+
+def parity_check(model_id: str, output_dir: Path) -> None:
+    import numpy as np
+    import onnxruntime as ort
+
+    model = AutoModelForMaskedLM.from_pretrained(
+        model_id, trust_remote_code=True, code_revision=CODE_REVISION
+    )
+    model.eval()
+    tokenizer = AutoTokenizer.from_pretrained(model_id)
+
+    documents = [
+        "Currently New York is rainy.",
+        "fastembed is a lightweight library for generating embeddings",
+        "hello world",
+    ]
+    features = tokenizer(
+        documents, padding=True, truncation=True, return_tensors="pt", return_token_type_ids=False
+    )
+
+    with torch.inference_mode():
+        torch_logits = model(**features).logits.numpy()
+
+    session = ort.InferenceSession(output_dir / "model.onnx")
+    onnx_logits = session.run(
+        ["logits"],
+        {
+            "input_ids": features["input_ids"].numpy(),
+            "attention_mask": features["attention_mask"].numpy(),
+        },
+    )[0]
+
+    max_diff = np.abs(torch_logits - onnx_logits).max()
+    print(f"max |torch - onnx| logits diff: {max_diff}")
+    assert max_diff < 1e-3, "onnx export does not match the torch model"
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--model-id", default=MODEL_ID)
+    parser.add_argument("--output-dir", default=f"models/{MODEL_ID.replace('/', '_')}", type=Path)
+    parser.add_argument("--opset", default=14, type=int)
+    args = parser.parse_args()
+
+    onnx_path = export(args.model_id, args.output_dir, args.opset)
+    print(f"Exported to {onnx_path}")
+    parity_check(args.model_id, args.output_dir)
+
+
+if __name__ == "__main__":
+    main()

fastembed/common/model_description.py

@@ -1,11 +1,17 @@
 from dataclasses import dataclass, field
-from typing import Optional, Any
+from enum import Enum
+from typing import Any
 
 
 @dataclass(frozen=True)
 class ModelSource:
-    hf: Optional[str] = None
-    url: Optional[str] = None
+    hf: str | None = None
+    url: str | None = 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:
@@ -27,8 +33,8 @@ class BaseModelDescription:
 
 @dataclass(frozen=True)
 class DenseModelDescription(BaseModelDescription):
-    dim: Optional[int] = None
-    tasks: Optional[dict[str, Any]] = None
+    dim: int | None = None
+    tasks: dict[str, Any] | None = field(default_factory=dict)
 
     def __post_init__(self) -> None:
         assert self.dim is not None, "dim is required for dense model description"
@@ -36,5 +42,12 @@ class DenseModelDescription(BaseModelDescription):
 
 @dataclass(frozen=True)
 class SparseModelDescription(BaseModelDescription):
-    requires_idf: Optional[bool] = None
-    vocab_size: Optional[int] = None
+    requires_idf: bool | None = None
+    vocab_size: int | None = None
+
+
+class PoolingType(str, Enum):
+    CLS = "CLS"
+    MEAN = "MEAN"
+    LAST_TOKEN = "LAST_TOKEN"
+    DISABLED = "DISABLED"

fastembed/common/model_management.py

@@ -1,10 +1,14 @@
 import os
 import time
+import gzip
 import json
 import shutil
 import tarfile
-from pathlib import Path
-from typing import Any, Optional, Union, TypeVar, Generic
+import tempfile
+import contextlib
+from copy import deepcopy
+from pathlib import Path, PureWindowsPath
+from typing import Any, TypeVar, Generic
 
 import requests
 from huggingface_hub import snapshot_download, model_info, list_repo_tree
@@ -20,6 +24,8 @@ from fastembed.common.model_description import BaseModelDescription
 
 T = TypeVar("T", bound=BaseModelDescription)
 
+_DOWNLOAD_CHUNK_SIZE = 256 * 1024
+
 
 class ModelManagement(Generic[T]):
     METADATA_FILE = "files_metadata.json"
@@ -33,6 +39,31 @@ class ModelManagement(Generic[T]):
         """
         raise NotImplementedError()
 
+    @classmethod
+    def add_custom_model(
+        cls,
+        *args: Any,
+        **kwargs: Any,
+    ) -> None:
+        """Add a custom model to the existing embedding classes based on the passed model descriptions
+
+        Model description dict should contain the fields same as in one of the model descriptions presented
+         in fastembed.common.model_description
+
+         E.g. for BaseModelDescription:
+              model: str
+              sources: ModelSource
+              model_file: str
+              description: str
+              license: str
+              size_in_GB: float
+              additional_files: list[str]
+
+        Returns:
+            None
+        """
+        raise NotImplementedError()
+
     @classmethod
     def _list_supported_models(cls) -> list[T]:
         raise NotImplementedError()
@@ -71,9 +102,7 @@ class ModelManagement(Generic[T]):
             str: The path to the downloaded file.
         """
 
-        if os.path.exists(output_path):
-            return output_path
-        response = requests.get(url, stream=True)
+        response = requests.get(url, stream=True, timeout=(10, 120))
 
         # Handle HTTP errors
         if response.status_code == 403:
@@ -81,6 +110,8 @@ class ModelManagement(Generic[T]):
                 "Authentication Error: You do not have permission to access this resource. "
                 "Please check your credentials."
             )
+        # Otherwise an error page gets written out as though it were the archive.
+        response.raise_for_status()
 
         # Get the total size of the file
         total_size_in_bytes = int(response.headers.get("content-length", 0))
@@ -98,7 +129,7 @@ class ModelManagement(Generic[T]):
             disable=not show_progress,
         ) as progress_bar:
             with open(output_path, "wb") as file:
-                for chunk in response.iter_content(chunk_size=1024):
+                for chunk in response.iter_content(chunk_size=_DOWNLOAD_CHUNK_SIZE):
                     if chunk:  # Filter out keep-alive new chunks
                         progress_bar.update(len(chunk))
                         file.write(chunk)
@@ -154,8 +185,8 @@ class ModelManagement(Generic[T]):
 
         def _collect_file_metadata(
             model_dir: Path, repo_files: list[RepoFile]
-        ) -> dict[str, dict[str, Union[int, str]]]:
-            meta: dict[str, dict[str, Union[int, str]]] = {}
+        ) -> dict[str, dict[str, int | str]]:
+            meta: dict[str, dict[str, int | str]] = {}
             file_info_map = {f.path: f for f in repo_files}
             for file_path in model_dir.rglob("*"):
                 if file_path.is_file() and file_path.name != cls.METADATA_FILE:
@@ -167,9 +198,7 @@ class ModelManagement(Generic[T]):
                         }
             return meta
 
-        def _save_file_metadata(
-            model_dir: Path, meta: dict[str, dict[str, Union[int, str]]]
-        ) -> None:
+        def _save_file_metadata(model_dir: Path, meta: dict[str, dict[str, int | str]]) -> None:
             try:
                 if not model_dir.exists():
                     model_dir.mkdir(parents=True, exist_ok=True)
@@ -199,11 +228,6 @@ class ModelManagement(Generic[T]):
                     logger.warning(
                         "Local file sizes do not match the metadata."
                     )  # do not raise, still make an attempt to load the model
-            else:
-                logger.warning(
-                    "Metadata file not found. Proceeding without checking local files."
-                )  # if users have downloaded models from hf manually, or they're updating from previous versions of
-                # fastembed
             result = snapshot_download(
                 repo_id=hf_source_repo,
                 allow_patterns=allow_patterns,
@@ -267,12 +291,18 @@ class ModelManagement(Generic[T]):
         """
         Decompresses a .tar.gz file to a cache directory.
 
+        Nothing is deleted on failure, since `cache_dir` may hold more than this archive.
+        Cleaning up a partial extraction is the caller's job.
+
         Args:
             targz_path (str): Path to the .tar.gz file.
             cache_dir (str): Path to the cache directory.
 
         Returns:
             cache_dir (str): Path to the cache directory.
+
+        Raises:
+            ValueError: If the archive is missing, corrupt, or holds an unsafe member.
         """
         # Check if targz_path exists and is a file
         if not os.path.isfile(targz_path):
@@ -285,60 +315,68 @@ class ModelManagement(Generic[T]):
         try:
             # Open the tar.gz file
             with tarfile.open(targz_path, "r:gz") as tar:
-                # Extract all files into the cache directory
-                tar.extractall(
-                    path=cache_dir,
-                )
-        except tarfile.TarError as e:
-            # If any error occurs while opening or extracting the tar.gz file,
-            # delete the cache directory (if it was created in this function)
-            # and raise the error again
-            if "tmp" in cache_dir:
-                shutil.rmtree(cache_dir)
-            raise ValueError(f"An error occurred while decompressing {targz_path}: {e}")
+                if hasattr(tarfile, "data_filter"):
+                    tar.extractall(path=cache_dir, filter="data")
+                else:
+                    # No PEP 706 filter before 3.10.12, so vet the members by hand.
+                    members = tar.getmembers()
+                    for member in members:
+                        cls._validate_tar_member(member)
+                    tar.extractall(path=cache_dir, members=members)
+                # tarfile stops at the end-of-archive marker, short of the gzip trailer, so
+                # the CRC is only checked if the rest of the stream is read.
+                while tar.fileobj.read(1 << 20):
+                    pass
+        except (tarfile.TarError, ValueError, EOFError, gzip.BadGzipFile) as e:
+            # gzip raises EOFError for a truncated stream and BadGzipFile for a corrupted one.
+            raise ValueError(f"An error occurred while decompressing {targz_path}: {e}") from e
 
         return cache_dir
 
+    @staticmethod
+    def _is_unsafe_tar_path(path: str) -> bool:
+        """Checks whether a tar member name or link target may escape the extraction dir.
+
+        Lexical on purpose: resolving against the extraction directory is unsound before
+        extraction, since `link/../escape` only escapes once an earlier member has been
+        written as a symlink. Any `..` component is therefore rejected outright.
+        """
+        # PureWindowsPath splits on both separators, so `root` covers POSIX "/evil" as
+        # well as "\\evil", which escapes on Windows without being absolute.
+        windows_path = PureWindowsPath(path)
+        return bool(windows_path.drive or windows_path.root) or ".." in windows_path.parts
+
+    @classmethod
+    def _validate_tar_member(cls, member: tarfile.TarInfo) -> None:
+        """Raises ValueError if a member could write outside the extraction directory."""
+        if cls._is_unsafe_tar_path(member.name):
+            raise ValueError(f"Unsafe tar member path: {member.name}")
+
+        if member.issym() or member.islnk():
+            if cls._is_unsafe_tar_path(member.linkname):
+                raise ValueError(f"Unsafe tar link target: {member.name} -> {member.linkname}")
+        elif not (member.isfile() or member.isdir()):
+            # Devices, fifos and the like have no place in a model archive.
+            raise ValueError(f"Unsupported tar member type: {member.name}")
+
     @classmethod
     def retrieve_model_gcs(
         cls,
         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
 
         # check if the model_dir and the model files are both present for macOS
         if model_dir.exists() and len(list(model_dir.glob("*"))) > 0:
             return model_dir
 
-        if model_tmp_dir.exists():
-            shutil.rmtree(model_tmp_dir)
-
-        cache_tmp_dir.mkdir(parents=True, exist_ok=True)
-
-        model_tar_gz = Path(cache_dir) / f"{fast_model_name}.tar.gz"
-
-        if model_tar_gz.exists():
-            model_tar_gz.unlink()
-
-        if not local_files_only:
-            cls.download_file_from_gcs(
-                source_url,
-                output_path=str(model_tar_gz),
-            )
-
-            cls.decompress_to_cache(targz_path=str(model_tar_gz), cache_dir=str(cache_tmp_dir))
-            assert model_tmp_dir.exists(), f"Could not find {model_tmp_dir} in {cache_tmp_dir}"
-
-            model_tar_gz.unlink()
-            # Rename from tmp to final name is atomic
-            model_tmp_dir.rename(model_dir)
-        else:
+        if local_files_only:
             logger.error(
                 f"Could not find the model tar.gz file at {model_dir} and local_files_only=True."
             )
@@ -346,6 +384,45 @@ class ModelManagement(Generic[T]):
                 f"Could not find the model tar.gz file at {model_dir} and local_files_only=True."
             )
 
+        if cache_tmp_dir.is_symlink():
+            raise ValueError(
+                f"{cache_tmp_dir} is a symlink, refusing to stage downloads through it"
+            )
+        cache_tmp_dir.mkdir(parents=True, exist_ok=True)
+
+        # The archive and everything extracted from it go in a directory of this attempt's own,
+        # so removing it undoes the attempt without touching any other download of the model.
+        staging_dir = Path(tempfile.mkdtemp(dir=cache_tmp_dir, prefix=f"{fast_model_name}-"))
+        try:
+            model_tar_gz = staging_dir / f"{fast_model_name}.tar.gz"
+            cls.download_file_from_gcs(
+                source_url,
+                output_path=str(model_tar_gz),
+            )
+
+            cls.decompress_to_cache(targz_path=str(model_tar_gz), cache_dir=str(staging_dir))
+
+            model_tmp_dir = staging_dir / fast_model_name
+            if not model_tmp_dir.is_dir() or model_tmp_dir.is_symlink():
+                raise ValueError(
+                    f"The archive from {source_url} has no {fast_model_name} directory"
+                )
+
+            # Replace a stale empty model_dir, which Windows will not rename onto. rmdir leaves
+            # anything else alone, including one another download has just filled.
+            with contextlib.suppress(OSError):
+                model_dir.rmdir()
+
+            try:
+                # Rename from the staging dir to the final name is atomic
+                model_tmp_dir.rename(model_dir)
+            except OSError:
+                # Another download of the same model finished first, so keep its copy.
+                if not (model_dir.is_dir() and any(model_dir.iterdir())):
+                    raise
+        finally:
+            shutil.rmtree(staging_dir, ignore_errors=True)
+
         return model_dir
 
     @classmethod
@@ -375,21 +452,47 @@ class ModelManagement(Generic[T]):
             Path: The path to the downloaded model directory.
         """
         local_files_only = kwargs.get("local_files_only", False)
-        specific_model_path: Optional[str] = kwargs.pop("specific_model_path", None)
+        hf_offline = os.environ.get("HF_HUB_OFFLINE", "").strip().upper()
+        if not local_files_only and hf_offline in {"1", "TRUE", "YES", "ON"}:
+            local_files_only = True
+            kwargs["local_files_only"] = True
+        specific_model_path: str | None = kwargs.pop("specific_model_path", None)
         if specific_model_path:
             return Path(specific_model_path)
         retries = 1 if local_files_only else retries
         hf_source = model.sources.hf
         url_source = model.sources.url
 
+        extra_patterns = [model.model_file]
+        extra_patterns.extend(model.additional_files)
+
+        if hf_source:
+            try:
+                cache_kwargs = deepcopy(kwargs)
+                cache_kwargs["local_files_only"] = True
+                resolved_path = Path(
+                    cls.download_files_from_huggingface(
+                        hf_source,
+                        cache_dir=cache_dir,
+                        extra_patterns=extra_patterns,
+                        **cache_kwargs,
+                    )
+                )
+                if (resolved_path / model.model_file).exists() and all(
+                    (resolved_path / file).exists() for file in extra_patterns
+                ):
+                    return resolved_path
+            except Exception:
+                pass
+            finally:
+                enable_progress_bars()
+
         sleep = 3.0
         while retries > 0:
             retries -= 1
 
-            if hf_source:
-                extra_patterns = [model.model_file]
-                extra_patterns.extend(model.additional_files)
-
+            if hf_source and not local_files_only:
+                # we have already tried loading with `local_files_only=True` via hf and we failed
                 try:
                     return Path(
                         cls.download_files_from_huggingface(
@@ -413,6 +516,7 @@ 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:
@@ -421,11 +525,12 @@ class ModelManagement(Generic[T]):
 
             if local_files_only:
                 logger.error("Could not find model in cache_dir")
+                break
             else:
                 logger.error(
                     f"Could not download model from either source, sleeping for {sleep} seconds, {retries} retries left."
                 )
-            time.sleep(sleep)
-            sleep *= 3
+                time.sleep(sleep)
+                sleep *= 3
 
         raise ValueError(f"Could not load model {model.model} from any source.")

fastembed/common/onnx_model.py

@@ -1,7 +1,7 @@
 import warnings
 from dataclasses import dataclass
 from pathlib import Path
-from typing import Any, Generic, Iterable, Optional, Sequence, Type, TypeVar
+from typing import Any, Generic, Iterable, Sequence, Type, TypeVar
 
 import numpy as np
 import onnxruntime as ort
@@ -9,7 +9,7 @@ import onnxruntime as ort
 from numpy.typing import NDArray
 from tokenizers import Tokenizer
 
-from fastembed.common.types import OnnxProvider, NumpyArray
+from fastembed.common.types import OnnxProvider, NumpyArray, Device
 from fastembed.parallel_processor import Worker
 
 # Holds type of the embedding result
@@ -19,21 +19,45 @@ T = TypeVar("T")
 @dataclass
 class OnnxOutputContext:
     model_output: NumpyArray
-    attention_mask: Optional[NDArray[np.int64]] = None
-    input_ids: Optional[NDArray[np.int64]] = None
+    attention_mask: NDArray[np.int64] | None = None
+    input_ids: NDArray[np.int64] | None = None
+    metadata: dict[str, Any] | None = None
 
 
 class OnnxModel(Generic[T]):
+    EXPOSED_SESSION_OPTIONS = ("enable_cpu_mem_arena",)
+
     @classmethod
     def _get_worker_class(cls) -> Type["EmbeddingWorker[T]"]:
         raise NotImplementedError("Subclasses must implement this method")
 
-    def _post_process_onnx_output(self, output: OnnxOutputContext) -> Iterable[T]:
+    def _get_worker_init_kwargs(self) -> dict[str, Any]:
+        """Additional kwargs a worker process needs to reconstruct this model.
+
+        Workers are started with `spawn`/`forkserver`, hence they don't inherit class-level state
+        which has been set up in runtime, e.g. models registered via `add_custom_model`.
+        Such state has to be shipped to the workers explicitly.
+
+        Returns:
+            dict[str, Any]: kwargs to pass to `_get_worker_class().init_embedding`.
+        """
+        return {}
+
+    def _post_process_onnx_output(self, output: OnnxOutputContext, **kwargs: Any) -> Iterable[T]:
+        """Post-process the ONNX model output to convert it into a usable format.
+
+        Args:
+            output (OnnxOutputContext): The raw output from the ONNX model.
+            **kwargs: Additional keyword arguments that may be needed by specific implementations.
+
+        Returns:
+            Iterable[T]: Post-processed output as an iterable of type T.
+        """
         raise NotImplementedError("Subclasses must implement this method")
 
     def __init__(self) -> None:
-        self.model: Optional[ort.InferenceSession] = None
-        self.tokenizer: Optional[Tokenizer] = None
+        self.model: ort.InferenceSession | None = None
+        self.tokenizer: Tokenizer | None = None
 
     def _preprocess_onnx_input(
         self, onnx_input: dict[str, NumpyArray], **kwargs: Any
@@ -47,24 +71,30 @@ class OnnxModel(Generic[T]):
         self,
         model_dir: Path,
         model_file: str,
-        threads: Optional[int],
-        providers: Optional[Sequence[OnnxProvider]] = None,
-        cuda: bool = False,
-        device_id: Optional[int] = None,
+        threads: int | None,
+        providers: Sequence[OnnxProvider] | None = None,
+        cuda: bool | Device = Device.AUTO,
+        device_id: int | None = None,
+        extra_session_options: dict[str, Any] | None = None,
     ) -> None:
         model_path = model_dir / model_file
         # List of Execution Providers: https://onnxruntime.ai/docs/execution-providers
+        available_providers = ort.get_available_providers()
+        cuda_available = "CUDAExecutionProvider" in available_providers
+        explicit_cuda = cuda is True or cuda == Device.CUDA
 
-        if cuda and providers is not None:
+        if explicit_cuda and providers is not None:
             warnings.warn(
-                f"`cuda` and `providers` are mutually exclusive parameters, cuda: {cuda}, providers: {providers}",
+                f"`cuda` and `providers` are mutually exclusive parameters, "
+                f"cuda: {cuda}, providers: {providers}. If you'd like to use providers, cuda should be one of "
+                f"[False, Device.CPU, Device.AUTO].",
                 category=UserWarning,
                 stacklevel=6,
             )
 
         if providers is not None:
             onnx_providers = list(providers)
-        elif cuda:
+        elif explicit_cuda or (cuda == Device.AUTO and cuda_available):
             if device_id is None:
                 onnx_providers = ["CUDAExecutionProvider"]
             else:
@@ -72,7 +102,6 @@ class OnnxModel(Generic[T]):
         else:
             onnx_providers = ["CPUExecutionProvider"]
 
-        available_providers = ort.get_available_providers()
         requested_provider_names: list[str] = []
         for provider in onnx_providers:
             # check providers available
@@ -90,6 +119,9 @@ class OnnxModel(Generic[T]):
             so.intra_op_num_threads = threads
             so.inter_op_num_threads = threads
 
+        if extra_session_options is not None:
+            self.add_extra_session_options(so, extra_session_options)
+
         self.model = ort.InferenceSession(
             str(model_path), providers=onnx_providers, sess_options=so
         )
@@ -104,6 +136,38 @@ class OnnxModel(Generic[T]):
                     RuntimeWarning,
                 )
 
+    @classmethod
+    def _select_exposed_session_options(cls, model_kwargs: dict[str, Any]) -> dict[str, Any]:
+        """A convenience method to select the exposed session options in models
+
+        Args:
+            model_kwargs (dict[str, Any]): The model kwargs.
+
+        Returns:
+            dict[str, Any]: a dict with filtered exposed session options.
+        """
+        return {k: v for k, v in model_kwargs.items() if k in cls.EXPOSED_SESSION_OPTIONS}
+
+    @classmethod
+    def add_extra_session_options(
+        cls, session_options: ort.SessionOptions, extra_options: dict[str, Any]
+    ) -> None:
+        """Add extra session options to the existing options object in-place
+
+        Args:
+            session_options (ort.SessionOptions): The existing session options object.
+            extra_options (dict[str, Any]): The extra session options available in cls.EXPOSED_SESSION_OPTIONS.
+
+        Returns:
+            None
+        """
+        for option in extra_options:
+            assert (
+                option in cls.EXPOSED_SESSION_OPTIONS
+            ), f"{option} is unknown or not exposed (exposed options: {cls.EXPOSED_SESSION_OPTIONS})"
+        if "enable_cpu_mem_arena" in extra_options:
+            session_options.enable_cpu_mem_arena = extra_options["enable_cpu_mem_arena"]
+
     def load_onnx_model(self) -> None:
         raise NotImplementedError("Subclasses must implement this method")
 

fastembed/common/preprocessor_utils.py

@@ -1,5 +1,6 @@
 import json
-from typing import Any
+import sys
+from typing import Any, Iterator
 from pathlib import Path
 
 from tokenizers import AddedToken, Tokenizer
@@ -8,9 +9,10 @@ from fastembed.image.transform.operators import Compose
 
 
 def load_special_tokens(model_dir: Path) -> dict[str, Any]:
+    """Read special_tokens_map.json, treating an absent file as an empty map."""
     tokens_map_path = model_dir / "special_tokens_map.json"
     if not tokens_map_path.exists():
-        raise ValueError(f"Could not find special_tokens_map.json in {model_dir}")
+        return {}
 
     with open(str(tokens_map_path)) as tokens_map_file:
         tokens_map = json.load(tokens_map_file)
@@ -18,11 +20,58 @@ def load_special_tokens(model_dir: Path) -> dict[str, Any]:
     return tokens_map
 
 
-def load_tokenizer(model_dir: Path) -> tuple[Tokenizer, dict[str, int]]:
-    config_path = model_dir / "config.json"
-    if not config_path.exists():
-        raise ValueError(f"Could not find config.json in {model_dir}")
+def iter_special_tokens(tokens_map: dict[str, Any]) -> Iterator[str | dict[str, Any]]:
+    """Yield the individual tokens declared in a special tokens map.
 
+    Most keys hold one token, but `additional_special_tokens` holds a list of them.
+    """
+    for value in tokens_map.values():
+        if isinstance(value, list):
+            yield from value
+        else:
+            yield value
+
+
+def _valid_context(value: Any) -> int | None:
+    """Return `value` if it can be used as a truncation limit, `None` otherwise.
+
+    Config files do not always carry a real limit: transformers writes `model_max_length` as
+    1e30 when the value is unknown, and some repos ship a 0 or a null. `enable_truncation`
+    raises an `OverflowError` on the former and silently produces empty encodings on the
+    latter, so both are rejected here rather than passed through.
+    """
+    if isinstance(value, bool) or not isinstance(value, int):
+        return None
+    if not 0 < value <= sys.maxsize:
+        return None
+    return value
+
+
+def _resolve_max_context(tokenizer_config: dict[str, Any], model_dir: Path) -> int:
+    """Pick the truncation limit, preferring the stricter of the two tokenizer config keys.
+
+    `config.json:max_position_embeddings` deliberately is not used as a fallback: it is the size
+    of the position table, not the usable context, and the two differ per architecture, e.g.
+    roberta reports 514 for a usable 512.
+    """
+    candidates = [
+        context
+        for context in (
+            _valid_context(tokenizer_config.get("model_max_length")),
+            _valid_context(tokenizer_config.get("max_length")),
+        )
+        if context is not None
+    ]
+    if not candidates:
+        raise ValueError(
+            f"Could not determine the maximum context length for {model_dir}. Set a positive "
+            "`model_max_length` or `max_length` in tokenizer_config.json."
+        )
+
+    return min(candidates)
+
+
+def load_tokenizer(model_dir: Path) -> tuple[Tokenizer, dict[str, int]]:
     tokenizer_path = model_dir / "tokenizer.json"
     if not tokenizer_path.exists():
         raise ValueError(f"Could not find tokenizer.json in {model_dir}")
@@ -31,43 +80,62 @@ def load_tokenizer(model_dir: Path) -> tuple[Tokenizer, dict[str, int]]:
     if not tokenizer_config_path.exists():
         raise ValueError(f"Could not find tokenizer_config.json in {model_dir}")
 
-    with open(str(config_path)) as config_file:
-        config = json.load(config_file)
+    # config.json is optional: transformers v5 no longer writes it for every model.
+    config_path = model_dir / "config.json"
+    config: dict[str, Any] = {}
+    if config_path.exists():
+        with open(str(config_path)) as config_file:
+            config = json.load(config_file)
 
     with open(str(tokenizer_config_path)) as tokenizer_config_file:
         tokenizer_config = json.load(tokenizer_config_file)
-        assert (
-            "model_max_length" in tokenizer_config or "max_length" in tokenizer_config
-        ), "Models without model_max_length or max_length are not supported."
-        if "model_max_length" not in tokenizer_config:
-            max_context = tokenizer_config["max_length"]
-        elif "max_length" not in tokenizer_config:
-            max_context = tokenizer_config["model_max_length"]
-        else:
-            max_context = min(tokenizer_config["model_max_length"], tokenizer_config["max_length"])
+
+    max_context = _resolve_max_context(tokenizer_config, model_dir)
 
     tokens_map = load_special_tokens(model_dir)
 
     tokenizer = Tokenizer.from_file(str(tokenizer_path))
     tokenizer.enable_truncation(max_length=max_context)
-    tokenizer.enable_padding(
-        pad_id=config.get("pad_token_id", 0), pad_token=tokenizer_config["pad_token"]
-    )
 
-    for token in tokens_map.values():
+    # Registered before the padding is resolved: the map may name a pad token that
+    # tokenizer.json does not carry, and it only gets an id once it is added.
+    for token in iter_special_tokens(tokens_map):
         if isinstance(token, str):
             tokenizer.add_special_tokens([token])
         elif isinstance(token, dict):
             tokenizer.add_special_tokens([AddedToken(**token)])
 
-    special_token_to_id: dict[str, int] = {}
+    # Padding is always normalized to batch-longest. A serialized fixed length shorter than the
+    # truncation limit leaves longer encodings untouched, which produces ragged batches, and a
+    # fixed length equal to it pads every batch to the maximum. Direction and pad token metadata
+    # are taken from the serialized settings, since some models pad on the left.
+    padding = tokenizer.padding or {}
+    pad_token = padding.get("pad_token") or tokenizer_config.get("pad_token")
+    if pad_token is None:
+        raise ValueError(f"Could not find a pad token for {model_dir}")
 
-    for token in tokens_map.values():
-        if isinstance(token, str):
-            special_token_to_id[token] = tokenizer.token_to_id(token)
-        elif isinstance(token, dict):
-            token_str = token.get("content", "")
-            special_token_to_id[token_str] = tokenizer.token_to_id(token_str)
+    # The vocabulary is the last resort, not a hardcoded 0: that silently disagrees with
+    # `pad_token` for every model whose pad token is not the first entry.
+    pad_id = padding.get("pad_id", config.get("pad_token_id"))
+    if pad_id is None:
+        pad_id = tokenizer.token_to_id(pad_token)
+    if pad_id is None:
+        raise ValueError(f"Could not resolve an id for the pad token {pad_token!r} in {model_dir}")
+
+    tokenizer.enable_padding(
+        direction=padding.get("direction", "right"),
+        pad_id=pad_id,
+        pad_type_id=padding.get("pad_type_id", 0),
+        pad_token=pad_token,
+        pad_to_multiple_of=padding.get("pad_to_multiple_of"),
+        length=None,
+    )
+
+    special_token_to_id = {
+        token.content: token_id
+        for token_id, token in tokenizer.get_added_tokens_decoder().items()
+        if token.special
+    }
 
     return tokenizer, special_token_to_id
 

fastembed/common/types.py

@@ -1,24 +1,27 @@
+from enum import Enum
 from pathlib import Path
-import sys
-from PIL import Image
-from typing import Any, Union
+from typing import Any, TypeAlias
+
 import numpy as np
 from numpy.typing import NDArray
-
-if sys.version_info >= (3, 10):
-    from typing import TypeAlias
-else:
-    from typing_extensions import TypeAlias
+from PIL import Image
 
 
-PathInput: TypeAlias = Union[str, Path]
-ImageInput: TypeAlias = Union[PathInput, Image.Image]
+class Device(str, Enum):
+    CPU = "cpu"
+    CUDA = "cuda"
+    AUTO = "auto"
 
-OnnxProvider: TypeAlias = Union[str, tuple[str, dict[Any, Any]]]
-NumpyArray = Union[
-    NDArray[np.float32],
-    NDArray[np.float16],
-    NDArray[np.int8],
-    NDArray[np.int64],
-    NDArray[np.int32],
-]
+
+PathInput: TypeAlias = str | Path
+ImageInput: TypeAlias = PathInput | Image.Image
+
+OnnxProvider: TypeAlias = str | tuple[str, dict[Any, Any]]
+NumpyArray: TypeAlias = (
+    NDArray[np.float64]
+    | NDArray[np.float32]
+    | NDArray[np.float16]
+    | NDArray[np.int8]
+    | NDArray[np.int64]
+    | NDArray[np.int32]
+)

fastembed/common/utils.py

@@ -5,9 +5,10 @@ import tempfile
 import unicodedata
 from pathlib import Path
 from itertools import islice
-from typing import Iterable, Optional, TypeVar
+from typing import Iterable, TypeVar
 
 import numpy as np
+from numpy.typing import NDArray
 
 from fastembed.common.types import NumpyArray
 
@@ -22,6 +23,25 @@ def normalize(input_array: NumpyArray, p: int = 2, dim: int = 1, eps: float = 1e
     return normalized_array
 
 
+def mean_pooling(input_array: NumpyArray, attention_mask: NDArray[np.int64]) -> NumpyArray:
+    input_mask_expanded = np.expand_dims(attention_mask, axis=-1).astype(np.int64)
+    input_mask_expanded = np.tile(input_mask_expanded, (1, 1, input_array.shape[-1]))
+    sum_embeddings = np.sum(input_array * input_mask_expanded, axis=1)
+    sum_mask = np.sum(input_mask_expanded, axis=1)
+    pooled_embeddings = sum_embeddings / np.maximum(sum_mask, 1e-9)
+    return pooled_embeddings
+
+
+def last_token_pooling(input_array: NumpyArray, attention_mask: NDArray[np.int64]) -> NumpyArray:
+    """Take the embedding of the last non-padding token of each sequence.
+
+    Locates the last position the attention mask marks as real, so it holds whichever
+    side the tokenizer pads on.
+    """
+    last_token_indices = attention_mask.shape[1] - 1 - np.argmax(attention_mask[:, ::-1], axis=1)
+    return input_array[np.arange(input_array.shape[0]), last_token_indices]
+
+
 def iter_batch(iterable: Iterable[T], size: int) -> Iterable[list[T]]:
     """
     >>> list(iter_batch([1,2,3,4,5], 3))
@@ -35,7 +55,7 @@ def iter_batch(iterable: Iterable[T], size: int) -> Iterable[list[T]]:
         yield b
 
 
-def define_cache_dir(cache_dir: Optional[str] = None) -> Path:
+def define_cache_dir(cache_dir: str | None = None) -> Path:
     """
     Define the cache directory for fastembed
     """

fastembed/config/dense_models.json

@@ -1,16 +0,0 @@
-{
-  "models": [
-    {
-      "model": "BAAI/bge-base-en",
-      "dim": 768,
-      "description": "Text embeddings, Unimodal (text), English...",
-      "license": "mit",
-      "size_in_GB": 0.42,
-      "sources": {
-        "hf": "Qdrant/fast-bge-base-en",
-        "url": "https://storage.googleapis.com/qdrant-fastembed/fast-bge-base-en.tar.gz"
-      },
-      "model_file": "model_optimized.onnx"
-    }
-  ]
-}

fastembed/embedding.py

@@ -1,4 +1,4 @@
-from typing import Optional, Any
+from typing import Any
 
 from loguru import logger
 
@@ -17,8 +17,8 @@ class JinaEmbedding(TextEmbedding):
     def __init__(
         self,
         model_name: str = "jinaai/jina-embeddings-v2-base-en",
-        cache_dir: Optional[str] = None,
-        threads: Optional[int] = None,
+        cache_dir: str | None = None,
+        threads: int | None = None,
         **kwargs: Any,
     ):
         super().__init__(model_name, cache_dir, threads, **kwargs)

fastembed/image/image_embedding.py

@@ -1,15 +1,21 @@
-from typing import Any, Iterable, Optional, Sequence, Type, Union
+from typing import Any, Iterable, Sequence, Type
 from dataclasses import asdict
 
-from fastembed.common.types import NumpyArray
+from fastembed.common.types import NumpyArray, Device
 from fastembed.common import ImageInput, OnnxProvider
 from fastembed.image.image_embedding_base import ImageEmbeddingBase
 from fastembed.image.onnx_embedding import OnnxImageEmbedding
+from fastembed.image.normalized_embedding import NormalizedEmbedding
+from fastembed.image.siglip_embedding import SiglipOnnxImageEmbedding
 from fastembed.common.model_description import DenseModelDescription
 
 
 class ImageEmbedding(ImageEmbeddingBase):
-    EMBEDDINGS_REGISTRY: list[Type[ImageEmbeddingBase]] = [OnnxImageEmbedding]
+    EMBEDDINGS_REGISTRY: list[Type[ImageEmbeddingBase]] = [
+        OnnxImageEmbedding,
+        NormalizedEmbedding,
+        SiglipOnnxImageEmbedding,
+    ]
 
     @classmethod
     def list_supported_models(cls) -> list[dict[str, Any]]:
@@ -48,11 +54,11 @@ class ImageEmbedding(ImageEmbeddingBase):
     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,
+        cache_dir: str | None = None,
+        threads: int | None = None,
+        providers: Sequence[OnnxProvider] | None = None,
+        cuda: bool | Device = Device.AUTO,
+        device_ids: list[int] | None = None,
         lazy_load: bool = False,
         **kwargs: Any,
     ):
@@ -77,11 +83,45 @@ class ImageEmbedding(ImageEmbeddingBase):
             "Please check the supported models using `ImageEmbedding.list_supported_models()`"
         )
 
+    @property
+    def embedding_size(self) -> int:
+        """Get the embedding size of the current model"""
+        if self._embedding_size is None:
+            self._embedding_size = self.get_embedding_size(self.model_name)
+        return self._embedding_size
+
+    @classmethod
+    def get_embedding_size(cls, model_name: str) -> int:
+        """Get the embedding size of the passed model
+
+        Args:
+            model_name (str): The name of the model to get embedding size for.
+
+        Returns:
+            int: The size of the embedding.
+
+        Raises:
+            ValueError: If the model name is not found in the supported models.
+        """
+        descriptions = cls._list_supported_models()
+        embedding_size: int | None = None
+        for description in descriptions:
+            if description.model.lower() == model_name.lower():
+                embedding_size = description.dim
+                break
+        if embedding_size is None:
+            model_names = [description.model for description in descriptions]
+            raise ValueError(
+                f"Embedding size for model {model_name} was None. "
+                f"Available model names: {model_names}"
+            )
+        return embedding_size
+
     def embed(
         self,
-        images: Union[ImageInput, Iterable[ImageInput]],
+        images: ImageInput | Iterable[ImageInput],
         batch_size: int = 16,
-        parallel: Optional[int] = None,
+        parallel: int | None = None,
         **kwargs: Any,
     ) -> Iterable[NumpyArray]:
         """

fastembed/image/image_embedding_base.py

@@ -1,4 +1,4 @@
-from typing import Iterable, Optional, Any, Union
+from typing import Iterable, Any
 
 from fastembed.common.model_description import DenseModelDescription
 from fastembed.common.types import NumpyArray
@@ -10,20 +10,21 @@ class ImageEmbeddingBase(ModelManagement[DenseModelDescription]):
     def __init__(
         self,
         model_name: str,
-        cache_dir: Optional[str] = None,
-        threads: Optional[int] = None,
+        cache_dir: str | None = None,
+        threads: int | None = None,
         **kwargs: Any,
     ):
         self.model_name = model_name
         self.cache_dir = cache_dir
         self.threads = threads
         self._local_files_only = kwargs.pop("local_files_only", False)
+        self._embedding_size: int | None = None
 
     def embed(
         self,
-        images: Union[ImageInput, Iterable[ImageInput]],
+        images: ImageInput | Iterable[ImageInput],
         batch_size: int = 16,
-        parallel: Optional[int] = None,
+        parallel: int | None = None,
         **kwargs: Any,
     ) -> Iterable[NumpyArray]:
         """
@@ -42,3 +43,13 @@ class ImageEmbeddingBase(ModelManagement[DenseModelDescription]):
             Iterable[NdArray]: The embeddings.
         """
         raise NotImplementedError()
+
+    @classmethod
+    def get_embedding_size(cls, model_name: str) -> int:
+        """Returns embedding size of the chosen model."""
+        raise NotImplementedError("Subclasses must implement this method")
+
+    @property
+    def embedding_size(self) -> int:
+        """Returns embedding size for the current model"""
+        raise NotImplementedError("Subclasses must implement this method")

fastembed/image/normalized_embedding.py

@@ -0,0 +1,69 @@
+from typing import Any, Iterable, Type
+
+
+from fastembed.common.types import NumpyArray
+from fastembed.common.onnx_model import OnnxOutputContext
+from fastembed.common.utils import normalize
+from fastembed.image.onnx_embedding import OnnxImageEmbedding
+from fastembed.image.onnx_image_model import ImageEmbeddingWorker
+from fastembed.common.model_description import DenseModelDescription, ModelSource
+
+supported_normalized_models: list[DenseModelDescription] = [
+    DenseModelDescription(
+        model="nomic-ai/nomic-embed-vision-v1.5",
+        dim=768,
+        description="Image embeddings, Multimodal (text&image), 2024 year",
+        license="apache-2.0",
+        size_in_GB=0.37,
+        sources=ModelSource(hf="nomic-ai/nomic-embed-vision-v1.5"),
+        model_file="onnx/model.onnx",
+    ),
+    DenseModelDescription(
+        model="nomic-ai/nomic-embed-vision-v1.5-Q",
+        dim=768,
+        description="Image embeddings, Multimodal (text&image), 2024 year",
+        license="apache-2.0",
+        size_in_GB=0.1,
+        sources=ModelSource(hf="nomic-ai/nomic-embed-vision-v1.5"),
+        model_file="onnx/model_quantized.onnx",
+    ),
+]
+
+
+class NormalizedEmbedding(OnnxImageEmbedding):
+    @classmethod
+    def _list_supported_models(cls) -> list[DenseModelDescription]:
+        """
+        Lists the supported models.
+
+        Returns:
+            list[DenseModelDescription]: A list of DenseModelDescription objects containing the model information.
+        """
+        return supported_normalized_models
+
+    @classmethod
+    def _get_worker_class(cls) -> Type["ImageEmbeddingWorker[NumpyArray]"]:
+        return NormalizedEmbeddingWorker
+
+    def _post_process_onnx_output(
+        self, output: OnnxOutputContext, **kwargs: Any
+    ) -> Iterable[NumpyArray]:
+        # The model emits last_hidden_state, which onnx_embed flattens to (batch, tokens * dim).
+        # Recover the token axis, take the CLS token (index 0) and normalize, matching the reference
+        # F.normalize(last_hidden_state[:, 0], p=2, dim=1).
+        dim = self.model_description.dim
+        assert dim is not None, "Model description is missing the embedding dim"
+        hidden_states = output.model_output.reshape(output.model_output.shape[0], -1, dim)
+        return normalize(hidden_states[:, 0])
+
+
+class NormalizedEmbeddingWorker(ImageEmbeddingWorker[NumpyArray]):
+    def init_embedding(
+        self, model_name: str, cache_dir: str, **kwargs: Any
+    ) -> NormalizedEmbedding:
+        return NormalizedEmbedding(
+            model_name=model_name,
+            cache_dir=cache_dir,
+            threads=1,
+            **kwargs,
+        )

fastembed/image/onnx_embedding.py

@@ -1,8 +1,7 @@
-from typing import Any, Iterable, Optional, Sequence, Type, Union
+from typing import Any, Iterable, Sequence, Type
 
-import numpy as np
 
-from fastembed.common.types import NumpyArray
+from fastembed.common.types import NumpyArray, Device
 from fastembed.common import ImageInput, OnnxProvider
 from fastembed.common.onnx_model import OnnxOutputContext
 from fastembed.common.utils import define_cache_dir, normalize
@@ -64,14 +63,14 @@ class OnnxImageEmbedding(ImageEmbeddingBase, OnnxImageModel[NumpyArray]):
     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,
+        cache_dir: str | None = None,
+        threads: int | None = None,
+        providers: Sequence[OnnxProvider] | None = None,
+        cuda: bool | Device = Device.AUTO,
+        device_ids: list[int] | None = None,
         lazy_load: bool = False,
-        device_id: Optional[int] = None,
-        specific_model_path: Optional[str] = None,
+        device_id: int | None = None,
+        specific_model_path: str | None = None,
         **kwargs: Any,
     ):
         """
@@ -83,10 +82,11 @@ class OnnxImageEmbedding(ImageEmbeddingBase, OnnxImageModel[NumpyArray]):
             threads (int, optional): The number of threads single onnxruntime session can use. Defaults to None.
             providers (Optional[Sequence[OnnxProvider]], optional): The list of onnxruntime providers to use.
                 Mutually exclusive with the `cuda` and `device_ids` arguments. Defaults to None.
-            cuda (bool, optional): Whether to use cuda for inference. Mutually exclusive with `providers`
-                Defaults to False.
+            cuda (Union[bool, Device], optional): Whether to use cuda for inference. Mutually exclusive with `providers`
+                Defaults to Device.AUTO.
             device_ids (Optional[list[int]], optional): The list of device ids to use for data parallel processing in
-                workers. Should be used with `cuda=True`, mutually exclusive with `providers`. Defaults to None.
+                workers. Should be used with `cuda` equals to `True`, `Device.AUTO` or `Device.CUDA`, mutually exclusive
+                with `providers`. Defaults to None.
             lazy_load (bool, optional): Whether to load the model during class initialization or on demand.
                 Should be set to True when using multiple-gpu and parallel encoding. Defaults to False.
             device_id (Optional[int], optional): The device id to use for loading the model in the worker process.
@@ -99,13 +99,14 @@ class OnnxImageEmbedding(ImageEmbeddingBase, OnnxImageModel[NumpyArray]):
         super().__init__(model_name, cache_dir, threads, **kwargs)
         self.providers = providers
         self.lazy_load = lazy_load
+        self._extra_session_options = self._select_exposed_session_options(kwargs)
 
         # List of device ids, that can be used for data parallel processing in workers
         self.device_ids = device_ids
         self.cuda = cuda
 
         # This device_id will be used if we need to load model in current process
-        self.device_id: Optional[int] = None
+        self.device_id: int | None = None
         if device_id is not None:
             self.device_id = device_id
         elif self.device_ids is not None:
@@ -113,11 +114,12 @@ class OnnxImageEmbedding(ImageEmbeddingBase, OnnxImageModel[NumpyArray]):
 
         self.model_description = self._get_model_description(model_name)
         self.cache_dir = str(define_cache_dir(cache_dir))
+        self._specific_model_path = specific_model_path
         self._model_dir = self.download_model(
             self.model_description,
             self.cache_dir,
             local_files_only=self._local_files_only,
-            specific_model_path=specific_model_path,
+            specific_model_path=self._specific_model_path,
         )
 
         if not self.lazy_load:
@@ -134,6 +136,7 @@ class OnnxImageEmbedding(ImageEmbeddingBase, OnnxImageModel[NumpyArray]):
             providers=self.providers,
             cuda=self.cuda,
             device_id=self.device_id,
+            extra_session_options=self._extra_session_options,
         )
 
     @classmethod
@@ -148,9 +151,9 @@ class OnnxImageEmbedding(ImageEmbeddingBase, OnnxImageModel[NumpyArray]):
 
     def embed(
         self,
-        images: Union[ImageInput, Iterable[ImageInput]],
+        images: ImageInput | Iterable[ImageInput],
         batch_size: int = 16,
-        parallel: Optional[int] = None,
+        parallel: int | None = None,
         **kwargs: Any,
     ) -> Iterable[NumpyArray]:
         """
@@ -178,6 +181,9 @@ class OnnxImageEmbedding(ImageEmbeddingBase, OnnxImageModel[NumpyArray]):
             providers=self.providers,
             cuda=self.cuda,
             device_ids=self.device_ids,
+            local_files_only=self._local_files_only,
+            specific_model_path=self._specific_model_path,
+            extra_session_options=self._extra_session_options,
             **kwargs,
         )
 
@@ -194,8 +200,10 @@ 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)
+    def _post_process_onnx_output(
+        self, output: OnnxOutputContext, **kwargs: Any
+    ) -> Iterable[NumpyArray]:
+        return normalize(output.model_output)
 
 
 class OnnxImageEmbeddingWorker(ImageEmbeddingWorker[NumpyArray]):

fastembed/image/onnx_image_model.py

@@ -2,13 +2,13 @@ import contextlib
 import os
 from multiprocessing import get_all_start_methods
 from pathlib import Path
-from typing import Any, Iterable, Optional, Sequence, Type, Union
+from typing import Any, Iterable, Sequence, Type
 
 import numpy as np
 from PIL import Image
 
 from fastembed.image.transform.operators import Compose
-from fastembed.common.types import NumpyArray
+from fastembed.common.types import NumpyArray, Device
 from fastembed.common import ImageInput, OnnxProvider
 from fastembed.common.onnx_model import EmbeddingWorker, OnnxModel, OnnxOutputContext, T
 from fastembed.common.preprocessor_utils import load_preprocessor
@@ -19,16 +19,27 @@ from fastembed.parallel_processor import ParallelWorkerPool
 
 
 class OnnxImageModel(OnnxModel[T]):
+    ONNX_OUTPUT_NAMES: list[str] | None = None
+
     @classmethod
     def _get_worker_class(cls) -> Type["ImageEmbeddingWorker[T]"]:
         raise NotImplementedError("Subclasses must implement this method")
 
-    def _post_process_onnx_output(self, output: OnnxOutputContext) -> Iterable[T]:
+    def _post_process_onnx_output(self, output: OnnxOutputContext, **kwargs: Any) -> Iterable[T]:
+        """Post-process the ONNX model output to convert it into a usable format.
+
+        Args:
+            output (OnnxOutputContext): The raw output from the ONNX model.
+            **kwargs: Additional keyword arguments that may be needed by specific implementations.
+
+        Returns:
+            Iterable[T]: Post-processed output as an iterable of type T.
+        """
         raise NotImplementedError("Subclasses must implement this method")
 
     def __init__(self) -> None:
         super().__init__()
-        self.processor: Optional[Compose] = None
+        self.processor: Compose | None = None
 
     def _preprocess_onnx_input(
         self, onnx_input: dict[str, NumpyArray], **kwargs: Any
@@ -42,10 +53,11 @@ class OnnxImageModel(OnnxModel[T]):
         self,
         model_dir: Path,
         model_file: str,
-        threads: Optional[int],
-        providers: Optional[Sequence[OnnxProvider]] = None,
-        cuda: bool = False,
-        device_id: Optional[int] = None,
+        threads: int | None,
+        providers: Sequence[OnnxProvider] | None = None,
+        cuda: bool | Device = Device.AUTO,
+        device_id: int | None = None,
+        extra_session_options: dict[str, Any] | None = None,
     ) -> None:
         super()._load_onnx_model(
             model_dir=model_dir,
@@ -54,6 +66,7 @@ class OnnxImageModel(OnnxModel[T]):
             providers=providers,
             cuda=cuda,
             device_id=device_id,
+            extra_session_options=extra_session_options,
         )
         self.processor = load_preprocessor(model_dir=model_dir)
 
@@ -65,16 +78,18 @@ class OnnxImageModel(OnnxModel[T]):
         return {input_name: encoded}
 
     def onnx_embed(self, images: list[ImageInput], **kwargs: Any) -> OnnxOutputContext:
-        with contextlib.ExitStack():
+        with contextlib.ExitStack() as stack:
             image_files = [
-                Image.open(image) if not isinstance(image, Image.Image) else image
+                stack.enter_context(Image.open(image))
+                if not isinstance(image, Image.Image)
+                else image
                 for image in images
             ]
             assert self.processor is not None, "Processor is not initialized"
             encoded = np.array(self.processor(image_files))
         onnx_input = self._build_onnx_input(encoded)
         onnx_input = self._preprocess_onnx_input(onnx_input)
-        model_output = self.model.run(None, onnx_input)  # type: ignore[union-attr]
+        model_output = self.model.run(self.ONNX_OUTPUT_NAMES, onnx_input)  # type: ignore[union-attr]
         embeddings = model_output[0].reshape(len(images), -1)
         return OnnxOutputContext(model_output=embeddings)
 
@@ -82,12 +97,15 @@ class OnnxImageModel(OnnxModel[T]):
         self,
         model_name: str,
         cache_dir: str,
-        images: Union[ImageInput, Iterable[ImageInput]],
+        images: ImageInput | Iterable[ImageInput],
         batch_size: int = 256,
-        parallel: Optional[int] = None,
-        providers: Optional[Sequence[OnnxProvider]] = None,
-        cuda: bool = False,
-        device_ids: Optional[list[int]] = None,
+        parallel: int | None = None,
+        providers: Sequence[OnnxProvider] | None = None,
+        cuda: bool | Device = Device.AUTO,
+        device_ids: list[int] | None = None,
+        local_files_only: bool = False,
+        specific_model_path: str | None = None,
+        extra_session_options: dict[str, Any] | None = None,
         **kwargs: Any,
     ) -> Iterable[T]:
         is_small = False
@@ -104,7 +122,7 @@ class OnnxImageModel(OnnxModel[T]):
                 self.load_onnx_model()
 
             for batch in iter_batch(images, 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()
@@ -114,9 +132,14 @@ class OnnxImageModel(OnnxModel[T]):
                 "model_name": model_name,
                 "cache_dir": cache_dir,
                 "providers": providers,
+                "local_files_only": local_files_only,
+                "specific_model_path": specific_model_path,
                 **kwargs,
             }
 
+            if extra_session_options is not None:
+                params.update(extra_session_options)
+
             pool = ParallelWorkerPool(
                 num_workers=parallel or 1,
                 worker=self._get_worker_class(),
@@ -125,7 +148,7 @@ class OnnxImageModel(OnnxModel[T]):
                 start_method=start_method,
             )
             for batch in pool.ordered_map(iter_batch(images, batch_size), **params):
-                yield from self._post_process_onnx_output(batch)  # type: ignore
+                yield from self._post_process_onnx_output(batch, **kwargs)  # type: ignore
 
 
 class ImageEmbeddingWorker(EmbeddingWorker[T]):

fastembed/image/siglip_embedding.py

@@ -0,0 +1,44 @@
+from typing import Any, Type
+
+from fastembed.common.model_description import DenseModelDescription, ModelSource
+from fastembed.image.onnx_embedding import OnnxImageEmbedding, OnnxImageEmbeddingWorker
+
+supported_siglip_models: list[DenseModelDescription] = [
+    DenseModelDescription(
+        model="google/siglip2-base-patch16-224",
+        dim=768,
+        description="Image embeddings, Multimodal (text&image), 2025 year",
+        license="apache-2.0",
+        size_in_GB=0.37,
+        sources=ModelSource(hf="onnx-community/siglip2-base-patch16-224-ONNX"),
+        model_file="onnx/vision_model.onnx",
+    ),
+]
+
+
+class SiglipOnnxImageEmbedding(OnnxImageEmbedding):
+    """SigLIP vision tower.
+
+    The exported graph returns both the per-patch `last_hidden_state` and the pooled
+    `pooler_output`; only the latter is the image embedding, so it must be selected explicitly.
+    """
+
+    ONNX_OUTPUT_NAMES = ["pooler_output"]
+
+    @classmethod
+    def _get_worker_class(cls) -> Type["OnnxImageEmbeddingWorker"]:
+        return SiglipImageEmbeddingWorker
+
+    @classmethod
+    def _list_supported_models(cls) -> list[DenseModelDescription]:
+        return supported_siglip_models
+
+
+class SiglipImageEmbeddingWorker(OnnxImageEmbeddingWorker):
+    def init_embedding(self, model_name: str, cache_dir: str, **kwargs: Any) -> OnnxImageEmbedding:
+        return SiglipOnnxImageEmbedding(
+            model_name=model_name,
+            cache_dir=cache_dir,
+            threads=1,
+            **kwargs,
+        )

fastembed/image/transform/functional.py

@@ -1,5 +1,3 @@
-from typing import Union
-
 import numpy as np
 from PIL import Image
 
@@ -15,7 +13,7 @@ def convert_to_rgb(image: Image.Image) -> Image.Image:
 
 
 def center_crop(
-    image: Union[Image.Image, NumpyArray],
+    image: Image.Image | NumpyArray,
     size: tuple[int, int],
 ) -> NumpyArray:
     if isinstance(image, np.ndarray):
@@ -64,43 +62,56 @@ def center_crop(
 
 def normalize(
     image: NumpyArray,
-    mean: Union[float, list[float]],
-    std: Union[float, list[float]],
+    mean: float | list[float],
+    std: float | list[float],
 ) -> NumpyArray:
-    num_channels = image.shape[1] if len(image.shape) == 4 else image.shape[0]
+    if image.ndim < 3:
+        raise ValueError(f"image must be (C, H, W) or (N, C, H, W), got shape {image.shape}")
+
+    # Channels sit on the third axis from the end, which covers (C, H, W) and
+    # (N, C, H, W) alike. Transposing instead reversed every axis, which put the
+    # batch dimension where the channels were meant to be.
+    num_channels = image.shape[-3]
 
     if not np.issubdtype(image.dtype, np.floating):
         image = image.astype(np.float32)
 
-    mean = mean if isinstance(mean, list) else [mean] * num_channels
+    mean_list = mean if isinstance(mean, list) else [mean] * num_channels
 
-    if len(mean) != num_channels:
+    if len(mean_list) != num_channels:
         raise ValueError(
             f"mean must have the same number of channels as the image, image has {num_channels} channels, got "
-            f"{len(mean)}"
+            f"{len(mean_list)}"
         )
 
-    mean_arr = np.array(mean, dtype=np.float32)
+    # (C, 1, 1) lines the channels up with the trailing (C, H, W) axes under numpy
+    # broadcasting, whatever batch dimensions lead them.
+    mean_arr = np.array(mean_list, dtype=np.float32).reshape(-1, 1, 1)
 
-    std = std if isinstance(std, list) else [std] * num_channels
-    if len(std) != num_channels:
+    std_list = std if isinstance(std, list) else [std] * num_channels
+    if len(std_list) != num_channels:
         raise ValueError(
-            f"std must have the same number of channels as the image, image has {num_channels} channels, got {len(std)}"
+            f"std must have the same number of channels as the image, image has {num_channels} channels, got {len(std_list)}"
         )
 
-    std_arr = np.array(std, dtype=np.float32)
+    std_arr = np.array(std_list, dtype=np.float32).reshape(-1, 1, 1)
 
-    image = ((image.T - mean_arr) / std_arr).T
-    return image
+    image_upd = (image - mean_arr) / std_arr
+    return image_upd
 
 
 def resize(
     image: Image.Image,
-    size: Union[int, tuple[int, int]],
-    resample: Union[int, Image.Resampling] = Image.Resampling.BILINEAR,
+    size: int | tuple[int, int],
+    resample: int | Image.Resampling = Image.Resampling.BILINEAR,
 ) -> Image.Image:
     if isinstance(size, tuple):
-        return image.resize(size, resample)
+        # fastembed keeps sizes as (height, width) — `Compose.from_config` builds the
+        # tuple as (size["height"], size["width"]) — while Pillow's resize takes
+        # (width, height). The two agree for square sizes, so this only shows up on a
+        # non-square image processor configuration.
+        height, width = size
+        return image.resize((width, height), resample)
 
     height, width = image.height, image.width
     short, long = (width, height) if width <= height else (height, width)
@@ -117,7 +128,7 @@ def rescale(image: NumpyArray, scale: float, dtype: type = np.float32) -> NumpyA
     return (image * scale).astype(dtype)
 
 
-def pil2ndarray(image: Union[Image.Image, NumpyArray]) -> NumpyArray:
+def pil2ndarray(image: Image.Image | NumpyArray) -> NumpyArray:
     if isinstance(image, Image.Image):
         return np.asarray(image).transpose((2, 0, 1))
     return image
@@ -126,7 +137,7 @@ def pil2ndarray(image: Union[Image.Image, NumpyArray]) -> NumpyArray:
 def pad2square(
     image: Image.Image,
     size: int,
-    fill_color: Union[str, int, tuple[int, ...]] = 0,
+    fill_color: str | int | tuple[int, ...] = 0,
 ) -> Image.Image:
     height, width = image.height, image.width
 
@@ -147,3 +158,77 @@ def pad2square(
     new_image = Image.new(mode="RGB", size=(size, size), color=fill_color)
     new_image.paste(image.crop((left, top, right, bottom)) if crop_required else image)
     return new_image
+
+
+def resize_longest_edge(
+    image: Image.Image,
+    max_size: int,
+    resample: int | Image.Resampling = Image.Resampling.LANCZOS,
+) -> Image.Image:
+    height, width = image.height, image.width
+    aspect_ratio = width / height
+
+    if width >= height:
+        # Width is longer
+        new_width = max_size
+        new_height = int(new_width / aspect_ratio)
+    else:
+        # Height is longer
+        new_height = max_size
+        new_width = int(new_height * aspect_ratio)
+
+    # Ensure even dimensions
+    if new_height % 2 != 0:
+        new_height += 1
+    if new_width % 2 != 0:
+        new_width += 1
+
+    return image.resize((new_width, new_height), resample)
+
+
+def crop_ndarray(
+    image: NumpyArray,
+    x1: int,
+    y1: int,
+    x2: int,
+    y2: int,
+    channel_first: bool = True,
+) -> NumpyArray:
+    if channel_first:
+        # (C, H, W) format
+        return image[:, y1:y2, x1:x2]
+    else:
+        # (H, W, C) format
+        return image[y1:y2, x1:x2, :]
+
+
+def resize_ndarray(
+    image: NumpyArray,
+    size: tuple[int, int],
+    resample: int | Image.Resampling = Image.Resampling.LANCZOS,
+    channel_first: bool = True,
+) -> NumpyArray:
+    # Convert to PIL-friendly format (H, W, C)
+    if channel_first:
+        img_hwc = image.transpose((1, 2, 0))
+    else:
+        img_hwc = image
+
+    # Handle different dtypes
+    if img_hwc.dtype == np.float32 or img_hwc.dtype == np.float64:
+        # Assume normalized, scale to 0-255 for PIL
+        img_hwc_scaled = (img_hwc * 255).astype(np.uint8)
+        pil_img = Image.fromarray(img_hwc_scaled, mode="RGB")
+        resized = pil_img.resize(size, resample)
+        result = np.array(resized).astype(np.float32) / 255.0
+    else:
+        # uint8 or similar
+        pil_img = Image.fromarray(img_hwc.astype(np.uint8), mode="RGB")
+        resized = pil_img.resize(size, resample)
+        result = np.array(resized)
+
+    # Convert back to original format
+    if channel_first:
+        result = result.transpose((2, 0, 1))
+
+    return result

fastembed/image/transform/operators.py

@@ -1,4 +1,5 @@
-from typing import Any, Union, Optional
+from typing import Any
+import math
 
 from PIL import Image
 
@@ -6,16 +7,19 @@ from fastembed.common.types import NumpyArray
 from fastembed.image.transform.functional import (
     center_crop,
     convert_to_rgb,
+    crop_ndarray,
     normalize,
     pil2ndarray,
     rescale,
     resize,
+    resize_longest_edge,
+    resize_ndarray,
     pad2square,
 )
 
 
 class Transform:
-    def __call__(self, images: list[Any]) -> Union[list[Image.Image], list[NumpyArray]]:
+    def __call__(self, images: list[Any]) -> list[Image.Image] | list[NumpyArray]:
         raise NotImplementedError("Subclasses must implement this method")
 
 
@@ -33,18 +37,28 @@ class CenterCrop(Transform):
 
 
 class Normalize(Transform):
-    def __init__(self, mean: Union[float, list[float]], std: Union[float, list[float]]):
+    def __init__(self, mean: float | list[float], std: float | list[float]):
         self.mean = mean
         self.std = std
 
-    def __call__(self, images: list[NumpyArray]) -> list[NumpyArray]:
-        return [normalize(image, mean=self.mean, std=self.std) for image in images]
+    def __call__(  # type: ignore[override]
+        self, images: list[NumpyArray] | list[list[NumpyArray]]
+    ) -> list[NumpyArray] | list[list[NumpyArray]]:
+        if images and isinstance(images[0], list):
+            # Nested structure from ImageSplitter
+            return [
+                [normalize(image, mean=self.mean, std=self.std) for image in img_patches]  # type: ignore[arg-type]
+                for img_patches in images
+            ]
+        else:
+            # Flat structure (backward compatibility)
+            return [normalize(image, mean=self.mean, std=self.std) for image in images]  # type: ignore[arg-type]
 
 
 class Resize(Transform):
     def __init__(
         self,
-        size: Union[int, tuple[int, int]],
+        size: int | tuple[int, int],
         resample: Image.Resampling = Image.Resampling.BICUBIC,
     ):
         self.size = size
@@ -58,12 +72,22 @@ class Rescale(Transform):
     def __init__(self, scale: float = 1 / 255):
         self.scale = scale
 
-    def __call__(self, images: list[NumpyArray]) -> list[NumpyArray]:
-        return [rescale(image, scale=self.scale) for image in images]
+    def __call__(  # type: ignore[override]
+        self, images: list[NumpyArray] | list[list[NumpyArray]]
+    ) -> list[NumpyArray] | list[list[NumpyArray]]:
+        if images and isinstance(images[0], list):
+            # Nested structure from ImageSplitter
+            return [
+                [rescale(image, scale=self.scale) for image in img_patches]  # type: ignore[arg-type]
+                for img_patches in images
+            ]
+        else:
+            # Flat structure (backward compatibility)
+            return [rescale(image, scale=self.scale) for image in images]  # type: ignore[arg-type]
 
 
 class PILtoNDarray(Transform):
-    def __call__(self, images: list[Union[Image.Image, NumpyArray]]) -> list[NumpyArray]:
+    def __call__(self, images: list[Image.Image | NumpyArray]) -> list[NumpyArray]:
         return [pil2ndarray(image) for image in images]
 
 
@@ -71,7 +95,7 @@ class PadtoSquare(Transform):
     def __init__(
         self,
         size: int,
-        fill_color: Union[str, int, tuple[int, ...]],
+        fill_color: str | int | tuple[int, ...],
     ):
         self.size = size
         self.fill_color = fill_color
@@ -82,13 +106,174 @@ class PadtoSquare(Transform):
         ]
 
 
+class ResizeLongestEdge(Transform):
+    """Resize images so the longest edge equals target size, preserving aspect ratio."""
+
+    def __init__(
+        self,
+        size: int,
+        resample: Image.Resampling = Image.Resampling.LANCZOS,
+    ):
+        self.size = size
+        self.resample = resample
+
+    def __call__(self, images: list[Image.Image]) -> list[Image.Image]:
+        return [resize_longest_edge(image, self.size, self.resample) for image in images]
+
+
+class ResizeForVisionEncoder(Transform):
+    """
+    Resize both dimensions to be multiples of vision_encoder_max_size.
+    Preserves aspect ratio approximately.
+    Works on numpy arrays in (C, H, W) format.
+    """
+
+    def __init__(
+        self,
+        max_size: int,
+        resample: Image.Resampling = Image.Resampling.LANCZOS,
+    ):
+        self.max_size = max_size
+        self.resample = resample
+
+    def __call__(self, images: list[NumpyArray]) -> list[NumpyArray]:
+        result = []
+        for image in images:
+            # Assume (C, H, W) format
+            _, height, width = image.shape
+
+            aspect_ratio = width / height
+
+            if width >= height:
+                # Calculate new width as multiple of max_size
+                new_width = math.ceil(width / self.max_size) * self.max_size
+                new_height = int(new_width / aspect_ratio)
+                new_height = math.ceil(new_height / self.max_size) * self.max_size
+            else:
+                # Calculate new height as multiple of max_size
+                new_height = math.ceil(height / self.max_size) * self.max_size
+                new_width = int(new_height * aspect_ratio)
+                new_width = math.ceil(new_width / self.max_size) * self.max_size
+
+            # Resize using the ndarray resize function
+            resized = resize_ndarray(
+                image,
+                size=(new_width, new_height),  # PIL expects (width, height)
+                resample=self.resample,
+                channel_first=True,
+            )
+            result.append(resized)
+
+        return result
+
+
+class ImageSplitter(Transform):
+    """
+    Split images into grid of patches plus a global view.
+
+    If image dimensions exceed max_size:
+    - Divide into ceil(H/max_size) x ceil(W/max_size) patches
+    - Each patch is cropped from the image
+    - Add a global view (original resized to max_size x max_size)
+
+    If image is smaller than max_size:
+    - Return single image unchanged
+
+    Works on numpy arrays in (C, H, W) format.
+    """
+
+    def __init__(
+        self,
+        max_size: int,
+        resample: Image.Resampling = Image.Resampling.LANCZOS,
+    ):
+        self.max_size = max_size
+        self.resample = resample
+
+    def __call__(self, images: list[NumpyArray]) -> list[list[NumpyArray]]:  # type: ignore[override]
+        result = []
+
+        for image in images:
+            # Assume (C, H, W) format
+            _, height, width = image.shape
+            max_height = max_width = self.max_size
+
+            frames = []
+
+            if height > max_height or width > max_width:
+                # Calculate the number of splits needed
+                num_splits_h = math.ceil(height / max_height)
+                num_splits_w = math.ceil(width / max_width)
+
+                # Calculate optimal patch dimensions
+                optimal_height = math.ceil(height / num_splits_h)
+                optimal_width = math.ceil(width / num_splits_w)
+
+                # Generate patches in grid order (row by row)
+                for r in range(num_splits_h):
+                    for c in range(num_splits_w):
+                        # Calculate crop coordinates
+                        start_x = c * optimal_width
+                        start_y = r * optimal_height
+                        end_x = min(start_x + optimal_width, width)
+                        end_y = min(start_y + optimal_height, height)
+
+                        # Crop the patch
+                        cropped = crop_ndarray(
+                            image, x1=start_x, y1=start_y, x2=end_x, y2=end_y, channel_first=True
+                        )
+                        frames.append(cropped)
+
+                # Add global view (resized to max_size x max_size)
+                global_view = resize_ndarray(
+                    image,
+                    size=(max_width, max_height),  # PIL expects (width, height)
+                    resample=self.resample,
+                    channel_first=True,
+                )
+                frames.append(global_view)
+            else:
+                # Image is small enough, no splitting needed
+                frames.append(image)
+
+            # Append (not extend) to preserve per-image grouping
+            result.append(frames)
+
+        return result
+
+
+class SquareResize(Transform):
+    """
+    Resize images to square dimensions (max_size x max_size).
+    Works on numpy arrays in (C, H, W) format.
+    """
+
+    def __init__(
+        self,
+        size: int,
+        resample: Image.Resampling = Image.Resampling.LANCZOS,
+    ):
+        self.size = size
+        self.resample = resample
+
+    def __call__(self, images: list[NumpyArray]) -> list[list[NumpyArray]]:  # type: ignore[override]
+        return [
+            [
+                resize_ndarray(
+                    image, size=(self.size, self.size), resample=self.resample, channel_first=True
+                )
+            ]
+            for image in images
+        ]
+
+
 class Compose:
     def __init__(self, transforms: list[Transform]):
         self.transforms = transforms
 
     def __call__(
-        self, images: Union[list[Image.Image], list[NumpyArray]]
-    ) -> Union[list[NumpyArray], list[Image.Image]]:
+        self, images: list[Image.Image] | list[NumpyArray]
+    ) -> list[NumpyArray] | list[Image.Image]:
         for transform in self.transforms:
             images = transform(images)
         return images
@@ -118,6 +303,7 @@ class Compose:
                 Valid size keys (nested):
                     - {"height", "width"}
                     - {"shortest_edge"}
+                    - {"longest_edge"}
 
         Returns:
             Compose: Image processor.
@@ -128,6 +314,7 @@ class Compose:
         cls._get_pad2square(transforms, config)
         cls._get_center_crop(transforms, config)
         cls._get_pil2ndarray(transforms, config)
+        cls._get_image_splitting(transforms, config)
         cls._get_rescale(transforms, config)
         cls._get_normalize(transforms, config)
         return cls(transforms=transforms)
@@ -196,6 +383,25 @@ class Compose:
                             resample=resample,
                         )
                     )
+        elif mode == "Idefics3ImageProcessor":
+            if config.get("do_resize", False):
+                size = config.get("size", {})
+                if "longest_edge" not in size:
+                    raise ValueError(
+                        "Size dictionary must contain 'longest_edge' key for Idefics3ImageProcessor"
+                    )
+
+                # Handle resample parameter - can be int enum or PIL.Image.Resampling
+                resample = config.get("resample", Image.Resampling.LANCZOS)
+                if isinstance(resample, int):
+                    resample = Image.Resampling(resample)
+
+                transforms.append(
+                    ResizeLongestEdge(
+                        size=size["longest_edge"],
+                        resample=resample,
+                    )
+                )
         else:
             raise ValueError(f"Preprocessor {mode} is not supported")
 
@@ -217,6 +423,8 @@ class Compose:
             pass
         elif mode == "JinaCLIPImageProcessor":
             pass
+        elif mode == "Idefics3ImageProcessor":
+            pass
         else:
             raise ValueError(f"Preprocessor {mode} is not supported")
 
@@ -224,6 +432,28 @@ class Compose:
     def _get_pil2ndarray(transforms: list[Transform], config: dict[str, Any]) -> None:
         transforms.append(PILtoNDarray())
 
+    @classmethod
+    def _get_image_splitting(cls, transforms: list[Transform], config: dict[str, Any]) -> None:
+        """
+        Add image splitting transforms for Idefics3.
+        Handles conditional logic: splitting vs square resize.
+        Must be called AFTER PILtoNDarray.
+        """
+        mode = config.get("image_processor_type", "CLIPImageProcessor")
+
+        if mode == "Idefics3ImageProcessor":
+            do_splitting = config.get("do_image_splitting", False)
+            max_size = config.get("max_image_size", {}).get("longest_edge", 512)
+            resample = config.get("resample", Image.Resampling.LANCZOS)
+            if isinstance(resample, int):
+                resample = Image.Resampling(resample)
+
+            if do_splitting:
+                transforms.append(ResizeForVisionEncoder(max_size, resample))
+                transforms.append(ImageSplitter(max_size, resample))
+            else:
+                transforms.append(SquareResize(max_size, resample))
+
     @staticmethod
     def _get_rescale(transforms: list[Transform], config: dict[str, Any]) -> None:
         if config.get("do_rescale", True):
@@ -253,7 +483,7 @@ class Compose:
             )
 
     @staticmethod
-    def _interpolation_resolver(resample: Optional[str] = None) -> Image.Resampling:
+    def _interpolation_resolver(resample: str | None = None) -> Image.Resampling:
         interpolation_map = {
             "nearest": Image.Resampling.NEAREST,
             "lanczos": Image.Resampling.LANCZOS,

fastembed/late_interaction/colbert.py

@@ -1,13 +1,14 @@
 import string
-from typing import Any, Iterable, Optional, Sequence, Type, Union
+from typing import Any, Iterable, Sequence, Type
 
 import numpy as np
-from tokenizers import Encoding
+from tokenizers import Encoding, Tokenizer
 
-from fastembed.common.types import NumpyArray
+from fastembed.common.preprocessor_utils import load_tokenizer
+from fastembed.common.types import NumpyArray, Device
 from fastembed.common import OnnxProvider
 from fastembed.common.onnx_model import OnnxOutputContext
-from fastembed.common.utils import define_cache_dir
+from fastembed.common.utils import define_cache_dir, iter_batch
 from fastembed.late_interaction.late_interaction_embedding_base import (
     LateInteractionTextEmbeddingBase,
 )
@@ -18,7 +19,7 @@ supported_colbert_models: list[DenseModelDescription] = [
     DenseModelDescription(
         model="colbert-ir/colbertv2.0",
         dim=128,
-        description="Late interaction model",
+        description="Text embeddings, Unimodal (text), English, 512 input tokens truncation, 2023 year",
         license="mit",
         size_in_GB=0.44,
         sources=ModelSource(hf="colbert-ir/colbertv2.0"),
@@ -27,7 +28,7 @@ supported_colbert_models: list[DenseModelDescription] = [
     DenseModelDescription(
         model="answerdotai/answerai-colbert-small-v1",
         dim=96,
-        description="Text embeddings, Unimodal (text), Multilingual (~100 languages), 512 input tokens truncation, 2024 year",
+        description="Text embeddings, Unimodal (text), English, 512 input tokens truncation, 2024 year",
         license="apache-2.0",
         size_in_GB=0.13,
         sources=ModelSource(hf="answerdotai/answerai-colbert-small-v1"),
@@ -43,26 +44,29 @@ class Colbert(LateInteractionTextEmbeddingBase, OnnxTextModel[NumpyArray]):
     MASK_TOKEN = "[MASK]"
 
     def _post_process_onnx_output(
-        self, output: OnnxOutputContext, is_doc: bool = True
+        self, output: OnnxOutputContext, is_doc: bool = True, **kwargs: Any
     ) -> Iterable[NumpyArray]:
         if not is_doc:
-            return output.model_output.astype(np.float32)
+            for embedding in output.model_output:
+                yield embedding
+        else:
+            if output.input_ids is None or output.attention_mask is None:
+                raise ValueError(
+                    "input_ids and attention_mask must be provided for document post-processing"
+                )
 
-        if output.input_ids is None or output.attention_mask is None:
-            raise ValueError(
-                "input_ids and attention_mask must be provided for document post-processing"
-            )
+            for i, token_sequence in enumerate(output.input_ids):
+                for j, token_id in enumerate(token_sequence):  # type: ignore
+                    if token_id in self.skip_list or token_id == self.pad_token_id:
+                        output.attention_mask[i, j] = 0
 
-        for i, token_sequence in enumerate(output.input_ids):
-            for j, token_id in enumerate(token_sequence):  # type: ignore
-                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)
+            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
 
-        output.model_output *= np.expand_dims(output.attention_mask, 2).astype(np.float32)
-        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)
+            for embedding, attention_mask in zip(output.model_output, output.attention_mask):
+                yield embedding[attention_mask == 1]
 
     def _preprocess_onnx_input(
         self, onnx_input: dict[str, NumpyArray], is_doc: bool = True, **kwargs: Any
@@ -84,29 +88,46 @@ class Colbert(LateInteractionTextEmbeddingBase, OnnxTextModel[NumpyArray]):
         )
 
     def _tokenize_query(self, query: str) -> list[Encoding]:
-        assert self.tokenizer is not None
-        encoded = self.tokenizer.encode_batch([query])
-        # colbert authors recommend to pad queries with [MASK] tokens for query augmentation to improve performance
-        if len(encoded[0].ids) < self.MIN_QUERY_LENGTH:
-            prev_padding = None
-            if self.tokenizer.padding:
-                prev_padding = self.tokenizer.padding
-            self.tokenizer.enable_padding(
-                pad_token=self.MASK_TOKEN,
-                pad_id=self.mask_token_id,
-                length=self.MIN_QUERY_LENGTH,
-            )
-            encoded = self.tokenizer.encode_batch([query])
-            if prev_padding is None:
-                self.tokenizer.no_padding()
-            else:
-                self.tokenizer.enable_padding(**prev_padding)
+        assert self.query_tokenizer is not None
+        encoded = self.query_tokenizer.encode_batch([query])
         return encoded
 
     def _tokenize_documents(self, documents: list[str]) -> list[Encoding]:
         encoded = self.tokenizer.encode_batch(documents)  # type: ignore[union-attr]
         return encoded
 
+    def token_count(
+        self,
+        texts: str | Iterable[str],
+        batch_size: int = 1024,
+        is_doc: bool = True,
+        include_extension: bool = False,
+        **kwargs: Any,
+    ) -> int:
+        if not hasattr(self, "model") or self.model is None:
+            self.load_onnx_model()  # loads the tokenizer as well
+        token_num = 0
+        texts = [texts] if isinstance(texts, str) else texts
+        tokenizer = self.tokenizer if is_doc else self.query_tokenizer
+        assert tokenizer is not None
+        for batch in iter_batch(texts, batch_size):
+            for tokens in tokenizer.encode_batch(batch):
+                if is_doc:
+                    token_num += sum(tokens.attention_mask)
+                else:
+                    attend_count = sum(tokens.attention_mask)
+                    if include_extension:
+                        token_num += max(attend_count, self.MIN_QUERY_LENGTH)
+
+                    else:
+                        token_num += attend_count
+            if include_extension:
+                token_num += len(
+                    batch
+                )  # add 1 for each cls.DOC_MARKER_TOKEN_ID or cls.QUERY_MARKER_TOKEN_ID
+
+        return token_num
+
     @classmethod
     def _list_supported_models(cls) -> list[DenseModelDescription]:
         """Lists the supported models.
@@ -119,14 +140,14 @@ class Colbert(LateInteractionTextEmbeddingBase, OnnxTextModel[NumpyArray]):
     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,
+        cache_dir: str | None = None,
+        threads: int | None = None,
+        providers: Sequence[OnnxProvider] | None = None,
+        cuda: bool | Device = Device.AUTO,
+        device_ids: list[int] | None = None,
         lazy_load: bool = False,
-        device_id: Optional[int] = None,
-        specific_model_path: Optional[str] = None,
+        device_id: int | None = None,
+        specific_model_path: str | None = None,
         **kwargs: Any,
     ):
         """
@@ -138,10 +159,11 @@ class Colbert(LateInteractionTextEmbeddingBase, OnnxTextModel[NumpyArray]):
             threads (int, optional): The number of threads single onnxruntime session can use. Defaults to None.
             providers (Optional[Sequence[OnnxProvider]], optional): The list of onnxruntime providers to use.
                 Mutually exclusive with the `cuda` and `device_ids` arguments. Defaults to None.
-            cuda (bool, optional): Whether to use cuda for inference. Mutually exclusive with `providers`
-                Defaults to False.
+            cuda (Union[bool, Device], optional): Whether to use cuda for inference. Mutually exclusive with `providers`
+                Defaults to Device.AUTO.
             device_ids (Optional[list[int]], optional): The list of device ids to use for data parallel processing in
-                workers. Should be used with `cuda=True`, mutually exclusive with `providers`. Defaults to None.
+                workers. Should be used with `cuda` equals to `True`, `Device.AUTO` or `Device.CUDA`, mutually exclusive
+                with `providers`. Defaults to None.
             lazy_load (bool, optional): Whether to load the model during class initialization or on demand.
                 Should be set to True when using multiple-gpu and parallel encoding. Defaults to False.
             device_id (Optional[int], optional): The device id to use for loading the model in the worker process.
@@ -154,13 +176,14 @@ class Colbert(LateInteractionTextEmbeddingBase, OnnxTextModel[NumpyArray]):
         super().__init__(model_name, cache_dir, threads, **kwargs)
         self.providers = providers
         self.lazy_load = lazy_load
+        self._extra_session_options = self._select_exposed_session_options(kwargs)
 
         # List of device ids, that can be used for data parallel processing in workers
         self.device_ids = device_ids
         self.cuda = cuda
 
         # This device_id will be used if we need to load model in current process
-        self.device_id: Optional[int] = None
+        self.device_id: int | None = None
         if device_id is not None:
             self.device_id = device_id
         elif self.device_ids is not None:
@@ -169,16 +192,19 @@ class Colbert(LateInteractionTextEmbeddingBase, OnnxTextModel[NumpyArray]):
         self.model_description = self._get_model_description(model_name)
         self.cache_dir = str(define_cache_dir(cache_dir))
 
+        self._specific_model_path = specific_model_path
         self._model_dir = self.download_model(
             self.model_description,
             self.cache_dir,
             local_files_only=self._local_files_only,
-            specific_model_path=specific_model_path,
+            specific_model_path=self._specific_model_path,
         )
-        self.mask_token_id: Optional[int] = None
-        self.pad_token_id: Optional[int] = None
+        self.mask_token_id: int | None = None
+        self.pad_token_id: int | None = None
         self.skip_list: set[int] = set()
 
+        self.query_tokenizer: Tokenizer | None = None
+
         if not self.lazy_load:
             self.load_onnx_model()
 
@@ -190,7 +216,10 @@ class Colbert(LateInteractionTextEmbeddingBase, OnnxTextModel[NumpyArray]):
             providers=self.providers,
             cuda=self.cuda,
             device_id=self.device_id,
+            extra_session_options=self._extra_session_options,
         )
+        self.query_tokenizer, _ = load_tokenizer(model_dir=self._model_dir)
+
         assert self.tokenizer is not None
         self.mask_token_id = self.special_token_to_id[self.MASK_TOKEN]
         self.pad_token_id = self.tokenizer.padding["pad_id"]
@@ -201,12 +230,18 @@ class Colbert(LateInteractionTextEmbeddingBase, OnnxTextModel[NumpyArray]):
         current_max_length = self.tokenizer.truncation["max_length"]
         # ensure not to overflow after adding document-marker
         self.tokenizer.enable_truncation(max_length=current_max_length - 1)
+        self.query_tokenizer.enable_truncation(max_length=current_max_length - 1)
+        self.query_tokenizer.enable_padding(
+            pad_token=self.MASK_TOKEN,
+            pad_id=self.mask_token_id,
+            length=self.MIN_QUERY_LENGTH,
+        )
 
     def embed(
         self,
-        documents: Union[str, Iterable[str]],
+        documents: str | Iterable[str],
         batch_size: int = 256,
-        parallel: Optional[int] = None,
+        parallel: int | None = None,
         **kwargs: Any,
     ) -> Iterable[NumpyArray]:
         """
@@ -233,10 +268,13 @@ class Colbert(LateInteractionTextEmbeddingBase, OnnxTextModel[NumpyArray]):
             providers=self.providers,
             cuda=self.cuda,
             device_ids=self.device_ids,
+            local_files_only=self._local_files_only,
+            specific_model_path=self._specific_model_path,
+            extra_session_options=self._extra_session_options,
             **kwargs,
         )
 
-    def query_embed(self, query: Union[str, Iterable[str]], **kwargs: Any) -> Iterable[NumpyArray]:
+    def query_embed(self, query: str | Iterable[str], **kwargs: Any) -> Iterable[NumpyArray]:
         if isinstance(query, str):
             query = [query]
 

fastembed/late_interaction/late_interaction_embedding_base.py

@@ -1,4 +1,4 @@
-from typing import Iterable, Optional, Union, Any
+from typing import Iterable, Any
 
 from fastembed.common.model_description import DenseModelDescription
 from fastembed.common.types import NumpyArray
@@ -9,20 +9,21 @@ class LateInteractionTextEmbeddingBase(ModelManagement[DenseModelDescription]):
     def __init__(
         self,
         model_name: str,
-        cache_dir: Optional[str] = None,
-        threads: Optional[int] = None,
+        cache_dir: str | None = None,
+        threads: int | None = None,
         **kwargs: Any,
     ):
         self.model_name = model_name
         self.cache_dir = cache_dir
         self.threads = threads
         self._local_files_only = kwargs.pop("local_files_only", False)
+        self._embedding_size: int | None = None
 
     def embed(
         self,
-        documents: Union[str, Iterable[str]],
+        documents: str | Iterable[str],
         batch_size: int = 256,
-        parallel: Optional[int] = None,
+        parallel: int | None = None,
         **kwargs: Any,
     ) -> Iterable[NumpyArray]:
         raise NotImplementedError()
@@ -42,7 +43,7 @@ class LateInteractionTextEmbeddingBase(ModelManagement[DenseModelDescription]):
         # This is model-specific, so that different models can have specialized implementations
         yield from self.embed(texts, **kwargs)
 
-    def query_embed(self, query: Union[str, Iterable[str]], **kwargs: Any) -> Iterable[NumpyArray]:
+    def query_embed(self, query: str | Iterable[str], **kwargs: Any) -> Iterable[NumpyArray]:
         """
         Embeds queries
 
@@ -58,3 +59,22 @@ class LateInteractionTextEmbeddingBase(ModelManagement[DenseModelDescription]):
             yield from self.embed([query], **kwargs)
         else:
             yield from self.embed(query, **kwargs)
+
+    @classmethod
+    def get_embedding_size(cls, model_name: str) -> int:
+        """Returns embedding size of the chosen model."""
+        raise NotImplementedError("Subclasses must implement this method")
+
+    @property
+    def embedding_size(self) -> int:
+        """Returns embedding size for the current model"""
+        raise NotImplementedError("Subclasses must implement this method")
+
+    def token_count(
+        self,
+        texts: str | Iterable[str],
+        batch_size: int = 1024,
+        **kwargs: Any,
+    ) -> int:
+        """Returns the number of tokens in the texts."""
+        raise NotImplementedError("Subclasses must implement this method")

fastembed/late_interaction/late_interaction_text_embedding.py

@@ -1,8 +1,8 @@
-from typing import Any, Iterable, Optional, Sequence, Type, Union
+from typing import Any, Iterable, Sequence, Type
 from dataclasses import asdict
 
 from fastembed.common.model_description import DenseModelDescription
-from fastembed.common.types import NumpyArray
+from fastembed.common.types import NumpyArray, Device
 from fastembed.common import OnnxProvider
 from fastembed.late_interaction.colbert import Colbert
 from fastembed.late_interaction.jina_colbert import JinaColbert
@@ -51,11 +51,11 @@ class LateInteractionTextEmbedding(LateInteractionTextEmbeddingBase):
     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,
+        cache_dir: str | None = None,
+        threads: int | None = None,
+        providers: Sequence[OnnxProvider] | None = None,
+        cuda: bool | Device = Device.AUTO,
+        device_ids: list[int] | None = None,
         lazy_load: bool = False,
         **kwargs: Any,
     ):
@@ -80,11 +80,45 @@ class LateInteractionTextEmbedding(LateInteractionTextEmbeddingBase):
             "Please check the supported models using `LateInteractionTextEmbedding.list_supported_models()`"
         )
 
+    @property
+    def embedding_size(self) -> int:
+        """Get the embedding size of the current model"""
+        if self._embedding_size is None:
+            self._embedding_size = self.get_embedding_size(self.model_name)
+        return self._embedding_size
+
+    @classmethod
+    def get_embedding_size(cls, model_name: str) -> int:
+        """Get the embedding size of the passed model
+
+        Args:
+            model_name (str): The name of the model to get embedding size for.
+
+        Returns:
+            int: The size of the embedding.
+
+        Raises:
+            ValueError: If the model name is not found in the supported models.
+        """
+        descriptions = cls._list_supported_models()
+        embedding_size: int | None = None
+        for description in descriptions:
+            if description.model.lower() == model_name.lower():
+                embedding_size = description.dim
+                break
+        if embedding_size is None:
+            model_names = [description.model for description in descriptions]
+            raise ValueError(
+                f"Embedding size for model {model_name} was None. "
+                f"Available model names: {model_names}"
+            )
+        return embedding_size
+
     def embed(
         self,
-        documents: Union[str, Iterable[str]],
+        documents: str | Iterable[str],
         batch_size: int = 256,
-        parallel: Optional[int] = None,
+        parallel: int | None = None,
         **kwargs: Any,
     ) -> Iterable[NumpyArray]:
         """
@@ -104,7 +138,7 @@ class LateInteractionTextEmbedding(LateInteractionTextEmbeddingBase):
         """
         yield from self.model.embed(documents, batch_size, parallel, **kwargs)
 
-    def query_embed(self, query: Union[str, Iterable[str]], **kwargs: Any) -> Iterable[NumpyArray]:
+    def query_embed(self, query: str | Iterable[str], **kwargs: Any) -> Iterable[NumpyArray]:
         """
         Embeds queries
 
@@ -117,3 +151,30 @@ class LateInteractionTextEmbedding(LateInteractionTextEmbeddingBase):
 
         # This is model-specific, so that different models can have specialized implementations
         yield from self.model.query_embed(query, **kwargs)
+
+    def token_count(
+        self,
+        texts: str | Iterable[str],
+        batch_size: int = 1024,
+        is_doc: bool = True,
+        include_extension: bool = False,
+        **kwargs: Any,
+    ) -> int:
+        """Returns the number of tokens in the texts.
+
+        Args:
+            texts (str | Iterable[str]): The list of texts to embed.
+            batch_size (int): Batch size for encoding
+            is_doc (bool): Whether the texts are documents (disable embedding a query with include_mask=True).
+            include_extension (bool): Turn on to count DOC / QUERY marker tokens, and [MASK] token in query mode.
+
+        Returns:
+            int: Sum of number of tokens in the texts.
+        """
+        return self.model.token_count(
+            texts,
+            batch_size=batch_size,
+            is_doc=is_doc,
+            include_extension=include_extension,
+            **kwargs,
+        )

fastembed/late_interaction/token_embeddings.py

@@ -0,0 +1,83 @@
+from dataclasses import asdict
+from typing import Iterable, Any, Type
+
+from fastembed.common.model_description import DenseModelDescription, ModelSource
+from fastembed.common.onnx_model import OnnxOutputContext
+from fastembed.common.types import NumpyArray
+from fastembed.late_interaction.late_interaction_embedding_base import (
+    LateInteractionTextEmbeddingBase,
+)
+from fastembed.text.onnx_embedding import OnnxTextEmbedding
+from fastembed.text.onnx_text_model import TextEmbeddingWorker
+
+
+supported_token_embeddings_models = [
+    DenseModelDescription(
+        model="jinaai/jina-embeddings-v2-small-en-tokens",
+        dim=512,
+        description="Text embeddings, Unimodal (text), English, 8192 input tokens truncation,"
+        " Prefixes for queries/documents: not necessary, 2023 year.",
+        license="apache-2.0",
+        size_in_GB=0.12,
+        sources=ModelSource(hf="xenova/jina-embeddings-v2-small-en"),
+        model_file="onnx/model.onnx",
+    ),
+]
+
+
+class TokenEmbeddingsModel(OnnxTextEmbedding, LateInteractionTextEmbeddingBase):
+    @classmethod
+    def _list_supported_models(cls) -> list[DenseModelDescription]:
+        """Lists the supported models.
+
+        Returns:
+            list[DenseModelDescription]: A list of DenseModelDescription objects containing the model information.
+        """
+        return supported_token_embeddings_models
+
+    @classmethod
+    def list_supported_models(cls) -> list[dict[str, Any]]:
+        """Lists the supported models.
+
+        Returns:
+            list[dict[str, Any]]: A list of dictionaries containing the model information.
+        """
+        return [asdict(model) for model in cls._list_supported_models()]
+
+    @classmethod
+    def _get_worker_class(cls) -> Type[TextEmbeddingWorker[NumpyArray]]:
+        return TokensEmbeddingWorker
+
+    def _post_process_onnx_output(
+        self, output: OnnxOutputContext, **kwargs: Any
+    ) -> Iterable[NumpyArray]:
+        # Size: (batch_size, sequence_length, hidden_size)
+        embeddings = output.model_output
+        # Size: (batch_size, sequence_length)
+        assert output.attention_mask is not None
+        masks = output.attention_mask
+
+        # For each document we only select those embeddings that are not masked out
+        for i in range(embeddings.shape[0]):
+            yield embeddings[i, masks[i] == 1]
+
+    def embed(
+        self,
+        documents: str | Iterable[str],
+        batch_size: int = 256,
+        parallel: int | None = None,
+        **kwargs: Any,
+    ) -> Iterable[NumpyArray]:
+        yield from super().embed(documents, batch_size=batch_size, parallel=parallel, **kwargs)
+
+
+class TokensEmbeddingWorker(TextEmbeddingWorker[NumpyArray]):
+    def init_embedding(
+        self, model_name: str, cache_dir: str, **kwargs: Any
+    ) -> TokenEmbeddingsModel:
+        return TokenEmbeddingsModel(
+            model_name=model_name,
+            cache_dir=cache_dir,
+            threads=1,
+            **kwargs,
+        )

fastembed/late_interaction_multimodal/colmodernvbert.py

@@ -0,0 +1,532 @@
+import contextlib
+from typing import Any, Iterable, Type, Optional, Sequence
+import json
+
+import numpy as np
+from tokenizers import Encoding
+from PIL import Image
+
+from fastembed.common import ImageInput
+from fastembed.common.model_description import DenseModelDescription, ModelSource
+from fastembed.common.onnx_model import OnnxOutputContext
+from fastembed.common.types import NumpyArray, OnnxProvider
+from fastembed.common.utils import define_cache_dir, iter_batch
+from fastembed.late_interaction_multimodal.late_interaction_multimodal_embedding_base import (
+    LateInteractionMultimodalEmbeddingBase,
+)
+from fastembed.late_interaction_multimodal.onnx_multimodal_model import (
+    OnnxMultimodalModel,
+    TextEmbeddingWorker,
+    ImageEmbeddingWorker,
+)
+
+supported_colmodernvbert_models: list[DenseModelDescription] = [
+    DenseModelDescription(
+        model="Qdrant/colmodernvbert",
+        dim=128,
+        description="The late-interaction version of ModernVBERT, CPU friendly, English, 2025.",
+        license="mit",
+        size_in_GB=1.0,
+        sources=ModelSource(hf="Qdrant/colmodernvbert"),
+        additional_files=["processor_config.json"],
+        model_file="model.onnx",
+    ),
+]
+
+
+class ColModernVBERT(LateInteractionMultimodalEmbeddingBase, OnnxMultimodalModel[NumpyArray]):
+    """
+    The ModernVBERT/colmodernvbert model implementation. This model uses
+    bidirectional attention, which proves to work better for retrieval.
+
+    See: https://huggingface.co/ModernVBERT/colmodernvbert
+    """
+
+    VISUAL_PROMPT_PREFIX = (
+        "<|begin_of_text|>User:<image>Describe the image.<end_of_utterance>\nAssistant:"
+    )
+    QUERY_AUGMENTATION_TOKEN = "<end_of_utterance>"
+
+    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,
+    ):
+        """
+        Args:
+            model_name (str): The name of the model to use.
+            cache_dir (str, optional): The path to the cache directory.
+                                       Can be set using the `FASTEMBED_CACHE_PATH` env variable.
+                                       Defaults to `fastembed_cache` in the system's temp directory.
+            threads (int, optional): The number of threads single onnxruntime session can use. Defaults to None.
+            providers (Optional[Sequence[OnnxProvider]], optional): The list of onnxruntime providers to use.
+                Mutually exclusive with the `cuda` and `device_ids` arguments. Defaults to None.
+            cuda (bool, optional): Whether to use cuda for inference. Mutually exclusive with `providers`
+                Defaults to False.
+            device_ids (Optional[list[int]], optional): The list of device ids to use for data parallel processing in
+                workers. Should be used with `cuda=True`, mutually exclusive with `providers`. Defaults to None.
+            lazy_load (bool, optional): Whether to load the model during class initialization or on demand.
+                Should be set to True when using multiple-gpu and parallel encoding. Defaults to False.
+            device_id (Optional[int], optional): The device id to use for loading the model in the worker process.
+
+        Raises:
+            ValueError: If the model_name is not in the format <org>/<model> e.g. BAAI/bge-base-en.
+        """
+        super().__init__(model_name, cache_dir, threads, **kwargs)
+        self.providers = providers
+        self.lazy_load = lazy_load
+        self._extra_session_options = self._select_exposed_session_options(kwargs)
+
+        # List of device ids, that can be used for data parallel processing in workers
+        self.device_ids = device_ids
+        self.cuda = cuda
+
+        # This device_id will be used if we need to load model in current process
+        self.device_id: Optional[int] = None
+        if device_id is not None:
+            self.device_id = device_id
+        elif self.device_ids is not None:
+            self.device_id = self.device_ids[0]
+
+        self.model_description = self._get_model_description(model_name)
+        self.cache_dir = str(define_cache_dir(cache_dir))
+
+        self._specific_model_path = specific_model_path
+        self._model_dir = self.download_model(
+            self.model_description,
+            self.cache_dir,
+            local_files_only=self._local_files_only,
+            specific_model_path=self._specific_model_path,
+        )
+        self.mask_token_id = None
+        self.pad_token_id = None
+        self.image_seq_len: Optional[int] = None
+        self.max_image_size: Optional[int] = None
+        self.image_size: Optional[int] = None
+
+        if not self.lazy_load:
+            self.load_onnx_model()
+
+    @classmethod
+    def _list_supported_models(cls) -> list[DenseModelDescription]:
+        """Lists the supported models.
+
+        Returns:
+            list[DenseModelDescription]: A list of DenseModelDescription objects containing the model information.
+        """
+        return supported_colmodernvbert_models
+
+    def load_onnx_model(self) -> None:
+        self._load_onnx_model(
+            model_dir=self._model_dir,
+            model_file=self.model_description.model_file,
+            threads=self.threads,
+            providers=self.providers,
+            cuda=self.cuda,
+            device_id=self.device_id,
+            extra_session_options=self._extra_session_options,
+        )
+
+        # Load image processing configuration
+        processor_config_path = self._model_dir / "processor_config.json"
+        with open(processor_config_path) as f:
+            processor_config = json.load(f)
+            self.image_seq_len = processor_config.get("image_seq_len", 64)
+
+        preprocessor_config_path = self._model_dir / "preprocessor_config.json"
+        with open(preprocessor_config_path) as f:
+            preprocessor_config = json.load(f)
+            self.max_image_size = preprocessor_config.get("max_image_size", {}).get(
+                "longest_edge", 512
+            )
+
+        # Load model configuration
+        config_path = self._model_dir / "config.json"
+        with open(config_path) as f:
+            model_config = json.load(f)
+            vision_config = model_config.get("vision_config", {})
+            self.image_size = vision_config.get("image_size", 512)
+
+    def _preprocess_onnx_text_input(
+        self, onnx_input: dict[str, NumpyArray], **kwargs: Any
+    ) -> dict[str, NumpyArray]:
+        """
+        Post-process the ONNX model output to convert it into a usable format.
+
+        Args:
+            output (OnnxOutputContext): The raw output from the ONNX model.
+
+        Returns:
+            Iterable[NumpyArray]: Post-processed output as NumPy arrays.
+        """
+        batch_size, seq_length = onnx_input["input_ids"].shape
+        empty_image_placeholder: NumpyArray = np.zeros(
+            (batch_size, seq_length, 3, self.image_size, self.image_size),
+            dtype=np.float32,  # type: ignore[type-var,arg-type,assignment]
+        )
+        onnx_input["pixel_values"] = empty_image_placeholder
+        return onnx_input
+
+    def _post_process_onnx_text_output(
+        self,
+        output: OnnxOutputContext,
+    ) -> Iterable[NumpyArray]:
+        """
+        Post-process the ONNX model output to convert it into a usable format.
+
+        Args:
+            output (OnnxOutputContext): The raw output from the ONNX model.
+
+        Returns:
+            Iterable[NumpyArray]: Post-processed output as NumPy arrays.
+        """
+        return output.model_output
+
+    def tokenize(self, documents: list[str], **kwargs: Any) -> list[Encoding]:
+        # Add query augmentation tokens (matching process_queries logic from colpali-engine)
+        augmented_queries = [doc + self.QUERY_AUGMENTATION_TOKEN * 10 for doc in documents]
+        encoded = self.tokenizer.encode_batch(augmented_queries)  # type: ignore[union-attr]
+        return encoded
+
+    def token_count(
+        self,
+        texts: str | Iterable[str],
+        batch_size: int = 1024,
+        include_extension: bool = False,
+        **kwargs: Any,
+    ) -> int:
+        if not hasattr(self, "model") or self.model is None:
+            self.load_onnx_model()  # loads the tokenizer as well
+        token_num = 0
+        texts = [texts] if isinstance(texts, str) else texts
+        assert self.tokenizer is not None
+        tokenize_func = self.tokenize if include_extension else self.tokenizer.encode_batch
+        for batch in iter_batch(texts, batch_size):
+            token_num += sum([sum(encoding.attention_mask) for encoding in tokenize_func(batch)])
+        return token_num
+
+    def onnx_embed_image(self, images: list[ImageInput], **kwargs: Any) -> OnnxOutputContext:
+        with contextlib.ExitStack() as stack:
+            image_files = [
+                stack.enter_context(Image.open(image))
+                if not isinstance(image, Image.Image)
+                else image
+                for image in images
+            ]
+            assert self.processor is not None, "Processor is not initialized"
+            processed = self.processor(image_files)
+            encoded, attention_mask, metadata = self._process_nested_patches(processed)  # type: ignore[arg-type]
+
+        onnx_input = {"pixel_values": encoded, "attention_mask": attention_mask}
+        onnx_input = self._preprocess_onnx_image_input(onnx_input, **kwargs)
+        model_output = self.model.run(None, onnx_input)  # type: ignore[union-attr]
+
+        return OnnxOutputContext(
+            model_output=model_output[0],
+            attention_mask=attention_mask,  # type: ignore[arg-type]
+            metadata=metadata,
+        )
+
+    @staticmethod
+    def _process_nested_patches(
+        processed: list[list[NumpyArray]],
+    ) -> tuple[NumpyArray, NumpyArray, dict[str, Any]]:
+        """
+        Process nested image patches (from ImageSplitter).
+
+        Args:
+            processed: List of patch lists, one per image [[img1_patches], [img2_patches], ...]
+
+        Returns:
+            tuple: (encoded array, attention_mask, metadata)
+                - encoded: (batch_size, max_patches, C, H, W)
+                - attention_mask: (batch_size, max_patches) with 1 for real patches, 0 for padding
+                - metadata: Dict with 'patch_counts' key
+        """
+        patch_counts = [len(patches) for patches in processed]
+        max_patches = max(patch_counts)
+
+        # Get dimensions from first patch
+        channels, height, width = processed[0][0].shape
+        batch_size = len(processed)
+
+        # Create padded array
+        encoded = np.zeros(
+            (batch_size, max_patches, channels, height, width), dtype=processed[0][0].dtype
+        )
+
+        # Create attention mask (1 for real patches, 0 for padding)
+        attention_mask = np.zeros((batch_size, max_patches), dtype=np.int64)
+
+        # Fill in patches and attention mask
+        for i, patches in enumerate(processed):
+            for j, patch in enumerate(patches):
+                encoded[i, j] = patch
+                attention_mask[i, j] = 1
+
+        metadata = {"patch_counts": patch_counts}
+        return encoded, attention_mask, metadata  # type: ignore[return-value]
+
+    def _preprocess_onnx_image_input(
+        self, onnx_input: dict[str, np.ndarray], **kwargs: Any
+    ) -> dict[str, NumpyArray]:
+        """
+        Add text input placeholders for image data, following Idefics3 processing logic.
+
+        Constructs input_ids dynamically based on the actual number of image patches,
+        using the same token expansion logic as Idefics3Processor.
+
+        Args:
+            onnx_input: Dict with 'pixel_values' (batch, num_patches, C, H, W)
+                        and 'attention_mask' (batch, num_patches) indicating real patches
+            **kwargs: Additional arguments
+
+        Returns:
+            Updated onnx_input with 'input_ids' and updated 'attention_mask' for token sequence
+        """
+        # The attention_mask in onnx_input has a shape of (batch_size, num_patches),
+        # and should be used to create an attention mask matching the input_ids shape.
+        patch_attention_mask = onnx_input["attention_mask"]
+        pixel_values = onnx_input["pixel_values"]
+
+        batch_size = pixel_values.shape[0]
+        batch_input_ids = []
+
+        # Build input_ids for each image based on its actual patch count
+        for i in range(batch_size):
+            # Count real patches (non-padded) from attention mask
+            patch_count = int(np.sum(patch_attention_mask[i]))
+
+            # Compute rows/cols from patch count
+            rows, cols = self._compute_rows_cols_from_patches(patch_count)
+
+            # Build input_ids for this image
+            input_ids = self._build_input_ids_for_image(rows, cols)
+            batch_input_ids.append(input_ids)
+
+        # Pad sequences to max length in batch
+        max_len = max(len(ids) for ids in batch_input_ids)
+
+        # Get padding config from tokenizer
+        padding_direction = self.tokenizer.padding["direction"]  # type: ignore[index,union-attr]
+        pad_token_id = self.tokenizer.padding["pad_id"]  # type: ignore[index,union-attr]
+
+        # Initialize with pad token
+        padded_input_ids = np.full((batch_size, max_len), pad_token_id, dtype=np.int64)
+        attention_mask = np.zeros((batch_size, max_len), dtype=np.int64)
+
+        for i, input_ids in enumerate(batch_input_ids):
+            seq_len = len(input_ids)
+            if padding_direction == "left":
+                # Left padding: place tokens at the END of the array
+                start_idx = max_len - seq_len
+                padded_input_ids[i, start_idx:] = input_ids
+                attention_mask[i, start_idx:] = 1
+            else:
+                # Right padding: place tokens at the START of the array
+                padded_input_ids[i, :seq_len] = input_ids
+                attention_mask[i, :seq_len] = 1
+
+        onnx_input["input_ids"] = padded_input_ids
+        # Update attention_mask with token-level data
+        onnx_input["attention_mask"] = attention_mask
+        return onnx_input
+
+    @staticmethod
+    def _compute_rows_cols_from_patches(patch_count: int) -> tuple[int, int]:
+        if patch_count <= 1:
+            return 0, 0
+
+        # Subtract 1 for the global image
+        grid_patches = patch_count - 1
+
+        # Find rows and cols (assume square or near-square grid)
+        rows = int(grid_patches**0.5)
+        cols = grid_patches // rows
+
+        # Verify the calculation
+        if rows * cols + 1 != patch_count:
+            # Handle non-square grids
+            for r in range(1, grid_patches + 1):
+                if grid_patches % r == 0:
+                    c = grid_patches // r
+                    if r * c + 1 == patch_count:
+                        return r, c
+            # Fallback: treat as unsplit
+            return 0, 0
+
+        return rows, cols
+
+    def _create_single_image_prompt_string(self) -> str:
+        return (
+            "<fake_token_around_image>"
+            + "<global-img>"
+            + "<image>" * self.image_seq_len  # type: ignore[operator]
+            + "<fake_token_around_image>"
+        )
+
+    def _create_split_image_prompt_string(self, rows: int, cols: int) -> str:
+        text_split_images = ""
+
+        # Add tokens for each patch in the grid
+        for n_h in range(rows):
+            for n_w in range(cols):
+                text_split_images += (
+                    "<fake_token_around_image>"
+                    + f"<row_{n_h + 1}_col_{n_w + 1}>"
+                    + "<image>" * self.image_seq_len  # type: ignore[operator]
+                )
+            text_split_images += "\n"
+
+        # Add global image at the end
+        text_split_images += (
+            "\n<fake_token_around_image>"
+            + "<global-img>"
+            + "<image>" * self.image_seq_len  # type: ignore[operator]
+            + "<fake_token_around_image>"
+        )
+
+        return text_split_images
+
+    def _build_input_ids_for_image(self, rows: int, cols: int) -> np.ndarray:
+        # Create the appropriate image prompt string
+        if rows == 0 and cols == 0:
+            image_prompt_tokens = self._create_single_image_prompt_string()
+        else:
+            image_prompt_tokens = self._create_split_image_prompt_string(rows, cols)
+
+        # Replace <image> in visual prompt with expanded tokens
+        # The visual prompt is: "<|begin_of_text|>User:<image>Describe the image.<end_of_utterance>\nAssistant:"
+        expanded_prompt = self.VISUAL_PROMPT_PREFIX.replace("<image>", image_prompt_tokens)
+
+        # Tokenize the complete prompt
+        encoded = self.tokenizer.encode(expanded_prompt)  # type: ignore[union-attr]
+
+        # Convert to numpy array
+        return np.array(encoded.ids, dtype=np.int64)
+
+    def _post_process_onnx_image_output(
+        self,
+        output: OnnxOutputContext,
+    ) -> Iterable[NumpyArray]:
+        """
+        Post-process the ONNX model output to convert it into a usable format.
+
+        Args:
+            output (OnnxOutputContext): The raw output from the ONNX model.
+
+        Returns:
+            Iterable[NumpyArray]: Post-processed output as NumPy arrays.
+        """
+        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
+        )
+
+    def embed_text(
+        self,
+        documents: str | Iterable[str],
+        batch_size: int = 256,
+        parallel: Optional[int] = None,
+        **kwargs: Any,
+    ) -> Iterable[NumpyArray]:
+        """
+        Encode a list of documents into list of embeddings.
+
+        Args:
+            documents: Iterator of documents or single document to embed
+            batch_size: Batch size for encoding -- higher values will use more memory, but be faster
+            parallel:
+                If > 1, data-parallel encoding will be used, recommended for offline encoding of large datasets.
+                If 0, use all available cores.
+                If None, don't use data-parallel processing, use default onnxruntime threading instead.
+
+        Returns:
+            List of embeddings, one per document
+        """
+        yield from self._embed_documents(
+            model_name=self.model_name,
+            cache_dir=str(self.cache_dir),
+            documents=documents,
+            batch_size=batch_size,
+            parallel=parallel,
+            providers=self.providers,
+            cuda=self.cuda,
+            device_ids=self.device_ids,
+            local_files_only=self._local_files_only,
+            specific_model_path=self._specific_model_path,
+            extra_session_options=self._extra_session_options,
+            **kwargs,
+        )
+
+    def embed_image(
+        self,
+        images: ImageInput | Iterable[ImageInput],
+        batch_size: int = 16,
+        parallel: Optional[int] = None,
+        **kwargs: Any,
+    ) -> Iterable[NumpyArray]:
+        """
+        Encode a list of images into list of embeddings.
+
+        Args:
+            images: Iterator of image paths or single image path to embed
+            batch_size: Batch size for encoding -- higher values will use more memory, but be faster
+            parallel:
+                If > 1, data-parallel encoding will be used, recommended for offline encoding of large datasets.
+                If 0, use all available cores.
+                If None, don't use data-parallel processing, use default onnxruntime threading instead.
+
+        Returns:
+            List of embeddings, one per document
+        """
+        yield from self._embed_images(
+            model_name=self.model_name,
+            cache_dir=str(self.cache_dir),
+            images=images,
+            batch_size=batch_size,
+            parallel=parallel,
+            providers=self.providers,
+            cuda=self.cuda,
+            device_ids=self.device_ids,
+            local_files_only=self._local_files_only,
+            specific_model_path=self._specific_model_path,
+            extra_session_options=self._extra_session_options,
+            **kwargs,
+        )
+
+    @classmethod
+    def _get_text_worker_class(cls) -> Type[TextEmbeddingWorker[NumpyArray]]:
+        return ColModernVBERTTextEmbeddingWorker
+
+    @classmethod
+    def _get_image_worker_class(cls) -> Type[ImageEmbeddingWorker[NumpyArray]]:
+        return ColModernVBERTImageEmbeddingWorker
+
+
+class ColModernVBERTTextEmbeddingWorker(TextEmbeddingWorker[NumpyArray]):
+    def init_embedding(self, model_name: str, cache_dir: str, **kwargs: Any) -> ColModernVBERT:
+        return ColModernVBERT(
+            model_name=model_name,
+            cache_dir=cache_dir,
+            threads=1,
+            **kwargs,
+        )
+
+
+class ColModernVBERTImageEmbeddingWorker(ImageEmbeddingWorker[NumpyArray]):
+    def init_embedding(self, model_name: str, cache_dir: str, **kwargs: Any) -> ColModernVBERT:
+        return ColModernVBERT(
+            model_name=model_name,
+            cache_dir=cache_dir,
+            threads=1,
+            **kwargs,
+        )

fastembed/late_interaction_multimodal/colpali.py

@@ -1,12 +1,12 @@
-from typing import Any, Iterable, Optional, Sequence, Type, Union
+from typing import Any, Iterable, Sequence, Type
 
 import numpy as np
 from tokenizers import Encoding
 
 from fastembed.common import OnnxProvider, ImageInput
 from fastembed.common.onnx_model import OnnxOutputContext
-from fastembed.common.types import NumpyArray
-from fastembed.common.utils import define_cache_dir
+from fastembed.common.types import NumpyArray, Device
+from fastembed.common.utils import define_cache_dir, iter_batch
 from fastembed.late_interaction_multimodal.late_interaction_multimodal_embedding_base import (
     LateInteractionMultimodalEmbeddingBase,
 )
@@ -46,14 +46,14 @@ class ColPali(LateInteractionMultimodalEmbeddingBase, OnnxMultimodalModel[NumpyA
     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,
+        cache_dir: str | None = None,
+        threads: int | None = None,
+        providers: Sequence[OnnxProvider] | None = None,
+        cuda: bool | Device = Device.AUTO,
+        device_ids: list[int] | None = None,
         lazy_load: bool = False,
-        device_id: Optional[int] = None,
-        specific_model_path: Optional[str] = None,
+        device_id: int | None = None,
+        specific_model_path: str | None = None,
         **kwargs: Any,
     ):
         """
@@ -65,10 +65,11 @@ class ColPali(LateInteractionMultimodalEmbeddingBase, OnnxMultimodalModel[NumpyA
             threads (int, optional): The number of threads single onnxruntime session can use. Defaults to None.
             providers (Optional[Sequence[OnnxProvider]], optional): The list of onnxruntime providers to use.
                 Mutually exclusive with the `cuda` and `device_ids` arguments. Defaults to None.
-            cuda (bool, optional): Whether to use cuda for inference. Mutually exclusive with `providers`
-                Defaults to False.
+            cuda (Union[bool, Device], optional): Whether to use cuda for inference. Mutually exclusive with `providers`
+                Defaults to Device.AUTO.
             device_ids (Optional[list[int]], optional): The list of device ids to use for data parallel processing in
-                workers. Should be used with `cuda=True`, mutually exclusive with `providers`. Defaults to None.
+                workers. Should be used with `cuda` equals to `True`, `Device.AUTO` or `Device.CUDA`, mutually exclusive
+                with `providers`. Defaults to None.
             lazy_load (bool, optional): Whether to load the model during class initialization or on demand.
                 Should be set to True when using multiple-gpu and parallel encoding. Defaults to False.
             device_id (Optional[int], optional): The device id to use for loading the model in the worker process.
@@ -80,13 +81,14 @@ class ColPali(LateInteractionMultimodalEmbeddingBase, OnnxMultimodalModel[NumpyA
         super().__init__(model_name, cache_dir, threads, **kwargs)
         self.providers = providers
         self.lazy_load = lazy_load
+        self._extra_session_options = self._select_exposed_session_options(kwargs)
 
         # List of device ids, that can be used for data parallel processing in workers
         self.device_ids = device_ids
         self.cuda = cuda
 
         # This device_id will be used if we need to load model in current process
-        self.device_id: Optional[int] = None
+        self.device_id: int | None = None
         if device_id is not None:
             self.device_id = device_id
         elif self.device_ids is not None:
@@ -95,11 +97,12 @@ class ColPali(LateInteractionMultimodalEmbeddingBase, OnnxMultimodalModel[NumpyA
         self.model_description = self._get_model_description(model_name)
         self.cache_dir = str(define_cache_dir(cache_dir))
 
+        self._specific_model_path = specific_model_path
         self._model_dir = self.download_model(
             self.model_description,
             self.cache_dir,
             local_files_only=self._local_files_only,
-            specific_model_path=specific_model_path,
+            specific_model_path=self._specific_model_path,
         )
         self.mask_token_id = None
         self.pad_token_id = None
@@ -124,6 +127,7 @@ class ColPali(LateInteractionMultimodalEmbeddingBase, OnnxMultimodalModel[NumpyA
             providers=self.providers,
             cuda=self.cuda,
             device_id=self.device_id,
+            extra_session_options=self._extra_session_options,
         )
 
     def _post_process_onnx_image_output(
@@ -142,7 +146,7 @@ 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 +161,7 @@ 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] = []
@@ -169,12 +173,29 @@ class ColPali(LateInteractionMultimodalEmbeddingBase, OnnxMultimodalModel[NumpyA
         encoded = self.tokenizer.encode_batch(texts_query)  # type: ignore[union-attr]
         return encoded
 
+    def token_count(
+        self,
+        texts: str | Iterable[str],
+        batch_size: int = 1024,
+        include_extension: bool = False,
+        **kwargs: Any,
+    ) -> int:
+        if not hasattr(self, "model") or self.model is None:
+            self.load_onnx_model()  # loads the tokenizer as well
+        token_num = 0
+        texts = [texts] if isinstance(texts, str) else texts
+        assert self.tokenizer is not None
+        tokenize_func = self.tokenize if include_extension else self.tokenizer.encode_batch
+        for batch in iter_batch(texts, batch_size):
+            token_num += sum([sum(encoding.attention_mask) for encoding in tokenize_func(batch)])
+        return token_num
+
     def _preprocess_onnx_text_input(
         self, onnx_input: dict[str, NumpyArray], **kwargs: Any
     ) -> dict[str, NumpyArray]:
         onnx_input["input_ids"] = np.array(
             [
-                self.QUERY_MARKER_TOKEN_ID + input_ids[2:].tolist()
+                self.QUERY_MARKER_TOKEN_ID + input_ids[2:].tolist()  # type: ignore[index]
                 for input_ids in onnx_input["input_ids"]
             ]
         )
@@ -197,20 +218,19 @@ class ColPali(LateInteractionMultimodalEmbeddingBase, OnnxMultimodalModel[NumpyA
         Returns:
             Dict[str, NumpyArray]: ONNX input with text placeholders.
         """
-
         onnx_input["input_ids"] = np.array(
-            [self.EMPTY_TEXT_PLACEHOLDER for _ in onnx_input["input_ids"]]
+            [self.EMPTY_TEXT_PLACEHOLDER for _ in onnx_input["pixel_values"]]
         )
         onnx_input["attention_mask"] = np.array(
-            [self.EVEN_ATTENTION_MASK for _ in onnx_input["input_ids"]]
+            [self.EVEN_ATTENTION_MASK for _ in onnx_input["pixel_values"]]
         )
         return onnx_input
 
     def embed_text(
         self,
-        documents: Union[str, Iterable[str]],
+        documents: str | Iterable[str],
         batch_size: int = 256,
-        parallel: Optional[int] = None,
+        parallel: int | None = None,
         **kwargs: Any,
     ) -> Iterable[NumpyArray]:
         """
@@ -236,14 +256,17 @@ class ColPali(LateInteractionMultimodalEmbeddingBase, OnnxMultimodalModel[NumpyA
             providers=self.providers,
             cuda=self.cuda,
             device_ids=self.device_ids,
+            local_files_only=self._local_files_only,
+            specific_model_path=self._specific_model_path,
+            extra_session_options=self._extra_session_options,
             **kwargs,
         )
 
     def embed_image(
         self,
-        images: Union[ImageInput, Iterable[ImageInput]],
+        images: ImageInput | Iterable[ImageInput],
         batch_size: int = 16,
-        parallel: Optional[int] = None,
+        parallel: int | None = None,
         **kwargs: Any,
     ) -> Iterable[NumpyArray]:
         """
@@ -269,6 +292,9 @@ class ColPali(LateInteractionMultimodalEmbeddingBase, OnnxMultimodalModel[NumpyA
             providers=self.providers,
             cuda=self.cuda,
             device_ids=self.device_ids,
+            local_files_only=self._local_files_only,
+            specific_model_path=self._specific_model_path,
+            extra_session_options=self._extra_session_options,
             **kwargs,
         )
 

fastembed/late_interaction_multimodal/late_interaction_multimodal_embedding.py

@@ -1,9 +1,10 @@
-from typing import Any, Iterable, Optional, Sequence, Type, Union
+from typing import Any, Iterable, Sequence, Type
 from dataclasses import asdict
 
 from fastembed.common import OnnxProvider, ImageInput
-from fastembed.common.types import NumpyArray
+from fastembed.common.types import NumpyArray, Device
 from fastembed.late_interaction_multimodal.colpali import ColPali
+from fastembed.late_interaction_multimodal.colmodernvbert import ColModernVBERT
 
 from fastembed.late_interaction_multimodal.late_interaction_multimodal_embedding_base import (
     LateInteractionMultimodalEmbeddingBase,
@@ -12,7 +13,10 @@ from fastembed.common.model_description import DenseModelDescription
 
 
 class LateInteractionMultimodalEmbedding(LateInteractionMultimodalEmbeddingBase):
-    EMBEDDINGS_REGISTRY: list[Type[LateInteractionMultimodalEmbeddingBase]] = [ColPali]
+    EMBEDDINGS_REGISTRY: list[Type[LateInteractionMultimodalEmbeddingBase]] = [
+        ColPali,
+        ColModernVBERT,
+    ]
 
     @classmethod
     def list_supported_models(cls) -> list[dict[str, Any]]:
@@ -54,11 +58,11 @@ class LateInteractionMultimodalEmbedding(LateInteractionMultimodalEmbeddingBase)
     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,
+        cache_dir: str | None = None,
+        threads: int | None = None,
+        providers: Sequence[OnnxProvider] | None = None,
+        cuda: bool | Device = Device.AUTO,
+        device_ids: list[int] | None = None,
         lazy_load: bool = False,
         **kwargs: Any,
     ):
@@ -83,11 +87,45 @@ class LateInteractionMultimodalEmbedding(LateInteractionMultimodalEmbeddingBase)
             "Please check the supported models using `LateInteractionMultimodalEmbedding.list_supported_models()`"
         )
 
+    @property
+    def embedding_size(self) -> int:
+        """Get the embedding size of the current model"""
+        if self._embedding_size is None:
+            self._embedding_size = self.get_embedding_size(self.model_name)
+        return self._embedding_size
+
+    @classmethod
+    def get_embedding_size(cls, model_name: str) -> int:
+        """Get the embedding size of the passed model
+
+        Args:
+            model_name (str): The name of the model to get embedding size for.
+
+        Returns:
+            int: The size of the embedding.
+
+        Raises:
+            ValueError: If the model name is not found in the supported models.
+        """
+        descriptions = cls._list_supported_models()
+        embedding_size: int | None = None
+        for description in descriptions:
+            if description.model.lower() == model_name.lower():
+                embedding_size = description.dim
+                break
+        if embedding_size is None:
+            model_names = [description.model for description in descriptions]
+            raise ValueError(
+                f"Embedding size for model {model_name} was None. "
+                f"Available model names: {model_names}"
+            )
+        return embedding_size
+
     def embed_text(
         self,
-        documents: Union[str, Iterable[str]],
+        documents: str | Iterable[str],
         batch_size: int = 256,
-        parallel: Optional[int] = None,
+        parallel: int | None = None,
         **kwargs: Any,
     ) -> Iterable[NumpyArray]:
         """
@@ -108,9 +146,9 @@ class LateInteractionMultimodalEmbedding(LateInteractionMultimodalEmbeddingBase)
 
     def embed_image(
         self,
-        images: Union[ImageInput, Iterable[ImageInput]],
+        images: ImageInput | Iterable[ImageInput],
         batch_size: int = 16,
-        parallel: Optional[int] = None,
+        parallel: int | None = None,
         **kwargs: Any,
     ) -> Iterable[NumpyArray]:
         """
@@ -128,3 +166,24 @@ class LateInteractionMultimodalEmbedding(LateInteractionMultimodalEmbeddingBase)
             List of embeddings, one per image
         """
         yield from self.model.embed_image(images, batch_size, parallel, **kwargs)
+
+    def token_count(
+        self,
+        texts: str | Iterable[str],
+        batch_size: int = 1024,
+        include_extension: bool = False,
+        **kwargs: Any,
+    ) -> int:
+        """Returns the number of tokens in the texts.
+
+        Args:
+            texts (str | Iterable[str]): The list of texts to embed.
+            batch_size (int): Batch size for encoding
+            include_extension (bool): Whether to include tokens added by preprocessing
+
+        Returns:
+            int: Sum of number of tokens in the texts.
+        """
+        return self.model.token_count(
+            texts, batch_size=batch_size, include_extension=include_extension, **kwargs
+        )

fastembed/late_interaction_multimodal/late_interaction_multimodal_embedding_base.py

@@ -1,4 +1,4 @@
-from typing import Iterable, Optional, Union, Any
+from typing import Iterable, Any
 
 
 from fastembed.common import ImageInput
@@ -11,20 +11,21 @@ class LateInteractionMultimodalEmbeddingBase(ModelManagement[DenseModelDescripti
     def __init__(
         self,
         model_name: str,
-        cache_dir: Optional[str] = None,
-        threads: Optional[int] = None,
+        cache_dir: str | None = None,
+        threads: int | None = None,
         **kwargs: Any,
     ):
         self.model_name = model_name
         self.cache_dir = cache_dir
         self.threads = threads
         self._local_files_only = kwargs.pop("local_files_only", False)
+        self._embedding_size: int | None = None
 
     def embed_text(
         self,
-        documents: Union[str, Iterable[str]],
+        documents: str | Iterable[str],
         batch_size: int = 256,
-        parallel: Optional[int] = None,
+        parallel: int | None = None,
         **kwargs: Any,
     ) -> Iterable[NumpyArray]:
         """
@@ -46,9 +47,9 @@ class LateInteractionMultimodalEmbeddingBase(ModelManagement[DenseModelDescripti
 
     def embed_image(
         self,
-        images: Union[ImageInput, Iterable[ImageInput]],
+        images: ImageInput | Iterable[ImageInput],
         batch_size: int = 16,
-        parallel: Optional[int] = None,
+        parallel: int | None = None,
         **kwargs: Any,
     ) -> Iterable[NumpyArray]:
         """
@@ -65,3 +66,21 @@ class LateInteractionMultimodalEmbeddingBase(ModelManagement[DenseModelDescripti
             List of embeddings, one per image
         """
         raise NotImplementedError()
+
+    @classmethod
+    def get_embedding_size(cls, model_name: str) -> int:
+        """Returns embedding size of the chosen model."""
+        raise NotImplementedError("Subclasses must implement this method")
+
+    @property
+    def embedding_size(self) -> int:
+        """Returns embedding size for the current model"""
+        raise NotImplementedError("Subclasses must implement this method")
+
+    def token_count(
+        self,
+        texts: str | Iterable[str],
+        **kwargs: Any,
+    ) -> int:
+        """Returns the number of tokens in the texts."""
+        raise NotImplementedError("Subclasses must implement this method")

fastembed/late_interaction_multimodal/onnx_multimodal_model.py

@@ -2,7 +2,7 @@ import contextlib
 import os
 from multiprocessing import get_all_start_methods
 from pathlib import Path
-from typing import Any, Iterable, Optional, Sequence, Type, Union
+from typing import Any, Iterable, Sequence, Type
 
 import numpy as np
 from PIL import Image
@@ -11,19 +11,19 @@ from tokenizers import Encoding, Tokenizer
 from fastembed.common import OnnxProvider, ImageInput
 from fastembed.common.onnx_model import EmbeddingWorker, OnnxModel, OnnxOutputContext, T
 from fastembed.common.preprocessor_utils import load_tokenizer, load_preprocessor
-from fastembed.common.types import NumpyArray
+from fastembed.common.types import NumpyArray, Device
 from fastembed.common.utils import iter_batch
 from fastembed.image.transform.operators import Compose
 from fastembed.parallel_processor import ParallelWorkerPool
 
 
 class OnnxMultimodalModel(OnnxModel[T]):
-    ONNX_OUTPUT_NAMES: Optional[list[str]] = None
+    ONNX_OUTPUT_NAMES: list[str] | None = None
 
     def __init__(self) -> None:
         super().__init__()
-        self.tokenizer: Optional[Tokenizer] = None
-        self.processor: Optional[Compose] = None
+        self.tokenizer: Tokenizer | None = None
+        self.processor: Compose | None = None
         self.special_token_to_id: dict[str, int] = {}
 
     def _preprocess_onnx_text_input(
@@ -60,10 +60,11 @@ class OnnxMultimodalModel(OnnxModel[T]):
         self,
         model_dir: Path,
         model_file: str,
-        threads: Optional[int],
-        providers: Optional[Sequence[OnnxProvider]] = None,
-        cuda: bool = False,
-        device_id: Optional[int] = None,
+        threads: int | None,
+        providers: Sequence[OnnxProvider] | None = None,
+        cuda: bool | Device = Device.AUTO,
+        device_id: int | None = None,
+        extra_session_options: dict[str, Any] | None = None,
     ) -> None:
         super()._load_onnx_model(
             model_dir=model_dir,
@@ -72,9 +73,10 @@ class OnnxMultimodalModel(OnnxModel[T]):
             providers=providers,
             cuda=cuda,
             device_id=device_id,
+            extra_session_options=extra_session_options,
         )
-        assert self.tokenizer is not None
         self.tokenizer, self.special_token_to_id = load_tokenizer(model_dir=model_dir)
+        assert self.tokenizer is not None
         self.processor = load_preprocessor(model_dir=model_dir)
 
     def load_onnx_model(self) -> None:
@@ -114,12 +116,15 @@ class OnnxMultimodalModel(OnnxModel[T]):
         self,
         model_name: str,
         cache_dir: str,
-        documents: Union[str, Iterable[str]],
+        documents: str | Iterable[str],
         batch_size: int = 256,
-        parallel: Optional[int] = None,
-        providers: Optional[Sequence[OnnxProvider]] = None,
-        cuda: bool = False,
-        device_ids: Optional[list[int]] = None,
+        parallel: int | None = None,
+        providers: Sequence[OnnxProvider] | None = None,
+        cuda: bool | Device = Device.AUTO,
+        device_ids: list[int] | None = None,
+        local_files_only: bool = False,
+        specific_model_path: str | None = None,
+        extra_session_options: dict[str, Any] | None = None,
         **kwargs: Any,
     ) -> Iterable[T]:
         is_small = False
@@ -146,9 +151,14 @@ class OnnxMultimodalModel(OnnxModel[T]):
                 "model_name": model_name,
                 "cache_dir": cache_dir,
                 "providers": providers,
+                "local_files_only": local_files_only,
+                "specific_model_path": specific_model_path,
                 **kwargs,
             }
 
+            if extra_session_options is not None:
+                params.update(extra_session_options)
+
             pool = ParallelWorkerPool(
                 num_workers=parallel or 1,
                 worker=self._get_text_worker_class(),
@@ -159,19 +169,17 @@ class OnnxMultimodalModel(OnnxModel[T]):
             for batch in pool.ordered_map(iter_batch(documents, batch_size), **params):
                 yield from self._post_process_onnx_text_output(batch)  # type: ignore
 
-    def _build_onnx_image_input(self, encoded: NumpyArray) -> dict[str, NumpyArray]:
-        input_name = self.model.get_inputs()[0].name  # type: ignore[union-attr]
-        return {input_name: encoded}
-
     def onnx_embed_image(self, images: list[ImageInput], **kwargs: Any) -> OnnxOutputContext:
-        with contextlib.ExitStack():
+        with contextlib.ExitStack() as stack:
             image_files = [
-                Image.open(image) if not isinstance(image, Image.Image) else image
+                stack.enter_context(Image.open(image))
+                if not isinstance(image, Image.Image)
+                else image
                 for image in images
             ]
             assert self.processor is not None, "Processor is not initialized"
             encoded = np.array(self.processor(image_files))
-        onnx_input = self._build_onnx_image_input(encoded)
+        onnx_input = {"pixel_values": encoded}
         onnx_input = self._preprocess_onnx_image_input(onnx_input, **kwargs)
         model_output = self.model.run(None, onnx_input)  # type: ignore[union-attr]
         embeddings = model_output[0].reshape(len(images), -1)
@@ -181,12 +189,15 @@ class OnnxMultimodalModel(OnnxModel[T]):
         self,
         model_name: str,
         cache_dir: str,
-        images: Union[Iterable[ImageInput], ImageInput],
+        images: Iterable[ImageInput] | ImageInput,
         batch_size: int = 256,
-        parallel: Optional[int] = None,
-        providers: Optional[Sequence[OnnxProvider]] = None,
-        cuda: bool = False,
-        device_ids: Optional[list[int]] = None,
+        parallel: int | None = None,
+        providers: Sequence[OnnxProvider] | None = None,
+        cuda: bool | Device = Device.AUTO,
+        device_ids: list[int] | None = None,
+        local_files_only: bool = False,
+        specific_model_path: str | None = None,
+        extra_session_options: dict[str, Any] | None = None,
         **kwargs: Any,
     ) -> Iterable[T]:
         is_small = False
@@ -213,9 +224,14 @@ class OnnxMultimodalModel(OnnxModel[T]):
                 "model_name": model_name,
                 "cache_dir": cache_dir,
                 "providers": providers,
+                "local_files_only": local_files_only,
+                "specific_model_path": specific_model_path,
                 **kwargs,
             }
 
+            if extra_session_options is not None:
+                params.update(extra_session_options)
+
             pool = ParallelWorkerPool(
                 num_workers=parallel or 1,
                 worker=self._get_image_worker_class(),

fastembed/model_loader.py

@@ -1,16 +0,0 @@
-from pathlib import Path
-import json
-from typing import Dict, List
-
-
-class ModelLoader:
-    def __init__(self):
-        self.config_dir = Path(__file__).parent / "configs"
-        self._models: Dict[str, List[Dict]] = {}
-
-    def load_models(self, model_type: str) -> List[Dict]:
-        if model_type not in self._models:
-            config_path = self.config_dir / f"{model_type}_models.json"
-            with open(config_path) as f:
-                self._models[model_type] = json.load(f)["models"]
-        return self._models[model_type]

fastembed/parallel_processor.py

@@ -8,8 +8,9 @@ from multiprocessing.context import BaseContext
 from multiprocessing.process import BaseProcess
 from multiprocessing.sharedctypes import Synchronized as BaseValue
 from queue import Empty
-from typing import Any, Iterable, Optional, Type
+from typing import Any, Iterable, Type
 
+from fastembed.common.types import Device
 
 # Single item should be processed in less than:
 processing_timeout = 10 * 60  # seconds
@@ -38,7 +39,7 @@ def _worker(
     output_queue: Queue,
     num_active_workers: BaseValue,
     worker_id: int,
-    kwargs: Optional[dict[str, Any]] = None,
+    kwargs: dict[str, Any] | None = None,
 ) -> None:
     """
     A worker that pulls data pints off the input queue, and places the execution result on the output queue.
@@ -93,21 +94,21 @@ class ParallelWorkerPool:
         self,
         num_workers: int,
         worker: Type[Worker],
-        start_method: Optional[str] = None,
-        device_ids: Optional[list[int]] = None,
-        cuda: bool = False,
+        start_method: str | None = None,
+        device_ids: list[int] | None = None,
+        cuda: bool | Device = Device.AUTO,
     ):
         self.worker_class = worker
         self.num_workers = num_workers
-        self.input_queue: Optional[Queue] = None
-        self.output_queue: Optional[Queue] = None
+        self.input_queue: Queue | None = None
+        self.output_queue: Queue | None = None
         self.ctx: BaseContext = get_context(start_method)
         self.processes: list[BaseProcess] = []
         self.queue_size = self.num_workers * max_internal_batch_size
         self.emergency_shutdown = False
         self.device_ids = device_ids
         self.cuda = cuda
-        self.num_active_workers: Optional[BaseValue] = None
+        self.num_active_workers: BaseValue | None = None
 
     def start(self, **kwargs: Any) -> None:
         self.input_queue = self.ctx.Queue(self.queue_size)
@@ -220,7 +221,7 @@ class ParallelWorkerPool:
                     f"Worker PID: {process.pid} terminated unexpectedly with code {process.exitcode}"
                 )
 
-    def join_or_terminate(self, timeout: Optional[int] = 1) -> None:
+    def join_or_terminate(self, timeout: int = 1) -> None:
         """
         Emergency shutdown
         @param timeout:

fastembed/postprocess/__init__.py

@@ -0,0 +1,3 @@
+from fastembed.postprocess.muvera import Muvera
+
+__all__ = ["Muvera"]

fastembed/postprocess/muvera.py

@@ -0,0 +1,362 @@
+import numpy as np
+
+from fastembed.common.types import NumpyArray
+from fastembed.late_interaction.late_interaction_embedding_base import (
+    LateInteractionTextEmbeddingBase,
+)
+from fastembed.late_interaction_multimodal.late_interaction_multimodal_embedding_base import (
+    LateInteractionMultimodalEmbeddingBase,
+)
+
+
+MultiVectorModel = LateInteractionTextEmbeddingBase | LateInteractionMultimodalEmbeddingBase
+MAX_HAMMING_DISTANCE = 65  # 64 bits + 1
+POPCOUNT_LUT = np.array([bin(x).count("1") for x in range(256)], dtype=np.uint8)
+
+
+def hamming_distance_matrix(ids: np.ndarray) -> np.ndarray:
+    """Compute full Hamming distance matrix
+
+    Args:
+    ids: shape (n,) - array of ids, only size of the array matters
+
+    Return:
+        np.ndarray (n, n) - hamming distance matrix
+    """
+    n = len(ids)
+    xor_vals = np.bitwise_xor(ids[:, None], ids[None, :])  # (n, n) uint64
+    bytes_view = xor_vals.view(np.uint8).reshape(n, n, 8)  # (n, n, 8)
+    return POPCOUNT_LUT[bytes_view].sum(axis=2)
+
+
+class SimHashProjection:
+    """
+    SimHash projection component for MUVERA clustering.
+
+    This class implements locality-sensitive hashing using random hyperplanes
+    to partition the vector space into 2^k_sim clusters. Each vector is assigned
+    to a cluster based on which side of k_sim random hyperplanes it falls on.
+
+    Attributes:
+        k_sim (int): Number of SimHash functions (hyperplanes)
+        dim (int): Dimensionality of input vectors
+        simhash_vectors (np.ndarray): Random hyperplane normal vectors of shape (dim, k_sim)
+    """
+
+    def __init__(self, k_sim: int, dim: int, random_generator: np.random.Generator):
+        """
+        Initialize SimHash projection with random hyperplanes.
+
+        Args:
+            k_sim (int): Number of SimHash functions, determines 2^k_sim clusters
+            dim (int): Dimensionality of input vectors
+            random_generator (np.random.Generator): Random number generator for reproducibility
+        """
+        self.k_sim = k_sim
+        self.dim = dim
+        # Generate k_sim random hyperplanes (normal vectors) from standard normal distribution
+        self.simhash_vectors = random_generator.normal(size=(dim, k_sim))
+
+    def get_cluster_ids(self, vectors: np.ndarray) -> np.ndarray:
+        """
+        Compute the cluster IDs for a given vector using SimHash.
+
+        The cluster ID is determined by computing the dot product of the vector
+        with each hyperplane normal vector, taking the sign, and interpreting
+        the resulting binary string as an integer.
+
+        Args:
+            vectors (np.ndarray): Input vectors of shape (n, dim,)
+
+        Returns:
+            np.ndarray: Cluster IDs in range [0, 2^k_sim - 1]
+
+        Raises:
+            AssertionError: If a vector shape doesn't match expected dimensionality
+        """
+        dot_product = (
+            vectors @ self.simhash_vectors
+        )  # (token_num, dim) x (dim, k_sim) -> (token_num, k_sim)
+        cluster_ids = (dot_product > 0) @ (1 << np.arange(self.k_sim))
+        return cluster_ids
+
+
+class Muvera:
+    """
+    MUVERA (Multi-Vector Retrieval Architecture) algorithm implementation.
+
+    This class creates Fixed Dimensional Encodings (FDEs) from variable-length
+    sequences of vectors by using SimHash clustering and random projections.
+    The process involves:
+    1. Clustering vectors using multiple SimHash projections
+    2. Computing cluster centers (with different strategies for docs vs queries)
+    3. Applying random projections for dimensionality reduction
+    4. Concatenating results from all projections
+
+    Attributes:
+        k_sim (int): Number of SimHash functions per projection
+        dim (int): Input vector dimensionality
+        dim_proj (int): Output dimensionality after random projection
+        r_reps (int): Number of random projection repetitions
+        random_seed (int): Random seed for consistent random matrix generation
+        simhash_projections (List[SimHashProjection]): SimHash instances for clustering
+        dim_reduction_projections (np.ndarray): Random projection matrices of shape (R_reps, d, d_proj)
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        k_sim: int = 5,
+        dim_proj: int = 16,
+        r_reps: int = 20,
+        random_seed: int = 42,
+    ):
+        """
+        Initialize MUVERA algorithm with specified parameters.
+
+        Args:
+            dim (int): Dimensionality of individual input vectors
+            k_sim (int, optional): Number of SimHash functions (creates 2^k_sim clusters).
+                                   Defaults to 5.
+            dim_proj (int, optional): Dimensionality after random projection (must be <= dim).
+                                    Defaults to 16.
+            r_reps (int, optional): Number of random projection repetitions for robustness.
+                                    Defaults to 20.
+            random_seed (int, optional): Seed for random number generator to ensure
+                                         reproducible results. Defaults to 42.
+
+        Raises:
+            ValueError: If dim_proj > dim (cannot project to higher dimensionality)
+        """
+        if dim_proj > dim:
+            raise ValueError(
+                f"Cannot project to a higher dimensionality (dim_proj={dim_proj} > dim={dim})"
+            )
+
+        self.k_sim = k_sim
+        self.dim = dim
+        self.dim_proj = dim_proj
+        self.r_reps = r_reps
+        # Create r_reps independent SimHash projections for robustness
+        generator = np.random.default_rng(random_seed)
+        self.simhash_projections = [
+            SimHashProjection(k_sim=self.k_sim, dim=self.dim, random_generator=generator)
+            for _ in range(r_reps)
+        ]
+        # Random projection matrices with entries from {-1, +1} for each repetition
+        self.dim_reduction_projections = generator.choice([-1, 1], size=(r_reps, dim, dim_proj))
+
+    @classmethod
+    def from_multivector_model(
+        cls,
+        model: MultiVectorModel,
+        k_sim: int = 5,
+        dim_proj: int = 16,
+        r_reps: int = 20,  # noqa[naming]
+        random_seed: int = 42,
+    ) -> "Muvera":
+        """
+        Create a Muvera instance from a multi-vector embedding model.
+
+        This class method provides a convenient way to initialize a MUVERA
+        that is compatible with a given multi-vector model by automatically extracting
+        the embedding dimensionality from the model.
+
+        Args:
+            model (MultiVectorModel): A late interaction text or multimodal embedding model
+                                    that provides multi-vector embeddings. Must have an
+                                    `embedding_size` attribute specifying the dimensionality
+                                    of individual vectors.
+            k_sim (int, optional): Number of SimHash functions (creates 2^k_sim clusters).
+                                   Defaults to 5.
+            dim_proj (int, optional): Dimensionality after random projection (must be <= model's
+                                    embedding_size). Defaults to 16.
+            r_reps (int, optional): Number of random projection repetitions for robustness.
+                                    Defaults to 20.
+            random_seed (int, optional): Seed for random number generator to ensure
+                                         reproducible results. Defaults to 42.
+
+        Returns:
+            Muvera: A configured MUVERA instance ready to process embeddings from the given model.
+
+        Raises:
+            ValueError: If dim_proj > model.embedding_size (cannot project to higher dimensionality)
+
+        Example:
+            >>> from fastembed import LateInteractionTextEmbedding
+            >>> model = LateInteractionTextEmbedding(model_name="colbert-ir/colbertv2.0")
+            >>> muvera = Muvera.from_multivector_model(
+            ...     model=model,
+            ...     k_sim=6,
+            ...     dim_proj=32
+            ... )
+            >>> # Now use postprocessor with embeddings from the model
+            >>> embeddings = np.array(list(model.embed(["sample text"])))
+            >>> fde = muvera.process_document(embeddings[0])
+        """
+        return cls(
+            dim=model.embedding_size,
+            k_sim=k_sim,
+            dim_proj=dim_proj,
+            r_reps=r_reps,
+            random_seed=random_seed,
+        )
+
+    def _get_output_dimension(self) -> int:
+        """
+        Get the output dimension of the MUVERA algorithm.
+
+        Returns:
+            int: Output dimension (r_reps * num_partitions * dim_proj) where b = 2^k_sim
+        """
+        num_partitions = 2**self.k_sim
+        return self.r_reps * num_partitions * self.dim_proj
+
+    @property
+    def embedding_size(self) -> int:
+        return self._get_output_dimension()
+
+    def process_document(self, vectors: NumpyArray) -> NumpyArray:
+        """
+        Encode a document's vectors into a Fixed Dimensional Encoding (FDE).
+
+        Uses document-specific settings: normalizes cluster centers by vector count
+        and fills empty clusters using Hamming distance-based selection.
+
+        Args:
+            vectors (NumpyArray): Document vectors of shape (n_tokens, dim)
+
+        Returns:
+            NumpyArray: Fixed dimensional encodings of shape (r_reps * b * dim_proj,)
+        """
+        return self.process(vectors, fill_empty_clusters=True, normalize_by_count=True)
+
+    def process_query(self, vectors: NumpyArray) -> NumpyArray:
+        """
+        Encode a query's vectors into a Fixed Dimensional Encoding (FDE).
+
+        Uses query-specific settings: no normalization by count and no empty
+        cluster filling to preserve query vector magnitudes.
+
+        Args:
+            vectors (NumpyArray]): Query vectors of shape (n_tokens, dim)
+
+        Returns:
+            NumpyArray: Fixed dimensional encoding of shape (r_reps * b * dim_proj,)
+        """
+        return self.process(vectors, fill_empty_clusters=False, normalize_by_count=False)
+
+    def process(
+        self,
+        vectors: NumpyArray,
+        fill_empty_clusters: bool = True,
+        normalize_by_count: bool = True,
+    ) -> NumpyArray:
+        """
+        Core encoding method that transforms variable-length vector sequences into FDEs.
+
+        The encoding process:
+        1. For each of r_reps random projections:
+           a. Assign vectors to clusters using SimHash
+           b. Compute cluster centers (sum of vectors in each cluster)
+           c. Optionally normalize by cluster size
+           d. Fill empty clusters using Hamming distance if requested
+           e. Apply random projection for dimensionality reduction
+           f. Flatten cluster centers into a vector
+        2. Concatenate all projection results
+
+        Args:
+            vectors (np.ndarray): Input vectors of shape (n_vectors, dim)
+            fill_empty_clusters (bool): Whether to fill empty clusters using nearest
+                                      vectors based on Hamming distance of cluster IDs
+            normalize_by_count (bool): Whether to normalize cluster centers by the
+                                     number of vectors assigned to each cluster
+
+        Returns:
+            np.ndarray: Fixed dimensional encoding of shape (r_reps * b * dim_proj)
+                        where B = 2^k_sim is the number of clusters
+
+        Raises:
+            AssertionError: If input vectors don't have expected dimensionality
+        """
+        assert (
+            vectors.shape[1] == self.dim
+        ), f"Expected vectors of shape (n, {self.dim}), got {vectors.shape}"
+
+        # Store results from each random projection
+        output_vectors = []
+
+        # num of space partitions in SimHash
+        num_partitions = 2**self.k_sim
+        cluster_center_ids = np.arange(num_partitions)
+        precomputed_hamming_matrix = (
+            hamming_distance_matrix(cluster_center_ids) if fill_empty_clusters else None
+        )
+
+        for projection_index, simhash in enumerate(self.simhash_projections):
+            # Initialize cluster centers and count vectors assigned to each cluster
+            cluster_centers = np.zeros((num_partitions, self.dim))
+            cluster_center_id_to_vectors: dict[int, list[int]] = {
+                cluster_center_id: [] for cluster_center_id in cluster_center_ids
+            }
+            cluster_vector_counts = None
+            empty_mask = None
+
+            # Assign each vector to its cluster and accumulate cluster centers
+            vector_cluster_ids = simhash.get_cluster_ids(vectors)
+            for cluster_id, (vec_idx, vec) in zip(vector_cluster_ids, enumerate(vectors)):
+                cluster_centers[cluster_id] += vec
+                cluster_center_id_to_vectors[cluster_id].append(vec_idx)
+
+            if normalize_by_count or fill_empty_clusters:
+                cluster_vector_counts = np.bincount(vector_cluster_ids, minlength=num_partitions)
+                empty_mask = cluster_vector_counts == 0
+
+            if normalize_by_count:
+                assert empty_mask is not None
+                assert cluster_vector_counts is not None
+                non_empty_mask = ~empty_mask
+                cluster_centers[non_empty_mask] /= cluster_vector_counts[non_empty_mask][:, None]
+
+            # Fill empty clusters using vectors with minimum Hamming distance
+            if fill_empty_clusters:
+                assert empty_mask is not None
+                assert precomputed_hamming_matrix is not None
+                masked_hamming = np.where(
+                    empty_mask[None, :], MAX_HAMMING_DISTANCE, precomputed_hamming_matrix
+                )
+                nearest_non_empty = np.argmin(masked_hamming, axis=1)
+                fill_vectors = np.array(
+                    [
+                        vectors[cluster_center_id_to_vectors[cluster_id][0]]
+                        for cluster_id in nearest_non_empty[empty_mask]
+                    ]
+                ).reshape(-1, self.dim)
+                cluster_centers[empty_mask] = fill_vectors
+
+            # Apply random projection for dimensionality reduction if needed
+            if self.dim_proj < self.dim:
+                dim_reduction_projection = self.dim_reduction_projections[
+                    projection_index
+                ]  # Get projection matrix for this repetition
+                projected_centers = (1 / np.sqrt(self.dim_proj)) * (
+                    cluster_centers @ dim_reduction_projection
+                )
+
+                # Flatten cluster centers into a single vector and add to output
+                output_vectors.append(projected_centers.flatten())
+                continue
+
+            # If no projection needed (dim_proj == dim), use original cluster centers
+            output_vectors.append(cluster_centers.flatten())
+
+        # Concatenate results from all R_reps projections into final FDE
+        return np.concatenate(output_vectors)
+
+
+if __name__ == "__main__":
+    v_arrs = np.random.randn(10, 100, 128)
+    muvera = Muvera(128, 4, 8, 20, 42)
+
+    for v_arr in v_arrs:
+        muvera.process(v_arr)  # type: ignore

fastembed/rerank/cross_encoder/custom_text_cross_encoder.py

@@ -0,0 +1,78 @@
+from typing import Sequence, Any, Type
+
+from fastembed.common import OnnxProvider
+from fastembed.common.model_description import BaseModelDescription
+from fastembed.common.types import Device
+from fastembed.rerank.cross_encoder.onnx_text_cross_encoder import OnnxTextCrossEncoder
+from fastembed.rerank.cross_encoder.onnx_text_model import TextRerankerWorker
+
+
+class CustomTextCrossEncoder(OnnxTextCrossEncoder):
+    SUPPORTED_MODELS: list[BaseModelDescription] = []
+
+    def __init__(
+        self,
+        model_name: str,
+        cache_dir: str | None = None,
+        threads: int | None = None,
+        providers: Sequence[OnnxProvider] | None = None,
+        cuda: bool | Device = Device.AUTO,
+        device_ids: list[int] | None = None,
+        lazy_load: bool = False,
+        device_id: int | None = None,
+        specific_model_path: str | None = 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 _get_worker_class(cls) -> Type[TextRerankerWorker]:
+        return CustomTextCrossEncoderWorker
+
+    def _get_worker_init_kwargs(self) -> dict[str, Any]:
+        return {"model_description": self.model_description}
+
+    @classmethod
+    def add_model(
+        cls,
+        model_description: BaseModelDescription,
+    ) -> None:
+        cls.SUPPORTED_MODELS.append(model_description)
+
+
+class CustomTextCrossEncoderWorker(TextRerankerWorker):
+    def init_embedding(
+        self,
+        model_name: str,
+        cache_dir: str,
+        model_description: BaseModelDescription | None = None,
+        **kwargs: Any,
+    ) -> CustomTextCrossEncoder:
+        if model_description is None:
+            raise ValueError(
+                "`model_description` is required to initialize a custom model in a worker "
+                "process, it is provided by `CustomTextCrossEncoder._get_worker_init_kwargs`"
+            )
+        # custom models live in a class-level registry, which spawned workers don't inherit
+        CustomTextCrossEncoder.add_model(model_description)
+        return CustomTextCrossEncoder(
+            model_name=model_name,
+            cache_dir=cache_dir,
+            threads=1,
+            **kwargs,
+        )

fastembed/rerank/cross_encoder/onnx_text_cross_encoder.py

@@ -1,9 +1,10 @@
-from typing import Any, Iterable, Optional, Sequence, Type
+from typing import Any, Iterable, Sequence, Type
 
 from loguru import logger
 
 from fastembed.common import OnnxProvider
 from fastembed.common.onnx_model import OnnxOutputContext
+from fastembed.common.types import Device
 from fastembed.common.utils import define_cache_dir
 from fastembed.rerank.cross_encoder.onnx_text_model import (
     OnnxCrossEncoderModel,
@@ -77,14 +78,14 @@ class OnnxTextCrossEncoder(TextCrossEncoderBase, OnnxCrossEncoderModel):
     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,
+        cache_dir: str | None = None,
+        threads: int | None = None,
+        providers: Sequence[OnnxProvider] | None = None,
+        cuda: bool | Device = Device.AUTO,
+        device_ids: list[int] | None = None,
         lazy_load: bool = False,
-        device_id: Optional[int] = None,
-        specific_model_path: Optional[str] = None,
+        device_id: int | None = None,
+        specific_model_path: str | None = None,
         **kwargs: Any,
     ):
         """
@@ -96,10 +97,11 @@ class OnnxTextCrossEncoder(TextCrossEncoderBase, OnnxCrossEncoderModel):
             threads (int, optional): The number of threads single onnxruntime session can use. Defaults to None.
             providers (Optional[Sequence[OnnxProvider]], optional): The list of onnxruntime providers to use.
                 Mutually exclusive with the `cuda` and `device_ids` arguments. Defaults to None.
-            cuda (bool, optional): Whether to use cuda for inference. Mutually exclusive with `providers`
-                Defaults to False.
+            cuda (Union[bool, Device], optional): Whether to use cuda for inference. Mutually exclusive with `providers`
+                Defaults to Device.AUTO.
             device_ids (Optional[list[int]], optional): The list of device ids to use for data parallel processing in
-                workers. Should be used with `cuda=True`, mutually exclusive with `providers`. Defaults to None.
+                workers. Should be used with `cuda` equals to `True`, `Device.AUTO` or `Device.CUDA`, mutually exclusive
+                with `providers`. Defaults to None.
             lazy_load (bool, optional): Whether to load the model during class initialization or on demand.
                 Should be set to True when using multiple-gpu and parallel encoding. Defaults to False.
             device_id (Optional[int], optional): The device id to use for loading the model in the worker process.
@@ -111,6 +113,7 @@ class OnnxTextCrossEncoder(TextCrossEncoderBase, OnnxCrossEncoderModel):
         super().__init__(model_name, cache_dir, threads, **kwargs)
         self.providers = providers
         self.lazy_load = lazy_load
+        self._extra_session_options = self._select_exposed_session_options(kwargs)
 
         # List of device ids, that can be used for data parallel processing in workers
         self.device_ids = device_ids
@@ -123,7 +126,7 @@ class OnnxTextCrossEncoder(TextCrossEncoderBase, OnnxCrossEncoderModel):
             )
 
         # This device_id will be used if we need to load model in current process
-        self.device_id: Optional[int] = None
+        self.device_id: int | None = None
         if device_id is not None:
             self.device_id = device_id
         elif self.device_ids is not None:
@@ -131,11 +134,12 @@ class OnnxTextCrossEncoder(TextCrossEncoderBase, OnnxCrossEncoderModel):
 
         self.model_description = self._get_model_description(model_name)
         self.cache_dir = str(define_cache_dir(cache_dir))
+        self._specific_model_path = specific_model_path
         self._model_dir = self.download_model(
             self.model_description,
             self.cache_dir,
             local_files_only=self._local_files_only,
-            specific_model_path=specific_model_path,
+            specific_model_path=self._specific_model_path,
         )
 
         if not self.lazy_load:
@@ -149,6 +153,7 @@ class OnnxTextCrossEncoder(TextCrossEncoderBase, OnnxCrossEncoderModel):
             providers=self.providers,
             cuda=self.cuda,
             device_id=self.device_id,
+            extra_session_options=self._extra_session_options,
         )
 
     def rerank(
@@ -177,7 +182,7 @@ class OnnxTextCrossEncoder(TextCrossEncoderBase, OnnxCrossEncoderModel):
         self,
         pairs: Iterable[tuple[str, str]],
         batch_size: int = 64,
-        parallel: Optional[int] = None,
+        parallel: int | None = None,
         **kwargs: Any,
     ) -> Iterable[float]:
         yield from self._rerank_pairs(
@@ -189,6 +194,9 @@ class OnnxTextCrossEncoder(TextCrossEncoderBase, OnnxCrossEncoderModel):
             providers=self.providers,
             cuda=self.cuda,
             device_ids=self.device_ids,
+            local_files_only=self._local_files_only,
+            specific_model_path=self._specific_model_path,
+            extra_session_options=self._extra_session_options,
             **kwargs,
         )
 
@@ -196,9 +204,25 @@ class OnnxTextCrossEncoder(TextCrossEncoderBase, OnnxCrossEncoderModel):
     def _get_worker_class(cls) -> Type[TextRerankerWorker]:
         return TextCrossEncoderWorker
 
-    def _post_process_onnx_output(self, output: OnnxOutputContext) -> Iterable[float]:
+    def _post_process_onnx_output(
+        self, output: OnnxOutputContext, **kwargs: Any
+    ) -> Iterable[float]:
         return (float(elem) for elem in output.model_output)
 
+    def token_count(
+        self, pairs: Iterable[tuple[str, str]], batch_size: int = 1024, **kwargs: Any
+    ) -> int:
+        """Returns the number of tokens in the pairs.
+
+        Args:
+            pairs: Iterable of tuples, where each tuple contains a query and a document to be tokenized
+            batch_size: Batch size for tokenizing
+
+        Returns:
+            token count: overall number of tokens in the pairs
+        """
+        return self._token_count(pairs, batch_size=batch_size, **kwargs)
+
 
 class TextCrossEncoderWorker(TextRerankerWorker):
     def init_embedding(

fastembed/rerank/cross_encoder/onnx_text_model.py

@@ -1,7 +1,7 @@
 import os
 from multiprocessing import get_all_start_methods
 from pathlib import Path
-from typing import Any, Iterable, Optional, Sequence, Type
+from typing import Any, Iterable, Sequence, Type
 
 import numpy as np
 from tokenizers import Encoding
@@ -12,14 +12,14 @@ from fastembed.common.onnx_model import (
     OnnxOutputContext,
     OnnxProvider,
 )
-from fastembed.common.types import NumpyArray
+from fastembed.common.types import NumpyArray, Device
 from fastembed.common.preprocessor_utils import load_tokenizer
 from fastembed.common.utils import iter_batch
 from fastembed.parallel_processor import ParallelWorkerPool
 
 
 class OnnxCrossEncoderModel(OnnxModel[float]):
-    ONNX_OUTPUT_NAMES: Optional[list[str]] = None
+    ONNX_OUTPUT_NAMES: list[str] | None = None
 
     @classmethod
     def _get_worker_class(cls) -> Type["TextRerankerWorker"]:
@@ -29,10 +29,11 @@ class OnnxCrossEncoderModel(OnnxModel[float]):
         self,
         model_dir: Path,
         model_file: str,
-        threads: Optional[int],
-        providers: Optional[Sequence[OnnxProvider]] = None,
-        cuda: bool = False,
-        device_id: Optional[int] = None,
+        threads: int | None,
+        providers: Sequence[OnnxProvider] | None = None,
+        cuda: bool | Device = Device.AUTO,
+        device_id: int | None = None,
+        extra_session_options: dict[str, Any] | None = None,
     ) -> None:
         super()._load_onnx_model(
             model_dir=model_dir,
@@ -41,6 +42,7 @@ class OnnxCrossEncoderModel(OnnxModel[float]):
             providers=providers,
             cuda=cuda,
             device_id=device_id,
+            extra_session_options=extra_session_options,
         )
         self.tokenizer, _ = load_tokenizer(model_dir=model_dir)
         assert self.tokenizer is not None
@@ -90,10 +92,13 @@ class OnnxCrossEncoderModel(OnnxModel[float]):
         cache_dir: str,
         pairs: Iterable[tuple[str, str]],
         batch_size: int,
-        parallel: Optional[int] = None,
-        providers: Optional[Sequence[OnnxProvider]] = None,
-        cuda: bool = False,
-        device_ids: Optional[list[int]] = None,
+        parallel: int | None = None,
+        providers: Sequence[OnnxProvider] | None = None,
+        cuda: bool | Device = Device.AUTO,
+        device_ids: list[int] | None = None,
+        local_files_only: bool = False,
+        specific_model_path: str | None = None,
+        extra_session_options: dict[str, Any] | None = None,
         **kwargs: Any,
     ) -> Iterable[float]:
         is_small = False
@@ -120,9 +125,15 @@ class OnnxCrossEncoderModel(OnnxModel[float]):
                 "model_name": model_name,
                 "cache_dir": cache_dir,
                 "providers": providers,
+                "local_files_only": local_files_only,
+                "specific_model_path": specific_model_path,
                 **kwargs,
+                **self._get_worker_init_kwargs(),
             }
 
+            if extra_session_options is not None:
+                params.update(extra_session_options)
+
             pool = ParallelWorkerPool(
                 num_workers=parallel or 1,
                 worker=self._get_worker_class(),
@@ -133,7 +144,18 @@ class OnnxCrossEncoderModel(OnnxModel[float]):
             for batch in pool.ordered_map(iter_batch(pairs, batch_size), **params):
                 yield from self._post_process_onnx_output(batch)  # type: ignore
 
-    def _post_process_onnx_output(self, output: OnnxOutputContext) -> Iterable[float]:
+    def _post_process_onnx_output(
+        self, output: OnnxOutputContext, **kwargs: Any
+    ) -> Iterable[float]:
+        """Post-process the ONNX model output to convert it into a usable format.
+
+        Args:
+            output (OnnxOutputContext): The raw output from the ONNX model.
+            **kwargs: Additional keyword arguments that may be needed by specific implementations.
+
+        Returns:
+            Iterable[float]: Post-processed output as an iterable of float values.
+        """
         raise NotImplementedError("Subclasses must implement this method")
 
     def _preprocess_onnx_input(
@@ -144,6 +166,20 @@ class OnnxCrossEncoderModel(OnnxModel[float]):
         """
         return onnx_input
 
+    def _token_count(
+        self, pairs: Iterable[tuple[str, str]], batch_size: int = 1024, **_: Any
+    ) -> int:
+        if not hasattr(self, "model") or self.model is None:
+            self.load_onnx_model()  # loads the tokenizer as well
+
+        token_num = 0
+        assert self.tokenizer is not None
+        for batch in iter_batch(pairs, batch_size):
+            for tokens in self.tokenizer.encode_batch(batch):
+                token_num += sum(tokens.attention_mask)
+
+        return token_num
+
 
 class TextRerankerWorker(EmbeddingWorker[float]):
     def __init__(

fastembed/rerank/cross_encoder/text_cross_encoder.py

@@ -1,15 +1,22 @@
-from typing import Any, Iterable, Optional, Sequence, Type
+from typing import Any, Iterable, Sequence, Type
 from dataclasses import asdict
 
 from fastembed.common import OnnxProvider
+from fastembed.common.types import Device
 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
@@ -47,11 +54,11 @@ class TextCrossEncoder(TextCrossEncoderBase):
     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,
+        cache_dir: str | None = None,
+        threads: int | None = None,
+        providers: Sequence[OnnxProvider] | None = None,
+        cuda: bool | Device = Device.AUTO,
+        device_ids: list[int] | None = None,
         lazy_load: bool = False,
         **kwargs: Any,
     ):
@@ -96,7 +103,7 @@ class TextCrossEncoder(TextCrossEncoderBase):
         self,
         pairs: Iterable[tuple[str, str]],
         batch_size: int = 64,
-        parallel: Optional[int] = None,
+        parallel: int | None = None,
         **kwargs: Any,
     ) -> Iterable[float]:
         """
@@ -124,3 +131,48 @@ 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: list[str] | None = 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 [],
+            )
+        )
+
+    def token_count(
+        self, pairs: Iterable[tuple[str, str]], batch_size: int = 1024, **kwargs: Any
+    ) -> int:
+        """Returns the number of tokens in the pairs.
+
+        Args:
+            pairs: Iterable of tuples, where each tuple contains a query and a document to be tokenized
+            batch_size: Batch size for tokenizing
+
+        Returns:
+            token count: overall number of tokens in the pairs
+        """
+        return self.model.token_count(pairs, batch_size=batch_size, **kwargs)

fastembed/rerank/cross_encoder/text_cross_encoder_base.py

@@ -1,4 +1,4 @@
-from typing import Any, Iterable, Optional
+from typing import Any, Iterable
 
 from fastembed.common.model_description import BaseModelDescription
 from fastembed.common.model_management import ModelManagement
@@ -8,8 +8,8 @@ class TextCrossEncoderBase(ModelManagement[BaseModelDescription]):
     def __init__(
         self,
         model_name: str,
-        cache_dir: Optional[str] = None,
-        threads: Optional[int] = None,
+        cache_dir: str | None = None,
+        threads: int | None = None,
         **kwargs: Any,
     ):
         self.model_name = model_name
@@ -41,7 +41,7 @@ class TextCrossEncoderBase(ModelManagement[BaseModelDescription]):
         self,
         pairs: Iterable[tuple[str, str]],
         batch_size: int = 64,
-        parallel: Optional[int] = None,
+        parallel: int | None = None,
         **kwargs: Any,
     ) -> Iterable[float]:
         """Rerank query-document pairs.
@@ -57,3 +57,7 @@ class TextCrossEncoderBase(ModelManagement[BaseModelDescription]):
             Iterable[float]: Scores for each individual pair
         """
         raise NotImplementedError("This method should be overridden by subclasses")
+
+    def token_count(self, pairs: Iterable[tuple[str, str]], **kwargs: Any) -> int:
+        """Returns the number of tokens in the pairs."""
+        raise NotImplementedError("This method should be overridden by subclasses")

fastembed/sparse/bm25.py

@@ -2,7 +2,7 @@ import os
 from collections import defaultdict
 from multiprocessing import get_all_start_methods
 from pathlib import Path
-from typing import Any, Iterable, Optional, Type, Union
+from typing import Any, Iterable, Type
 
 import mmh3
 import numpy as np
@@ -21,13 +21,9 @@ from fastembed.sparse.sparse_embedding_base import (
 from fastembed.sparse.utils.tokenizer import SimpleTokenizer
 from fastembed.common.model_description import SparseModelDescription, ModelSource
 
+
 supported_languages = [
     "arabic",
-    "azerbaijani",
-    "basque",
-    "bengali",
-    "catalan",
-    "chinese",
     "danish",
     "dutch",
     "english",
@@ -35,21 +31,15 @@ supported_languages = [
     "french",
     "german",
     "greek",
-    "hebrew",
-    "hinglish",
     "hungarian",
-    "indonesian",
     "italian",
-    "kazakh",
-    "nepali",
     "norwegian",
     "portuguese",
     "romanian",
     "russian",
-    "slovene",
     "spanish",
     "swedish",
-    "tajik",
+    "tamil",
     "turkish",
 ]
 
@@ -101,14 +91,14 @@ class Bm25(SparseTextEmbeddingBase):
     def __init__(
         self,
         model_name: str,
-        cache_dir: Optional[str] = None,
+        cache_dir: str | None = None,
         k: float = 1.2,
         b: float = 0.75,
         avg_len: float = 256.0,
         language: str = "english",
         token_max_length: int = 40,
         disable_stemmer: bool = False,
-        specific_model_path: Optional[str] = None,
+        specific_model_path: str | None = None,
         **kwargs: Any,
     ):
         super().__init__(model_name, cache_dir, **kwargs)
@@ -125,11 +115,12 @@ class Bm25(SparseTextEmbeddingBase):
         model_description = self._get_model_description(model_name)
         self.cache_dir = str(define_cache_dir(cache_dir))
 
+        self._specific_model_path = specific_model_path
         self._model_dir = self.download_model(
             model_description,
             self.cache_dir,
             local_files_only=self._local_files_only,
-            specific_model_path=specific_model_path,
+            specific_model_path=self._specific_model_path,
         )
 
         self.token_max_length = token_max_length
@@ -167,9 +158,11 @@ class Bm25(SparseTextEmbeddingBase):
         self,
         model_name: str,
         cache_dir: str,
-        documents: Union[str, Iterable[str]],
+        documents: str | Iterable[str],
         batch_size: int = 256,
-        parallel: Optional[int] = None,
+        parallel: int | None = None,
+        local_files_only: bool = False,
+        specific_model_path: str | None = None,
     ) -> Iterable[SparseEmbedding]:
         is_small = False
 
@@ -198,6 +191,8 @@ class Bm25(SparseTextEmbeddingBase):
                 "language": self.language,
                 "token_max_length": self.token_max_length,
                 "disable_stemmer": self.disable_stemmer,
+                "local_files_only": local_files_only,
+                "specific_model_path": specific_model_path,
             }
             pool = ParallelWorkerPool(
                 num_workers=parallel or 1,
@@ -210,9 +205,9 @@ class Bm25(SparseTextEmbeddingBase):
 
     def embed(
         self,
-        documents: Union[str, Iterable[str]],
+        documents: str | Iterable[str],
         batch_size: int = 256,
-        parallel: Optional[int] = None,
+        parallel: int | None = None,
         **kwargs: Any,
     ) -> Iterable[SparseEmbedding]:
         """
@@ -236,6 +231,8 @@ class Bm25(SparseTextEmbeddingBase):
             documents=documents,
             batch_size=batch_size,
             parallel=parallel,
+            local_files_only=self._local_files_only,
+            specific_model_path=self._specific_model_path,
         )
 
     def _stem(self, tokens: list[str]) -> list[str]:
@@ -271,6 +268,15 @@ class Bm25(SparseTextEmbeddingBase):
             embeddings.append(SparseEmbedding.from_dict(token_id2value))
         return embeddings
 
+    def token_count(self, texts: str | Iterable[str], **kwargs: Any) -> int:
+        token_num = 0
+        texts = [texts] if isinstance(texts, str) else texts
+        for text in texts:
+            document = remove_non_alphanumeric(text)
+            tokens = self.tokenizer.tokenize(document)
+            token_num += len(tokens)
+        return token_num
+
     def _term_frequency(self, tokens: list[str]) -> dict[int, float]:
         """Calculate the term frequency part of the BM25 formula.
 
@@ -305,9 +311,7 @@ class Bm25(SparseTextEmbeddingBase):
     def compute_token_id(cls, token: str) -> int:
         return abs(mmh3.hash(token))
 
-    def query_embed(
-        self, query: Union[str, Iterable[str]], **kwargs: Any
-    ) -> Iterable[SparseEmbedding]:
+    def query_embed(self, query: str | Iterable[str], **kwargs: Any) -> Iterable[SparseEmbedding]:
         """To emulate BM25 behaviour, we don't need to use weights in the query, and
         it's enough to just hash the tokens and assign a weight of 1.0 to them.
         """

fastembed/sparse/bm42.py

@@ -1,7 +1,7 @@
 import math
 import string
 from pathlib import Path
-from typing import Any, Iterable, Optional, Sequence, Type, Union
+from typing import Any, Iterable, Sequence, Type
 
 import mmh3
 import numpy as np
@@ -9,6 +9,7 @@ from py_rust_stemmers import SnowballStemmer
 
 from fastembed.common import OnnxProvider
 from fastembed.common.onnx_model import OnnxOutputContext
+from fastembed.common.types import Device
 from fastembed.common.utils import define_cache_dir
 from fastembed.sparse.sparse_embedding_base import (
     SparseEmbedding,
@@ -31,9 +32,17 @@ supported_bm42_models: list[SparseModelDescription] = [
     ),
 ]
 
-MODEL_TO_LANGUAGE = {
+
+_MODEL_TO_LANGUAGE = {
     "Qdrant/bm42-all-minilm-l6-v2-attentions": "english",
 }
+MODEL_TO_LANGUAGE = {
+    model_name.lower(): language for model_name, language in _MODEL_TO_LANGUAGE.items()
+}
+
+
+def get_language_by_model_name(model_name: str) -> str:
+    return MODEL_TO_LANGUAGE[model_name.lower()]
 
 
 class Bm42(SparseTextEmbeddingBase, OnnxTextModel[SparseEmbedding]):
@@ -57,15 +66,15 @@ class Bm42(SparseTextEmbeddingBase, OnnxTextModel[SparseEmbedding]):
     def __init__(
         self,
         model_name: str,
-        cache_dir: Optional[str] = None,
-        threads: Optional[int] = None,
-        providers: Optional[Sequence[OnnxProvider]] = None,
+        cache_dir: str | None = None,
+        threads: int | None = None,
+        providers: Sequence[OnnxProvider] | None = None,
         alpha: float = 0.5,
-        cuda: bool = False,
-        device_ids: Optional[list[int]] = None,
+        cuda: bool | Device = Device.AUTO,
+        device_ids: list[int] | None = None,
         lazy_load: bool = False,
-        device_id: Optional[int] = None,
-        specific_model_path: Optional[str] = None,
+        device_id: int | None = None,
+        specific_model_path: str | None = None,
         **kwargs: Any,
     ):
         """
@@ -79,10 +88,11 @@ class Bm42(SparseTextEmbeddingBase, OnnxTextModel[SparseEmbedding]):
             alpha (float, optional): Parameter, that defines the importance of the token weight in the document
                 versus the importance of the token frequency in the corpus. Defaults to 0.5, based on empirical testing.
                 It is recommended to only change this parameter based on training data for a specific dataset.
-            cuda (bool, optional): Whether to use cuda for inference. Mutually exclusive with `providers`
-                Defaults to False.
+            cuda (Union[bool, Device], optional): Whether to use cuda for inference. Mutually exclusive with `providers`
+                Defaults to Device.AUTO.
             device_ids (Optional[list[int]], optional): The list of device ids to use for data parallel processing in
-                workers. Should be used with `cuda=True`, mutually exclusive with `providers`. Defaults to None.
+                workers. Should be used with `cuda` equals to `True`, `Device.AUTO` or `Device.CUDA`, mutually exclusive
+                with `providers`. Defaults to None.
             lazy_load (bool, optional): Whether to load the model during class initialization or on demand.
                 Should be set to True when using multiple-gpu and parallel encoding. Defaults to False.
             device_id (Optional[int], optional): The device id to use for loading the model in the worker process.
@@ -95,13 +105,14 @@ class Bm42(SparseTextEmbeddingBase, OnnxTextModel[SparseEmbedding]):
         super().__init__(model_name, cache_dir, threads, **kwargs)
         self.providers = providers
         self.lazy_load = lazy_load
+        self._extra_session_options = self._select_exposed_session_options(kwargs)
 
         # List of device ids, that can be used for data parallel processing in workers
         self.device_ids = device_ids
         self.cuda = cuda
 
         # This device_id will be used if we need to load model in current process
-        self.device_id: Optional[int] = None
+        self.device_id: int | None = None
         if device_id is not None:
             self.device_id = device_id
         elif self.device_ids is not None:
@@ -110,11 +121,12 @@ class Bm42(SparseTextEmbeddingBase, OnnxTextModel[SparseEmbedding]):
         self.model_description = self._get_model_description(model_name)
         self.cache_dir = str(define_cache_dir(cache_dir))
 
+        self._specific_model_path = specific_model_path
         self._model_dir = self.download_model(
             self.model_description,
             self.cache_dir,
             local_files_only=self._local_files_only,
-            specific_model_path=specific_model_path,
+            specific_model_path=self._specific_model_path,
         )
 
         self.invert_vocab: dict[int, str] = {}
@@ -123,7 +135,7 @@ class Bm42(SparseTextEmbeddingBase, OnnxTextModel[SparseEmbedding]):
         self.special_tokens_ids: set[int] = set()
         self.punctuation = set(string.punctuation)
         self.stopwords = set(self._load_stopwords(self._model_dir))
-        self.stemmer = SnowballStemmer(MODEL_TO_LANGUAGE[model_name])
+        self.stemmer = SnowballStemmer(get_language_by_model_name(self.model_name))
         self.alpha = alpha
 
         if not self.lazy_load:
@@ -137,6 +149,7 @@ class Bm42(SparseTextEmbeddingBase, OnnxTextModel[SparseEmbedding]):
             providers=self.providers,
             cuda=self.cuda,
             device_id=self.device_id,
+            extra_session_options=self._extra_session_options,
         )
 
         for token, idx in self.tokenizer.get_vocab().items():  # type: ignore[union-attr]
@@ -217,7 +230,9 @@ class Bm42(SparseTextEmbeddingBase, OnnxTextModel[SparseEmbedding]):
 
         return new_vector
 
-    def _post_process_onnx_output(self, output: OnnxOutputContext) -> Iterable[SparseEmbedding]:
+    def _post_process_onnx_output(
+        self, output: OnnxOutputContext, **kwargs: Any
+    ) -> Iterable[SparseEmbedding]:
         if output.input_ids is None:
             raise ValueError("input_ids must be provided for document post-processing")
 
@@ -269,9 +284,9 @@ class Bm42(SparseTextEmbeddingBase, OnnxTextModel[SparseEmbedding]):
 
     def embed(
         self,
-        documents: Union[str, Iterable[str]],
+        documents: str | Iterable[str],
         batch_size: int = 256,
-        parallel: Optional[int] = None,
+        parallel: int | None = None,
         **kwargs: Any,
     ) -> Iterable[SparseEmbedding]:
         """
@@ -299,6 +314,9 @@ class Bm42(SparseTextEmbeddingBase, OnnxTextModel[SparseEmbedding]):
             cuda=self.cuda,
             device_ids=self.device_ids,
             alpha=self.alpha,
+            local_files_only=self._local_files_only,
+            specific_model_path=self._specific_model_path,
+            extra_session_options=self._extra_session_options,
         )
 
     @classmethod
@@ -309,9 +327,7 @@ class Bm42(SparseTextEmbeddingBase, OnnxTextModel[SparseEmbedding]):
             result[token_id] = 1.0
         return result
 
-    def query_embed(
-        self, query: Union[str, Iterable[str]], **kwargs: Any
-    ) -> Iterable[SparseEmbedding]:
+    def query_embed(self, query: str | Iterable[str], **kwargs: Any) -> Iterable[SparseEmbedding]:
         """
         To emulate BM25 behaviour, we don't need to use smart weights in the query, and
         it's enough to just hash the tokens and assign a weight of 1.0 to them.
@@ -336,6 +352,13 @@ class Bm42(SparseTextEmbeddingBase, OnnxTextModel[SparseEmbedding]):
     def _get_worker_class(cls) -> Type[TextEmbeddingWorker[SparseEmbedding]]:
         return Bm42TextEmbeddingWorker
 
+    def token_count(
+        self, texts: str | Iterable[str], batch_size: int = 1024, **kwargs: Any
+    ) -> int:
+        if not hasattr(self, "model") or self.model is None:
+            self.load_onnx_model()  # loads the tokenizer as well
+        return self._token_count(texts, batch_size=batch_size, **kwargs)
+
 
 class Bm42TextEmbeddingWorker(TextEmbeddingWorker[SparseEmbedding]):
     def init_embedding(self, model_name: str, cache_dir: str, **kwargs: Any) -> Bm42:

fastembed/sparse/if_splade.py

@@ -0,0 +1,246 @@
+import json
+from typing import Any, Iterable, Sequence, Type
+
+import numpy as np
+
+from fastembed.common import OnnxProvider
+from fastembed.common.model_description import ModelSource, SparseModelDescription
+from fastembed.common.onnx_model import OnnxOutputContext
+from fastembed.common.preprocessor_utils import load_tokenizer
+from fastembed.common.types import Device
+from fastembed.common.utils import define_cache_dir, iter_batch
+from fastembed.sparse.sparse_embedding_base import (
+    SparseEmbedding,
+    SparseTextEmbeddingBase,
+)
+from fastembed.text.onnx_text_model import OnnxTextModel, TextEmbeddingWorker
+
+IDF_FILE = "idf.json"
+
+supported_if_splade_models: list[SparseModelDescription] = [
+    SparseModelDescription(
+        model="opensearch-project/opensearch-neural-sparse-encoding-doc-v3-gte",
+        vocab_size=30522,
+        description="Inference-free SPLADE model. Documents are expanded with an ONNX encoder at index "
+        "time, queries are encoded with a tokenizer and an IDF lookup table only, "
+        "without any model inference.",
+        license="apache-2.0",
+        size_in_GB=0.55,
+        sources=ModelSource(hf="Qdrant/opensearch-neural-sparse-encoding-doc-v3-gte"),
+        model_file="model.onnx",
+        additional_files=[IDF_FILE],
+        requires_idf=None,
+    ),
+]
+
+
+class IfSplade(SparseTextEmbeddingBase, OnnxTextModel[SparseEmbedding]):
+    """Inference-free (asymmetric) SPLADE model.
+
+    Documents are encoded with a neural encoder which expands them into a sparse vocabulary-sized
+    vector, while queries are encoded by tokenizing the text and looking up a precomputed IDF
+    weight per token — no neural inference happens at query time.
+
+    Query and document embeddings are compared with a dot product.
+    Special tokens are excluded from both document and query embeddings.
+    """
+
+    def _post_process_onnx_output(
+        self, output: OnnxOutputContext, **kwargs: Any
+    ) -> Iterable[SparseEmbedding]:
+        if output.attention_mask is None:
+            raise ValueError("attention_mask must be provided for document post-processing")
+
+        # Max-pool token logits over the sequence, masking out the padding
+        pooled = np.max(
+            output.model_output * np.expand_dims(output.attention_mask, axis=-1), axis=1
+        )
+        # v3 models of the opensearch-neural-sparse family use a double log activation,
+        # log(1 + log(1 + relu(x))), to increase sparsity of document embeddings
+        scores = np.log1p(np.log1p(np.maximum(pooled, 0.0)))
+
+        if self.special_tokens_ids:
+            scores[:, list(self.special_tokens_ids)] = 0.0
+
+        for row_scores in scores:
+            indices = row_scores.nonzero()[0]
+            yield SparseEmbedding(values=row_scores[indices], indices=indices)
+
+    def token_count(
+        self, texts: str | Iterable[str], batch_size: int = 1024, **kwargs: Any
+    ) -> int:
+        # unlike `OnnxTextModel._token_count`, does not require the onnx model to be loaded
+        token_num = 0
+        texts = [texts] if isinstance(texts, str) else texts
+        for batch in iter_batch(texts, batch_size):
+            for tokens in self.tokenizer.encode_batch(batch):  # type: ignore[union-attr]
+                token_num += sum(tokens.attention_mask)
+        return token_num
+
+    @classmethod
+    def _list_supported_models(cls) -> list[SparseModelDescription]:
+        """Lists the supported models.
+
+        Returns:
+            list[SparseModelDescription]: A list of SparseModelDescription objects containing the model information.
+        """
+        return supported_if_splade_models
+
+    def __init__(
+        self,
+        model_name: str,
+        cache_dir: str | None = None,
+        threads: int | None = None,
+        providers: Sequence[OnnxProvider] | None = None,
+        cuda: bool | Device = Device.AUTO,
+        device_ids: list[int] | None = None,
+        lazy_load: bool = False,
+        device_id: int | None = None,
+        specific_model_path: str | None = None,
+        **kwargs: Any,
+    ):
+        """
+        Args:
+            model_name (str): The name of the model to use.
+            cache_dir (str, optional): The path to the cache directory.
+                                       Can be set using the `FASTEMBED_CACHE_PATH` env variable.
+                                       Defaults to `fastembed_cache` in the system's temp directory.
+            threads (int, optional): The number of threads single onnxruntime session can use. Defaults to None.
+            providers (Optional[Sequence[OnnxProvider]], optional): The list of onnxruntime providers to use.
+                Mutually exclusive with the `cuda` and `device_ids` arguments. Defaults to None.
+            cuda (Union[bool, Device], optional): Whether to use cuda for inference. Mutually exclusive with `providers`
+                Defaults to Device.
+            device_ids (Optional[list[int]], optional): The list of device ids to use for data parallel processing in
+                workers. Should be used with `cuda` equals to `True`, `Device.AUTO` or `Device.CUDA`, mutually exclusive
+                with `providers`. Defaults to None.
+            lazy_load (bool, optional): Whether to load the model during class initialization or on demand.
+                Should be set to True when using multiple-gpu and parallel encoding. Defaults to False.
+            device_id (Optional[int], optional): The device id to use for loading the model in the worker process.
+            specific_model_path (Optional[str], optional): The specific path to the onnx model dir if it should be imported from somewhere else
+
+        Raises:
+            ValueError: If the model_name is not in the format <org>/<model> e.g. BAAI/bge-base-en.
+        """
+        super().__init__(model_name, cache_dir, threads, **kwargs)
+        self.providers = providers
+        self.lazy_load = lazy_load
+        self._extra_session_options = self._select_exposed_session_options(kwargs)
+
+        # List of device ids, that can be used for data parallel processing in workers
+        self.device_ids = device_ids
+        self.cuda = cuda
+
+        # This device_id will be used if we need to load model in current process
+        self.device_id: int | None = None
+        if device_id is not None:
+            self.device_id = device_id
+        elif self.device_ids is not None:
+            self.device_id = self.device_ids[0]
+
+        self.model_description = self._get_model_description(model_name)
+        self.cache_dir = str(define_cache_dir(cache_dir))
+
+        self._specific_model_path = specific_model_path
+        self._model_dir = self.download_model(
+            self.model_description,
+            self.cache_dir,
+            local_files_only=self._local_files_only,
+            specific_model_path=self._specific_model_path,
+        )
+
+        # The tokenizer and the idf table are lightweight and are required for query embedding,
+        # which does not involve any model inference, so they are loaded eagerly, while
+        # `lazy_load` only defers the initialization of the onnx model
+        self.tokenizer, self.special_token_to_id = load_tokenizer(model_dir=self._model_dir)
+        self.special_tokens_ids: set[int] = set(self.special_token_to_id.values())
+        self._token_id_to_idf = self._load_idf()
+
+        if not self.lazy_load:
+            self.load_onnx_model()
+
+    def load_onnx_model(self) -> None:
+        self._load_onnx_model(
+            model_dir=self._model_dir,
+            model_file=self.model_description.model_file,
+            threads=self.threads,
+            providers=self.providers,
+            cuda=self.cuda,
+            device_id=self.device_id,
+            extra_session_options=self._extra_session_options,
+        )
+
+    def _load_idf(self) -> dict[int, float]:
+        with open(self._model_dir / IDF_FILE) as f:
+            token_to_idf: dict[str, float] = json.load(f)
+
+        vocab: dict[str, int] = self.tokenizer.get_vocab()  # type: ignore[union-attr]
+        return {vocab[token]: idf for token, idf in token_to_idf.items() if token in vocab}
+
+    def embed(
+        self,
+        documents: str | Iterable[str],
+        batch_size: int = 256,
+        parallel: int | None = None,
+        **kwargs: Any,
+    ) -> Iterable[SparseEmbedding]:
+        """
+        Encode a list of documents into list of embeddings.
+
+        Args:
+            documents: Iterator of documents or single document to embed
+            batch_size: Batch size for encoding -- higher values will use more memory, but be faster
+            parallel:
+                If > 1, data-parallel encoding will be used, recommended for offline encoding of large datasets.
+                If 0, use all available cores.
+                If None, don't use data-parallel processing, use default onnxruntime threading instead.
+
+        Returns:
+            List of embeddings, one per document
+        """
+        yield from self._embed_documents(
+            model_name=self.model_name,
+            cache_dir=str(self.cache_dir),
+            documents=documents,
+            batch_size=batch_size,
+            parallel=parallel,
+            providers=self.providers,
+            cuda=self.cuda,
+            device_ids=self.device_ids,
+            local_files_only=self._local_files_only,
+            specific_model_path=self._specific_model_path,
+            extra_session_options=self._extra_session_options,
+            **kwargs,
+        )
+
+    def query_embed(self, query: str | Iterable[str], **kwargs: Any) -> Iterable[SparseEmbedding]:
+        """
+        Encode a list of queries into list of sparse embeddings without any model inference.
+
+        A query is tokenized, and each unique token is assigned its IDF weight from
+        a precomputed lookup table shipped with the model. Special tokens are ignored.
+        """
+        if isinstance(query, str):
+            query = [query]
+
+        for text in query:
+            token_ids = set(self.tokenizer.encode(text).ids) - self.special_tokens_ids  # type: ignore[union-attr]
+            embedding = {
+                token_id: self._token_id_to_idf[token_id]
+                for token_id in sorted(token_ids)
+                if token_id in self._token_id_to_idf
+            }
+            yield SparseEmbedding.from_dict(embedding)
+
+    @classmethod
+    def _get_worker_class(cls) -> Type[TextEmbeddingWorker[SparseEmbedding]]:
+        return IfSpladeEmbeddingWorker
+
+
+class IfSpladeEmbeddingWorker(TextEmbeddingWorker[SparseEmbedding]):
+    def init_embedding(self, model_name: str, cache_dir: str, **kwargs: Any) -> IfSplade:
+        return IfSplade(
+            model_name=model_name,
+            cache_dir=cache_dir,
+            threads=1,
+            **kwargs,
+        )

fastembed/sparse/minicoil.py

@@ -0,0 +1,372 @@
+from pathlib import Path
+
+from typing import Any, Sequence, Iterable, Type
+
+import numpy as np
+from numpy.typing import NDArray
+from py_rust_stemmers import SnowballStemmer
+from tokenizers import Tokenizer
+
+from fastembed.common.model_description import SparseModelDescription, ModelSource
+from fastembed.common.onnx_model import OnnxOutputContext
+from fastembed.common import OnnxProvider
+from fastembed.common.types import Device
+from fastembed.common.utils import define_cache_dir
+from fastembed.sparse.sparse_embedding_base import (
+    SparseEmbedding,
+    SparseTextEmbeddingBase,
+)
+from fastembed.sparse.utils.minicoil_encoder import Encoder
+from fastembed.sparse.utils.sparse_vectors_converter import SparseVectorConverter, WordEmbedding
+from fastembed.sparse.utils.vocab_resolver import VocabResolver, VocabTokenizer
+from fastembed.text.onnx_text_model import OnnxTextModel, TextEmbeddingWorker
+
+
+MINICOIL_MODEL_FILE = "minicoil.triplet.model.npy"
+MINICOIL_VOCAB_FILE = "minicoil.triplet.model.vocab"
+STOPWORDS_FILE = "stopwords.txt"
+
+
+supported_minicoil_models: list[SparseModelDescription] = [
+    SparseModelDescription(
+        model="Qdrant/minicoil-v1",
+        vocab_size=19125,
+        description="Sparse embedding model, that resolves semantic meaning of the words, "
+        "while keeping exact keyword match behavior. "
+        "Based on jinaai/jina-embeddings-v2-small-en-tokens",
+        license="apache-2.0",
+        size_in_GB=0.09,
+        sources=ModelSource(hf="Qdrant/minicoil-v1"),
+        model_file="onnx/model.onnx",
+        additional_files=[
+            STOPWORDS_FILE,
+            MINICOIL_MODEL_FILE,
+            MINICOIL_VOCAB_FILE,
+        ],
+        requires_idf=True,
+    ),
+]
+
+_MODEL_TO_LANGUAGE = {
+    "Qdrant/minicoil-v1": "english",
+}
+MODEL_TO_LANGUAGE = {
+    model_name.lower(): language for model_name, language in _MODEL_TO_LANGUAGE.items()
+}
+
+
+def get_language_by_model_name(model_name: str) -> str:
+    return MODEL_TO_LANGUAGE[model_name.lower()]
+
+
+class MiniCOIL(SparseTextEmbeddingBase, OnnxTextModel[SparseEmbedding]):
+    """
+        MiniCOIL is a sparse embedding model, that resolves semantic meaning of the words,
+        while keeping exact keyword match behavior.
+
+        Each vocabulary token is converted into 4d component of a sparse vector, which is then weighted by the token frequency in the corpus.
+        If the token is not found in the corpus, it is treated exactly like in BM25.
+    `
+        The model is based on `jinaai/jina-embeddings-v2-small-en-tokens`
+    """
+
+    def __init__(
+        self,
+        model_name: str,
+        cache_dir: str | None = None,
+        threads: int | None = None,
+        providers: Sequence[OnnxProvider] | None = None,
+        k: float = 1.2,
+        b: float = 0.75,
+        avg_len: float = 150.0,
+        cuda: bool | Device = Device.AUTO,
+        device_ids: list[int] | None = None,
+        lazy_load: bool = False,
+        device_id: int | None = None,
+        specific_model_path: str | None = None,
+        **kwargs: Any,
+    ):
+        """
+        Args:
+            model_name (str): The name of the model to use.
+            cache_dir (str, optional): The path to the cache directory.
+                                       Can be set using the `FASTEMBED_CACHE_PATH` env variable.
+                                       Defaults to `fastembed_cache` in the system's temp directory.
+            threads (int, optional): The number of threads single onnxruntime session can use. Defaults to None.
+            providers (Optional[Sequence[OnnxProvider]], optional): The providers to use for onnxruntime.
+            k (float, optional): The k parameter in the BM25 formula. Defines the saturation of the term frequency.
+                I.e. defines how fast the moment when additional terms stop to increase the score. Defaults to 1.2.
+            b (float, optional): The b parameter in the BM25 formula. Defines the importance of the document length.
+                Defaults to 0.75.
+            avg_len (float, optional): The average length of the documents in the corpus. Defaults to 150.0.
+            cuda (Union[bool, Device], optional): Whether to use cuda for inference. Mutually exclusive with `providers`
+                Defaults to Device.AUTO.
+            device_ids (Optional[list[int]], optional): The list of device ids to use for data parallel processing in
+                workers. Should be used with `cuda` equals to `True`, `Device.AUTO` or `Device.CUDA`, mutually exclusive
+                with `providers`. Defaults to None.
+            lazy_load (bool, optional): Whether to load the model during class initialization or on demand.
+                Should be set to True when using multiple-gpu and parallel encoding. Defaults to False.
+            device_id (Optional[int], optional): The device id to use for loading the model in the worker process.
+            specific_model_path (Optional[str], optional): The specific path to the onnx model dir if it should be imported from somewhere else
+
+        Raises:
+            ValueError: If the model_name is not in the format <org>/<model> e.g. BAAI/bge-base-en.
+        """
+
+        super().__init__(model_name, cache_dir, threads, **kwargs)
+        self.providers = providers
+        self.lazy_load = lazy_load
+        self.device_ids = device_ids
+        self.cuda = cuda
+        self.device_id = device_id
+        self._extra_session_options = self._select_exposed_session_options(kwargs)
+
+        self.k = k
+        self.b = b
+        self.avg_len = avg_len
+
+        # Initialize class attributes
+        self.tokenizer: Tokenizer | None = None
+        self.invert_vocab: dict[int, str] = {}
+        self.special_tokens: set[str] = set()
+        self.special_tokens_ids: set[int] = set()
+        self.stopwords: set[str] = set()
+        self.vocab_resolver: VocabResolver | None = None
+        self.encoder: Encoder | None = None
+        self.output_dim: int | None = None
+        self.sparse_vector_converter: SparseVectorConverter | None = None
+
+        self.model_description = self._get_model_description(model_name)
+        self.cache_dir = str(define_cache_dir(cache_dir))
+        self._specific_model_path = specific_model_path
+        self._model_dir = self.download_model(
+            self.model_description,
+            self.cache_dir,
+            local_files_only=self._local_files_only,
+            specific_model_path=self._specific_model_path,
+        )
+
+        if not self.lazy_load:
+            self.load_onnx_model()
+
+    def load_onnx_model(self) -> None:
+        self._load_onnx_model(
+            model_dir=self._model_dir,
+            model_file=self.model_description.model_file,
+            threads=self.threads,
+            providers=self.providers,
+            cuda=self.cuda,
+            device_id=self.device_id,
+            extra_session_options=self._extra_session_options,
+        )
+
+        assert self.tokenizer is not None
+
+        for token, idx in self.tokenizer.get_vocab().items():  # type: ignore[union-attr]
+            self.invert_vocab[idx] = token
+        self.special_tokens = set(self.special_token_to_id.keys())
+        self.special_tokens_ids = set(self.special_token_to_id.values())
+        self.stopwords = set(self._load_stopwords(self._model_dir))
+
+        stemmer = SnowballStemmer(get_language_by_model_name(self.model_name))
+
+        self.vocab_resolver = VocabResolver(
+            tokenizer=VocabTokenizer(self.tokenizer),
+            stopwords=self.stopwords,
+            stemmer=stemmer,
+        )
+        self.vocab_resolver.load_json_vocab(str(self._model_dir / MINICOIL_VOCAB_FILE))
+
+        weights = np.load(str(self._model_dir / MINICOIL_MODEL_FILE), mmap_mode="r")
+        self.encoder = Encoder(weights)
+        self.output_dim = self.encoder.output_dim
+
+        self.sparse_vector_converter = SparseVectorConverter(
+            stopwords=self.stopwords,
+            stemmer=stemmer,
+            k=self.k,
+            b=self.b,
+            avg_len=self.avg_len,
+        )
+
+    def token_count(
+        self, texts: str | Iterable[str], batch_size: int = 1024, **kwargs: Any
+    ) -> int:
+        return self._token_count(texts, batch_size=batch_size, **kwargs)
+
+    def embed(
+        self,
+        documents: str | Iterable[str],
+        batch_size: int = 256,
+        parallel: int | None = None,
+        **kwargs: Any,
+    ) -> Iterable[SparseEmbedding]:
+        """
+        Encode a list of documents into list of embeddings.
+        We use mean pooling with attention so that the model can handle variable-length inputs.
+
+        Args:
+            documents: Iterator of documents or single document to embed
+            batch_size: Batch size for encoding -- higher values will use more memory, but be faster
+            parallel:
+                If > 1, data-parallel encoding will be used, recommended for offline encoding of large datasets.
+                If 0, use all available cores.
+                If None, don't use data-parallel processing, use default onnxruntime threading instead.
+
+        Returns:
+            List of embeddings, one per document
+        """
+        yield from self._embed_documents(
+            model_name=self.model_name,
+            cache_dir=str(self.cache_dir),
+            documents=documents,
+            batch_size=batch_size,
+            parallel=parallel,
+            providers=self.providers,
+            cuda=self.cuda,
+            device_ids=self.device_ids,
+            k=self.k,
+            b=self.b,
+            avg_len=self.avg_len,
+            is_query=False,
+            local_files_only=self._local_files_only,
+            specific_model_path=self._specific_model_path,
+            extra_session_options=self._extra_session_options,
+            **kwargs,
+        )
+
+    def query_embed(self, query: str | Iterable[str], **kwargs: Any) -> Iterable[SparseEmbedding]:
+        """
+        Encode a list of queries into list of embeddings.
+        """
+        yield from self._embed_documents(
+            model_name=self.model_name,
+            cache_dir=str(self.cache_dir),
+            documents=query,
+            providers=self.providers,
+            cuda=self.cuda,
+            device_ids=self.device_ids,
+            k=self.k,
+            b=self.b,
+            avg_len=self.avg_len,
+            is_query=True,
+            local_files_only=self._local_files_only,
+            specific_model_path=self._specific_model_path,
+            **kwargs,
+        )
+
+    @classmethod
+    def _load_stopwords(cls, model_dir: Path) -> list[str]:
+        stopwords_path = model_dir / STOPWORDS_FILE
+        if not stopwords_path.exists():
+            return []
+
+        with open(stopwords_path, "r") as f:
+            return f.read().splitlines()
+
+    @classmethod
+    def _list_supported_models(cls) -> list[SparseModelDescription]:
+        """Lists the supported models.
+
+        Returns:
+            list[SparseModelDescription]: A list of SparseModelDescription objects containing the model information.
+        """
+        return supported_minicoil_models
+
+    def _post_process_onnx_output(
+        self, output: OnnxOutputContext, is_query: bool = False, **kwargs: Any
+    ) -> Iterable[SparseEmbedding]:
+        if output.input_ids is None:
+            raise ValueError("input_ids must be provided for document post-processing")
+
+        assert self.vocab_resolver is not None
+        assert self.encoder is not None
+        assert self.sparse_vector_converter is not None
+
+        # Size: (batch_size, sequence_length, hidden_size)
+        embeddings = output.model_output
+        # Size: (batch_size, sequence_length)
+        assert output.attention_mask is not None
+        masks = output.attention_mask
+
+        vocab_size = self.vocab_resolver.vocab_size()
+        embedding_size = self.encoder.output_dim
+
+        # For each document we only select those embeddings that are not masked out
+
+        for i in range(embeddings.shape[0]):
+            # Size: (sequence_length, hidden_size)
+            token_embeddings = embeddings[i, masks[i] == 1]
+
+            # Size: (sequence_length)
+            token_ids: NDArray[np.int64] = output.input_ids[i, masks[i] == 1]
+
+            word_ids_array, counts, oov, forms = self.vocab_resolver.resolve_tokens(token_ids)
+
+            # Size: (1, words)
+            word_ids_array_expanded: NDArray[np.int64] = np.expand_dims(word_ids_array, axis=0)
+
+            # Size: (1, words, embedding_size)
+            token_embeddings_array: NDArray[np.float32] = np.expand_dims(token_embeddings, axis=0)
+
+            assert word_ids_array_expanded.shape[1] == token_embeddings_array.shape[1]
+
+            # Size of word_ids_mapping: (unique_words, 2) - [vocab_id, batch_id]
+            # Size of embeddings: (unique_words, embedding_size)
+            ids_mapping, minicoil_embeddings = self.encoder.forward(
+                word_ids_array_expanded, token_embeddings_array
+            )
+
+            # Size of counts: (unique_words)
+            words_ids: list[int] = ids_mapping[:, 0].tolist()  # type: ignore[assignment]
+
+            sentence_result: dict[str, WordEmbedding] = {}
+
+            words = [self.vocab_resolver.lookup_word(word_id) for word_id in words_ids]
+
+            for word, word_id, emb in zip(words, words_ids, minicoil_embeddings.tolist()):  # type: ignore[arg-type]
+                if word_id == 0:
+                    continue
+
+                sentence_result[word] = WordEmbedding(
+                    word=word,
+                    forms=forms[word],
+                    count=int(counts[word_id]),
+                    word_id=int(word_id),
+                    embedding=emb,  # type: ignore[arg-type]
+                )
+
+            for oov_word, count in oov.items():
+                # {
+                #     "word": oov_word,
+                #     "forms": [oov_word],
+                #     "count": int(count),
+                #     "word_id": -1,
+                #     "embedding": [1]
+                # }
+                sentence_result[oov_word] = WordEmbedding(
+                    word=oov_word, forms=[oov_word], count=int(count), word_id=-1, embedding=[1]
+                )
+
+            if not is_query:
+                yield self.sparse_vector_converter.embedding_to_vector(
+                    sentence_result, vocab_size=vocab_size, embedding_size=embedding_size
+                )
+            else:
+                yield self.sparse_vector_converter.embedding_to_vector_query(
+                    sentence_result, vocab_size=vocab_size, embedding_size=embedding_size
+                )
+
+    @classmethod
+    def _get_worker_class(cls) -> Type["MiniCoilTextEmbeddingWorker"]:
+        return MiniCoilTextEmbeddingWorker
+
+
+class MiniCoilTextEmbeddingWorker(TextEmbeddingWorker[SparseEmbedding]):
+    def init_embedding(self, model_name: str, cache_dir: str, **kwargs: Any) -> MiniCOIL:
+        return MiniCOIL(
+            model_name=model_name,
+            cache_dir=cache_dir,
+            threads=1,
+            **kwargs,
+        )

fastembed/sparse/sparse_embedding_base.py

@@ -1,5 +1,5 @@
 from dataclasses import dataclass
-from typing import Iterable, Optional, Union, Any
+from typing import Iterable, Any
 
 import numpy as np
 from numpy.typing import NDArray
@@ -12,7 +12,7 @@ from fastembed.common.model_management import ModelManagement
 @dataclass
 class SparseEmbedding:
     values: NumpyArray
-    indices: Union[NDArray[np.int64], NDArray[np.int32]]
+    indices: NDArray[np.int64] | NDArray[np.int32]
 
     def as_object(self) -> dict[str, NumpyArray]:
         return {
@@ -35,8 +35,8 @@ class SparseTextEmbeddingBase(ModelManagement[SparseModelDescription]):
     def __init__(
         self,
         model_name: str,
-        cache_dir: Optional[str] = None,
-        threads: Optional[int] = None,
+        cache_dir: str | None = None,
+        threads: int | None = None,
         **kwargs: Any,
     ):
         self.model_name = model_name
@@ -46,9 +46,9 @@ class SparseTextEmbeddingBase(ModelManagement[SparseModelDescription]):
 
     def embed(
         self,
-        documents: Union[str, Iterable[str]],
+        documents: str | Iterable[str],
         batch_size: int = 256,
-        parallel: Optional[int] = None,
+        parallel: int | None = None,
         **kwargs: Any,
     ) -> Iterable[SparseEmbedding]:
         raise NotImplementedError()
@@ -68,9 +68,7 @@ class SparseTextEmbeddingBase(ModelManagement[SparseModelDescription]):
         # This is model-specific, so that different models can have specialized implementations
         yield from self.embed(texts, **kwargs)
 
-    def query_embed(
-        self, query: Union[str, Iterable[str]], **kwargs: Any
-    ) -> Iterable[SparseEmbedding]:
+    def query_embed(self, query: str | Iterable[str], **kwargs: Any) -> Iterable[SparseEmbedding]:
         """
         Embeds queries
 
@@ -86,3 +84,7 @@ class SparseTextEmbeddingBase(ModelManagement[SparseModelDescription]):
             yield from self.embed([query], **kwargs)
         else:
             yield from self.embed(query, **kwargs)
+
+    def token_count(self, texts: str | Iterable[str], **kwargs: Any) -> int:
+        """Returns the number of tokens in the texts."""
+        raise NotImplementedError("Subclasses must implement this method")

fastembed/sparse/sparse_text_embedding.py

@@ -1,9 +1,12 @@
-from typing import Any, Iterable, Optional, Sequence, Type, Union
+from typing import Any, Iterable, Sequence, Type
 from dataclasses import asdict
 
 from fastembed.common import OnnxProvider
+from fastembed.common.types import Device
 from fastembed.sparse.bm25 import Bm25
 from fastembed.sparse.bm42 import Bm42
+from fastembed.sparse.if_splade import IfSplade
+from fastembed.sparse.minicoil import MiniCOIL
 from fastembed.sparse.sparse_embedding_base import (
     SparseEmbedding,
     SparseTextEmbeddingBase,
@@ -14,7 +17,13 @@ from fastembed.common.model_description import SparseModelDescription
 
 
 class SparseTextEmbedding(SparseTextEmbeddingBase):
-    EMBEDDINGS_REGISTRY: list[Type[SparseTextEmbeddingBase]] = [SpladePP, Bm42, Bm25]
+    EMBEDDINGS_REGISTRY: list[Type[SparseTextEmbeddingBase]] = [
+        SpladePP,
+        Bm42,
+        Bm25,
+        MiniCOIL,
+        IfSplade,
+    ]
 
     @classmethod
     def list_supported_models(cls) -> list[dict[str, Any]]:
@@ -52,16 +61,16 @@ class SparseTextEmbedding(SparseTextEmbeddingBase):
     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,
+        cache_dir: str | None = None,
+        threads: int | None = None,
+        providers: Sequence[OnnxProvider] | None = None,
+        cuda: bool | Device = Device.AUTO,
+        device_ids: list[int] | None = None,
         lazy_load: bool = False,
         **kwargs: Any,
     ):
         super().__init__(model_name, cache_dir, threads, **kwargs)
-        if model_name == "prithvida/Splade_PP_en_v1":
+        if model_name.lower() == "prithvida/Splade_PP_en_v1".lower():
             warnings.warn(
                 "The right spelling is prithivida/Splade_PP_en_v1. "
                 "Support of this name will be removed soon, please fix the model_name",
@@ -92,9 +101,9 @@ class SparseTextEmbedding(SparseTextEmbeddingBase):
 
     def embed(
         self,
-        documents: Union[str, Iterable[str]],
+        documents: str | Iterable[str],
         batch_size: int = 256,
-        parallel: Optional[int] = None,
+        parallel: int | None = None,
         **kwargs: Any,
     ) -> Iterable[SparseEmbedding]:
         """
@@ -114,9 +123,7 @@ class SparseTextEmbedding(SparseTextEmbeddingBase):
         """
         yield from self.model.embed(documents, batch_size, parallel, **kwargs)
 
-    def query_embed(
-        self, query: Union[str, Iterable[str]], **kwargs: Any
-    ) -> Iterable[SparseEmbedding]:
+    def query_embed(self, query: str | Iterable[str], **kwargs: Any) -> Iterable[SparseEmbedding]:
         """
         Embeds queries
 
@@ -127,3 +134,17 @@ class SparseTextEmbedding(SparseTextEmbeddingBase):
             Iterable[SparseEmbedding]: The sparse embeddings.
         """
         yield from self.model.query_embed(query, **kwargs)
+
+    def token_count(
+        self, texts: str | Iterable[str], batch_size: int = 1024, **kwargs: Any
+    ) -> int:
+        """Returns the number of tokens in the texts.
+
+        Args:
+            texts (str | Iterable[str]): The list of texts to embed.
+            batch_size (int): Batch size for encoding
+
+        Returns:
+            int: Sum of number of tokens in the texts.
+        """
+        return self.model.token_count(texts, batch_size=batch_size, **kwargs)

fastembed/sparse/splade_pp.py

@@ -1,8 +1,9 @@
-from typing import Any, Iterable, Optional, Sequence, Type, Union
+from typing import Any, Iterable, Sequence, Type
 
 import numpy as np
 from fastembed.common import OnnxProvider
 from fastembed.common.onnx_model import OnnxOutputContext
+from fastembed.common.types import Device
 from fastembed.common.utils import define_cache_dir
 from fastembed.sparse.sparse_embedding_base import (
     SparseEmbedding,
@@ -18,7 +19,7 @@ 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(
@@ -27,14 +28,16 @@ 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",
     ),
 ]
 
 
 class SpladePP(SparseTextEmbeddingBase, OnnxTextModel[SparseEmbedding]):
-    def _post_process_onnx_output(self, output: OnnxOutputContext) -> Iterable[SparseEmbedding]:
+    def _post_process_onnx_output(
+        self, output: OnnxOutputContext, **kwargs: Any
+    ) -> Iterable[SparseEmbedding]:
         if output.attention_mask is None:
             raise ValueError("attention_mask must be provided for document post-processing")
 
@@ -51,6 +54,11 @@ class SpladePP(SparseTextEmbeddingBase, OnnxTextModel[SparseEmbedding]):
             scores = row_scores[indices]
             yield SparseEmbedding(values=scores, indices=indices)
 
+    def token_count(
+        self, texts: str | Iterable[str], batch_size: int = 1024, **kwargs: Any
+    ) -> int:
+        return self._token_count(texts, batch_size=batch_size, **kwargs)
+
     @classmethod
     def _list_supported_models(cls) -> list[SparseModelDescription]:
         """Lists the supported models.
@@ -63,14 +71,14 @@ class SpladePP(SparseTextEmbeddingBase, OnnxTextModel[SparseEmbedding]):
     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,
+        cache_dir: str | None = None,
+        threads: int | None = None,
+        providers: Sequence[OnnxProvider] | None = None,
+        cuda: bool | Device = Device.AUTO,
+        device_ids: list[int] | None = None,
         lazy_load: bool = False,
-        device_id: Optional[int] = None,
-        specific_model_path: Optional[str] = None,
+        device_id: int | None = None,
+        specific_model_path: str | None = None,
         **kwargs: Any,
     ):
         """
@@ -82,10 +90,11 @@ class SpladePP(SparseTextEmbeddingBase, OnnxTextModel[SparseEmbedding]):
             threads (int, optional): The number of threads single onnxruntime session can use. Defaults to None.
             providers (Optional[Sequence[OnnxProvider]], optional): The list of onnxruntime providers to use.
                 Mutually exclusive with the `cuda` and `device_ids` arguments. Defaults to None.
-            cuda (bool, optional): Whether to use cuda for inference. Mutually exclusive with `providers`
-                Defaults to False.
+            cuda (Union[bool, Device], optional): Whether to use cuda for inference. Mutually exclusive with `providers`
+                Defaults to Device.
             device_ids (Optional[list[int]], optional): The list of device ids to use for data parallel processing in
-                workers. Should be used with `cuda=True`, mutually exclusive with `providers`. Defaults to None.
+                workers. Should be used with `cuda` equals to `True`, `Device.AUTO` or `Device.CUDA`, mutually exclusive
+                with `providers`. Defaults to None.
             lazy_load (bool, optional): Whether to load the model during class initialization or on demand.
                 Should be set to True when using multiple-gpu and parallel encoding. Defaults to False.
             device_id (Optional[int], optional): The device id to use for loading the model in the worker process.
@@ -97,13 +106,14 @@ class SpladePP(SparseTextEmbeddingBase, OnnxTextModel[SparseEmbedding]):
         super().__init__(model_name, cache_dir, threads, **kwargs)
         self.providers = providers
         self.lazy_load = lazy_load
+        self._extra_session_options = self._select_exposed_session_options(kwargs)
 
         # List of device ids, that can be used for data parallel processing in workers
         self.device_ids = device_ids
         self.cuda = cuda
 
         # This device_id will be used if we need to load model in current process
-        self.device_id: Optional[int] = None
+        self.device_id: int | None = None
         if device_id is not None:
             self.device_id = device_id
         elif self.device_ids is not None:
@@ -112,11 +122,12 @@ class SpladePP(SparseTextEmbeddingBase, OnnxTextModel[SparseEmbedding]):
         self.model_description = self._get_model_description(model_name)
         self.cache_dir = str(define_cache_dir(cache_dir))
 
+        self._specific_model_path = specific_model_path
         self._model_dir = self.download_model(
             self.model_description,
             self.cache_dir,
             local_files_only=self._local_files_only,
-            specific_model_path=specific_model_path,
+            specific_model_path=self._specific_model_path,
         )
 
         if not self.lazy_load:
@@ -130,13 +141,14 @@ class SpladePP(SparseTextEmbeddingBase, OnnxTextModel[SparseEmbedding]):
             providers=self.providers,
             cuda=self.cuda,
             device_id=self.device_id,
+            extra_session_options=self._extra_session_options,
         )
 
     def embed(
         self,
-        documents: Union[str, Iterable[str]],
+        documents: str | Iterable[str],
         batch_size: int = 256,
-        parallel: Optional[int] = None,
+        parallel: int | None = None,
         **kwargs: Any,
     ) -> Iterable[SparseEmbedding]:
         """
@@ -163,6 +175,9 @@ class SpladePP(SparseTextEmbeddingBase, OnnxTextModel[SparseEmbedding]):
             providers=self.providers,
             cuda=self.cuda,
             device_ids=self.device_ids,
+            local_files_only=self._local_files_only,
+            specific_model_path=self._specific_model_path,
+            extra_session_options=self._extra_session_options,
             **kwargs,
         )
 

fastembed/sparse/utils/minicoil_encoder.py

@@ -0,0 +1,146 @@
+"""
+Pure numpy implementation of encoder model for a single word.
+
+This model is not trainable, and should only be used for inference.
+"""
+
+import numpy as np
+from fastembed.common.types import NumpyArray
+
+
+class Encoder:
+    """
+    Encoder(768, 4, 10000)
+
+    Will look like this:
+
+
+                                         Per-word
+                                         Encoder Matrix
+     ┌─────────────────────┐
+     │ Token Embedding(768)├──────┐      (10k, 768, 4)
+     └─────────────────────┘      │         ┌─────────┐
+                                  │         │         │
+     ┌─────────────────────┐      │       ┌─┴───────┐ │
+     │                     │      │       │         │ │
+     └─────────────────────┘      │     ┌─┴───────┐ │ │      ┌─────────┐
+                                  └────►│         │ │ ├─────►│Tanh     │
+     ┌─────────────────────┐            │         │ │ │      └─────────┘
+     │                     │            │         │ ├─┘
+     └─────────────────────┘            │         ├─┘
+                                        │         │
+     ┌─────────────────────┐            └─────────┘
+     │                     │
+     └─────────────────────┘
+
+     Final linear transformation is accompanied by a non-linear activation function: Tanh.
+
+     Tanh is used to ensure that the output is in the range [-1, 1].
+     It would be easier to visually interpret the output of the model, assuming that each dimension
+     would need to encode a type of semantic cluster.
+    """
+
+    def __init__(
+        self,
+        weights: NumpyArray,
+    ):
+        self.weights = weights
+        self.vocab_size, self.input_dim, self.output_dim = weights.shape
+
+        self.encoder_weights: NumpyArray = weights
+
+        # Activation function
+        self.activation = np.tanh
+
+    @staticmethod
+    def convert_vocab_ids(vocab_ids: NumpyArray) -> NumpyArray:
+        """
+        Convert vocab_ids of shape (batch_size, seq_len) into (batch_size, seq_len, 2)
+        by appending batch_id alongside each vocab_id.
+        """
+        batch_size, seq_len = vocab_ids.shape
+        batch_ids = np.arange(batch_size, dtype=vocab_ids.dtype).reshape(batch_size, 1)
+        batch_ids = np.repeat(batch_ids, seq_len, axis=1)
+        # Stack vocab_ids and batch_ids along the last dimension
+        combined: NumpyArray = np.stack((vocab_ids, batch_ids), axis=2).astype(np.int32)
+        return combined
+
+    @classmethod
+    def avg_by_vocab_ids(
+        cls, vocab_ids: NumpyArray, embeddings: NumpyArray
+    ) -> tuple[NumpyArray, NumpyArray]:
+        """
+        Takes:
+            vocab_ids: (batch_size, seq_len) int array
+            embeddings: (batch_size, seq_len, input_dim) float array
+
+        Returns:
+            unique_flattened_vocab_ids: (total_unique, 2) array of [vocab_id, batch_id]
+            unique_flattened_embeddings: (total_unique, input_dim) averaged embeddings
+        """
+        input_dim = embeddings.shape[2]
+
+        # Flatten vocab_ids and embeddings
+        # flattened_vocab_ids: (batch_size*seq_len, 2)
+        flattened_vocab_ids = cls.convert_vocab_ids(vocab_ids).reshape(-1, 2)
+
+        # flattened_embeddings: (batch_size*seq_len, input_dim)
+        flattened_embeddings = embeddings.reshape(-1, input_dim)
+
+        # Find unique (vocab_id, batch_id) pairs
+        unique_flattened_vocab_ids, inverse_indices = np.unique(
+            flattened_vocab_ids, axis=0, return_inverse=True
+        )
+
+        # Prepare arrays to accumulate sums
+        unique_count = unique_flattened_vocab_ids.shape[0]
+        unique_flattened_embeddings = np.zeros((unique_count, input_dim), dtype=np.float32)
+        unique_flattened_count = np.zeros(unique_count, dtype=np.int32)
+
+        # Use np.add.at to accumulate sums based on inverse indices
+        np.add.at(unique_flattened_embeddings, inverse_indices, flattened_embeddings)
+        np.add.at(unique_flattened_count, inverse_indices, 1)
+
+        # Compute averages
+        unique_flattened_embeddings /= unique_flattened_count[:, None]
+
+        return unique_flattened_vocab_ids.astype(np.int32), unique_flattened_embeddings.astype(
+            np.float32
+        )
+
+    def forward(
+        self, vocab_ids: NumpyArray, embeddings: NumpyArray
+    ) -> tuple[NumpyArray, NumpyArray]:
+        """
+        Args:
+            vocab_ids: (batch_size, seq_len) int array
+            embeddings: (batch_size, seq_len, input_dim) float array
+
+        Returns:
+            unique_flattened_vocab_ids_and_batch_ids: (total_unique, 2)
+            unique_flattened_encoded: (total_unique, output_dim)
+        """
+        # Average embeddings for duplicate vocab_ids
+        unique_flattened_vocab_ids_and_batch_ids, unique_flattened_embeddings = (
+            self.avg_by_vocab_ids(vocab_ids, embeddings)
+        )
+
+        # Select the encoder weights for each unique vocab_id
+        unique_flattened_vocab_ids = unique_flattened_vocab_ids_and_batch_ids[:, 0].astype(
+            np.int32
+        )
+
+        # unique_encoder_weights: (total_unique, input_dim, output_dim)
+        unique_encoder_weights = self.encoder_weights[unique_flattened_vocab_ids]
+
+        # Compute linear transform: (total_unique, output_dim)
+        # Using Einstein summation for matrix multiplication:
+        # 'bi,bio->bo' means: for each "b" (batch element), multiply embeddings (b,i) by weights (b,i,o) -> (b,o)
+        unique_flattened_encoded = np.einsum(
+            "bi,bio->bo", unique_flattened_embeddings, unique_encoder_weights
+        )
+
+        # Apply Tanh activation and ensure float32 type
+        unique_flattened_encoded = self.activation(unique_flattened_encoded).astype(np.float32)
+
+        return unique_flattened_vocab_ids_and_batch_ids.astype(np.int32), unique_flattened_encoded

fastembed/sparse/utils/sparse_vectors_converter.py

@@ -0,0 +1,244 @@
+import copy
+from dataclasses import dataclass
+
+import mmh3
+import numpy as np
+from py_rust_stemmers import SnowballStemmer
+
+from fastembed.common.utils import get_all_punctuation, remove_non_alphanumeric
+from fastembed.sparse.sparse_embedding_base import SparseEmbedding
+
+GAP = 32000
+INT32_MAX = 2**31 - 1
+
+
+@dataclass
+class WordEmbedding:
+    word: str
+    forms: list[str]
+    count: int
+    word_id: int
+    embedding: list[float]
+
+
+class SparseVectorConverter:
+    def __init__(
+        self,
+        stopwords: set[str],
+        stemmer: SnowballStemmer,
+        k: float = 1.2,
+        b: float = 0.75,
+        avg_len: float = 150.0,
+    ):
+        punctuation = set(get_all_punctuation())
+        special_tokens = {"[CLS]", "[SEP]", "[PAD]", "[UNK]", "[MASK]"}
+
+        self.stemmer = stemmer
+        self.unwanted_tokens = punctuation | special_tokens | stopwords
+
+        self.k = k
+        self.b = b
+        self.avg_len = avg_len
+
+    @classmethod
+    def unkn_word_token_id(
+        cls, word: str, shift: int
+    ) -> int:  # 2-3 words can collide in 1 index with this mapping, not considering mm3 collisions
+        token_hash = abs(mmh3.hash(word))
+
+        range_size = INT32_MAX - shift
+        remapped_hash = shift + (token_hash % range_size)
+
+        return remapped_hash
+
+    def bm25_tf(self, num_occurrences: int, sentence_len: int) -> float:
+        res = num_occurrences * (self.k + 1)
+        res /= num_occurrences + self.k * (1 - self.b + self.b * sentence_len / self.avg_len)
+        return res
+
+    @classmethod
+    def normalize_vector(cls, vector: list[float]) -> list[float]:
+        norm = sum([x**2 for x in vector]) ** 0.5
+        if norm < 1e-8:
+            return vector
+        return [x / norm for x in vector]
+
+    def clean_words(
+        self, sentence_embedding: dict[str, WordEmbedding], token_max_length: int = 40
+    ) -> dict[str, WordEmbedding]:
+        """
+        Clean miniCOIL-produced sentence_embedding, as unknown to the miniCOIL's stemmer tokens should fully resemble
+        our BM25 token representation.
+
+        sentence_embedding = {"9°": {"word": "9°", "word_id": -1, "count": 2, "embedding": [1], "forms": ["9°"]},
+                "9": {"word": "9", "word_id": -1, "count": 2, "embedding": [1], "forms": ["9"]},
+                "bat": {"word": "bat", "word_id": 2, "count": 3, "embedding": [0.2, 0.1, -0.2, -0.2], "forms": ["bats", "bat"]},
+                "9°9": {"word": "9°9", "word_id": -1, "count": 1, "embedding": [1], "forms": ["9°9"]},
+                "screech": {"word": "screech", "word_id": -1, "count": 1, "embedding": [1], "forms": ["screech"]},
+                "screeched": {"word": "screeched", "word_id": -1, "count": 1, "embedding": [1], "forms": ["screeched"]}
+                }
+        cleaned_embedding_ground_truth = {
+                "9": {"word": "9", "word_id": -1, "count": 6, "embedding": [1], "forms": ["9°", "9", "9°9", "9°9"]},
+                "bat": {"word": "bat", "word_id": 2, "count": 3, "embedding": [0.2, 0.1, -0.2, -0.2], "forms": ["bats", "bat"]},
+                "screech": {"word": "screech", "word_id": -1, "count": 2, "embedding": [1], "forms": ["screech", "screeched"]}
+                }
+        """
+
+        new_sentence_embedding: dict[str, WordEmbedding] = {}
+
+        for word, embedding in sentence_embedding.items():
+            # embedding = {
+            #     "word": "vector",
+            #     "forms": ["vector", "vectors"],
+            #     "count": 2,
+            #     "word_id": 1231,
+            #     "embedding": [0.1, 0.2, 0.3, 0.4]
+            # }
+            if embedding.word_id > 0:
+                # Known word, no need to clean
+                new_sentence_embedding[word] = embedding
+            else:
+                # Unknown word
+                if word in self.unwanted_tokens:
+                    continue
+
+                # Example complex word split:
+                # word = `word^vec`
+                word_cleaned = remove_non_alphanumeric(word).strip()
+                # word_cleaned = `word vec`
+
+                if len(word_cleaned) > 0:
+                    # Subwords: ['word', 'vec']
+                    for subword in word_cleaned.split():
+                        stemmed_subword: str = self.stemmer.stem_word(subword)
+                        if (
+                            len(stemmed_subword) <= token_max_length
+                            and stemmed_subword not in self.unwanted_tokens
+                        ):
+                            if stemmed_subword not in new_sentence_embedding:
+                                new_sentence_embedding[stemmed_subword] = copy.deepcopy(embedding)
+                                new_sentence_embedding[stemmed_subword].word = stemmed_subword
+                            else:
+                                new_sentence_embedding[stemmed_subword].count += embedding.count
+                                new_sentence_embedding[stemmed_subword].forms += embedding.forms
+
+        return new_sentence_embedding
+
+    def embedding_to_vector(
+        self,
+        sentence_embedding: dict[str, WordEmbedding],
+        embedding_size: int,
+        vocab_size: int,
+    ) -> SparseEmbedding:
+        """
+        Convert miniCOIL sentence embedding to Qdrant sparse vector
+
+        Example input:
+
+        ```
+        {
+            "vector": WordEmbedding({ // Vocabulary word, encoded with miniCOIL normally
+                "word": "vector",
+                "forms": ["vector", "vectors"],
+                "count": 2,
+                "word_id": 1231,
+                "embedding": [0.1, 0.2, 0.3, 0.4]
+            }),
+            "axiotic": WordEmbedding({ // Out-of-vocabulary word, fallback to BM25
+                "word": "axiotic",
+                "forms": ["axiotics"],
+                "count": 1,
+                "word_id": -1,
+            })
+        }
+        ```
+
+        """
+
+        indices: list[int] = []
+        values: list[float] = []
+
+        # Example:
+        # vocab_size = 10000
+        # embedding_size = 4
+        # GAP = 32000
+        #
+        # We want to start random words section from the bucket, that is guaranteed to not
+        # include any vocab words.
+        # We need (vocab_size * embedding_size) slots for vocab words.
+        # Therefore we need (vocab_size * embedding_size) // GAP + 1 buckets for vocab words.
+        # Therefore, we can start random words from bucket (vocab_size * embedding_size) // GAP + 1 + 1
+
+        # ID at which the scope of OOV words starts
+        unknown_words_shift = ((vocab_size * embedding_size) // GAP + 2) * GAP
+        sentence_embedding_cleaned = self.clean_words(sentence_embedding)
+
+        # Calculate sentence length after cleaning
+        sentence_len = 0
+        for embedding in sentence_embedding_cleaned.values():
+            sentence_len += embedding.count
+
+        for embedding in sentence_embedding_cleaned.values():
+            word_id = embedding.word_id
+            num_occurrences = embedding.count
+            tf = self.bm25_tf(num_occurrences, sentence_len)
+            if (
+                word_id > 0
+            ):  # miniCOIL starts with ID 1, we generally won't have word_id == 0 (UNK), as we don't add
+                # these words to sentence_embedding
+                embedding_values = embedding.embedding
+                normalized_embedding = self.normalize_vector(embedding_values)
+
+                for val_id, value in enumerate(normalized_embedding):
+                    indices.append(
+                        word_id * embedding_size + val_id
+                    )  # since miniCOIL IDs start with 1
+                    values.append(value * tf)
+            else:
+                indices.append(self.unkn_word_token_id(embedding.word, unknown_words_shift))
+                values.append(tf)
+
+        return SparseEmbedding(
+            indices=np.array(indices, dtype=np.int32),
+            values=np.array(values, dtype=np.float32),
+        )
+
+    def embedding_to_vector_query(
+        self,
+        sentence_embedding: dict[str, WordEmbedding],
+        embedding_size: int,
+        vocab_size: int,
+    ) -> SparseEmbedding:
+        """
+        Same as `embedding_to_vector`, but no TF
+        """
+
+        indices: list[int] = []
+        values: list[float] = []
+
+        # ID at which the scope of OOV words starts
+        unknown_words_shift = ((vocab_size * embedding_size) // GAP + 2) * GAP
+
+        sentence_embedding_cleaned = self.clean_words(sentence_embedding)
+
+        for embedding in sentence_embedding_cleaned.values():
+            word_id = embedding.word_id
+            tf = 1.0
+
+            if word_id >= 0:  # miniCOIL starts with ID 1
+                embedding_values = embedding.embedding
+                normalized_embedding = self.normalize_vector(embedding_values)
+
+                for val_id, value in enumerate(normalized_embedding):
+                    indices.append(
+                        word_id * embedding_size + val_id
+                    )  # since miniCOIL IDs start with 1
+                    values.append(value * tf)
+            else:
+                indices.append(self.unkn_word_token_id(embedding.word, unknown_words_shift))
+                values.append(tf)
+
+        return SparseEmbedding(
+            indices=np.array(indices, dtype=np.int32),
+            values=np.array(values, dtype=np.float32),
+        )

fastembed/sparse/utils/vocab_resolver.py

@@ -0,0 +1,202 @@
+from collections import defaultdict
+from typing import Iterable
+
+from py_rust_stemmers import SnowballStemmer
+import numpy as np
+from tokenizers import Tokenizer
+from numpy.typing import NDArray
+
+from fastembed.common.types import NumpyArray
+
+
+class VocabTokenizerBase:
+    def tokenize(self, sentence: str) -> NumpyArray:
+        raise NotImplementedError()
+
+    def convert_ids_to_tokens(self, token_ids: NumpyArray) -> list[str]:
+        raise NotImplementedError()
+
+
+class VocabTokenizer(VocabTokenizerBase):
+    def __init__(self, tokenizer: Tokenizer):
+        self.tokenizer = tokenizer
+
+    def tokenize(self, sentence: str) -> NumpyArray:
+        return np.array(self.tokenizer.encode(sentence).ids)
+
+    def convert_ids_to_tokens(self, token_ids: NumpyArray) -> list[str]:
+        return [self.tokenizer.id_to_token(token_id) for token_id in token_ids]
+
+
+class VocabResolver:
+    def __init__(self, tokenizer: VocabTokenizerBase, stopwords: set[str], stemmer: SnowballStemmer):
+        # Word to id mapping
+        self.vocab: dict[str, int] = {}
+        # Id to word mapping
+        self.words: list[str] = []
+        # Lemma to word mapping
+        self.stem_mapping: dict[str, str] = {}
+        self.tokenizer: VocabTokenizerBase = tokenizer
+        self.stemmer = stemmer
+        self.stopwords: set[str] = stopwords
+
+    def tokenize(self, sentence: str) -> NumpyArray:
+        return self.tokenizer.tokenize(sentence)
+
+    def lookup_word(self, word_id: int) -> str:
+        if word_id == 0:
+            return "UNK"
+        return self.words[word_id - 1]
+
+    def convert_ids_to_tokens(self, token_ids: NumpyArray) -> list[str]:
+        return self.tokenizer.convert_ids_to_tokens(token_ids)
+
+    def vocab_size(self) -> int:
+        # We need +1 for UNK token
+        return len(self.vocab) + 1
+
+    def save_vocab(self, path: str) -> None:
+        with open(path, "w") as f:
+            for word in self.words:
+                f.write(word + "\n")
+
+    def save_json_vocab(self, path: str) -> None:
+        import json
+
+        with open(path, "w") as f:
+            json.dump({"vocab": self.words, "stem_mapping": self.stem_mapping}, f, indent=2)
+
+    def load_json_vocab(self, path: str) -> None:
+        import json
+
+        with open(path, "r") as f:
+            data = json.load(f)
+            self.words = data["vocab"]
+            self.vocab = {word: idx + 1 for idx, word in enumerate(self.words)}
+            self.stem_mapping = data["stem_mapping"]
+
+    def add_word(self, word: str) -> None:
+        if word not in self.vocab:
+            self.vocab[word] = len(self.vocab) + 1
+            self.words.append(word)
+            stem = self.stemmer.stem_word(word)
+            if stem not in self.stem_mapping:
+                self.stem_mapping[stem] = word
+            else:
+                existing_word = self.stem_mapping[stem]
+                if len(existing_word) > len(word):
+                    # Prefer shorter words for the same stem
+                    # Example: "swim" is preferred over "swimming"
+                    self.stem_mapping[stem] = word
+
+    def load_vocab(self, path: str) -> None:
+        with open(path, "r") as f:
+            for line in f:
+                self.add_word(line.strip())
+
+    @classmethod
+    def _reconstruct_bpe(
+        cls, bpe_tokens: Iterable[tuple[int, str]]
+    ) -> list[tuple[str, list[int]]]:
+        result: list[tuple[str, list[int]]] = []
+        acc: str = ""
+        acc_idx: list[int] = []
+
+        continuing_subword_prefix = "##"
+        continuing_subword_prefix_len = len(continuing_subword_prefix)
+
+        for idx, token in bpe_tokens:
+            if token.startswith(continuing_subword_prefix):
+                acc += token[continuing_subword_prefix_len:]
+                acc_idx.append(idx)
+            else:
+                if acc:
+                    result.append((acc, acc_idx))
+                    acc_idx = []
+                acc = token
+                acc_idx.append(idx)
+
+        if acc:
+            result.append((acc, acc_idx))
+        return result
+
+    def resolve_tokens(
+        self, token_ids: NDArray[np.int64]
+    ) -> tuple[NDArray[np.int64], dict[int, int], dict[str, int], dict[str, list[str]]]:
+        """
+        Mark known tokens (including composed tokens) with vocab ids.
+
+        Args:
+            token_ids: (seq_len) - list of ids of tokens
+                Example:
+                    [
+                        101,  3897, 19332, 12718, 23348,
+                        1010,  1996,  7151,  2296, 4845,
+                        2359,  2005,  4234,  1010,  4332,
+                        2871,  3191,  2062, 102
+                    ]
+
+            returns:
+                - token_ids with vocab ids
+                    [
+                        0,  151, 151, 0, 0,
+                        912,  0,  0,  0, 332,
+                        332,  332,  0,  7121,  191,
+                        0,  0,  332, 0
+                    ]
+                - counts of each token
+                    {
+                        151: 1,
+                        332: 3,
+                        7121: 1,
+                        191: 1,
+                        912: 1
+                    }
+                - oov counts of each token
+                    {
+                        "the": 1,
+                        "a": 1,
+                        "[CLS]": 1,
+                        "[SEP]": 1,
+                        ...
+                    }
+                - forms of each token
+                    {
+                        "hello": ["hello"],
+                        "world": ["worlds", "world", "worlding"],
+                    }
+
+        """
+        tokens = self.convert_ids_to_tokens(token_ids)
+        tokens_mapping = self._reconstruct_bpe(enumerate(tokens))
+
+        counts: dict[int, int] = defaultdict(int)
+        oov_count: dict[str, int] = defaultdict(int)
+
+        forms: dict[str, list[str]] = defaultdict(list)
+
+        for token, mapped_token_ids in tokens_mapping:
+            vocab_id = 0
+            if token in self.stopwords:
+                vocab_id = 0
+            elif token in self.vocab:
+                vocab_id = self.vocab[token]
+                forms[token].append(token)
+            elif token in self.stem_mapping:
+                vocab_id = self.vocab[self.stem_mapping[token]]
+                forms[self.stem_mapping[token]].append(token)
+            else:
+                stem = self.stemmer.stem_word(token)
+                if stem in self.stem_mapping:
+                    vocab_id = self.vocab[self.stem_mapping[stem]]
+                    forms[self.stem_mapping[stem]].append(token)
+
+            for token_id in mapped_token_ids:
+                token_ids[token_id] = vocab_id
+
+            if vocab_id == 0:
+                oov_count[token] += 1
+            else:
+                counts[vocab_id] += 1
+        return token_ids, counts, oov_count, forms
+

fastembed/text/builtin_sentence_embedding.py

@@ -0,0 +1,69 @@
+from typing import Any, Iterable, Type
+
+
+from fastembed.common.types import NumpyArray
+from fastembed.common.onnx_model import OnnxOutputContext
+from fastembed.text.onnx_embedding import OnnxTextEmbedding, OnnxTextEmbeddingWorker
+from fastembed.common.model_description import DenseModelDescription, ModelSource
+
+
+supported_builtin_sentence_embedding_models: list[DenseModelDescription] = [
+    DenseModelDescription(
+        model="google/embeddinggemma-300m",
+        dim=768,
+        description=(
+            "Text embeddings, Unimodal (text), multilingual, 2048 input tokens truncation, "
+            "Prefixes for queries/documents: `task: search result | query: {content}` for query, "
+            "`title: {title | 'none'} | text: {content}` for documents, 2025 year."
+        ),
+        license="gemma",
+        size_in_GB=1.24,
+        sources=ModelSource(
+            hf="onnx-community/embeddinggemma-300m-ONNX",
+        ),
+        model_file="onnx/model.onnx",
+        additional_files=["onnx/model.onnx_data"],
+    ),
+]
+
+
+class BuiltinSentenceEmbedding(OnnxTextEmbedding):
+    """Builtin Sentence Embedding uses built-in pooling and normalization of underlying onnx models"""
+
+    @classmethod
+    def _get_worker_class(cls) -> Type[OnnxTextEmbeddingWorker]:
+        return BuiltinSentenceEmbeddingWorker
+
+    @classmethod
+    def _list_supported_models(cls) -> list[DenseModelDescription]:
+        """Lists the supported models.
+
+        Returns:
+            list[DenseModelDescription]: A list of DenseModelDescription objects containing the model information.
+        """
+        return supported_builtin_sentence_embedding_models
+
+    def _post_process_onnx_output(
+        self, output: OnnxOutputContext, **kwargs: Any
+    ) -> Iterable[NumpyArray]:
+        return output.model_output
+
+    def _run_model(
+        self, onnx_input: dict[str, Any], onnx_output_names: list[str] | None = None
+    ) -> NumpyArray:
+        return self.model.run(onnx_output_names, onnx_input)[1]  # type: ignore[union-attr]
+
+
+class BuiltinSentenceEmbeddingWorker(OnnxTextEmbeddingWorker):
+    def init_embedding(
+        self,
+        model_name: str,
+        cache_dir: str,
+        **kwargs: Any,
+    ) -> OnnxTextEmbedding:
+        return BuiltinSentenceEmbedding(
+            model_name=model_name,
+            cache_dir=cache_dir,
+            threads=1,
+            **kwargs,
+        )

fastembed/text/clip_embedding.py

@@ -35,7 +35,9 @@ class CLIPOnnxEmbedding(OnnxTextEmbedding):
         """
         return supported_clip_models
 
-    def _post_process_onnx_output(self, output: OnnxOutputContext) -> Iterable[NumpyArray]:
+    def _post_process_onnx_output(
+        self, output: OnnxOutputContext, **kwargs: Any
+    ) -> Iterable[NumpyArray]:
         return output.model_output
 
 

fastembed/text/custom_text_embedding.py

@@ -0,0 +1,144 @@
+from typing import Sequence, Any, Iterable, Type
+from dataclasses import dataclass
+
+import numpy as np
+from numpy.typing import NDArray
+
+from fastembed.common import OnnxProvider
+from fastembed.common.model_description import (
+    PoolingType,
+    DenseModelDescription,
+)
+from fastembed.common.onnx_model import OnnxOutputContext
+from fastembed.common.types import NumpyArray, Device
+from fastembed.common.utils import normalize, mean_pooling, last_token_pooling
+from fastembed.text.onnx_embedding import OnnxTextEmbedding
+from fastembed.text.onnx_text_model import TextEmbeddingWorker
+
+
+@dataclass(frozen=True)
+class PostprocessingConfig:
+    pooling: PoolingType
+    normalization: bool
+
+
+class CustomTextEmbedding(OnnxTextEmbedding):
+    SUPPORTED_MODELS: list[DenseModelDescription] = []
+    POSTPROCESSING_MAPPING: dict[str, PostprocessingConfig] = {}
+
+    def __init__(
+        self,
+        model_name: str,
+        cache_dir: str | None = None,
+        threads: int | None = None,
+        providers: Sequence[OnnxProvider] | None = None,
+        cuda: bool | Device = Device.AUTO,
+        device_ids: list[int] | None = None,
+        lazy_load: bool = False,
+        device_id: int | None = None,
+        specific_model_path: str | None = 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,
+        )
+        postprocessing_config = self.POSTPROCESSING_MAPPING[self.model_description.model]
+        self._pooling = postprocessing_config.pooling
+        self._normalization = postprocessing_config.normalization
+
+    @classmethod
+    def _list_supported_models(cls) -> list[DenseModelDescription]:
+        return cls.SUPPORTED_MODELS
+
+    @classmethod
+    def _get_worker_class(cls) -> Type["TextEmbeddingWorker[NumpyArray]"]:
+        return CustomTextEmbeddingWorker
+
+    def _get_worker_init_kwargs(self) -> dict[str, Any]:
+        return {
+            "model_description": self.model_description,
+            "postprocessing_config": self.POSTPROCESSING_MAPPING[self.model_description.model],
+        }
+
+    def _post_process_onnx_output(
+        self, output: OnnxOutputContext, **kwargs: Any
+    ) -> Iterable[NumpyArray]:
+        return self._normalize(self._pool(output.model_output, output.attention_mask))
+
+    def _pool(
+        self, embeddings: NumpyArray, attention_mask: NDArray[np.int64] | None = None
+    ) -> NumpyArray:
+        if self._pooling == PoolingType.CLS:
+            return embeddings[:, 0]
+
+        if self._pooling == PoolingType.MEAN:
+            if attention_mask is None:
+                raise ValueError("attention_mask must be provided for mean pooling")
+            return mean_pooling(embeddings, attention_mask)
+
+        if self._pooling == PoolingType.LAST_TOKEN:
+            if attention_mask is None:
+                raise ValueError("attention_mask must be provided for last token pooling")
+            return last_token_pooling(embeddings, attention_mask)
+
+        if self._pooling == PoolingType.DISABLED:
+            return embeddings
+
+        raise ValueError(
+            f"Unsupported pooling type {self._pooling}. "
+            f"Supported types are: {PoolingType.CLS}, {PoolingType.MEAN}, "
+            f"{PoolingType.LAST_TOKEN}, {PoolingType.DISABLED}."
+        )
+
+    def _normalize(self, embeddings: NumpyArray) -> NumpyArray:
+        return normalize(embeddings) if self._normalization else embeddings
+
+    @classmethod
+    def add_model(
+        cls,
+        model_description: DenseModelDescription,
+        pooling: PoolingType,
+        normalization: bool,
+    ) -> None:
+        cls.SUPPORTED_MODELS.append(model_description)
+        cls.POSTPROCESSING_MAPPING[model_description.model] = PostprocessingConfig(
+            pooling=pooling, normalization=normalization
+        )
+
+
+class CustomTextEmbeddingWorker(TextEmbeddingWorker[NumpyArray]):
+    def init_embedding(
+        self,
+        model_name: str,
+        cache_dir: str,
+        model_description: DenseModelDescription | None = None,
+        postprocessing_config: PostprocessingConfig | None = None,
+        **kwargs: Any,
+    ) -> CustomTextEmbedding:
+        if model_description is None or postprocessing_config is None:
+            raise ValueError(
+                "`model_description` and `postprocessing_config` are required to initialize a "
+                "custom model in a worker process, they are provided by "
+                "`CustomTextEmbedding._get_worker_init_kwargs`"
+            )
+        # custom models live in a class-level registry, which spawned workers don't inherit
+        CustomTextEmbedding.add_model(
+            model_description,
+            pooling=postprocessing_config.pooling,
+            normalization=postprocessing_config.normalization,
+        )
+        return CustomTextEmbedding(
+            model_name=model_name,
+            cache_dir=cache_dir,
+            threads=1,
+            **kwargs,
+        )

fastembed/text/last_token_normalized_embedding.py

@@ -0,0 +1,94 @@
+from typing import Any, Iterable, Type
+
+import onnxruntime as ort
+
+from fastembed.common.types import NumpyArray
+from fastembed.common.onnx_model import OnnxOutputContext
+from fastembed.common.utils import last_token_pooling, normalize
+from fastembed.text.onnx_embedding import OnnxTextEmbedding, OnnxTextEmbeddingWorker
+from fastembed.common.model_description import DenseModelDescription, ModelSource
+
+supported_last_token_normalized_models: list[DenseModelDescription] = [
+    DenseModelDescription(
+        model="Qwen/Qwen3-Embedding-0.6B",
+        dim=1024,
+        description=(
+            "Text embeddings, Unimodal (text), multilingual, 32768 input tokens truncation, "
+            "Prefixes for queries/documents: `Instruct: {task_description}\\nQuery:{query}` "
+            "for queries, none for documents, 2025 year."
+        ),
+        license="apache-2.0",
+        size_in_GB=2.38,
+        sources=ModelSource(hf="Qdrant/Qwen3-Embedding-0.6B-onnx"),
+        model_file="onnx/model.onnx",
+        additional_files=["onnx/model.onnx_data"],
+    ),
+    DenseModelDescription(
+        model="Qwen/Qwen3-Embedding-0.6B-Q",
+        dim=1024,
+        description=(
+            "Text embeddings, Unimodal (text), multilingual, 32768 input tokens truncation, "
+            "Prefixes for queries/documents: `Instruct: {task_description}\\nQuery:{query}` "
+            "for queries, none for documents, int8 weights, requires onnxruntime>=1.23, "
+            "2025 year."
+        ),
+        license="apache-2.0",
+        size_in_GB=1.12,
+        sources=ModelSource(hf="Qdrant/Qwen3-Embedding-0.6B-onnx"),
+        model_file="onnx/model_quantized.onnx",
+    ),
+]
+
+
+class LastTokenNormalizedEmbedding(OnnxTextEmbedding):
+    """Decoder-based embedding models, which pool the last non-padding token and normalize it"""
+
+    @classmethod
+    def _get_worker_class(cls) -> Type[OnnxTextEmbeddingWorker]:
+        return LastTokenNormalizedEmbeddingWorker
+
+    def load_onnx_model(self) -> None:
+        try:
+            super().load_onnx_model()
+        except Exception as e:
+            # int8 weights are stored as 8-bit MatMulNBits, which onnxruntime only
+            # implements since 1.23; older versions fail with "nbits_ == 4 was false"
+            if "nbits" not in str(e).lower():
+                raise
+            raise RuntimeError(
+                f"Could not load {self.model_name}: its int8 weights require "
+                f"onnxruntime>=1.23, but onnxruntime {ort.__version__} is installed. "
+                f"Either upgrade onnxruntime or use a non-quantized model."
+            ) from e
+
+    @classmethod
+    def _list_supported_models(cls) -> list[DenseModelDescription]:
+        """Lists the supported models.
+
+        Returns:
+            list[DenseModelDescription]: A list of DenseModelDescription objects containing the model information.
+        """
+        return supported_last_token_normalized_models
+
+    def _post_process_onnx_output(
+        self, output: OnnxOutputContext, **kwargs: Any
+    ) -> Iterable[NumpyArray]:
+        if output.attention_mask is None:
+            raise ValueError("attention_mask must be provided for last token pooling")
+
+        return normalize(last_token_pooling(output.model_output, output.attention_mask))
+
+
+class LastTokenNormalizedEmbeddingWorker(OnnxTextEmbeddingWorker):
+    def init_embedding(
+        self,
+        model_name: str,
+        cache_dir: str,
+        **kwargs: Any,
+    ) -> OnnxTextEmbedding:
+        return LastTokenNormalizedEmbedding(
+            model_name=model_name,
+            cache_dir=cache_dir,
+            threads=1,
+            **kwargs,
+        )

fastembed/text/multitask_embedding.py

@@ -1,8 +1,9 @@
 from enum import Enum
-from typing import Any, Type, Iterable, Union, Optional
+from typing import Any, Type, Iterable
 
 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 +45,9 @@ 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: int | None = None, **kwargs: Any):
         super().__init__(*args, **kwargs)
-        self.current_task_id: Union[Task, int] = self.PASSAGE_TASK
+        self.default_task_id: Task | int = task_id if task_id is not None else self.PASSAGE_TASK
 
     @classmethod
     def _get_worker_class(cls) -> Type[OnnxTextEmbeddingWorker]:
@@ -57,30 +58,34 @@ 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: int | Task | None = 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(
         self,
-        documents: Union[str, Iterable[str]],
+        documents: str | Iterable[str],
         batch_size: int = 256,
-        parallel: Optional[int] = None,
-        task_id: int = PASSAGE_TASK,
+        parallel: int | None = None,
+        task_id: int | None = 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)
+    def query_embed(self, query: str | Iterable[str], **kwargs: Any) -> Iterable[NumpyArray]:
+        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 +95,15 @@ 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/text/onnx_embedding.py

@@ -1,12 +1,11 @@
-from typing import Any, Iterable, Optional, Sequence, Type, Union
+from typing import Any, Iterable, Sequence, Type
 
-import numpy as np
-from fastembed.common.types import NumpyArray, OnnxProvider
+from fastembed.common.model_description import DenseModelDescription, ModelSource
 from fastembed.common.onnx_model import OnnxOutputContext
+from fastembed.common.types import Device, NumpyArray, OnnxProvider
 from fastembed.common.utils import define_cache_dir, normalize
 from fastembed.text.onnx_text_model import OnnxTextModel, TextEmbeddingWorker
 from fastembed.text.text_embedding_base import TextEmbeddingBase
-from fastembed.common.model_description import DenseModelDescription, ModelSource
 
 supported_onnx_models: list[DenseModelDescription] = [
     DenseModelDescription(
@@ -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",
     ),
@@ -34,8 +34,9 @@ supported_onnx_models: list[DenseModelDescription] = [
         license="mit",
         size_in_GB=0.21,
         sources=ModelSource(
-            hf="qdrant/bge-base-en-v1.5-onnx-q",
+            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",
     ),
@@ -75,7 +77,7 @@ supported_onnx_models: list[DenseModelDescription] = [
         ),
         license="mit",
         size_in_GB=0.067,
-        sources=ModelSource(hf="qdrant/bge-small-en-v1.5-onnx-q"),
+        sources=ModelSource(hf="Qdrant/bge-small-en-v1.5-onnx-Q"),
         model_file="model_optimized.onnx",
     ),
     DenseModelDescription(
@@ -90,21 +92,10 @@ 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",
     ),
-    DenseModelDescription(
-        model="thenlper/gte-large",
-        dim=1024,
-        description=(
-            "Text embeddings, Unimodal (text), English, 512 input tokens truncation, "
-            "Prefixes for queries/documents: not necessary, 2023 year."
-        ),
-        license="mit",
-        size_in_GB=1.20,
-        sources=ModelSource(hf="qdrant/gte-large-onnx"),
-        model_file="model.onnx",
-    ),
     DenseModelDescription(
         model="mixedbread-ai/mxbai-embed-large-v1",
         dim=1024,
@@ -189,6 +180,42 @@ supported_onnx_models: list[DenseModelDescription] = [
         sources=ModelSource(hf="jinaai/jina-clip-v1"),
         model_file="onnx/text_model.onnx",
     ),
+    DenseModelDescription(
+        model="minishlab/potion-base-8M",
+        dim=256,
+        description=(
+            "Text embeddings, Unimodal (text), English, 512 input tokens truncation, "
+            "Prefixes for queries/documents: not necessary, 2024 year."
+        ),
+        license="mit",
+        size_in_GB=0.030,
+        sources=ModelSource(hf="minishlab/potion-base-8m-onnx"),
+        model_file="model.onnx",
+    ),
+    DenseModelDescription(
+        model="minishlab/potion-retrieval-32M",
+        dim=512,
+        description=(
+            "Text embeddings, Unimodal (text), English, 512 input tokens truncation, "
+            "Prefixes for queries/documents: not necessary, 2025 year."
+        ),
+        license="mit",
+        size_in_GB=0.129,
+        sources=ModelSource(hf="minishlab/potion-retrieval-32m-onnx"),
+        model_file="model.onnx",
+    ),
+    DenseModelDescription(
+        model="minishlab/potion-multilingual-128M",
+        dim=256,
+        description=(
+            "Text embeddings, Unimodal (text), Multilingual, 512 input tokens truncation, "
+            "Prefixes for queries/documents: not necessary, 2025 year."
+        ),
+        license="mit",
+        size_in_GB=0.512,
+        sources=ModelSource(hf="minishlab/potion-multilingual-128m-onnx"),
+        model_file="model.onnx",
+    ),
 ]
 
 
@@ -208,14 +235,14 @@ class OnnxTextEmbedding(TextEmbeddingBase, OnnxTextModel[NumpyArray]):
     def __init__(
         self,
         model_name: str = "BAAI/bge-small-en-v1.5",
-        cache_dir: Optional[str] = None,
-        threads: Optional[int] = None,
-        providers: Optional[Sequence[OnnxProvider]] = None,
-        cuda: bool = False,
-        device_ids: Optional[list[int]] = None,
+        cache_dir: str | None = None,
+        threads: int | None = None,
+        providers: Sequence[OnnxProvider] | None = None,
+        cuda: bool | Device = Device.AUTO,
+        device_ids: list[int] | None = None,
         lazy_load: bool = False,
-        device_id: Optional[int] = None,
-        specific_model_path: Optional[str] = None,
+        device_id: int | None = None,
+        specific_model_path: str | None = None,
         **kwargs: Any,
     ):
         """
@@ -227,10 +254,11 @@ class OnnxTextEmbedding(TextEmbeddingBase, OnnxTextModel[NumpyArray]):
             threads (int, optional): The number of threads single onnxruntime session can use. Defaults to None.
             providers (Optional[Sequence[OnnxProvider]], optional): The list of onnxruntime providers to use.
                 Mutually exclusive with the `cuda` and `device_ids` arguments. Defaults to None.
-            cuda (bool, optional): Whether to use cuda for inference. Mutually exclusive with `providers`
-                Defaults to False.
+            cuda (Union[bool, Device], optional): Whether to use cuda for inference. Mutually exclusive with `providers`
+                Defaults to Device.AUTO.
             device_ids (Optional[list[int]], optional): The list of device ids to use for data parallel processing in
-                workers. Should be used with `cuda=True`, mutually exclusive with `providers`. Defaults to None.
+                workers. Should be used with `cuda` equals to `True`, `Device.AUTO` or `Device.CUDA`, mutually exclusive
+                with `providers`. Defaults to None.
             lazy_load (bool, optional): Whether to load the model during class initialization or on demand.
                 Should be set to True when using multiple-gpu and parallel encoding. Defaults to False.
             device_id (Optional[int], optional): The device id to use for loading the model in the worker process.
@@ -242,13 +270,13 @@ class OnnxTextEmbedding(TextEmbeddingBase, OnnxTextModel[NumpyArray]):
         super().__init__(model_name, cache_dir, threads, **kwargs)
         self.providers = providers
         self.lazy_load = lazy_load
-
+        self._extra_session_options = self._select_exposed_session_options(kwargs)
         # List of device ids, that can be used for data parallel processing in workers
         self.device_ids = device_ids
         self.cuda = cuda
 
         # This device_id will be used if we need to load model in current process
-        self.device_id: Optional[int] = None
+        self.device_id: int | None = None
         if device_id is not None:
             self.device_id = device_id
         elif self.device_ids is not None:
@@ -256,11 +284,12 @@ class OnnxTextEmbedding(TextEmbeddingBase, OnnxTextModel[NumpyArray]):
 
         self.model_description = self._get_model_description(model_name)
         self.cache_dir = str(define_cache_dir(cache_dir))
+        self._specific_model_path = specific_model_path
         self._model_dir = self.download_model(
             self.model_description,
             self.cache_dir,
             local_files_only=self._local_files_only,
-            specific_model_path=specific_model_path,
+            specific_model_path=self._specific_model_path,
         )
 
         if not self.lazy_load:
@@ -268,9 +297,9 @@ class OnnxTextEmbedding(TextEmbeddingBase, OnnxTextModel[NumpyArray]):
 
     def embed(
         self,
-        documents: Union[str, Iterable[str]],
+        documents: str | Iterable[str],
         batch_size: int = 256,
-        parallel: Optional[int] = None,
+        parallel: int | None = None,
         **kwargs: Any,
     ) -> Iterable[NumpyArray]:
         """
@@ -297,6 +326,9 @@ class OnnxTextEmbedding(TextEmbeddingBase, OnnxTextModel[NumpyArray]):
             providers=self.providers,
             cuda=self.cuda,
             device_ids=self.device_ids,
+            local_files_only=self._local_files_only,
+            specific_model_path=self._specific_model_path,
+            extra_session_options=self._extra_session_options,
             **kwargs,
         )
 
@@ -312,15 +344,18 @@ class OnnxTextEmbedding(TextEmbeddingBase, OnnxTextModel[NumpyArray]):
         """
         return onnx_input
 
-    def _post_process_onnx_output(self, output: OnnxOutputContext) -> Iterable[NumpyArray]:
+    def _post_process_onnx_output(
+        self, output: OnnxOutputContext, **kwargs: Any
+    ) -> Iterable[NumpyArray]:
         embeddings = output.model_output
+
         if embeddings.ndim == 3:  # (batch_size, seq_len, embedding_dim)
             processed_embeddings = embeddings[:, 0]
         elif embeddings.ndim == 2:  # (batch_size, embedding_dim)
             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(
@@ -330,8 +365,14 @@ class OnnxTextEmbedding(TextEmbeddingBase, OnnxTextModel[NumpyArray]):
             providers=self.providers,
             cuda=self.cuda,
             device_id=self.device_id,
+            extra_session_options=self._extra_session_options,
         )
 
+    def token_count(
+        self, texts: str | Iterable[str], batch_size: int = 1024, **kwargs: Any
+    ) -> int:
+        return self._token_count(texts, batch_size=batch_size, **kwargs)
+
 
 class OnnxTextEmbeddingWorker(TextEmbeddingWorker[NumpyArray]):
     def init_embedding(

fastembed/text/onnx_text_model.py

@@ -1,13 +1,13 @@
 import os
 from multiprocessing import get_all_start_methods
 from pathlib import Path
-from typing import Any, Iterable, Optional, Sequence, Type, Union
+from typing import Any, Iterable, Sequence, Type
 
 import numpy as np
 from numpy.typing import NDArray
 from tokenizers import Encoding, Tokenizer
 
-from fastembed.common.types import NumpyArray, OnnxProvider
+from fastembed.common.types import NumpyArray, OnnxProvider, Device
 from fastembed.common.onnx_model import EmbeddingWorker, OnnxModel, OnnxOutputContext, T
 from fastembed.common.preprocessor_utils import load_tokenizer
 from fastembed.common.utils import iter_batch
@@ -15,23 +15,32 @@ from fastembed.parallel_processor import ParallelWorkerPool
 
 
 class OnnxTextModel(OnnxModel[T]):
-    ONNX_OUTPUT_NAMES: Optional[list[str]] = None
+    ONNX_OUTPUT_NAMES: list[str] | None = None
 
     @classmethod
     def _get_worker_class(cls) -> Type["TextEmbeddingWorker[T]"]:
         raise NotImplementedError("Subclasses must implement this method")
 
-    def _post_process_onnx_output(self, output: OnnxOutputContext) -> Iterable[T]:
+    def _post_process_onnx_output(self, output: OnnxOutputContext, **kwargs: Any) -> Iterable[T]:
+        """Post-process the ONNX model output to convert it into a usable format.
+
+        Args:
+            output (OnnxOutputContext): The raw output from the ONNX model.
+            **kwargs: Additional keyword arguments that may be needed by specific implementations.
+
+        Returns:
+            Iterable[T]: Post-processed output as an iterable of type T.
+        """
         raise NotImplementedError("Subclasses must implement this method")
 
     def __init__(self) -> None:
         super().__init__()
-        self.tokenizer: Optional[Tokenizer] = None
+        self.tokenizer: Tokenizer | None = None
         self.special_token_to_id: dict[str, int] = {}
 
     def _preprocess_onnx_input(
         self, onnx_input: dict[str, NumpyArray], **kwargs: Any
-    ) -> dict[str, Union[NumpyArray, NDArray[np.int64]]]:
+    ) -> dict[str, NumpyArray | NDArray[np.int64]]:
         """
         Preprocess the onnx input.
         """
@@ -41,10 +50,11 @@ class OnnxTextModel(OnnxModel[T]):
         self,
         model_dir: Path,
         model_file: str,
-        threads: Optional[int],
-        providers: Optional[Sequence[OnnxProvider]] = None,
-        cuda: bool = False,
-        device_id: Optional[int] = None,
+        threads: int | None,
+        providers: Sequence[OnnxProvider] | None = None,
+        cuda: bool | Device = Device.AUTO,
+        device_id: int | None = None,
+        extra_session_options: dict[str, Any] | None = None,
     ) -> None:
         super()._load_onnx_model(
             model_dir=model_dir,
@@ -53,6 +63,7 @@ class OnnxTextModel(OnnxModel[T]):
             providers=providers,
             cuda=cuda,
             device_id=device_id,
+            extra_session_options=extra_session_options,
         )
         self.tokenizer, self.special_token_to_id = load_tokenizer(model_dir=model_dir)
 
@@ -81,24 +92,34 @@ class OnnxTextModel(OnnxModel[T]):
                 [np.zeros(len(e), dtype=np.int64) for e in input_ids], dtype=np.int64
             )
         onnx_input = self._preprocess_onnx_input(onnx_input, **kwargs)
+        model_output = self._run_model(
+            onnx_input=onnx_input, onnx_output_names=self.ONNX_OUTPUT_NAMES
+        )
 
-        model_output = self.model.run(self.ONNX_OUTPUT_NAMES, onnx_input)  # type: ignore[union-attr]
         return OnnxOutputContext(
-            model_output=model_output[0],
+            model_output=model_output,
             attention_mask=onnx_input.get("attention_mask", attention_mask),
             input_ids=onnx_input.get("input_ids", input_ids),
         )
 
+    def _run_model(
+        self, onnx_input: dict[str, Any], onnx_output_names: list[str] | None = None
+    ) -> NumpyArray:
+        return self.model.run(onnx_output_names, onnx_input)[0]  # type: ignore[union-attr]
+
     def _embed_documents(
         self,
         model_name: str,
         cache_dir: str,
-        documents: Union[str, Iterable[str]],
+        documents: str | Iterable[str],
         batch_size: int = 256,
-        parallel: Optional[int] = None,
-        providers: Optional[Sequence[OnnxProvider]] = None,
-        cuda: bool = False,
-        device_ids: Optional[list[int]] = None,
+        parallel: int | None = None,
+        providers: Sequence[OnnxProvider] | None = None,
+        cuda: bool | Device = Device.AUTO,
+        device_ids: list[int] | None = None,
+        local_files_only: bool = False,
+        specific_model_path: str | None = None,
+        extra_session_options: dict[str, Any] | None = None,
         **kwargs: Any,
     ) -> Iterable[T]:
         is_small = False
@@ -115,7 +136,9 @@ 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), **kwargs
+                )
         else:
             if parallel == 0:
                 parallel = os.cpu_count()
@@ -125,9 +148,15 @@ class OnnxTextModel(OnnxModel[T]):
                 "model_name": model_name,
                 "cache_dir": cache_dir,
                 "providers": providers,
+                "local_files_only": local_files_only,
+                "specific_model_path": specific_model_path,
                 **kwargs,
+                **self._get_worker_init_kwargs(),
             }
 
+            if extra_session_options is not None:
+                params.update(extra_session_options)
+
             pool = ParallelWorkerPool(
                 num_workers=parallel or 1,
                 worker=self._get_worker_class(),
@@ -136,7 +165,20 @@ class OnnxTextModel(OnnxModel[T]):
                 start_method=start_method,
             )
             for batch in pool.ordered_map(iter_batch(documents, batch_size), **params):
-                yield from self._post_process_onnx_output(batch)  # type: ignore
+                yield from self._post_process_onnx_output(batch, **kwargs)  # type: ignore
+
+    def _token_count(self, texts: str | Iterable[str], batch_size: int = 1024, **_: Any) -> int:
+        if not hasattr(self, "model") or self.model is None:
+            self.load_onnx_model()  # loads the tokenizer as well
+
+        token_num = 0
+        assert self.tokenizer is not None
+        texts = [texts] if isinstance(texts, str) else texts
+        for batch in iter_batch(texts, batch_size):
+            for tokens in self.tokenizer.encode_batch(batch):
+                token_num += sum(tokens.attention_mask)
+
+        return token_num
 
 
 class TextEmbeddingWorker(EmbeddingWorker[T]):

fastembed/text/pooled_embedding.py

@@ -1,9 +1,11 @@
 from typing import Any, Iterable, Type
 
 import numpy as np
+from numpy.typing import NDArray
 
 from fastembed.common.types import NumpyArray
 from fastembed.common.onnx_model import OnnxOutputContext
+from fastembed.common.utils import mean_pooling
 from fastembed.text.onnx_embedding import OnnxTextEmbedding, OnnxTextEmbeddingWorker
 from fastembed.common.model_description import DenseModelDescription, ModelSource
 
@@ -80,6 +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"],
@@ -93,16 +96,10 @@ class PooledEmbedding(OnnxTextEmbedding):
         return PooledEmbeddingWorker
 
     @classmethod
-    def mean_pooling(cls, model_output: NumpyArray, attention_mask: NumpyArray) -> NumpyArray:
-        token_embeddings = model_output.astype(np.float32)
-        attention_mask = attention_mask.astype(np.float32)
-        input_mask_expanded = np.expand_dims(attention_mask, axis=-1)
-        input_mask_expanded = np.tile(input_mask_expanded, (1, 1, token_embeddings.shape[-1]))
-        input_mask_expanded = input_mask_expanded.astype(np.float32)
-        sum_embeddings = np.sum(token_embeddings * input_mask_expanded, axis=1)
-        sum_mask = np.sum(input_mask_expanded, axis=1)
-        pooled_embeddings = sum_embeddings / np.maximum(sum_mask, 1e-9)
-        return pooled_embeddings
+    def mean_pooling(
+        cls, model_output: NumpyArray, attention_mask: NDArray[np.int64]
+    ) -> NumpyArray:
+        return mean_pooling(model_output, attention_mask)
 
     @classmethod
     def _list_supported_models(cls) -> list[DenseModelDescription]:
@@ -113,13 +110,15 @@ class PooledEmbedding(OnnxTextEmbedding):
         """
         return supported_pooled_models
 
-    def _post_process_onnx_output(self, output: OnnxOutputContext) -> Iterable[NumpyArray]:
+    def _post_process_onnx_output(
+        self, output: OnnxOutputContext, **kwargs: Any
+    ) -> Iterable[NumpyArray]:
         if output.attention_mask is None:
             raise ValueError("attention_mask must be provided for document post-processing")
 
         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/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",
     ),
@@ -57,9 +57,9 @@ supported_pooled_normalized_models: list[DenseModelDescription] = [
             "Prefixes for queries/documents: not necessary, 2024 year."
         ),
         license="apache-2.0",
-        size_in_GB=0.32,
+        size_in_GB=0.64,
         sources=ModelSource(hf="jinaai/jina-embeddings-v2-base-de"),
-        model_file="onnx/model_fp16.onnx",
+        model_file="onnx/model.onnx",
     ),
     DenseModelDescription(
         model="jinaai/jina-embeddings-v2-base-code",
@@ -109,6 +109,18 @@ supported_pooled_normalized_models: list[DenseModelDescription] = [
         sources=ModelSource(hf="thenlper/gte-base"),
         model_file="onnx/model.onnx",
     ),
+    DenseModelDescription(
+        model="thenlper/gte-large",
+        dim=1024,
+        description=(
+            "Text embeddings, Unimodal (text), English, 512 input tokens truncation, "
+            "Prefixes for queries/documents: not necessary, 2023 year."
+        ),
+        license="mit",
+        size_in_GB=1.20,
+        sources=ModelSource(hf="qdrant/gte-large-onnx"),
+        model_file="model.onnx",
+    ),
 ]
 
 
@@ -126,13 +138,15 @@ class PooledNormalizedEmbedding(PooledEmbedding):
         """
         return supported_pooled_normalized_models
 
-    def _post_process_onnx_output(self, output: OnnxOutputContext) -> Iterable[NumpyArray]:
+    def _post_process_onnx_output(
+        self, output: OnnxOutputContext, **kwargs: Any
+    ) -> Iterable[NumpyArray]:
         if output.attention_mask is None:
             raise ValueError("attention_mask must be provided for document post-processing")
 
         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/text/siglip_embedding.py

@@ -0,0 +1,71 @@
+from typing import Any, Type
+
+from fastembed.common.model_description import DenseModelDescription, ModelSource
+from fastembed.text.onnx_embedding import OnnxTextEmbedding, OnnxTextEmbeddingWorker
+
+supported_siglip_models: list[DenseModelDescription] = [
+    DenseModelDescription(
+        model="google/siglip2-base-patch16-224",
+        dim=768,
+        description=(
+            "Text embeddings, Multimodal (text&image), multilingual, 64 input tokens truncation, "
+            "Prefixes for queries/documents: not necessary, 2025 year"
+        ),
+        license="apache-2.0",
+        size_in_GB=1.13,
+        sources=ModelSource(hf="onnx-community/siglip2-base-patch16-224-ONNX"),
+        model_file="onnx/text_model.onnx",
+    ),
+]
+
+
+class SiglipOnnxTextEmbedding(OnnxTextEmbedding):
+    """SigLIP text tower.
+
+    SigLIP always pools the hidden state at the last sequence position, whether or not it is
+    padding, so every batch must be padded to the model's fixed training length (rather than to
+    the longest sequence in the batch) or the resulting embeddings become dependent on what else
+    is in the batch.
+
+    The exported graph also returns both `last_hidden_state` and `pooler_output`; only the
+    latter is the text embedding, so it must be selected explicitly.
+    """
+
+    ONNX_OUTPUT_NAMES = ["pooler_output"]
+
+    @classmethod
+    def _get_worker_class(cls) -> Type[OnnxTextEmbeddingWorker]:
+        return SiglipTextEmbeddingWorker
+
+    @classmethod
+    def _list_supported_models(cls) -> list[DenseModelDescription]:
+        return supported_siglip_models
+
+    def load_onnx_model(self) -> None:
+        super().load_onnx_model()
+        if self.tokenizer is not None:
+            truncation = self.tokenizer.truncation
+            padding = self.tokenizer.padding
+            if truncation and padding and padding.get("length") is None:
+                self.tokenizer.enable_padding(
+                    direction=padding["direction"],
+                    pad_id=padding["pad_id"],
+                    pad_type_id=padding["pad_type_id"],
+                    pad_token=padding["pad_token"],
+                    length=truncation["max_length"],
+                )
+
+
+class SiglipTextEmbeddingWorker(OnnxTextEmbeddingWorker):
+    def init_embedding(
+        self,
+        model_name: str,
+        cache_dir: str,
+        **kwargs: Any,
+    ) -> OnnxTextEmbedding:
+        return SiglipOnnxTextEmbedding(
+            model_name=model_name,
+            cache_dir=cache_dir,
+            threads=1,
+            **kwargs,
+        )

fastembed/text/text_embedding.py

@@ -1,24 +1,32 @@
 import warnings
-from typing import Any, Iterable, Optional, Sequence, Type, Union
+from typing import Any, Iterable, Sequence, Type
 from dataclasses import asdict
 
-from fastembed.common.types import NumpyArray, OnnxProvider
+from fastembed.common.types import NumpyArray, OnnxProvider, Device
 from fastembed.text.clip_embedding import CLIPOnnxEmbedding
+from fastembed.text.custom_text_embedding import CustomTextEmbedding
 from fastembed.text.pooled_normalized_embedding import PooledNormalizedEmbedding
 from fastembed.text.pooled_embedding import PooledEmbedding
 from fastembed.text.multitask_embedding import JinaEmbeddingV3
+from fastembed.text.builtin_sentence_embedding import BuiltinSentenceEmbedding
+from fastembed.text.last_token_normalized_embedding import LastTokenNormalizedEmbedding
 from fastembed.text.onnx_embedding import OnnxTextEmbedding
+from fastembed.text.siglip_embedding import SiglipOnnxTextEmbedding
 from fastembed.text.text_embedding_base import TextEmbeddingBase
-from fastembed.common.model_description import DenseModelDescription
+from fastembed.common.model_description import DenseModelDescription, ModelSource, PoolingType
 
 
 class TextEmbedding(TextEmbeddingBase):
     EMBEDDINGS_REGISTRY: list[Type[TextEmbeddingBase]] = [
         OnnxTextEmbedding,
         CLIPOnnxEmbedding,
+        SiglipOnnxTextEmbedding,
         PooledNormalizedEmbedding,
         PooledEmbedding,
         JinaEmbeddingV3,
+        BuiltinSentenceEmbedding,
+        LastTokenNormalizedEmbedding,
+        CustomTextEmbedding,
     ]
 
     @classmethod
@@ -37,43 +45,61 @@ class TextEmbedding(TextEmbeddingBase):
             result.extend(embedding._list_supported_models())
         return result
 
+    @classmethod
+    def add_custom_model(
+        cls,
+        model: str,
+        pooling: PoolingType,
+        normalization: bool,
+        sources: ModelSource,
+        dim: int,
+        model_file: str = "onnx/model.onnx",
+        description: str = "",
+        license: str = "",
+        size_in_gb: float = 0.0,
+        additional_files: list[str] | None = None,
+    ) -> None:
+        registered_models = cls._list_supported_models()
+        for registered_model in registered_models:
+            if model.lower() == registered_model.model.lower():
+                raise ValueError(
+                    f"Model {model} is already registered in TextEmbedding, if you still want to add this model, "
+                    f"please use another model name"
+                )
+
+        CustomTextEmbedding.add_model(
+            DenseModelDescription(
+                model=model,
+                sources=sources,
+                dim=dim,
+                model_file=model_file,
+                description=description,
+                license=license,
+                size_in_GB=size_in_gb,
+                additional_files=additional_files or [],
+            ),
+            pooling=pooling,
+            normalization=normalization,
+        )
+
     def __init__(
         self,
         model_name: str = "BAAI/bge-small-en-v1.5",
-        cache_dir: Optional[str] = None,
-        threads: Optional[int] = None,
-        providers: Optional[Sequence[OnnxProvider]] = None,
-        cuda: bool = False,
-        device_ids: Optional[list[int]] = None,
+        cache_dir: str | None = None,
+        threads: int | None = None,
+        providers: Sequence[OnnxProvider] | None = None,
+        cuda: bool | Device = Device.AUTO,
+        device_ids: list[int] | None = None,
         lazy_load: bool = False,
         **kwargs: Any,
     ):
         super().__init__(model_name, cache_dir, threads, **kwargs)
-        if model_name == "nomic-ai/nomic-embed-text-v1.5-Q":
+        if model_name.lower() == "jinaai/jina-embeddings-v2-base-de":
             warnings.warn(
-                "The model 'nomic-ai/nomic-embed-text-v1.5-Q' has been updated on HuggingFace. "
-                "Please review the latest documentation and release notes to ensure compatibility with your workflow. ",
+                "The model 'jinaai/jina-embeddings-v2-base-de' used to run with fp16 model, but due to onnxruntime updates, now it runs with the original fp32 model.",
                 UserWarning,
                 stacklevel=2,
             )
-        if model_name == "sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2":
-            warnings.warn(
-                "The model 'sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2' has been updated to "
-                "include a mean pooling layer. Please ensure your usage aligns with the new functionality. "
-                "Support for the previous version without mean pooling will be removed as of version 0.5.2.",
-                UserWarning,
-                stacklevel=2,
-            )
-        if model_name in {
-            "sentence-transformers/paraphrase-multilingual-mpnet-base-v2",
-            "intfloat/multilingual-e5-large",
-        }:
-            warnings.warn(
-                f"{model_name} has been updated as of fastembed 0.5.2, outputs are now average pooled.",
-                UserWarning,
-                stacklevel=2,
-            )
-
         for EMBEDDING_MODEL_TYPE in self.EMBEDDINGS_REGISTRY:
             supported_models = EMBEDDING_MODEL_TYPE._list_supported_models()
             if any(model_name.lower() == model.model.lower() for model in supported_models):
@@ -94,11 +120,45 @@ class TextEmbedding(TextEmbeddingBase):
             "Please check the supported models using `TextEmbedding.list_supported_models()`"
         )
 
+    @property
+    def embedding_size(self) -> int:
+        """Get the embedding size of the current model"""
+        if self._embedding_size is None:
+            self._embedding_size = self.get_embedding_size(self.model_name)
+        return self._embedding_size
+
+    @classmethod
+    def get_embedding_size(cls, model_name: str) -> int:
+        """Get the embedding size of the passed model
+
+        Args:
+            model_name (str): The name of the model to get embedding size for.
+
+        Returns:
+            int: The size of the embedding.
+
+        Raises:
+            ValueError: If the model name is not found in the supported models.
+        """
+        descriptions = cls._list_supported_models()
+        embedding_size: int | None = None
+        for description in descriptions:
+            if description.model.lower() == model_name.lower():
+                embedding_size = description.dim
+                break
+        if embedding_size is None:
+            model_names = [description.model for description in descriptions]
+            raise ValueError(
+                f"Embedding size for model {model_name} was None. "
+                f"Available model names: {model_names}"
+            )
+        return embedding_size
+
     def embed(
         self,
-        documents: Union[str, Iterable[str]],
+        documents: str | Iterable[str],
         batch_size: int = 256,
-        parallel: Optional[int] = None,
+        parallel: int | None = None,
         **kwargs: Any,
     ) -> Iterable[NumpyArray]:
         """
@@ -118,7 +178,7 @@ class TextEmbedding(TextEmbeddingBase):
         """
         yield from self.model.embed(documents, batch_size, parallel, **kwargs)
 
-    def query_embed(self, query: Union[str, Iterable[str]], **kwargs: Any) -> Iterable[NumpyArray]:
+    def query_embed(self, query: str | Iterable[str], **kwargs: Any) -> Iterable[NumpyArray]:
         """
         Embeds queries
 
@@ -144,3 +204,17 @@ class TextEmbedding(TextEmbeddingBase):
         """
         # This is model-specific, so that different models can have specialized implementations
         yield from self.model.passage_embed(texts, **kwargs)
+
+    def token_count(
+        self, texts: str | Iterable[str], batch_size: int = 1024, **kwargs: Any
+    ) -> int:
+        """Returns the number of tokens in the texts.
+
+        Args:
+            texts (str | Iterable[str]): The list of texts to embed.
+            batch_size (int): Batch size for encoding
+
+        Returns:
+            int: Sum of number of tokens in the texts.
+        """
+        return self.model.token_count(texts, batch_size=batch_size, **kwargs)

fastembed/text/text_embedding_base.py

@@ -1,4 +1,4 @@
-from typing import Iterable, Optional, Union, Any
+from typing import Iterable, Any
 
 from fastembed.common.model_description import DenseModelDescription
 from fastembed.common.types import NumpyArray
@@ -9,20 +9,21 @@ class TextEmbeddingBase(ModelManagement[DenseModelDescription]):
     def __init__(
         self,
         model_name: str,
-        cache_dir: Optional[str] = None,
-        threads: Optional[int] = None,
+        cache_dir: str | None = None,
+        threads: int | None = None,
         **kwargs: Any,
     ):
         self.model_name = model_name
         self.cache_dir = cache_dir
         self.threads = threads
         self._local_files_only = kwargs.pop("local_files_only", False)
+        self._embedding_size: int | None = None
 
     def embed(
         self,
-        documents: Union[str, Iterable[str]],
+        documents: str | Iterable[str],
         batch_size: int = 256,
-        parallel: Optional[int] = None,
+        parallel: int | None = None,
         **kwargs: Any,
     ) -> Iterable[NumpyArray]:
         raise NotImplementedError()
@@ -42,7 +43,7 @@ class TextEmbeddingBase(ModelManagement[DenseModelDescription]):
         # This is model-specific, so that different models can have specialized implementations
         yield from self.embed(texts, **kwargs)
 
-    def query_embed(self, query: Union[str, Iterable[str]], **kwargs: Any) -> Iterable[NumpyArray]:
+    def query_embed(self, query: str | Iterable[str], **kwargs: Any) -> Iterable[NumpyArray]:
         """
         Embeds queries
 
@@ -58,3 +59,17 @@ class TextEmbeddingBase(ModelManagement[DenseModelDescription]):
             yield from self.embed([query], **kwargs)
         else:
             yield from self.embed(query, **kwargs)
+
+    @classmethod
+    def get_embedding_size(cls, model_name: str) -> int:
+        """Returns embedding size of the passed model."""
+        raise NotImplementedError("Subclasses must implement this method")
+
+    @property
+    def embedding_size(self) -> int:
+        """Returns embedding size for the current model"""
+        raise NotImplementedError("Subclasses must implement this method")
+
+    def token_count(self, texts: str | Iterable[str], **kwargs: Any) -> int:
+        """Returns the number of tokens in the texts."""
+        raise NotImplementedError("Subclasses must implement this method")

mkdocs.yml

@@ -13,6 +13,7 @@ copyright: |
 theme:
   name: material
   logo: assets/favicon.png
+  favicon: assets/favicon.png
   custom_dir: docs/overrides
   icon:
     repo: fontawesome/brands/github

pyproject.toml

@@ -1,9 +1,9 @@
 [tool.poetry]
-name = "fastembed"
-version = "0.5.1"
+name = "fastembed-gpu"
+version = "0.8.1"
 description = "Fast, light, accurate library built for retrieval embedding generation"
 authors = ["Qdrant Team <info@qdrant.tech>", "NirantK <nirant.bits@gmail.com>"]
-license = "Apache License"
+license = "Apache-2.0"
 readme = "README.md"
 packages = [{include = "fastembed"}]
 homepage = "https://github.com/qdrant/fastembed"
@@ -11,46 +11,56 @@ repository = "https://github.com/qdrant/fastembed"
 keywords = ["vector", "embedding", "neural", "search", "qdrant", "sentence-transformers"]
 
 [tool.poetry.dependencies]
-python = ">=3.9.0"
+python = ">=3.10.0"
 numpy = [
-    { version = ">=1.21", python = ">=3.10,<3.12" },
-    { version = ">=1.26", python = ">=3.12,<3.13" },
-    { version = ">=2.1.0", python = ">=3.13" },
-    { version = ">=1.21,<2.1.0", python = "<3.10" },
+    { version = ">=1.21,<2.3.0", python = "3.10" },
+    { version = ">=1.21", python = "3.11" },
+    { version = ">=1.26", python = "3.12" },
+    { version = ">=2.1.0", python = "3.13" },
+    { version = ">=2.3.0", python = ">=3.14" },
 ]
-onnxruntime = [
-    { version = ">1.20.0", python = ">=3.13" },
-    { version = ">=1.17.0,<1.20.0", python = "<3.10" },
-    { version = ">=1.17.0,!=1.20.0", python = ">=3.10,<3.13" },
+onnxruntime-gpu = [
+    { version = ">=1.17.0,!=1.20.0,<1.24", python = "3.10" },
+    { version = ">=1.17.0,!=1.20.0,!=1.24.0,!=1.24.1", python = ">=3.11,<3.13" },
+    { version = ">1.21.0,!=1.24.0,!=1.24.1", python = "3.13" },
+    { version = ">=1.24.2", python = ">=3.14" },
 ]
 tqdm = "^4.66"
 requests = "^2.31"
 tokenizers = ">=0.15,<1.0"
-huggingface-hub = ">=0.20,<1.0"
+huggingface-hub = ">=0.20,<2.0"
 loguru = "^0.7.2"
-pillow = ">=10.3.0,<12.0.0"
-mmh3 = "^4.1.0"
+pillow = [
+    { version = ">=10.3.0,<13.0", python = ">=3.10,<3.13" },
+    { version = ">=11.0.0,<13.0", python = "3.13" },
+    { version = ">=12.0.0,<13.0", python = ">=3.14" },
+]
+mmh3 = ">=4.1.0,<6.0.0"
 py-rust-stemmers = "^0.1.0"
 
 [tool.poetry.group.test.dependencies]
-pytest = "^7.4.2"
+pytest = ">=7.4.2,<10.0.0"
 ruff = ">=0.3.1,<1.0"
 
 [tool.poetry.group.dev.dependencies]
 notebook = ">=7.0.2"
-pre-commit = "^3.6.2"
-onnx = ">=1.15.0"
+pre-commit = ">=3.6.2,<5.0.0"
+onnx = [
+    { version = ">=1.15.0", python = ">=3.10,<3.13" },
+    { version = ">=1.18.0", python = "3.13" },
+    { version = ">=1.20.0", python = ">=3.14" },
+]
 
 [tool.poetry.group.docs.dependencies]
 mkdocs-material = "^9.5.10"
-mkdocstrings = "^0.24.0"
-pillow = ">=10.3.0,<12.0.0"
+mkdocstrings = ">=0.24,<1.1"
+pillow = ">=10.3.0,<13.0.0"
 cairosvg = "^2.7.1"
 mknotebooks = "^0.8.0"
 
 [tool.poetry.group.types.dependencies]
 pyright = ">=1.1.293"
-mypy = "^1.0.0"
+mypy = ">=1,<3"
 
 [build-system]
 requires = ["poetry-core"]

tests/test_attention_embeddings.py

@@ -1,4 +1,5 @@
 import os
+from contextlib import contextmanager
 
 import numpy as np
 import pytest
@@ -7,98 +8,119 @@ from fastembed import SparseTextEmbedding
 from tests.utils import delete_model_cache
 
 
-@pytest.mark.parametrize("model_name", ["Qdrant/bm42-all-minilm-l6-v2-attentions", "Qdrant/bm25"])
-def test_attention_embeddings(model_name: str) -> None:
+_MODELS_TO_CACHE = ("Qdrant/bm42-all-minilm-l6-v2-attentions", "Qdrant/bm25")
+MODELS_TO_CACHE = tuple([x.lower() for x in _MODELS_TO_CACHE])
+
+
+@pytest.fixture(scope="module")
+def model_cache():
     is_ci = os.getenv("CI")
-    model = SparseTextEmbedding(model_name=model_name)
+    cache = {}
 
-    output = list(
-        model.query_embed(
-            [
-                "I must not fear. Fear is the mind-killer.",
-            ]
-        )
-    )
+    @contextmanager
+    def get_model(model_name: str):
+        lowercase_model_name = model_name.lower()
+        if lowercase_model_name not in cache:
+            cache[lowercase_model_name] = SparseTextEmbedding(lowercase_model_name)
+        yield cache[lowercase_model_name]
+        if lowercase_model_name not in MODELS_TO_CACHE:
+            print("deleting model")
+            model_inst = cache.pop(lowercase_model_name)
+            if is_ci:
+                delete_model_cache(model_inst.model._model_dir)
+            del model_inst
 
-    assert len(output) == 1
-
-    for result in output:
-        assert len(result.indices) == len(result.values)
-        assert np.allclose(result.values, np.ones(len(result.values)))
-
-    quotes = [
-        "I must not fear. Fear is the mind-killer.",
-        "All animals are equal, but some animals are more equal than others.",
-        "It was a pleasure to burn.",
-        "The sky above the port was the color of television, tuned to a dead channel.",
-        "In the beginning, the universe was created."
-        " This has made a lot of people very angry and been widely regarded as a bad move.",
-        "It's a truth universally acknowledged that a zombie in possession of brains must be in want of more brains.",
-        "War is peace. Freedom is slavery. Ignorance is strength.",
-        "We're not in Infinity; we're in the suburbs.",
-        "I was a thousand times more evil than thou!",
-        "History is merely a list of surprises... It can only prepare us to be surprised yet again.",
-        ".",  # Empty string
-    ]
-
-    output = list(model.embed(quotes))
-
-    assert len(output) == len(quotes)
-
-    for result in output[:-1]:
-        assert len(result.indices) == len(result.values)
-        assert len(result.indices) > 0
-
-    assert len(output[-1].indices) == 0
-
-    # Test support for unknown languages
-    output = list(
-        model.query_embed(
-            [
-                "привет мир!",
-            ]
-        )
-    )
-
-    assert len(output) == 1
-
-    for result in output:
-        assert len(result.indices) == len(result.values)
-        assert len(result.indices) == 2
+    yield get_model
 
     if is_ci:
-        delete_model_cache(model.model._model_dir)
+        for name, model in cache.items():
+            delete_model_cache(model.model._model_dir)
+    cache.clear()
 
 
 @pytest.mark.parametrize("model_name", ["Qdrant/bm42-all-minilm-l6-v2-attentions", "Qdrant/bm25"])
-def test_parallel_processing(model_name: str) -> None:
-    is_ci = os.getenv("CI")
+def test_attention_embeddings(model_cache, model_name: str) -> None:
+    with model_cache(model_name) as model:
+        output = list(
+            model.query_embed(
+                [
+                    "I must not fear. Fear is the mind-killer.",
+                ]
+            )
+        )
 
-    model = SparseTextEmbedding(model_name=model_name)
+        assert len(output) == 1
 
-    docs = ["hello world", "attention embedding", "Mangez-vous vraiment des grenouilles?"] * 100
-    embeddings = list(model.embed(docs, batch_size=10, parallel=2))
+        for result in output:
+            assert len(result.indices) == len(result.values)
+            assert np.allclose(result.values, np.ones(len(result.values)))
 
-    embeddings_2 = list(model.embed(docs, batch_size=10, parallel=None))
+        quotes = [
+            "I must not fear. Fear is the mind-killer.",
+            "All animals are equal, but some animals are more equal than others.",
+            "It was a pleasure to burn.",
+            "The sky above the port was the color of television, tuned to a dead channel.",
+            "In the beginning, the universe was created."
+            " This has made a lot of people very angry and been widely regarded as a bad move.",
+            "It's a truth universally acknowledged that a zombie in possession of brains must be in want of more brains.",
+            "War is peace. Freedom is slavery. Ignorance is strength.",
+            "We're not in Infinity; we're in the suburbs.",
+            "I was a thousand times more evil than thou!",
+            "History is merely a list of surprises... It can only prepare us to be surprised yet again.",
+            ".",  # Empty string
+        ]
 
-    embeddings_3 = list(model.embed(docs, batch_size=10, parallel=0))
+        output = list(model.embed(quotes))
 
-    assert len(embeddings) == len(docs)
+        assert len(output) == len(quotes)
 
-    for emb_1, emb_2, emb_3 in zip(embeddings, embeddings_2, embeddings_3):
-        assert np.allclose(emb_1.indices, emb_2.indices)
-        assert np.allclose(emb_1.indices, emb_3.indices)
-        assert np.allclose(emb_1.values, emb_2.values)
-        assert np.allclose(emb_1.values, emb_3.values)
+        for result in output[:-1]:
+            assert len(result.indices) == len(result.values)
+            assert len(result.indices) > 0
 
-    if is_ci:
-        delete_model_cache(model.model._model_dir)
+        assert len(output[-1].indices) == 0
+
+        # Test support for unknown languages
+        output = list(
+            model.query_embed(
+                [
+                    "привет мир!",
+                ]
+            )
+        )
+
+        assert len(output) == 1
+
+        for result in output:
+            assert len(result.indices) == len(result.values)
+            assert len(result.indices) == 2
+
+
+@pytest.mark.parametrize("model_name", ["Qdrant/bm42-all-minilm-l6-v2-attentions", "Qdrant/bm25"])
+def test_parallel_processing(model_cache, model_name: str) -> None:
+    with model_cache(model_name) as model:
+        docs = [
+            "hello world",
+            "attention embedding",
+            "Mangez-vous vraiment des grenouilles?",
+        ] * 100
+        embeddings = list(model.embed(docs, batch_size=10, parallel=2))
+
+        embeddings_2 = list(model.embed(docs, batch_size=10, parallel=None))
+
+        embeddings_3 = list(model.embed(docs, batch_size=10, parallel=0))
+
+        assert len(embeddings) == len(docs)
+
+        for emb_1, emb_2, emb_3 in zip(embeddings, embeddings_2, embeddings_3):
+            assert np.allclose(emb_1.indices, emb_2.indices)
+            assert np.allclose(emb_1.indices, emb_3.indices)
+            assert np.allclose(emb_1.values, emb_2.values)
+            assert np.allclose(emb_1.values, emb_3.values)
 
 
 @pytest.mark.parametrize("model_name", ["Qdrant/bm25"])
-def test_multilanguage(model_name: str) -> None:
-    is_ci = os.getenv("CI")
-
+def test_multilanguage(model_cache, model_name: str) -> None:
     docs = ["Mangez-vous vraiment des grenouilles?", "Je suis au lit"]
 
     model = SparseTextEmbedding(model_name=model_name, language="french")
@@ -109,39 +131,30 @@ def test_multilanguage(model_name: str) -> None:
     assert embeddings[1].values.shape == (1,)
     assert embeddings[1].indices.shape == (1,)
 
-    model = SparseTextEmbedding(model_name=model_name, language="english")
-    embeddings = list(model.embed(docs))[:2]
-    assert embeddings[0].values.shape == (5,)
-    assert embeddings[0].indices.shape == (5,)
+    with model_cache(model_name) as model:  # language = "english"
+        embeddings = list(model.embed(docs))[:2]
+        assert embeddings[0].values.shape == (5,)
+        assert embeddings[0].indices.shape == (5,)
 
-    assert embeddings[1].values.shape == (4,)
-    assert embeddings[1].indices.shape == (4,)
-
-    if is_ci:
-        delete_model_cache(model.model._model_dir)
+        assert embeddings[1].values.shape == (4,)
+        assert embeddings[1].indices.shape == (4,)
 
 
 @pytest.mark.parametrize("model_name", ["Qdrant/bm25"])
-def test_special_characters(model_name: str) -> None:
-    is_ci = os.getenv("CI")
-
-    docs = [
-        "Über den größten Flüssen Österreichs äußern sich Experten häufig: Öko-Systeme müssen geschützt werden!",
-        "L'élève français s'écrie : « Où est mon crayon ? J'ai besoin de finir cet exercice avant la récréation!",
-        "Într-o zi însorită, Ștefan și Ioana au mâncat mămăligă cu brânză și au băut țuică la cabană.",
-        "Üzgün öğretmen öğrencilere seslendi: Lütfen gürültü yapmayın, sınavınızı bitirmeye çalışıyorum!",
-        "Ο Ξενοφών είπε: «Ψάχνω για ένα ωραίο δώρο για τη γιαγιά μου. Ίσως ένα φυτό ή ένα βιβλίο;»",
-        "Hola! ¿Cómo estás? Estoy muy emocionado por el cumpleaños de mi hermano, ¡va a ser increíble! También quiero comprar un pastel de chocolate con fresas y un regalo especial: un libro titulado «Cien años de soledad",
-    ]
-
-    model = SparseTextEmbedding(model_name=model_name, language="english")
-    embeddings = list(model.embed(docs))
-    for idx, shape in enumerate([14, 18, 15, 10, 15]):
-        assert embeddings[idx].values.shape == (shape,)
-        assert embeddings[idx].indices.shape == (shape,)
-
-    if is_ci:
-        delete_model_cache(model.model._model_dir)
+def test_special_characters(model_cache, model_name: str) -> None:
+    with model_cache(model_name) as model:
+        docs = [
+            "Über den größten Flüssen Österreichs äußern sich Experten häufig: Öko-Systeme müssen geschützt werden!",
+            "L'élève français s'écrie : « Où est mon crayon ? J'ai besoin de finir cet exercice avant la récréation!",
+            "Într-o zi însorită, Ștefan și Ioana au mâncat mămăligă cu brânză și au băut țuică la cabană.",
+            "Üzgün öğretmen öğrencilere seslendi: Lütfen gürültü yapmayın, sınavınızı bitirmeye çalışıyorum!",
+            "Ο Ξενοφών είπε: «Ψάχνω για ένα ωραίο δώρο για τη γιαγιά μου. Ίσως ένα φυτό ή ένα βιβλίο;»",
+            "Hola! ¿Cómo estás? Estoy muy emocionado por el cumpleaños de mi hermano, ¡va a ser increíble! También quiero comprar un pastel de chocolate con fresas y un regalo especial: un libro titulado «Cien años de soledad",
+        ]
+        embeddings = list(model.embed(docs))
+        for idx, shape in enumerate([14, 18, 15, 10, 15]):
+            assert embeddings[idx].values.shape == (shape,)
+            assert embeddings[idx].indices.shape == (shape,)
 
 
 @pytest.mark.parametrize("model_name", ["Qdrant/bm42-all-minilm-l6-v2-attentions"])

tests/test_common.py

@@ -1,3 +1,5 @@
+import numpy as np
+
 from fastembed import (
     TextEmbedding,
     SparseTextEmbedding,
@@ -5,6 +7,7 @@ from fastembed import (
     LateInteractionMultimodalEmbedding,
     LateInteractionTextEmbedding,
 )
+from fastembed.common.utils import last_token_pooling
 
 
 def test_text_list_supported_models():
@@ -28,3 +31,31 @@ def test_text_list_supported_models():
         assert "model_file" in description and description["model_file"]
         assert "sources" in description and description["sources"]
         assert "hf" in description["sources"] or "url" in description["sources"]
+
+
+def test_last_token_pooling():
+    token_embeddings = np.array(
+        [
+            [[1.0, 1.0], [2.0, 2.0], [9.0, 9.0], [9.0, 9.0]],  # 2 real tokens, then padding
+            [[3.0, 3.0], [4.0, 4.0], [5.0, 5.0], [6.0, 6.0]],  # no padding
+        ]
+    )
+    attention_mask = np.array([[1, 1, 0, 0], [1, 1, 1, 1]], dtype=np.int64)
+
+    pooled = last_token_pooling(token_embeddings, attention_mask)
+
+    assert np.allclose(pooled, [[2.0, 2.0], [6.0, 6.0]])
+
+
+def test_last_token_pooling_with_left_padding():
+    token_embeddings = np.array(
+        [
+            [[9.0, 9.0], [9.0, 9.0], [1.0, 1.0], [2.0, 2.0]],  # padding, then 2 real tokens
+            [[3.0, 3.0], [4.0, 4.0], [5.0, 5.0], [6.0, 6.0]],  # no padding
+        ]
+    )
+    attention_mask = np.array([[0, 0, 1, 1], [1, 1, 1, 1]], dtype=np.int64)
+
+    pooled = last_token_pooling(token_embeddings, attention_mask)
+
+    assert np.allclose(pooled, [[2.0, 2.0], [6.0, 6.0]])

tests/test_custom_models.py

@@ -0,0 +1,305 @@
+import itertools
+import os
+import numpy as np
+import pytest
+
+from fastembed.common.model_description import (
+    PoolingType,
+    ModelSource,
+    DenseModelDescription,
+    BaseModelDescription,
+)
+from fastembed.common.onnx_model import OnnxOutputContext
+from fastembed.common.utils import normalize, mean_pooling, last_token_pooling
+from fastembed.text.custom_text_embedding import CustomTextEmbedding, PostprocessingConfig
+from fastembed.rerank.cross_encoder.custom_text_cross_encoder import CustomTextCrossEncoder
+from fastembed.rerank.cross_encoder import TextCrossEncoder
+from fastembed.text.text_embedding import TextEmbedding
+from tests.utils import delete_model_cache
+
+
+@pytest.fixture(autouse=True)
+def restore_custom_models_fixture():
+    CustomTextEmbedding.SUPPORTED_MODELS = []
+    CustomTextEmbedding.POSTPROCESSING_MAPPING = {}
+    CustomTextCrossEncoder.SUPPORTED_MODELS = []
+    yield
+    CustomTextEmbedding.SUPPORTED_MODELS = []
+    CustomTextEmbedding.POSTPROCESSING_MAPPING = {}
+    CustomTextCrossEncoder.SUPPORTED_MODELS = []
+
+
+def test_text_custom_model():
+    is_ci = os.getenv("CI")
+    custom_model_name = "intfloat/multilingual-e5-small"
+    canonical_vector = np.array(
+        [3.1317e-02, 3.0939e-02, -3.5117e-02, -6.7274e-02, 8.5084e-02], dtype=np.float32
+    )
+    pooling = PoolingType.MEAN
+    normalization = True
+    dim = 384
+    size_in_gb = 0.47
+    source = ModelSource(hf=custom_model_name)
+
+    TextEmbedding.add_custom_model(
+        custom_model_name,
+        pooling=pooling,
+        normalization=normalization,
+        sources=source,
+        dim=dim,
+        size_in_gb=size_in_gb,
+    )
+
+    assert CustomTextEmbedding.SUPPORTED_MODELS[0] == DenseModelDescription(
+        model=custom_model_name,
+        sources=source,
+        model_file="onnx/model.onnx",
+        description="",
+        license="",
+        size_in_GB=size_in_gb,
+        additional_files=[],
+        dim=dim,
+        tasks={},
+    )
+    assert CustomTextEmbedding.POSTPROCESSING_MAPPING[custom_model_name] == PostprocessingConfig(
+        pooling=pooling, normalization=normalization
+    )
+
+    model = TextEmbedding(custom_model_name)
+    docs = ["hello world", "flag embedding"]
+    embeddings = list(model.embed(docs))
+    embeddings = np.stack(embeddings, axis=0)
+    assert embeddings.shape == (2, dim)
+
+    assert np.allclose(embeddings[0, : canonical_vector.shape[0]], canonical_vector, atol=1e-3)
+
+    if is_ci:
+        delete_model_cache(model.model._model_dir)
+
+
+def test_text_custom_model_parallel_processing():
+    is_ci = os.getenv("CI")
+    custom_model_name = "intfloat/multilingual-e5-small"
+    dim = 384
+
+    TextEmbedding.add_custom_model(
+        custom_model_name,
+        pooling=PoolingType.MEAN,
+        normalization=True,
+        sources=ModelSource(hf=custom_model_name),
+        dim=dim,
+        size_in_gb=0.47,
+    )
+
+    model = TextEmbedding(custom_model_name)
+    docs = ["hello world", "flag embedding"] * 50
+    embeddings = np.stack(list(model.embed(docs, batch_size=10, parallel=2)), axis=0)
+
+    assert embeddings.shape == (len(docs), dim)
+
+    if is_ci:
+        delete_model_cache(model.model._model_dir)
+
+
+def test_cross_encoder_custom_model():
+    is_ci = os.getenv("CI")
+    custom_model_name = "Xenova/ms-marco-MiniLM-L-4-v2"
+    size_in_gb = 0.08
+    source = ModelSource(hf=custom_model_name)
+    canonical_vector = np.array([-5.7170815, -11.112114], dtype=np.float32)
+
+    TextCrossEncoder.add_custom_model(
+        custom_model_name,
+        model_file="onnx/model.onnx",
+        sources=source,
+        size_in_gb=size_in_gb,
+    )
+
+    assert CustomTextCrossEncoder.SUPPORTED_MODELS[0] == BaseModelDescription(
+        model=custom_model_name,
+        sources=source,
+        model_file="onnx/model.onnx",
+        description="",
+        license="",
+        size_in_GB=size_in_gb,
+    )
+
+    model = TextCrossEncoder(custom_model_name)
+    pairs = [
+        ("What is AI?", "Artificial intelligence is ..."),
+        ("What is ML?", "Machine learning is ..."),
+    ]
+    scores = list(model.rerank_pairs(pairs))
+
+    embeddings = np.stack(scores, axis=0)
+    assert embeddings.shape == (2,)
+    assert np.allclose(embeddings, canonical_vector, atol=1e-3)
+    if is_ci:
+        delete_model_cache(model.model._model_dir)
+
+
+def test_cross_encoder_custom_model_parallel_processing():
+    is_ci = os.getenv("CI")
+    custom_model_name = "Xenova/ms-marco-MiniLM-L-4-v2"
+
+    TextCrossEncoder.add_custom_model(
+        custom_model_name,
+        model_file="onnx/model.onnx",
+        sources=ModelSource(hf=custom_model_name),
+        size_in_gb=0.08,
+    )
+
+    model = TextCrossEncoder(custom_model_name)
+    pairs = [("What is AI?", "Artificial intelligence is ...")] * 50
+    scores = np.stack(list(model.rerank_pairs(pairs, batch_size=10, parallel=2)), axis=0)
+
+    assert scores.shape == (len(pairs),)
+
+    if is_ci:
+        delete_model_cache(model.model._model_dir)
+
+
+def test_mock_add_custom_models():
+    dim = 5
+    size_in_gb = 0.1
+    source = ModelSource(hf="artificial")
+
+    num_tokens = 10
+    dummy_pooled_embedding = np.random.random((1, dim)).astype(np.float32)
+    dummy_token_embedding = np.random.random((1, num_tokens, dim)).astype(np.float32)
+    dummy_attention_mask = np.ones((1, num_tokens)).astype(np.int64)
+
+    dummy_token_output = OnnxOutputContext(
+        model_output=dummy_token_embedding, attention_mask=dummy_attention_mask
+    )
+    dummy_pooled_output = OnnxOutputContext(model_output=dummy_pooled_embedding)
+    input_data = {
+        f"{PoolingType.MEAN.lower()}-normalized": dummy_token_output,
+        f"{PoolingType.MEAN.lower()}": dummy_token_output,
+        f"{PoolingType.CLS.lower()}-normalized": dummy_token_output,
+        f"{PoolingType.CLS.lower()}": dummy_token_output,
+        f"{PoolingType.LAST_TOKEN.lower()}-normalized": dummy_token_output,
+        f"{PoolingType.LAST_TOKEN.lower()}": dummy_token_output,
+        f"{PoolingType.DISABLED.lower()}-normalized": dummy_pooled_output,
+        f"{PoolingType.DISABLED.lower()}": dummy_pooled_output,
+    }
+
+    expected_output = {
+        f"{PoolingType.MEAN.lower()}-normalized": normalize(
+            mean_pooling(dummy_token_embedding, dummy_attention_mask)
+        ),
+        f"{PoolingType.MEAN.lower()}": mean_pooling(dummy_token_embedding, dummy_attention_mask),
+        f"{PoolingType.CLS.lower()}-normalized": normalize(dummy_token_embedding[:, 0]),
+        f"{PoolingType.CLS.lower()}": dummy_token_embedding[:, 0],
+        f"{PoolingType.LAST_TOKEN.lower()}-normalized": normalize(
+            last_token_pooling(dummy_token_embedding, dummy_attention_mask)
+        ),
+        f"{PoolingType.LAST_TOKEN.lower()}": last_token_pooling(
+            dummy_token_embedding, dummy_attention_mask
+        ),
+        f"{PoolingType.DISABLED.lower()}-normalized": normalize(dummy_pooled_embedding),
+        f"{PoolingType.DISABLED.lower()}": dummy_pooled_embedding,
+    }
+
+    for pooling, normalization in itertools.product(
+        (PoolingType.MEAN, PoolingType.CLS, PoolingType.LAST_TOKEN, PoolingType.DISABLED),
+        (True, False),
+    ):
+        model_name = f"{pooling.name.lower()}{'-normalized' if normalization else ''}"
+        TextEmbedding.add_custom_model(
+            model_name,
+            pooling=pooling,
+            normalization=normalization,
+            sources=source,
+            dim=dim,
+            size_in_gb=size_in_gb,
+        )
+
+        custom_text_embedding = CustomTextEmbedding(
+            model_name,
+            lazy_load=True,
+            specific_model_path="./",  # disable model downloading and loading
+        )
+
+        post_processed_output = next(
+            iter(custom_text_embedding._post_process_onnx_output(input_data[model_name]))
+        )
+        assert np.allclose(post_processed_output, expected_output[model_name], atol=1e-3)
+
+
+def test_custom_text_model_lookup_is_case_insensitive():
+    model_name = "Org/Model"
+
+    TextEmbedding.add_custom_model(
+        model_name,
+        pooling=PoolingType.MEAN,
+        normalization=True,
+        sources=ModelSource(hf="artificial"),
+        dim=5,
+        size_in_gb=0.1,
+    )
+
+    model = TextEmbedding("org/model", lazy_load=True, specific_model_path="./")
+
+    assert isinstance(model.model, CustomTextEmbedding)
+    assert model.model._pooling == PoolingType.MEAN
+    assert model.model._normalization is True
+
+
+def test_do_not_add_existing_model():
+    existing_base_model = "sentence-transformers/all-MiniLM-L6-v2"
+    custom_model_name = "intfloat/multilingual-e5-small"
+
+    with pytest.raises(ValueError, match=f"Model {existing_base_model} is already registered"):
+        TextEmbedding.add_custom_model(
+            existing_base_model,
+            pooling=PoolingType.MEAN,
+            normalization=True,
+            sources=ModelSource(hf=existing_base_model),
+            dim=384,
+            size_in_gb=0.47,
+        )
+
+    TextEmbedding.add_custom_model(
+        custom_model_name,
+        pooling=PoolingType.MEAN,
+        normalization=False,
+        sources=ModelSource(hf=existing_base_model),
+        dim=384,
+        size_in_gb=0.47,
+    )
+
+    with pytest.raises(ValueError, match=f"Model {custom_model_name} is already registered"):
+        TextEmbedding.add_custom_model(
+            custom_model_name,
+            pooling=PoolingType.MEAN,
+            normalization=True,
+            sources=ModelSource(hf=custom_model_name),
+            dim=384,
+            size_in_gb=0.47,
+        )
+
+
+def test_do_not_add_existing_cross_encoder():
+    existing_base_model = "Xenova/ms-marco-MiniLM-L-6-v2"
+    custom_model_name = "Xenova/ms-marco-MiniLM-L-4-v2"
+
+    with pytest.raises(ValueError, match=f"Model {existing_base_model} is already registered"):
+        TextCrossEncoder.add_custom_model(
+            existing_base_model,
+            sources=ModelSource(hf=existing_base_model),
+            size_in_gb=0.08,
+        )
+
+    TextCrossEncoder.add_custom_model(
+        custom_model_name,
+        sources=ModelSource(hf=existing_base_model),
+        size_in_gb=0.08,
+    )
+
+    with pytest.raises(ValueError, match=f"Model {custom_model_name} is already registered"):
+        TextCrossEncoder.add_custom_model(
+            custom_model_name,
+            sources=ModelSource(hf=custom_model_name),
+            size_in_gb=0.08,
+        )

tests/test_image_onnx_embeddings.py

@@ -1,4 +1,6 @@
 import os
+import platform
+from contextlib import contextmanager
 from io import BytesIO
 
 import numpy as np
@@ -8,7 +10,7 @@ from PIL import Image
 
 from fastembed import ImageEmbedding
 from tests.config import TEST_MISC_DIR
-from tests.utils import delete_model_cache
+from tests.utils import delete_model_cache, should_test_model
 
 CANONICAL_VECTOR_VALUES = {
     "Qdrant/clip-ViT-B-32-vision": np.array([-0.0098, 0.0128, -0.0274, 0.002, -0.0059]),
@@ -24,88 +26,124 @@ CANONICAL_VECTOR_VALUES = {
     "jinaai/jina-clip-v1": np.array(
         [-0.029, 0.0216, 0.0396, 0.0283, -0.0023, 0.0151, 0.011, -0.0235, 0.0251, -0.0343]
     ),
+    "nomic-ai/nomic-embed-vision-v1.5": np.array(
+        [0.0048, -0.0254, 0.0067, -0.0296, -0.0435, -0.0123, 0.0024, -0.0361, -0.0703, -0.0186]
+    ),
+    "nomic-ai/nomic-embed-vision-v1.5-Q": np.array(
+        [-0.0011, -0.0477, 0.0024, -0.049, -0.0458, -0.0314, 0.017, -0.0383, -0.0537, -0.021]
+    ),
+    "google/siglip2-base-patch16-224": np.array(
+        [-0.02095927, -0.0075177, -0.00144479, -0.0080948, 0.05031789]
+    ),
 }
 
+_MODELS_TO_CACHE = ("Qdrant/clip-ViT-B-32-vision",)
+MODELS_TO_CACHE = tuple([x.lower() for x in _MODELS_TO_CACHE])
 
-def test_embedding() -> None:
+
+@pytest.fixture(scope="module")
+def model_cache():
     is_ci = os.getenv("CI")
+    cache = {}
+
+    @contextmanager
+    def get_model(model_name: str):
+        lowercase_model_name = model_name.lower()
+        if lowercase_model_name not in cache:
+            cache[lowercase_model_name] = ImageEmbedding(lowercase_model_name)
+        yield cache[lowercase_model_name]
+        if lowercase_model_name not in MODELS_TO_CACHE:
+            model_inst = cache.pop(lowercase_model_name)
+            if is_ci:
+                delete_model_cache(model_inst.model._model_dir)
+            del model_inst
+
+    yield get_model
+
+    if is_ci:
+        for name, model in cache.items():
+            delete_model_cache(model.model._model_dir)
+    cache.clear()
+
+
+@pytest.mark.parametrize("model_name", ["Qdrant/clip-ViT-B-32-vision"])
+def test_embedding(model_cache, model_name: str) -> None:
+    is_ci = os.getenv("CI")
+    is_mac = platform.system() == "Darwin"
+    is_manual = os.getenv("GITHUB_EVENT_NAME") == "workflow_dispatch"
 
     for model_desc in ImageEmbedding._list_supported_models():
-        if not is_ci and model_desc.size_in_GB > 1:
+        # quantized int8 ops diverge on macOS; canonical vector is generated on linux/amd64 (CI)
+        if is_mac and model_desc.model == "nomic-ai/nomic-embed-vision-v1.5-Q":
+            continue
+        if not should_test_model(model_desc, model_name, is_ci, is_manual):
             continue
 
         dim = model_desc.dim
 
-        model = ImageEmbedding(model_name=model_desc.model)
+        with model_cache(model_desc.model) as model:
+            images = [
+                TEST_MISC_DIR / "image.jpeg",
+                str(TEST_MISC_DIR / "small_image.jpeg"),
+                Image.open((TEST_MISC_DIR / "small_image.jpeg")),
+                Image.open(BytesIO(requests.get("https://qdrant.tech/img/logo.png").content)),
+            ]
+            embeddings = list(model.embed(images))
+            embeddings = np.stack(embeddings, axis=0)
+            assert embeddings.shape == (len(images), dim)
 
-        images = [
+            canonical_vector = CANONICAL_VECTOR_VALUES[model_desc.model]
+
+            assert np.allclose(
+                embeddings[0, : canonical_vector.shape[0]], canonical_vector, atol=1e-3
+            ), model_desc.model
+
+            assert np.allclose(embeddings[1], embeddings[2]), model_desc.model
+
+
+@pytest.mark.parametrize("n_dims,model_name", [(512, "Qdrant/clip-ViT-B-32-vision")])
+def test_batch_embedding(model_cache, n_dims: int, model_name: str) -> None:
+    with model_cache(model_name) as model:
+        n_images = 32
+        test_images = [
             TEST_MISC_DIR / "image.jpeg",
             str(TEST_MISC_DIR / "small_image.jpeg"),
-            Image.open((TEST_MISC_DIR / "small_image.jpeg")),
-            Image.open(BytesIO(requests.get("https://qdrant.tech/img/logo.png").content)),
+            Image.open(TEST_MISC_DIR / "small_image.jpeg"),
         ]
-        embeddings = list(model.embed(images))
+        images = test_images * n_images
+
+        embeddings = list(model.embed(images, batch_size=10))
         embeddings = np.stack(embeddings, axis=0)
-        assert embeddings.shape == (len(images), dim)
+        assert np.allclose(embeddings[1], embeddings[2])
 
-        canonical_vector = CANONICAL_VECTOR_VALUES[model_desc.model]
+        canonical_vector = CANONICAL_VECTOR_VALUES[model_name]
 
-        assert np.allclose(
-            embeddings[0, : canonical_vector.shape[0]], canonical_vector, atol=1e-3
-        ), model_desc.model
-
-        assert np.allclose(embeddings[1], embeddings[2]), model_desc.model
-
-        if is_ci:
-            delete_model_cache(model.model._model_dir)
+        assert embeddings.shape == (len(test_images) * n_images, n_dims)
+        assert np.allclose(embeddings[0, : canonical_vector.shape[0]], canonical_vector, atol=1e-3)
 
 
 @pytest.mark.parametrize("n_dims,model_name", [(512, "Qdrant/clip-ViT-B-32-vision")])
-def test_batch_embedding(n_dims: int, model_name: str) -> None:
-    is_ci = os.getenv("CI")
-    model = ImageEmbedding(model_name=model_name)
-    n_images = 32
-    test_images = [
-        TEST_MISC_DIR / "image.jpeg",
-        str(TEST_MISC_DIR / "small_image.jpeg"),
-        Image.open(TEST_MISC_DIR / "small_image.jpeg"),
-    ]
-    images = test_images * n_images
+def test_parallel_processing(model_cache, n_dims: int, model_name: str) -> None:
+    with model_cache(model_name) as model:
+        n_images = 32
+        test_images = [
+            TEST_MISC_DIR / "image.jpeg",
+            str(TEST_MISC_DIR / "small_image.jpeg"),
+            Image.open(TEST_MISC_DIR / "small_image.jpeg"),
+        ]
+        images = test_images * n_images
+        embeddings = list(model.embed(images, batch_size=10, parallel=2))
+        embeddings = np.stack(embeddings, axis=0)
 
-    embeddings = list(model.embed(images, batch_size=10))
-    embeddings = np.stack(embeddings, axis=0)
+        embeddings_2 = list(model.embed(images, batch_size=10, parallel=None))
+        embeddings_2 = np.stack(embeddings_2, axis=0)
 
-    assert embeddings.shape == (len(test_images) * n_images, n_dims)
-    if is_ci:
-        delete_model_cache(model.model._model_dir)
+        embeddings_3 = list(model.embed(images, batch_size=10, parallel=0))
+        embeddings_3 = np.stack(embeddings_3, axis=0)
 
-
-@pytest.mark.parametrize("n_dims,model_name", [(512, "Qdrant/clip-ViT-B-32-vision")])
-def test_parallel_processing(n_dims: int, model_name: str) -> None:
-    is_ci = os.getenv("CI")
-    model = ImageEmbedding(model_name=model_name)
-
-    n_images = 32
-    test_images = [
-        TEST_MISC_DIR / "image.jpeg",
-        str(TEST_MISC_DIR / "small_image.jpeg"),
-        Image.open(TEST_MISC_DIR / "small_image.jpeg"),
-    ]
-    images = test_images * n_images
-    embeddings = list(model.embed(images, batch_size=10, parallel=2))
-    embeddings = np.stack(embeddings, axis=0)
-
-    embeddings_2 = list(model.embed(images, batch_size=10, parallel=None))
-    embeddings_2 = np.stack(embeddings_2, axis=0)
-
-    embeddings_3 = list(model.embed(images, batch_size=10, parallel=0))
-    embeddings_3 = np.stack(embeddings_3, axis=0)
-
-    assert embeddings.shape == (n_images * len(test_images), n_dims)
-    assert np.allclose(embeddings, embeddings_2, atol=1e-3)
-    assert np.allclose(embeddings, embeddings_3, atol=1e-3)
-    if is_ci:
-        delete_model_cache(model.model._model_dir)
+        assert embeddings.shape == (n_images * len(test_images), n_dims)
+        assert np.allclose(embeddings, embeddings_2, atol=1e-3)
+        assert np.allclose(embeddings, embeddings_3, atol=1e-3)
 
 
 @pytest.mark.parametrize("model_name", ["Qdrant/clip-ViT-B-32-vision"])
@@ -121,3 +159,31 @@ def test_lazy_load(model_name: str) -> None:
     assert hasattr(model.model, "model")
     if is_ci:
         delete_model_cache(model.model._model_dir)
+
+
+def test_get_embedding_size() -> None:
+    assert ImageEmbedding.get_embedding_size(model_name="Qdrant/clip-ViT-B-32-vision") == 512
+    assert ImageEmbedding.get_embedding_size(model_name="Qdrant/clip-vit-b-32-vision") == 512
+
+
+def test_embedding_size() -> None:
+    is_ci = os.getenv("CI")
+    model_name = "Qdrant/clip-ViT-B-32-vision"
+    model = ImageEmbedding(model_name=model_name, lazy_load=True)
+    assert model.embedding_size == 512
+
+    model_name = "Qdrant/clip-vit-b-32-vision"
+    model = ImageEmbedding(model_name=model_name, lazy_load=True)
+    assert model.embedding_size == 512
+    if is_ci:
+        delete_model_cache(model.model._model_dir)
+
+
+@pytest.mark.parametrize("model_name", ["Qdrant/clip-ViT-B-32-vision"])
+def test_session_options(model_cache, model_name) -> None:
+    with model_cache(model_name) as default_model:
+        default_session_options = default_model.model.model.get_session_options()
+        assert default_session_options.enable_cpu_mem_arena is True
+        model = ImageEmbedding(model_name=model_name, enable_cpu_mem_arena=False)
+        session_options = model.model.model.get_session_options()
+        assert session_options.enable_cpu_mem_arena is False

tests/test_image_transform.py

@@ -0,0 +1,90 @@
+import numpy as np
+import pytest
+from PIL import Image
+
+from fastembed.image.transform.functional import normalize, resize
+
+
+@pytest.mark.parametrize(
+    ("size", "expected"),
+    [
+        ((100, 200), (200, 100)),  # the bug: a non-square size came back transposed
+        ((224, 224), (224, 224)),  # the square path every shipped model takes
+    ],
+)
+def test_resize_tuple_is_height_width(size: tuple[int, int], expected: tuple[int, int]) -> None:
+    """A ``(height, width)`` size must reach Pillow as ``(width, height)``."""
+    resized = resize(Image.new("RGB", (300, 300)), size=size)
+
+    assert resized.size == expected  # PIL reports (width, height)
+
+
+def test_resize_int_keeps_shortest_edge_behaviour() -> None:
+    """The int branch already emitted Pillow order; it must not be disturbed."""
+    landscape = Image.new("RGB", (400, 200))
+    portrait = Image.new("RGB", (200, 400))
+
+    # size sets the shortest edge, and the aspect ratio is preserved.
+    assert resize(landscape, size=100).size == (200, 100)
+    assert resize(portrait, size=100).size == (100, 200)
+
+
+@pytest.mark.parametrize(
+    ("mean", "std"),
+    [
+        ([0.1, 0.2, 0.3], [0.5, 0.6, 0.7]),  # per-channel, as every model config gives it
+        (0.5, 0.25),  # scalar, expanded to one value per channel
+    ],
+)
+def test_normalize_chw_is_channel_wise(
+    mean: list[float] | float, std: list[float] | float
+) -> None:
+    """Each channel must be normalized by its own mean/std, not by any other axis."""
+    rng = np.random.default_rng(0)
+    image = rng.random((3, 5, 7)).astype(np.float32)
+    means = mean if isinstance(mean, list) else [mean] * 3
+    stds = std if isinstance(std, list) else [std] * 3
+
+    result = normalize(image, mean=mean, std=std)
+
+    for c in range(3):
+        assert np.allclose(result[c], (image[c] - means[c]) / stds[c], atol=1e-6)
+
+
+@pytest.mark.parametrize("batch_size", [2, 3])
+def test_normalize_batched_matches_per_image(batch_size: int) -> None:
+    """A batch must give exactly what the (C, H, W) path gives image by image.
+
+    batch_size 2 used to raise, since transposing reversed every axis; batch_size 3
+    matched the channel count and silently normalized along the batch axis instead.
+    """
+    rng = np.random.default_rng(2)
+    batch = rng.random((batch_size, 3, 4, 4)).astype(np.float32)
+    mean, std = [0.1, 0.2, 0.3], [0.5, 0.6, 0.7]
+
+    result = normalize(batch, mean=mean, std=std)
+
+    per_image = np.stack([normalize(image, mean=mean, std=std) for image in batch])
+    assert result.shape == batch.shape
+    assert np.array_equal(result, per_image)
+
+
+def test_normalize_rejects_input_without_a_channel_axis() -> None:
+    """Every pipeline runs ConvertToRGB first, so normalize only ever sees (C, H, W)."""
+    with pytest.raises(ValueError, match=r"must be \(C, H, W\)"):
+        normalize(np.zeros((4, 6), dtype=np.float32), mean=0.5, std=0.25)
+
+
+@pytest.mark.parametrize(
+    ("mean", "std", "expected"),
+    [
+        ([0.1, 0.2], [1.0, 1.0, 1.0], "mean must"),
+        ([0.1, 0.2, 0.3], [1.0, 1.0], "std must"),
+    ],
+)
+def test_normalize_channel_count_mismatch_raises(
+    mean: list[float], std: list[float], expected: str
+) -> None:
+    image = np.zeros((3, 4, 4), dtype=np.float32)
+    with pytest.raises(ValueError, match=expected):
+        normalize(image, mean=mean, std=std)

tests/test_late_interaction_embeddings.py

@@ -1,4 +1,5 @@
 import os
+from contextlib import contextmanager
 
 import pytest
 import numpy as np
@@ -6,7 +7,7 @@ import numpy as np
 from fastembed.late_interaction.late_interaction_text_embedding import (
     LateInteractionTextEmbedding,
 )
-from tests.utils import delete_model_cache
+from tests.utils import delete_model_cache, should_test_model
 
 # vectors are abridged and rounded for brevity
 CANONICAL_COLUMN_VALUES = {
@@ -150,82 +151,124 @@ CANONICAL_QUERY_VALUES = {
     ),
 }
 
+_MODELS_TO_CACHE = ("answerdotai/answerai-colbert-small-v1",)
+MODELS_TO_CACHE = tuple([x.lower() for x in _MODELS_TO_CACHE])
+
+
+@pytest.fixture(scope="module")
+def model_cache():
+    is_ci = os.getenv("CI")
+    cache = {}
+
+    @contextmanager
+    def get_model(model_name: str):
+        lowercase_model_name = model_name.lower()
+        if lowercase_model_name not in cache:
+            cache[lowercase_model_name] = LateInteractionTextEmbedding(lowercase_model_name)
+        yield cache[lowercase_model_name]
+        if lowercase_model_name not in MODELS_TO_CACHE:
+            model_inst = cache.pop(lowercase_model_name)
+            if is_ci:
+                delete_model_cache(model_inst.model._model_dir)
+            del model_inst
+
+    yield get_model
+
+    if is_ci:
+        for name, model in cache.items():
+            delete_model_cache(model.model._model_dir)
+    cache.clear()
+
+
 docs = ["Hello World"]
 
 
-def test_batch_embedding():
-    is_ci = os.getenv("CI")
+@pytest.mark.parametrize("model_name", ["answerdotai/answerai-colbert-small-v1"])
+def test_batch_embedding(model_cache, model_name: str):
     docs_to_embed = docs * 10
 
-    for model_name, expected_result in CANONICAL_COLUMN_VALUES.items():
-        print("evaluating", model_name)
-        model = LateInteractionTextEmbedding(model_name=model_name)
+    with model_cache(model_name) as model:
         result = list(model.embed(docs_to_embed, batch_size=6))
+        expected_result = CANONICAL_COLUMN_VALUES[model_name]
 
         for value in result:
             token_num, abridged_dim = expected_result.shape
             assert np.allclose(value[:, :abridged_dim], expected_result, atol=2e-3)
 
-        if is_ci:
-            delete_model_cache(model.model._model_dir)
+
+@pytest.mark.parametrize("model_name", ["answerdotai/answerai-colbert-small-v1"])
+def test_batch_inference_size_same_as_single_inference(model_cache, model_name: str):
+    with model_cache(model_name) as model:
+        docs_to_embed = [
+            "short document",
+            "A bit longer document, which should not affect the size",
+        ]
+        result = list(model.embed(docs_to_embed, batch_size=1))
+        result_2 = list(model.embed(docs_to_embed, batch_size=2))
+        assert len(result[0]) == len(result_2[0])
 
 
-def test_single_embedding():
+@pytest.mark.parametrize("model_name", ["answerdotai/answerai-colbert-small-v1"])
+def test_single_embedding(model_cache, model_name: str):
     is_ci = os.getenv("CI")
+    is_manual = os.getenv("GITHUB_EVENT_NAME") == "workflow_dispatch"
     docs_to_embed = docs
 
-    for model_name, expected_result in CANONICAL_COLUMN_VALUES.items():
+    for model_desc in LateInteractionTextEmbedding._list_supported_models():
+        if not should_test_model(model_desc, model_name, is_ci, is_manual):
+            continue
+
         print("evaluating", model_name)
-        model = LateInteractionTextEmbedding(model_name=model_name)
-        result = next(iter(model.embed(docs_to_embed, batch_size=6)))
-        token_num, abridged_dim = expected_result.shape
-        assert np.allclose(result[:, :abridged_dim], expected_result, atol=2e-3)
-
-        if is_ci:
-            delete_model_cache(model.model._model_dir)
+        with model_cache(model_desc.model) as model:
+            whole_result = list(model.embed(docs_to_embed, batch_size=6))
+            assert len(whole_result) == 1
+            result = whole_result[0]
+            expected_result = CANONICAL_COLUMN_VALUES[model_desc.model]
+            token_num, abridged_dim = expected_result.shape
+            assert np.allclose(result[:, :abridged_dim], expected_result, atol=2e-3)
 
 
-def test_single_embedding_query():
+@pytest.mark.parametrize("model_name", ["answerdotai/answerai-colbert-small-v1"])
+def test_single_embedding_query(model_cache, model_name: str):
     is_ci = os.getenv("CI")
+    is_manual = os.getenv("GITHUB_EVENT_NAME") == "workflow_dispatch"
     queries_to_embed = docs
 
-    for model_name, expected_result in CANONICAL_QUERY_VALUES.items():
-        print("evaluating", model_name)
-        model = LateInteractionTextEmbedding(model_name=model_name)
-        result = next(iter(model.query_embed(queries_to_embed)))
-        token_num, abridged_dim = expected_result.shape
-        assert np.allclose(result[:, :abridged_dim], expected_result, atol=2e-3)
+    for model_desc in LateInteractionTextEmbedding._list_supported_models():
+        if not should_test_model(model_desc, model_name, is_ci, is_manual):
+            continue
 
-        if is_ci:
-            delete_model_cache(model.model._model_dir)
+        print("evaluating", model_desc.model)
+        with model_cache(model_desc.model) as model:
+            whole_result = list(model.query_embed(queries_to_embed))
+            assert len(whole_result) == 1
+            result = whole_result[0]
+            expected_result = CANONICAL_QUERY_VALUES[model_desc.model]
+            token_num, abridged_dim = expected_result.shape
+            assert np.allclose(result[:, :abridged_dim], expected_result, atol=2e-3)
 
 
-def test_parallel_processing():
-    is_ci = os.getenv("CI")
-    model = LateInteractionTextEmbedding(model_name="colbert-ir/colbertv2.0")
-    token_dim = 128
-    docs = ["hello world", "flag embedding"] * 100
-    embeddings = list(model.embed(docs, batch_size=10, parallel=2))
-    embeddings = np.stack(embeddings, axis=0)
+@pytest.mark.parametrize("token_dim,model_name", [(96, "answerdotai/answerai-colbert-small-v1")])
+def test_parallel_processing(model_cache, token_dim: int, model_name: str):
+    with model_cache(model_name) as model:
+        docs = ["hello world", "flag embedding"] * 100
+        embeddings = list(model.embed(docs, batch_size=10, parallel=2))
 
-    embeddings_2 = list(model.embed(docs, batch_size=10, parallel=None))
-    embeddings_2 = np.stack(embeddings_2, axis=0)
+        embeddings_2 = list(model.embed(docs, batch_size=10, parallel=None))
 
-    embeddings_3 = list(model.embed(docs, batch_size=10, parallel=0))
-    embeddings_3 = np.stack(embeddings_3, axis=0)
+        # embeddings_3 = list(model.embed(docs, batch_size=10, parallel=0))  # inherits OnnxTextModel which
+        #         # is tested in TextEmbedding, disabling it here to reduce number of requests to hf
+        #         # multiprocessing is enough to test with `parallel=2`, and `parallel=None` is okay to tests since it reuses
+        #         # model from cache
 
-    assert embeddings.shape[0] == len(docs) and embeddings.shape[-1] == token_dim
-    assert np.allclose(embeddings, embeddings_2, atol=1e-3)
-    assert np.allclose(embeddings, embeddings_3, atol=1e-3)
+        assert len(embeddings) == len(docs) and embeddings[0].shape[-1] == token_dim
 
-    if is_ci:
-        delete_model_cache(model.model._model_dir)
+        for i in range(len(embeddings)):
+            assert np.allclose(embeddings[i], embeddings_2[i], atol=1e-3)
+            # assert np.allclose(embeddings[i], embeddings_3[i], atol=1e-3)
 
 
-@pytest.mark.parametrize(
-    "model_name",
-    ["colbert-ir/colbertv2.0"],
-)
+@pytest.mark.parametrize("model_name", ["answerdotai/answerai-colbert-small-v1"])
 def test_lazy_load(model_name: str):
     is_ci = os.getenv("CI")
 
@@ -244,3 +287,58 @@ def test_lazy_load(model_name: str):
 
     if is_ci:
         delete_model_cache(model.model._model_dir)
+
+
+def test_get_embedding_size():
+    model_name = "answerdotai/answerai-colbert-small-v1"
+    assert LateInteractionTextEmbedding.get_embedding_size(model_name) == 96
+
+    model_name = "answerdotai/answerai-ColBERT-small-v1"
+    assert LateInteractionTextEmbedding.get_embedding_size(model_name) == 96
+
+
+def test_embedding_size():
+    is_ci = os.getenv("CI")
+    model_name = "answerdotai/answerai-colbert-small-v1"
+    model = LateInteractionTextEmbedding(model_name=model_name, lazy_load=True)
+    assert model.embedding_size == 96
+
+    model_name = "answerdotai/answerai-ColBERT-small-v1"
+    model = LateInteractionTextEmbedding(model_name=model_name, lazy_load=True)
+    assert model.embedding_size == 96
+    if is_ci:
+        delete_model_cache(model.model._model_dir)
+
+
+@pytest.mark.parametrize("model_name", ["answerdotai/answerai-ColBERT-small-v1"])
+def test_session_options(model_cache, model_name) -> None:
+    with model_cache(model_name) as default_model:
+        default_session_options = default_model.model.model.get_session_options()
+        assert default_session_options.enable_cpu_mem_arena is True
+        model = LateInteractionTextEmbedding(model_name=model_name, enable_cpu_mem_arena=False)
+        session_options = model.model.model.get_session_options()
+        assert session_options.enable_cpu_mem_arena is False
+
+
+@pytest.mark.parametrize("model_name", ["answerdotai/answerai-colbert-small-v1"])
+def test_token_count(model_cache, model_name) -> None:
+    with model_cache(model_name) as model:
+        documents = ["short doc", "it is a long document to check attention mask for paddings"]
+        short_doc_token_count = model.token_count(documents[0])
+        long_doc_token_count = model.token_count(documents[1])
+        documents_token_count = model.token_count(documents)
+        assert short_doc_token_count + long_doc_token_count == documents_token_count
+        # 2 is 2*DOC_MARKER_TOKEN_ID for each document
+        assert short_doc_token_count + long_doc_token_count + 2 == model.token_count(
+            documents, include_extension=True
+        )
+        assert short_doc_token_count + long_doc_token_count == model.token_count(
+            documents, batch_size=1
+        )
+        assert short_doc_token_count + long_doc_token_count == model.token_count(
+            documents, is_doc=False
+        )
+        # query min length is 32
+        assert model.token_count(documents, is_doc=False, include_extension=True) == 64
+        very_long_query = "It's a very long query which definitely contains more than 32 tokens and we're using it to check whether the method can handle large query properly without cutting it to 32 tokens"
+        assert model.token_count(very_long_query, is_doc=False, include_extension=True) > 32

tests/test_late_interaction_multimodal.py

@@ -1,25 +1,36 @@
 import os
+from contextlib import contextmanager
 
+import pytest
 from PIL import Image
 import numpy as np
 
 from fastembed import LateInteractionMultimodalEmbedding
 from tests.config import TEST_MISC_DIR
-
+from tests.utils import delete_model_cache
 
 # vectors are abridged and rounded for brevity
 CANONICAL_IMAGE_VALUES = {
     "Qdrant/colpali-v1.3-fp16": np.array(
         [
-            [
-                [-0.0345, -0.022, 0.0567, -0.0518, -0.0782, 0.1714, -0.1738],
-                [-0.1181, -0.099, 0.0268, 0.0774, 0.0228, 0.0563, -0.1021],
-                [-0.117, -0.0683, 0.0371, 0.0921, 0.0107, 0.0659, -0.0666],
-                [-0.1393, -0.0948, 0.037, 0.0951, -0.0126, 0.0678, -0.087],
-                [-0.0957, -0.081, 0.0404, 0.052, 0.0409, 0.0335, -0.064],
-                [-0.0626, -0.0445, 0.056, 0.0592, -0.0229, 0.0409, -0.0301],
-                [-0.1299, -0.0691, 0.1097, 0.0728, 0.0123, 0.0519, 0.0122],
-            ]
+            [-0.0345, -0.022, 0.0567, -0.0518, -0.0782, 0.1714, -0.1738],
+            [-0.1181, -0.099, 0.0268, 0.0774, 0.0228, 0.0563, -0.1021],
+            [-0.117, -0.0683, 0.0371, 0.0921, 0.0107, 0.0659, -0.0666],
+            [-0.1393, -0.0948, 0.037, 0.0951, -0.0126, 0.0678, -0.087],
+            [-0.0957, -0.081, 0.0404, 0.052, 0.0409, 0.0335, -0.064],
+            [-0.0626, -0.0445, 0.056, 0.0592, -0.0229, 0.0409, -0.0301],
+            [-0.1299, -0.0691, 0.1097, 0.0728, 0.0123, 0.0519, 0.0122],
+        ]
+    ),
+    "Qdrant/colmodernvbert": np.array(
+        [
+            [0.11614, -0.15793, -0.11194, 0.0688, 0.08001, 0.10575, -0.07871],
+            [0.10094, -0.13301, -0.12069, 0.10932, 0.04645, 0.09884, 0.04048],
+            [0.13106, -0.18613, -0.13469, 0.10566, 0.03659, 0.07712, -0.03916],
+            [0.09754, -0.09596, -0.04839, 0.14991, 0.05692, 0.10569, -0.08349],
+            [0.02576, -0.15651, -0.09977, 0.09707, 0.13412, 0.09994, -0.09931],
+            [-0.06741, -0.1787, -0.19677, -0.07618, 0.13102, -0.02131, -0.02437],
+            [-0.02776, -0.10187, -0.13793, 0.03835, 0.04766, 0.04701, -0.15635],
         ]
     ),
 }
@@ -36,6 +47,17 @@ CANONICAL_QUERY_VALUES = {
             [-0.0165, -0.0106, 0.1672, -0.0768, 0.0389, -0.0038, 0.1137],
         ]
     ),
+    "Qdrant/colmodernvbert": np.array(
+        [
+            [0.05, 0.06557, 0.04026, 0.14981, 0.1842, 0.0263, -0.18706],
+            [-0.05664, -0.14028, 0.00649, -0.02849, 0.09034, -0.01494, 0.10693],
+            [-0.10147, -0.00716, 0.09084, -0.08236, -0.01849, -0.00972, -0.00461],
+            [-0.1233, -0.10814, -0.02337, -0.00329, 0.05984, 0.09934, 0.09846],
+            [-0.07053, -0.13119, -0.06487, 0.01508, 0.07459, 0.07655, 0.14821],
+            [0.00526, -0.13842, -0.05837, -0.02721, 0.13009, 0.05076, 0.17962],
+            [0.00924, -0.14383, -0.03057, -0.03691, 0.11718, 0.037, 0.13344],
+        ]
+    ),
 }
 
 queries = ["hello world", "flag embedding"]
@@ -45,40 +67,99 @@ images = [
     Image.open((TEST_MISC_DIR / "image.jpeg")),
 ]
 
+_MODELS_TO_CACHE = ("Qdrant/colmodernvbert",)
+MODELS_TO_CACHE = tuple(model_name.lower() for model_name in _MODELS_TO_CACHE)
 
-def test_batch_embedding():
+
+@pytest.fixture(scope="module")
+def model_cache():
     is_ci = os.getenv("CI")
+    cache = {}
 
-    if not is_ci:
-        for model_name, expected_result in CANONICAL_IMAGE_VALUES.items():
-            print("evaluating", model_name)
-            model = LateInteractionMultimodalEmbedding(model_name=model_name)
+    @contextmanager
+    def get_model(model_name: str):
+        lowercase_model_name = model_name.lower()
+        if lowercase_model_name not in cache:
+            cache[lowercase_model_name] = LateInteractionMultimodalEmbedding(lowercase_model_name)
+        yield cache[lowercase_model_name]
+        if lowercase_model_name not in MODELS_TO_CACHE:
+            model_inst = cache.pop(lowercase_model_name)
+            if is_ci:
+                delete_model_cache(model_inst.model._model_dir)
+            del model_inst
+
+    yield get_model
+
+    if is_ci:
+        for _, model in cache.items():
+            delete_model_cache(model.model._model_dir)
+    cache.clear()
+
+
+def test_batch_embedding(model_cache):
+    for model_name, expected_result in CANONICAL_IMAGE_VALUES.items():
+        if model_name.lower() == "Qdrant/colpali-v1.3-fp16".lower() and os.getenv("CI"):
+            continue  # colpali is too large for ci
+
+        print("evaluating", model_name)
+        with model_cache(model_name) as model:
             result = list(model.embed_image(images, batch_size=2))
 
             for value in result:
-                batch_size, token_num, abridged_dim = expected_result.shape
-                assert np.allclose(value[:token_num, :abridged_dim], expected_result, atol=1e-3)
+                token_num, abridged_dim = expected_result.shape
+                assert np.allclose(value[:token_num, :abridged_dim], expected_result, atol=2e-3)
 
 
-def test_single_embedding():
-    is_ci = os.getenv("CI")
-    if not is_ci:
-        for model_name, expected_result in CANONICAL_IMAGE_VALUES.items():
-            print("evaluating", model_name)
-            model = LateInteractionMultimodalEmbedding(model_name=model_name)
+def test_single_embedding(model_cache):
+    for model_name, expected_result in CANONICAL_IMAGE_VALUES.items():
+        if model_name.lower() == "Qdrant/colpali-v1.3-fp16".lower() and os.getenv("CI"):
+            continue  # colpali is too large for ci
+        print("evaluating", model_name)
+        with model_cache(model_name) as model:
             result = next(iter(model.embed_image(images, batch_size=6)))
-            batch_size, token_num, abridged_dim = expected_result.shape
-            assert np.allclose(result[:token_num, :abridged_dim], expected_result, atol=2e-3)
-
-
-def test_single_embedding_query():
-    is_ci = os.getenv("CI")
-    if not is_ci:
-        queries_to_embed = queries
-
-        for model_name, expected_result in CANONICAL_QUERY_VALUES.items():
-            print("evaluating", model_name)
-            model = LateInteractionMultimodalEmbedding(model_name=model_name)
-            result = next(iter(model.embed_text(queries_to_embed)))
             token_num, abridged_dim = expected_result.shape
             assert np.allclose(result[:token_num, :abridged_dim], expected_result, atol=2e-3)
+
+
+def test_single_embedding_query(model_cache):
+    for model_name, expected_result in CANONICAL_QUERY_VALUES.items():
+        if model_name.lower() == "Qdrant/colpali-v1.3-fp16".lower() and os.getenv("CI"):
+            continue  # colpali is too large for ci
+        print("evaluating", model_name)
+        with model_cache(model_name) as model:
+            result = next(iter(model.embed_text(queries)))
+            token_num, abridged_dim = expected_result.shape
+            assert np.allclose(result[:token_num, :abridged_dim], expected_result, atol=2e-3)
+
+
+def test_get_embedding_size():
+    model_name = "Qdrant/colpali-v1.3-fp16"
+    assert LateInteractionMultimodalEmbedding.get_embedding_size(model_name) == 128
+
+    model_name = "Qdrant/ColPali-v1.3-fp16"
+    assert LateInteractionMultimodalEmbedding.get_embedding_size(model_name) == 128
+
+    model_name = "Qdrant/colmodernvbert"
+    assert LateInteractionMultimodalEmbedding.get_embedding_size(model_name) == 128
+
+
+def test_embedding_size():
+    model_name = "Qdrant/colmodernvbert"
+    model = LateInteractionMultimodalEmbedding(model_name=model_name, lazy_load=True)
+    assert model.embedding_size == 128
+
+
+def test_token_count(model_cache) -> None:
+    model_name = "Qdrant/colmodernvbert"
+    with model_cache(model_name) as model:
+        documents = ["short doc", "it is a long document to check attention mask for paddings"]
+        short_doc_token_count = model.token_count(documents[0])
+        long_doc_token_count = model.token_count(documents[1])
+        documents_token_count = model.token_count(documents)
+        assert short_doc_token_count + long_doc_token_count == documents_token_count
+        assert short_doc_token_count + long_doc_token_count == model.token_count(
+            documents, batch_size=1
+        )
+        assert short_doc_token_count + long_doc_token_count < model.token_count(
+            documents, include_extension=True
+        )

tests/test_multi_gpu.py

@@ -1,5 +1,5 @@
 import pytest
-from typing import Optional
+
 from fastembed import (
     TextEmbedding,
     SparseTextEmbedding,
@@ -14,7 +14,7 @@ CACHE_DIR = "../model_cache"
 
 @pytest.mark.skip(reason="Requires a multi-gpu server")
 @pytest.mark.parametrize("device_id", [None, 0, 1])
-def test_gpu_via_providers(device_id: Optional[int]) -> None:
+def test_gpu_via_providers(device_id: int | None) -> None:
     docs = ["hello world", "flag embedding"]
 
     device_id = device_id if device_id is not None else 0
@@ -86,7 +86,7 @@ def test_gpu_via_providers(device_id: Optional[int]) -> None:
 
 @pytest.mark.skip(reason="Requires a multi-gpu server")
 @pytest.mark.parametrize("device_ids", [None, [0], [1], [0, 1]])
-def test_gpu_cuda_device_ids(device_ids: Optional[list[int]]) -> None:
+def test_gpu_cuda_device_ids(device_ids: list[int] | None) -> None:
     docs = ["hello world", "flag embedding"]
     device_id = device_ids[0] if device_ids else 0
     embedding_model = TextEmbedding(
@@ -171,7 +171,7 @@ def test_gpu_cuda_device_ids(device_ids: Optional[list[int]]) -> None:
 @pytest.mark.parametrize(
     "device_ids,parallel", [(None, None), (None, 2), ([1], None), ([1], 1), ([1], 2), ([0, 1], 2)]
 )
-def test_multi_gpu_parallel_inference(device_ids: Optional[list[int]], parallel: int) -> None:
+def test_multi_gpu_parallel_inference(device_ids: list[int] | None, parallel: int) -> None:
     docs = ["hello world", "flag embedding"] * 100
     batch_size = 5
 

tests/test_postprocess.py

@@ -0,0 +1,38 @@
+import numpy as np
+
+from fastembed import LateInteractionTextEmbedding
+from fastembed.postprocess import Muvera
+
+CANONICAL_VALUES = [-2.61810007e-04, 1.89005750e00, -2.32070747e00]
+CANONICAL_QUERY_VALUES = [
+    -0.85783903,
+    1.1077204,
+    -0.09522747,
+]  # part of the values are zeros, should be compared with the result of nonzero mask
+
+DIM = 128
+K_SIM = 5
+DIM_PROJ = 16
+R_REPS = 20
+
+
+def test_single_input():
+    model = LateInteractionTextEmbedding("colbert-ir/colbertv2.0", lazy_load=True)
+    random_generator = np.random.default_rng(42)
+    multivector = random_generator.random((10, 128))
+
+    for muvera in (
+        Muvera(dim=DIM, k_sim=K_SIM, dim_proj=DIM_PROJ, r_reps=R_REPS, random_seed=42),
+        Muvera.from_multivector_model(model, k_sim=K_SIM, dim_proj=DIM_PROJ, r_reps=R_REPS),
+    ):
+        fde = muvera.process(multivector)
+        assert fde.shape[0] == muvera.embedding_size
+        assert np.allclose(fde[:3], CANONICAL_VALUES)
+
+        fde_doc = muvera.process_document(multivector)
+        assert fde_doc.shape[0] == muvera.embedding_size
+        assert np.allclose(fde, fde_doc)
+
+        fde_query = muvera.process_query(multivector)
+        assert fde_query.shape[0] == muvera.embedding_size
+        assert np.allclose(fde_query[np.nonzero(fde_query)][:3], CANONICAL_QUERY_VALUES)

tests/test_preprocessor_utils.py

@@ -0,0 +1,333 @@
+import itertools
+import json
+import os
+import shutil
+
+from pathlib import Path
+from typing import Any
+
+import numpy as np
+import pytest
+
+from tokenizers import Tokenizer
+
+from fastembed.common.preprocessor_utils import load_tokenizer
+from fastembed.text.text_embedding import TextEmbedding
+from tests.utils import delete_model_cache
+
+# transformers writes its VERY_LARGE_INTEGER in place of `model_max_length` when the real
+# value is unknown, which is more than `enable_truncation` can accept
+HF_SENTINEL = int(1e30)
+
+# a lightweight model whose config files serve as a realistic starting point for the cases below
+BASE_MODEL = "BAAI/bge-small-en-v1.5"
+TOKENIZER_FILES = (
+    "config.json",
+    "tokenizer.json",
+    "tokenizer_config.json",
+    "special_tokens_map.json",
+)
+
+
+def _patch_json(path: Path, overrides: dict[str, Any], drop: tuple[str, ...] = ()) -> None:
+    with open(path) as source:
+        content = json.load(source)
+
+    content.update(overrides)
+    for key in drop:
+        content.pop(key, None)
+
+    with open(path, "w") as target:
+        json.dump(content, target)
+
+
+def _set_serialized_padding(path: Path, padding: dict[str, Any] | None) -> None:
+    """Rewrite tokenizer.json through the tokenizers library, so the format stays authoritative."""
+    tokenizer = Tokenizer.from_file(str(path))
+    if padding is None:
+        tokenizer.no_padding()
+    else:
+        tokenizer.enable_padding(**padding)
+    tokenizer.save(str(path))
+
+
+@pytest.fixture(scope="module")
+def make_model_dir(tmp_path_factory):
+    """Build model directories from a real model's config files, with targeted overrides.
+
+    `load_tokenizer` reads only the four files in `TOKENIZER_FILES`, so the onnx weights are
+    never copied.
+    """
+    is_ci = os.getenv("CI")
+    base_model = TextEmbedding(BASE_MODEL)
+    source_dir = Path(base_model.model._model_dir)
+    counter = itertools.count()
+
+    def factory(
+        tokenizer_config: dict[str, Any] | None = None,
+        config: dict[str, Any] | None = None,
+        padding: dict[str, Any] | None = None,
+        drop_from_tokenizer_config: tuple[str, ...] = (),
+        drop_from_config: tuple[str, ...] = (),
+        drop_files: tuple[str, ...] = (),
+    ) -> Path:
+        model_dir = tmp_path_factory.mktemp(f"model_dir_{next(counter)}")
+        for file_name in TOKENIZER_FILES:
+            if file_name in drop_files:
+                continue
+            shutil.copy(source_dir / file_name, model_dir / file_name)
+
+        if "tokenizer_config.json" not in drop_files:
+            _patch_json(
+                model_dir / "tokenizer_config.json",
+                tokenizer_config or {},
+                drop_from_tokenizer_config,
+            )
+        if "config.json" not in drop_files:
+            _patch_json(model_dir / "config.json", config or {}, drop_from_config)
+        if padding is not None:
+            _set_serialized_padding(model_dir / "tokenizer.json", padding)
+
+        return model_dir
+
+    yield factory
+
+    if is_ci:
+        delete_model_cache(base_model.model._model_dir)
+
+
+def test_fixed_padding_is_relaxed_to_batch_longest(make_model_dir) -> None:
+    """Fixed padding shorter than the truncation limit leaves longer encodings ragged."""
+    model_dir = make_model_dir(
+        tokenizer_config={"model_max_length": 512, "max_length": None},
+        padding={"length": 128, "pad_id": 0, "pad_token": "[PAD]", "direction": "right"},
+    )
+
+    tokenizer, _ = load_tokenizer(model_dir)
+
+    assert tokenizer.padding["length"] is None
+    assert tokenizer.truncation["max_length"] == 512
+
+    encoded = tokenizer.encode_batch(["hello world", "retrieval " * 200])
+    # ragged encodings make this raise, the same way onnx_embed does
+    input_ids = np.array([encoding.ids for encoding in encoded])
+
+    assert input_ids.shape == (2, len(encoded[0].ids))
+    assert input_ids.shape[1] > 128
+
+
+def test_batch_longest_padding_does_not_pad_to_the_truncation_limit(make_model_dir) -> None:
+    model_dir = make_model_dir(
+        tokenizer_config={"model_max_length": 512, "max_length": None},
+        padding={"length": 128, "pad_id": 0, "pad_token": "[PAD]", "direction": "right"},
+    )
+
+    tokenizer, _ = load_tokenizer(model_dir)
+    encoded = tokenizer.encode_batch(["hello world", "hello"])
+
+    assert len(encoded[0].ids) == len(encoded[1].ids) < 128
+
+
+def test_serialized_left_padding_is_preserved(make_model_dir) -> None:
+    """ColModernVBERT pads on the left, normalizing the length must not reset the direction."""
+    model_dir = make_model_dir(
+        padding={"length": None, "pad_id": 0, "pad_token": "[PAD]", "direction": "left"},
+    )
+
+    tokenizer, _ = load_tokenizer(model_dir)
+
+    assert tokenizer.padding["direction"] == "left"
+    assert tokenizer.padding["length"] is None
+
+    encoded = tokenizer.encode_batch(["hello world and then some", "hello"])
+    assert encoded[1].ids[0] == 0
+    assert encoded[1].attention_mask[0] == 0
+
+
+def test_serialized_pad_to_multiple_of_is_preserved(make_model_dir) -> None:
+    """Everything the tokenizer declared is kept; only the fixed length is overridden."""
+    model_dir = make_model_dir(
+        padding={"length": 128, "pad_id": 0, "pad_token": "[PAD]", "pad_to_multiple_of": 8},
+    )
+
+    tokenizer, _ = load_tokenizer(model_dir)
+
+    assert tokenizer.padding["length"] is None
+    assert tokenizer.padding["pad_to_multiple_of"] == 8
+
+    encoded = tokenizer.encode_batch(["hello world", "hello"])
+    assert len(encoded[0].ids) % 8 == 0
+
+
+def test_serialized_pad_id_takes_precedence_over_config(make_model_dir) -> None:
+    mask_pad_id = 103  # [MASK] in the bert-base vocab, any id other than the config's works
+    model_dir = make_model_dir(
+        config={"pad_token_id": 7},
+        padding={"length": None, "pad_id": mask_pad_id, "pad_token": "[MASK]"},
+    )
+
+    tokenizer, _ = load_tokenizer(model_dir)
+
+    assert tokenizer.padding["pad_id"] == mask_pad_id
+    assert tokenizer.padding["pad_token"] == "[MASK]"
+
+
+def test_pad_token_falls_back_to_tokenizer_config(make_model_dir) -> None:
+    model_dir = make_model_dir(config={"pad_token_id": 3}, padding=None)
+
+    tokenizer, _ = load_tokenizer(model_dir)
+
+    assert tokenizer.padding["pad_token"] == "[PAD]"
+    assert tokenizer.padding["pad_id"] == 3
+
+
+def test_missing_pad_token_raises(make_model_dir) -> None:
+    model_dir = make_model_dir(drop_from_tokenizer_config=("pad_token",))
+
+    with pytest.raises(ValueError, match="Could not find a pad token"):
+        load_tokenizer(model_dir)
+
+
+@pytest.mark.parametrize(
+    "model_max_length,max_length,expected",
+    [
+        (512, 128, 128),  # both usable, the stricter one wins
+        (128, 512, 128),
+        (512, None, 512),
+        (None, 256, 256),
+        (HF_SENTINEL, 128, 128),  # qdrant/gte-large-onnx
+        (0, 256, 256),
+        (512, 0, 512),
+    ],
+)
+def test_max_context_resolution(make_model_dir, model_max_length, max_length, expected) -> None:
+    model_dir = make_model_dir(
+        tokenizer_config={"model_max_length": model_max_length, "max_length": max_length},
+    )
+
+    tokenizer, _ = load_tokenizer(model_dir)
+
+    assert tokenizer.truncation["max_length"] == expected
+
+
+@pytest.mark.parametrize(
+    "model_max_length,max_length",
+    [
+        (HF_SENTINEL, None),  # transformers' placeholder is not a limit
+        (0, None),  # a zero would truncate everything away
+        (None, 0),
+        (None, None),
+        ("512", None),  # not an integer
+    ],
+)
+def test_unusable_max_context_raises(make_model_dir, model_max_length, max_length) -> None:
+    model_dir = make_model_dir(
+        tokenizer_config={"model_max_length": model_max_length, "max_length": max_length},
+    )
+
+    with pytest.raises(ValueError, match="Could not determine the maximum context length"):
+        load_tokenizer(model_dir)
+
+
+def test_absent_max_context_keys_raise(make_model_dir) -> None:
+    model_dir = make_model_dir(
+        drop_from_tokenizer_config=("model_max_length", "max_length"),
+    )
+
+    with pytest.raises(ValueError, match="Could not determine the maximum context length"):
+        load_tokenizer(model_dir)
+
+
+@pytest.fixture(scope="module")
+def token_id(make_model_dir):
+    """Resolve vocabulary ids by name, so the cases below carry no magic numbers."""
+    return Tokenizer.from_file(str(make_model_dir() / "tokenizer.json")).token_to_id
+
+
+@pytest.mark.parametrize(
+    "dropped",
+    [("config.json",), ("special_tokens_map.json",), ("config.json", "special_tokens_map.json")],
+    ids=["no-config", "no-special-tokens-map", "neither"],
+)
+def test_optional_files_do_not_change_what_is_loaded(make_model_dir, dropped) -> None:
+    """Both files are redundant: everything they carry is already in the tokenizer."""
+    baseline, baseline_specials = load_tokenizer(make_model_dir())
+
+    tokenizer, specials = load_tokenizer(make_model_dir(drop_files=dropped))
+
+    assert specials == baseline_specials
+    assert tokenizer.padding == baseline.padding
+    assert tokenizer.encode("hello world").ids == baseline.encode("hello world").ids
+
+
+@pytest.mark.parametrize("missing", ("tokenizer.json", "tokenizer_config.json"))
+def test_the_remaining_files_are_still_required(make_model_dir, missing) -> None:
+    """Relaxing the optional two must not relax the two that carry irreplaceable data."""
+    model_dir = make_model_dir(drop_files=(missing,))
+
+    with pytest.raises(ValueError, match=f"Could not find {missing}"):
+        load_tokenizer(model_dir)
+
+
+@pytest.mark.parametrize(
+    "model_files",
+    [
+        pytest.param({"drop_from_config": ("pad_token_id",)}, id="config-omits-pad-token-id"),
+        pytest.param({"drop_files": ("config.json",)}, id="config-is-absent"),
+    ],
+)
+def test_pad_id_falls_back_to_the_vocabulary(make_model_dir, token_id, model_files) -> None:
+    """Last link of the chain; a hardcoded 0 would silently disagree with `pad_token`."""
+    expected = token_id("[SEP]")
+    assert expected != 0, "a pad token whose id is 0 would pass even without a lookup"
+    model_dir = make_model_dir(tokenizer_config={"pad_token": "[SEP]"}, **model_files)
+
+    tokenizer, _ = load_tokenizer(model_dir)
+
+    assert tokenizer.padding["pad_id"] == expected
+
+
+def test_pad_token_that_resolves_nowhere_raises(make_model_dir) -> None:
+    """Without config.json a pad token outside the vocabulary has no id left to fall back on."""
+    model_dir = make_model_dir(
+        tokenizer_config={"pad_token": "[NOT_IN_VOCAB]"},
+        drop_files=("config.json",),
+    )
+
+    with pytest.raises(ValueError, match="Could not resolve an id for the pad token"):
+        load_tokenizer(model_dir)
+
+
+def test_pad_token_named_only_in_the_map_resolves(make_model_dir) -> None:
+    """The map is read first, so it can name a pad token tokenizer.json does not carry."""
+    model_dir = make_model_dir(
+        tokenizer_config={"pad_token": "<|mypad|>"},
+        drop_files=("config.json",),
+    )
+    _patch_json(model_dir / "special_tokens_map.json", {"pad_token": "<|mypad|>"})
+
+    tokenizer, specials = load_tokenizer(model_dir)
+
+    assert tokenizer.padding["pad_token"] == "<|mypad|>"
+    assert tokenizer.padding["pad_id"] == specials["<|mypad|>"]
+
+
+@pytest.mark.parametrize(
+    "additional",
+    [
+        pytest.param(["<|list_str|>"], id="list-of-strings"),
+        pytest.param([{"content": "<|list_str|>"}], id="list-of-added-token-dicts"),
+    ],
+)
+def test_list_valued_map_entries_are_registered(make_model_dir, additional) -> None:
+    """`additional_special_tokens` holds a list, which the str/dict dispatch alone drops.
+
+    Real repos ship both spellings, and their tokens are in tokenizer.json already, so
+    only a token living nowhere else shows the drop.
+    """
+    model_dir = make_model_dir()
+    _patch_json(model_dir / "special_tokens_map.json", {"additional_special_tokens": additional})
+
+    _, specials = load_tokenizer(model_dir)
+
+    assert "<|list_str|>" in specials

tests/test_sparse_embeddings.py

@@ -1,14 +1,16 @@
 import os
+from contextlib import contextmanager
 
 import pytest
 import numpy as np
 
 from fastembed.sparse.bm25 import Bm25
 from fastembed.sparse.sparse_text_embedding import SparseTextEmbedding
-from tests.utils import delete_model_cache
+from tests.utils import delete_model_cache, should_test_model
+
 
 CANONICAL_COLUMN_VALUES = {
-    "prithvida/Splade_PP_en_v1": {
+    "prithivida/Splade_PP_en_v1": {
         "indices": [
             2040,
             2047,
@@ -43,112 +45,249 @@ CANONICAL_COLUMN_VALUES = {
             2.1904349327087402,
             1.0531445741653442,
         ],
-    }
+    },
+    "Qdrant/minicoil-v1": {
+        "indices": [80, 81, 82, 83, 6664, 6665, 6666, 6667],
+        "values": [
+            0.52634597,
+            0.8711344,
+            1.2264385,
+            0.52123857,
+            0.974713,
+            -0.97803956,
+            -0.94312465,
+            -0.12508166,
+        ],
+    },
+    # first 15 non-zero dimensions of the embedding
+    "opensearch-project/opensearch-neural-sparse-encoding-doc-v3-gte": {
+        "indices": [
+            999,
+            1010,
+            1011,
+            1024,
+            1028,
+            1029,
+            1045,
+            1074,
+            1993,
+            2017,
+            2033,
+            2054,
+            2073,
+            2080,
+            2088,
+        ],
+        "values": [
+            0.16544909,
+            0.00529129,
+            0.0392109,
+            0.12337475,
+            0.09640586,
+            0.05325737,
+            0.09611791,
+            0.03159865,
+            0.01349991,
+            0.09392473,
+            0.01928805,
+            0.05238346,
+            0.05515401,
+            0.03156782,
+            0.98263124,
+        ],
+    },
 }
 
+CANONICAL_QUERY_VALUES = {
+    "opensearch-project/opensearch-neural-sparse-encoding-doc-v3-gte": {
+        "indices": [2088, 7592],
+        "values": [3.42086864, 6.93775654],
+    },
+    "Qdrant/minicoil-v1": {
+        "indices": [80, 81, 82, 83, 6664, 6665, 6666, 6667],
+        "values": [
+            0.31389374,
+            0.5195128,
+            0.7314033,
+            0.3108479,
+            0.5812834,
+            -0.5832673,
+            -0.5624452,
+            -0.0745942,
+        ],
+    },
+}
+
+
+_MODELS_TO_CACHE = (
+    "prithivida/Splade_PP_en_v1",
+    "Qdrant/minicoil-v1",
+    "Qdrant/bm25",
+    "Qdrant/bm42-all-minilm-l6-v2-attentions",
+    "opensearch-project/opensearch-neural-sparse-encoding-doc-v3-gte",
+)
+MODELS_TO_CACHE = tuple([x.lower() for x in _MODELS_TO_CACHE])
+
+
+@pytest.fixture(scope="module")
+def model_cache():
+    is_ci = os.getenv("CI")
+    cache = {}
+
+    @contextmanager
+    def get_model(model_name: str):
+        lowercase_model_name = model_name.lower()
+        if lowercase_model_name not in cache:
+            cache[lowercase_model_name] = SparseTextEmbedding(lowercase_model_name)
+        yield cache[lowercase_model_name]
+        if lowercase_model_name not in MODELS_TO_CACHE:
+            model_inst = cache.pop(lowercase_model_name)
+            if is_ci:
+                delete_model_cache(model_inst.model._model_dir)
+            del model_inst
+
+    yield get_model
+
+    if is_ci:
+        for name, model in cache.items():
+            delete_model_cache(model.model._model_dir)
+    cache.clear()
+
+
 docs = ["Hello World"]
 
 
-def test_batch_embedding() -> None:
-    is_ci = os.getenv("CI")
+@pytest.mark.parametrize(
+    "model_name",
+    ["prithivida/Splade_PP_en_v1", "Qdrant/minicoil-v1"],
+)
+def test_batch_embedding(model_cache, model_name: str) -> None:
     docs_to_embed = docs * 10
 
-    for model_name, expected_result in CANONICAL_COLUMN_VALUES.items():
-        model = SparseTextEmbedding(model_name=model_name)
+    with model_cache(model_name) as model:
         result = next(iter(model.embed(docs_to_embed, batch_size=6)))
+        expected_result = CANONICAL_COLUMN_VALUES[model_name]
         assert result.indices.tolist() == expected_result["indices"]
 
         for i, value in enumerate(result.values):
             assert pytest.approx(value, abs=0.001) == expected_result["values"][i]
-        if is_ci:
-            delete_model_cache(model.model._model_dir)
 
 
-def test_single_embedding() -> None:
+def test_single_embedding(model_cache) -> None:
     is_ci = os.getenv("CI")
-    for model_name, expected_result in CANONICAL_COLUMN_VALUES.items():
-        model = SparseTextEmbedding(model_name=model_name)
+    is_manual = os.getenv("GITHUB_EVENT_NAME") == "workflow_dispatch"
 
-        passage_result = next(iter(model.embed(docs, batch_size=6)))
-        query_result = next(iter(model.query_embed(docs)))
-        for result in [passage_result, query_result]:
-            assert result.indices.tolist() == expected_result["indices"]
+    for model_desc in SparseTextEmbedding._list_supported_models():
+        if (
+            model_desc.model not in CANONICAL_COLUMN_VALUES
+        ):  # attention models and bm25 are also parts of
+            # SparseTextEmbedding, however, they have their own tests
+            continue
+        if not should_test_model(model_desc, model_desc.model, is_ci, is_manual):
+            continue
+        with model_cache(model_desc.model) as model:
+            passage_result = next(iter(model.embed(docs, batch_size=6)))
+            query_result = next(iter(model.query_embed(docs)))
+            expected_result = CANONICAL_COLUMN_VALUES[model_desc.model]
+            expected_query_result = CANONICAL_QUERY_VALUES.get(model_desc.model, expected_result)
 
-            for i, value in enumerate(result.values):
+            # canonical values might contain only a prefix of the non-zero dimensions
+            num_dims = len(expected_result["indices"])
+            assert passage_result.indices.tolist()[:num_dims] == expected_result["indices"]
+            for i, value in enumerate(passage_result.values[:num_dims]):
                 assert pytest.approx(value, abs=0.001) == expected_result["values"][i]
-        if is_ci:
-            delete_model_cache(model.model._model_dir)
+
+            num_query_dims = len(expected_query_result["indices"])
+            assert (
+                query_result.indices.tolist()[:num_query_dims] == expected_query_result["indices"]
+            )
+            for i, value in enumerate(query_result.values[:num_query_dims]):
+                assert pytest.approx(value, abs=0.001) == expected_query_result["values"][i]
 
 
-def test_parallel_processing() -> None:
-    is_ci = os.getenv("CI")
-    model = SparseTextEmbedding(model_name="prithivida/Splade_PP_en_v1")
-    docs = ["hello world", "flag embedding"] * 30
-    sparse_embeddings_duo = list(model.embed(docs, batch_size=10, parallel=2))
-    sparse_embeddings_all = list(model.embed(docs, batch_size=10, parallel=0))
-    sparse_embeddings = list(model.embed(docs, batch_size=10, parallel=None))
+@pytest.mark.parametrize(
+    "model_name",
+    ["prithivida/Splade_PP_en_v1", "Qdrant/minicoil-v1"],
+)
+def test_parallel_processing(model_cache, model_name: str) -> None:
+    with model_cache(model_name) as model:
+        docs = ["hello world", "flag embedding"] * 30
+        sparse_embeddings_duo = list(model.embed(docs, batch_size=10, parallel=2))
+        # sparse_embeddings_all = list(model.embed(docs, batch_size=10, parallel=0))  # inherits OnnxTextModel which
+        # is tested in TextEmbedding, disabling it here to reduce number of requests to hf
+        # multiprocessing is enough to test with `parallel=2`, and `parallel=None` is okay to tests since it reuses
+        # model from cache
+        sparse_embeddings = list(model.embed(docs, batch_size=10, parallel=None))
 
-    assert (
-        len(sparse_embeddings)
-        == len(sparse_embeddings_duo)
-        == len(sparse_embeddings_all)
-        == len(docs)
-    )
-
-    for sparse_embedding, sparse_embedding_duo, sparse_embedding_all in zip(
-        sparse_embeddings, sparse_embeddings_duo, sparse_embeddings_all
-    ):
         assert (
-            sparse_embedding.indices.tolist()
-            == sparse_embedding_duo.indices.tolist()
-            == sparse_embedding_all.indices.tolist()
+            len(sparse_embeddings)
+            == len(sparse_embeddings_duo)
+            # == len(sparse_embeddings_all)
+            == len(docs)
         )
-        assert np.allclose(sparse_embedding.values, sparse_embedding_duo.values, atol=1e-3)
-        assert np.allclose(sparse_embedding.values, sparse_embedding_all.values, atol=1e-3)
 
-    if is_ci:
-        delete_model_cache(model.model._model_dir)
+        for (
+            sparse_embedding,
+            sparse_embedding_duo,
+            # sparse_embedding_all
+        ) in zip(
+            sparse_embeddings,
+            sparse_embeddings_duo,
+            # sparse_embeddings_all
+        ):
+            assert (
+                sparse_embedding.indices.tolist() == sparse_embedding_duo.indices.tolist()
+                # == sparse_embedding_all.indices.tolist()
+            )
+            assert np.allclose(sparse_embedding.values, sparse_embedding_duo.values, atol=1e-3)
+            # assert np.allclose(sparse_embedding.values, sparse_embedding_all.values, atol=1e-3)
 
 
-@pytest.fixture
-def bm25_instance() -> None:
-    ci = os.getenv("CI", True)
-    model = Bm25("Qdrant/bm25", language="english")
-    yield model
-    if ci:
-        delete_model_cache(model._model_dir)
+def test_stem_with_stopwords_and_punctuation(model_cache) -> None:
+    with model_cache("Qdrant/bm25") as model:
+        bm25_instance = model.model
+        # Setup
+        original_stopwords = bm25_instance.stopwords.copy()
+        original_punctuation = bm25_instance.punctuation.copy()
+
+        bm25_instance.stopwords = {"the", "is", "a"}
+        bm25_instance.punctuation = {".", ",", "!"}
+
+        # Test data
+        tokens = ["The", "quick", "brown", "fox", "is", "a", "test", "sentence", ".", "!"]
+
+        # Execute
+        result = bm25_instance._stem(tokens)
+
+        # Assert
+        expected = ["quick", "brown", "fox", "test", "sentenc"]
+        assert result == expected, f"Expected {expected}, but got {result}"
+
+        bm25_instance.stopwords = original_stopwords
+        bm25_instance.punctuation = original_punctuation
 
 
-def test_stem_with_stopwords_and_punctuation(bm25_instance: Bm25) -> None:
-    # Setup
-    bm25_instance.stopwords = {"the", "is", "a"}
-    bm25_instance.punctuation = {".", ",", "!"}
+def test_stem_case_insensitive_stopwords(model_cache) -> None:
+    with model_cache("Qdrant/bm25") as model:
+        bm25_instance = model.model
+        original_stopwords = bm25_instance.stopwords.copy()
+        original_punctuation = bm25_instance.punctuation.copy()
 
-    # Test data
-    tokens = ["The", "quick", "brown", "fox", "is", "a", "test", "sentence", ".", "!"]
+        # Setup
+        bm25_instance.stopwords = {"the", "is", "a"}
+        bm25_instance.punctuation = {".", ",", "!"}
 
-    # Execute
-    result = bm25_instance._stem(tokens)
+        # Test data
+        tokens = ["THE", "Quick", "Brown", "Fox", "IS", "A", "Test", "Sentence", ".", "!"]
 
-    # Assert
-    expected = ["quick", "brown", "fox", "test", "sentenc"]
-    assert result == expected, f"Expected {expected}, but got {result}"
+        # Execute
+        result = bm25_instance._stem(tokens)
 
-
-def test_stem_case_insensitive_stopwords(bm25_instance: Bm25) -> None:
-    # Setup
-    bm25_instance.stopwords = {"the", "is", "a"}
-    bm25_instance.punctuation = {".", ",", "!"}
-
-    # Test data
-    tokens = ["THE", "Quick", "Brown", "Fox", "IS", "A", "Test", "Sentence", ".", "!"]
-
-    # Execute
-    result = bm25_instance._stem(tokens)
-
-    # Assert
-    expected = ["quick", "brown", "fox", "test", "sentenc"]
-    assert result == expected, f"Expected {expected}, but got {result}"
+        # Assert
+        expected = ["quick", "brown", "fox", "test", "sentenc"]
+        assert result == expected, f"Expected {expected}, but got {result}"
+        bm25_instance.stopwords = original_stopwords
+        bm25_instance.punctuation = original_punctuation
 
 
 @pytest.mark.parametrize("disable_stemmer", [True, False])
@@ -172,10 +311,23 @@ def test_disable_stemmer_behavior(disable_stemmer: bool) -> None:
     assert result == expected, f"Expected {expected}, but got {result}"
 
 
-@pytest.mark.parametrize(
-    "model_name",
-    ["prithivida/Splade_PP_en_v1"],
-)
+def test_if_splade_query_embed_is_inference_free() -> None:
+    is_ci = os.getenv("CI")
+    model = SparseTextEmbedding(
+        model_name="opensearch-project/opensearch-neural-sparse-encoding-doc-v3-gte",
+        lazy_load=True,
+    )
+    embeddings = list(model.query_embed(["hello world", "flag embedding"]))
+    # queries are embedded with a tokenizer and an idf lookup table only,
+    # the onnx model must stay unloaded
+    assert not hasattr(model.model, "model")
+    assert all(len(embedding.indices) > 0 for embedding in embeddings)
+
+    if is_ci:
+        delete_model_cache(model.model._model_dir)
+
+
+@pytest.mark.parametrize("model_name", ["prithivida/Splade_PP_en_v1"])
 def test_lazy_load(model_name: str) -> None:
     is_ci = os.getenv("CI")
     model = SparseTextEmbedding(model_name=model_name, lazy_load=True)
@@ -193,3 +345,42 @@ def test_lazy_load(model_name: str) -> None:
 
     if is_ci:
         delete_model_cache(model.model._model_dir)
+
+
+@pytest.mark.parametrize(
+    "model_name",
+    [
+        "prithivida/Splade_PP_en_v1",
+        "Qdrant/minicoil-v1",
+        "Qdrant/bm42-all-minilm-l6-v2-attentions",
+    ],
+)
+def test_session_options(model_cache, model_name) -> None:
+    with model_cache(model_name) as default_model:
+        default_session_options = default_model.model.model.get_session_options()
+        assert default_session_options.enable_cpu_mem_arena is True
+        model = SparseTextEmbedding(model_name=model_name, enable_cpu_mem_arena=False)
+        session_options = model.model.model.get_session_options()
+        assert session_options.enable_cpu_mem_arena is False
+
+
+@pytest.mark.parametrize(
+    "model_name",
+    [
+        "prithivida/Splade_PP_en_v1",
+        "Qdrant/minicoil-v1",
+        "Qdrant/bm42-all-minilm-l6-v2-attentions",
+        "Qdrant/bm25",
+    ],
+)
+def test_token_count(model_cache, model_name) -> None:
+    with model_cache(model_name) as model:
+        documents = [
+            "Name me a couple of cities were the capitals of Germany?",
+            "Berlin is the current capital of Germany, Bonn is a former capital of Germany.",
+        ]
+        first_doc_token_count = model.token_count(documents[0])
+        second_doc_token_count = model.token_count(documents[1])
+        doc_token_count = model.token_count(documents)
+        assert first_doc_token_count + second_doc_token_count == doc_token_count
+        assert doc_token_count == model.token_count(documents, batch_size=1)

tests/test_text_cross_encoder.py

@@ -1,10 +1,11 @@
 import os
+from contextlib import contextmanager
 
 import numpy as np
 import pytest
 
 from fastembed.rerank.cross_encoder import TextCrossEncoder
-from tests.utils import delete_model_cache
+from tests.utils import delete_model_cache, should_test_model
 
 CANONICAL_SCORE_VALUES = {
     "Xenova/ms-marco-MiniLM-L-6-v2": np.array([8.500708, -2.541011]),
@@ -15,73 +16,84 @@ CANONICAL_SCORE_VALUES = {
     "jinaai/jina-reranker-v2-base-multilingual": np.array([1.6533, -1.6455]),
 }
 
-SELECTED_MODELS = {
-    "Xenova": "Xenova/ms-marco-MiniLM-L-6-v2",
-    "BAAI": "BAAI/bge-reranker-base",
-    "jinaai": "jinaai/jina-reranker-v1-tiny-en",
-}
+
+_MODELS_TO_CACHE = ("Xenova/ms-marco-MiniLM-L-6-v2",)
+MODELS_TO_CACHE = tuple([x.lower() for x in _MODELS_TO_CACHE])
 
 
-@pytest.mark.parametrize(
-    "model_name",
-    [model_name for model_name in CANONICAL_SCORE_VALUES],
-)
-def test_rerank(model_name: str) -> None:
+@pytest.fixture(scope="module")
+def model_cache():
     is_ci = os.getenv("CI")
+    cache = {}
 
-    model = TextCrossEncoder(model_name=model_name)
+    @contextmanager
+    def get_model(model_name: str):
+        lowercase_model_name = model_name.lower()
+        if lowercase_model_name not in cache:
+            cache[lowercase_model_name] = TextCrossEncoder(lowercase_model_name)
+        yield cache[lowercase_model_name]
+        if lowercase_model_name not in MODELS_TO_CACHE:
+            model_inst = cache.pop(lowercase_model_name)
+            if is_ci:
+                delete_model_cache(model_inst.model._model_dir)
+            del model_inst
 
-    query = "What is the capital of France?"
-    documents = ["Paris is the capital of France.", "Berlin is the capital of Germany."]
-    scores = np.array(list(model.rerank(query, documents)))
+    yield get_model
 
-    pairs = [(query, doc) for doc in documents]
-    scores2 = np.array(list(model.rerank_pairs(pairs)))
-    assert np.allclose(
-        scores, scores2, atol=1e-5
-    ), f"Model: {model_name}, Scores: {scores}, Scores2: {scores2}"
-
-    canonical_scores = CANONICAL_SCORE_VALUES[model_name]
-    assert np.allclose(
-        scores, canonical_scores, atol=1e-3
-    ), f"Model: {model_name}, Scores: {scores}, Expected: {canonical_scores}"
     if is_ci:
-        delete_model_cache(model.model._model_dir)
+        for name, model in cache.items():
+            delete_model_cache(model.model._model_dir)
+    cache.clear()
 
 
-@pytest.mark.parametrize(
-    "model_name",
-    [model_name for model_name in SELECTED_MODELS.values()],
-)
-def test_batch_rerank(model_name: str) -> None:
+@pytest.mark.parametrize("model_name", ["Xenova/ms-marco-MiniLM-L-6-v2"])
+def test_rerank(model_cache, model_name: str) -> None:
     is_ci = os.getenv("CI")
+    is_manual = os.getenv("GITHUB_EVENT_NAME") == "workflow_dispatch"
 
-    model = TextCrossEncoder(model_name=model_name)
+    for model_desc in TextCrossEncoder._list_supported_models():
+        if not should_test_model(model_desc, model_name, is_ci, is_manual):
+            continue
 
-    query = "What is the capital of France?"
-    documents = ["Paris is the capital of France.", "Berlin is the capital of Germany."] * 50
-    scores = np.array(list(model.rerank(query, documents, batch_size=10)))
+        with model_cache(model_desc.model) as model:
+            query = "What is the capital of France?"
+            documents = ["Paris is the capital of France.", "Berlin is the capital of Germany."]
+            scores = np.array(list(model.rerank(query, documents)))
 
-    pairs = [(query, doc) for doc in documents]
-    scores2 = np.array(list(model.rerank_pairs(pairs)))
-    assert np.allclose(
-        scores, scores2, atol=1e-5
-    ), f"Model: {model_name}, Scores: {scores}, Scores2: {scores2}"
+            pairs = [(query, doc) for doc in documents]
+            scores2 = np.array(list(model.rerank_pairs(pairs)))
+            assert np.allclose(
+                scores, scores2, atol=1e-5
+            ), f"Model: {model_desc.model}, Scores: {scores}, Scores2: {scores2}"
 
-    canonical_scores = np.tile(CANONICAL_SCORE_VALUES[model_name], 50)
-
-    assert scores.shape == canonical_scores.shape, f"Unexpected shape for model {model_name}"
-    assert np.allclose(
-        scores, canonical_scores, atol=1e-3
-    ), f"Model: {model_name}, Scores: {scores}, Expected: {canonical_scores}"
-    if is_ci:
-        delete_model_cache(model.model._model_dir)
+            canonical_scores = CANONICAL_SCORE_VALUES[model_desc.model]
+            assert np.allclose(
+                scores, canonical_scores, atol=1e-3
+            ), f"Model: {model_desc.model}, Scores: {scores}, Expected: {canonical_scores}"
 
 
-@pytest.mark.parametrize(
-    "model_name",
-    ["Xenova/ms-marco-MiniLM-L-6-v2"],
-)
+@pytest.mark.parametrize("model_name", ["Xenova/ms-marco-MiniLM-L-6-v2"])
+def test_batch_rerank(model_cache, model_name: str) -> None:
+    with model_cache(model_name) as model:
+        query = "What is the capital of France?"
+        documents = ["Paris is the capital of France.", "Berlin is the capital of Germany."] * 50
+        scores = np.array(list(model.rerank(query, documents, batch_size=10)))
+
+        pairs = [(query, doc) for doc in documents]
+        scores2 = np.array(list(model.rerank_pairs(pairs)))
+        assert np.allclose(
+            scores, scores2, atol=1e-5
+        ), f"Model: {model_name}, Scores: {scores}, Scores2: {scores2}"
+
+        canonical_scores = np.tile(CANONICAL_SCORE_VALUES[model_name], 50)
+
+        assert scores.shape == canonical_scores.shape, f"Unexpected shape for model {model_name}"
+        assert np.allclose(
+            scores, canonical_scores, atol=1e-3
+        ), f"Model: {model_name}, Scores: {scores}, Expected: {canonical_scores}"
+
+
+@pytest.mark.parametrize("model_name", ["Xenova/ms-marco-MiniLM-L-6-v2"])
 def test_lazy_load(model_name: str) -> None:
     is_ci = os.getenv("CI")
     model = TextCrossEncoder(model_name=model_name, lazy_load=True)
@@ -95,25 +107,45 @@ def test_lazy_load(model_name: str) -> None:
         delete_model_cache(model.model._model_dir)
 
 
-@pytest.mark.parametrize(
-    "model_name",
-    [model_name for model_name in SELECTED_MODELS.values()],
-)
-def test_rerank_pairs_parallel(model_name: str) -> None:
-    is_ci = os.getenv("CI")
+@pytest.mark.parametrize("model_name", ["Xenova/ms-marco-MiniLM-L-6-v2"])
+def test_rerank_pairs_parallel(model_cache, model_name: str) -> None:
+    with model_cache(model_name) as model:
+        query = "What is the capital of France?"
+        documents = ["Paris is the capital of France.", "Berlin is the capital of Germany."] * 10
+        pairs = [(query, doc) for doc in documents]
+        scores_parallel = np.array(list(model.rerank_pairs(pairs, parallel=2, batch_size=10)))
+        scores_sequential = np.array(list(model.rerank_pairs(pairs, batch_size=10)))
+        assert np.allclose(
+            scores_parallel, scores_sequential, atol=1e-5
+        ), f"Model: {model_name}, Scores (Parallel): {scores_parallel}, Scores (Sequential): {scores_sequential}"
+        canonical_scores = CANONICAL_SCORE_VALUES[model_name]
+        assert np.allclose(
+            scores_parallel[: len(canonical_scores)], canonical_scores, atol=1e-3
+        ), f"Model: {model_name}, Scores (Parallel): {scores_parallel}, Expected: {canonical_scores}"
 
-    model = TextCrossEncoder(model_name=model_name)
-    query = "What is the capital of France?"
-    documents = ["Paris is the capital of France.", "Berlin is the capital of Germany."] * 10
-    pairs = [(query, doc) for doc in documents]
-    scores_parallel = np.array(list(model.rerank_pairs(pairs, parallel=2, batch_size=10)))
-    scores_sequential = np.array(list(model.rerank_pairs(pairs, batch_size=10)))
-    assert np.allclose(
-        scores_parallel, scores_sequential, atol=1e-5
-    ), f"Model: {model_name}, Scores (Parallel): {scores_parallel}, Scores (Sequential): {scores_sequential}"
-    canonical_scores = CANONICAL_SCORE_VALUES[model_name]
-    assert np.allclose(
-        scores_parallel[: len(canonical_scores)], canonical_scores, atol=1e-3
-    ), f"Model: {model_name}, Scores (Parallel): {scores_parallel}, Expected: {canonical_scores}"
-    if is_ci:
-        delete_model_cache(model.model._model_dir)
+
+@pytest.mark.parametrize("model_name", ["Xenova/ms-marco-MiniLM-L-6-v2"])
+def test_token_count(model_cache, model_name: str) -> None:
+    with model_cache(model_name) as model:
+        pairs = [
+            ("What is the capital of France?", "Paris is the capital of France."),
+            (
+                "Name me a couple of cities were the capitals of Germany?",
+                "Berlin is the current capital of Germany, Bonn is a former capital of Germany.",
+            ),
+        ]
+        first_pair_token_count = model.token_count([pairs[0]])
+        second_pair_token_count = model.token_count([pairs[1]])
+        pairs_token_count = model.token_count(pairs)
+        assert first_pair_token_count + second_pair_token_count == pairs_token_count
+        assert pairs_token_count == model.token_count(pairs, batch_size=1)
+
+
+@pytest.mark.parametrize("model_name", ["Xenova/ms-marco-MiniLM-L-6-v2"])
+def test_session_options(model_cache, model_name) -> None:
+    with model_cache(model_name) as default_model:
+        default_session_options = default_model.model.model.get_session_options()
+        assert default_session_options.enable_cpu_mem_arena is True
+        model = TextCrossEncoder(model_name=model_name, enable_cpu_mem_arena=False)
+        session_options = model.model.model.get_session_options()
+        assert session_options.enable_cpu_mem_arena is False

tests/test_text_multitask_embeddings.py

@@ -4,7 +4,7 @@ import numpy as np
 import pytest
 
 from fastembed import TextEmbedding
-from fastembed.text.multitask_embedding import Task
+from fastembed.text.multitask_embedding import JinaEmbeddingV3, Task
 from tests.utils import delete_model_cache
 
 
@@ -60,52 +60,43 @@ CANONICAL_VECTOR_VALUES = {
 docs = ["Hello World", "Follow the white rabbit."]
 
 
-def test_batch_embedding():
+@pytest.mark.parametrize("dim,model_name", [(1024, "jinaai/jina-embeddings-v3")])
+def test_batch_embedding(dim: int, model_name: str):
     is_ci = os.getenv("CI")
+    is_manual = os.getenv("GITHUB_EVENT_NAME") == "workflow_dispatch"
+    if is_ci and not is_manual:
+        pytest.skip("Skipping multitask models in CI non-manual mode")
+
     docs_to_embed = docs * 10
     default_task = Task.RETRIEVAL_PASSAGE
 
-    for model_desc in TextEmbedding._list_supported_models():
-        if not is_ci and model_desc.size_in_GB > 1:
-            continue
+    model = TextEmbedding(model_name=model_name)
 
-        model_name = model_desc.model
-        dim = model_desc.dim
+    embeddings = list(model.embed(documents=docs_to_embed, batch_size=6))
+    embeddings = np.stack(embeddings, axis=0)
 
-        if model_name not in CANONICAL_VECTOR_VALUES.keys():
-            continue
+    assert embeddings.shape == (len(docs_to_embed), dim)
 
-        model = TextEmbedding(model_name=model_name)
+    canonical_vector = CANONICAL_VECTOR_VALUES[model_name][default_task]["vectors"]
+    assert np.allclose(
+        embeddings[: len(docs), : canonical_vector.shape[1]], canonical_vector, atol=1e-4
+    ), model_name
 
-        print(f"evaluating {model_name} default task")
-
-        embeddings = list(model.embed(documents=docs_to_embed, batch_size=6))
-        embeddings = np.stack(embeddings, axis=0)
-
-        assert embeddings.shape == (len(docs_to_embed), dim)
-
-        canonical_vector = CANONICAL_VECTOR_VALUES[model_name][default_task]["vectors"]
-        assert np.allclose(
-            embeddings[: len(docs), : canonical_vector.shape[1]], canonical_vector, atol=1e-4
-        ), model_desc.model
-
-        if is_ci:
-            delete_model_cache(model.model._model_dir)
+    if is_ci:
+        delete_model_cache(model.model._model_dir)
 
 
 def test_single_embedding():
     is_ci = os.getenv("CI")
+    is_manual = os.getenv("GITHUB_EVENT_NAME") == "workflow_dispatch"
+    if is_ci and not is_manual:
+        pytest.skip("Skipping multitask models in CI non-manual mode")
 
-    for model_desc in TextEmbedding._list_supported_models():
-        if not is_ci and model_desc.size_in_GB > 1:
-            continue
-
+    for model_desc in JinaEmbeddingV3._list_supported_models():
+        # todo: once we add more models, we should not test models >1GB size locally
         model_name = model_desc.model
         dim = model_desc.dim
 
-        if model_name not in CANONICAL_VECTOR_VALUES.keys():
-            continue
-
         model = TextEmbedding(model_name=model_name)
 
         for task in CANONICAL_VECTOR_VALUES[model_name]:
@@ -118,27 +109,42 @@ def test_single_embedding():
 
             canonical_vector = task["vectors"]
             assert np.allclose(
-                embeddings[: len(docs), : canonical_vector.shape[1]], canonical_vector, atol=1e-4
+                embeddings[:, : canonical_vector.shape[1]], canonical_vector, atol=1e-4
             ), model_desc.model
 
+        classification_embeddings = list(model.embed(documents=docs, task_id=Task.CLASSIFICATION))
+        classification_embeddings = np.stack(classification_embeddings, axis=0)
+
+        assert classification_embeddings.shape == (len(docs), dim)
+
+        model = TextEmbedding(model_name=model_name, task_id=Task.CLASSIFICATION)
+        default_embeddings = list(model.embed(documents=docs))
+        default_embeddings = np.stack(default_embeddings, axis=0)
+
+        assert default_embeddings.shape == (len(docs), dim)
+
+        assert np.allclose(
+            classification_embeddings,
+            default_embeddings,
+            atol=1e-4,
+        ), model_desc.model
         if is_ci:
             delete_model_cache(model.model._model_dir)
 
 
 def test_single_embedding_query():
     is_ci = os.getenv("CI")
+    is_manual = os.getenv("GITHUB_EVENT_NAME") == "workflow_dispatch"
+    if is_ci and not is_manual:
+        pytest.skip("Skipping multitask models in CI non-manual mode")
+
     task_id = Task.RETRIEVAL_QUERY
 
-    for model_desc in TextEmbedding._list_supported_models():
-        if not is_ci and model_desc.size_in_GB > 1:
-            continue
-
+    for model_desc in JinaEmbeddingV3._list_supported_models():
+        # todo: once we add more models, we should not test models >1GB size locally
         model_name = model_desc.model
         dim = model_desc.dim
 
-        if model_name not in CANONICAL_VECTOR_VALUES.keys():
-            continue
-
         model = TextEmbedding(model_name=model_name)
 
         print(f"evaluating {model_name} query_embed task_id: {task_id}")
@@ -150,7 +156,7 @@ def test_single_embedding_query():
 
         canonical_vector = CANONICAL_VECTOR_VALUES[model_name][task_id]["vectors"]
         assert np.allclose(
-            embeddings[: len(docs), : canonical_vector.shape[1]], canonical_vector, atol=1e-4
+            embeddings[:, : canonical_vector.shape[1]], canonical_vector, atol=1e-4
         ), model_desc.model
 
         if is_ci:
@@ -159,18 +165,18 @@ def test_single_embedding_query():
 
 def test_single_embedding_passage():
     is_ci = os.getenv("CI")
+    is_manual = os.getenv("GITHUB_EVENT_NAME") == "workflow_dispatch"
+    if is_ci and not is_manual:
+        pytest.skip("Skipping multitask models in CI non-manual mode")
+
     task_id = Task.RETRIEVAL_PASSAGE
 
-    for model_desc in TextEmbedding._list_supported_models():
-        if not is_ci and model_desc.size_in_GB > 1:
-            continue
+    for model_desc in JinaEmbeddingV3._list_supported_models():
+        # todo: once we add more models, we should not test models >1GB size locally
 
         model_name = model_desc.model
         dim = model_desc.dim
 
-        if model_name not in CANONICAL_VECTOR_VALUES.keys():
-            continue
-
         model = TextEmbedding(model_name=model_name)
 
         print(f"evaluating {model_name} passage_embed task_id: {task_id}")
@@ -182,21 +188,22 @@ def test_single_embedding_passage():
 
         canonical_vector = CANONICAL_VECTOR_VALUES[model_name][task_id]["vectors"]
         assert np.allclose(
-            embeddings[: len(docs), : canonical_vector.shape[1]], canonical_vector, atol=1e-4
+            embeddings[:, : canonical_vector.shape[1]], canonical_vector, atol=1e-4
         ), model_desc.model
 
         if is_ci:
             delete_model_cache(model.model._model_dir)
 
 
-def test_parallel_processing():
+@pytest.mark.parametrize("dim,model_name", [(1024, "jinaai/jina-embeddings-v3")])
+def test_parallel_processing(dim: int, model_name: str):
     is_ci = os.getenv("CI")
+    is_manual = os.getenv("GITHUB_EVENT_NAME") == "workflow_dispatch"
+    if is_ci and not is_manual:
+        pytest.skip("Skipping in CI non-manual mode")
 
     docs = ["Hello World", "Follow the white rabbit."] * 10
 
-    model_name = "jinaai/jina-embeddings-v3"
-    dim = 1024
-
     model = TextEmbedding(model_name=model_name)
 
     task_id = Task.SEPARATION
@@ -216,33 +223,14 @@ def test_parallel_processing():
         delete_model_cache(model.model._model_dir)
 
 
-def test_task_assignment():
-    is_ci = os.getenv("CI")
-
-    for model_desc in TextEmbedding._list_supported_models():
-        if not is_ci and model_desc.size_in_GB > 1:
-            continue
-
-        model_name = model_desc.model
-        if model_name not in CANONICAL_VECTOR_VALUES.keys():
-            continue
-
-        model = TextEmbedding(model_name=model_name)
-
-        for i, task_id in enumerate(Task):
-            _ = list(model.embed(documents=docs, batch_size=1, task_id=i))
-            assert model.model.current_task_id == task_id
-
-        if is_ci:
-            delete_model_cache(model.model._model_dir)
-
-
-@pytest.mark.parametrize(
-    "model_name",
-    ["jinaai/jina-embeddings-v3"],
-)
+@pytest.mark.parametrize("model_name", ["jinaai/jina-embeddings-v3"])
 def test_lazy_load(model_name: str):
     is_ci = os.getenv("CI")
+    is_manual = os.getenv("GITHUB_EVENT_NAME") == "workflow_dispatch"
+
+    if is_ci and not is_manual:
+        pytest.skip("Skipping in CI non-manual mode")
+
     model = TextEmbedding(model_name=model_name, lazy_load=True)
     assert not hasattr(model.model, "model")
 

tests/test_text_onnx_embeddings.py

@@ -1,11 +1,14 @@
 import os
 import platform
+from contextlib import contextmanager
 
 import numpy as np
 import pytest
 
+from fastembed.text.last_token_normalized_embedding import LastTokenNormalizedEmbedding
+from fastembed.text.onnx_embedding import OnnxTextEmbedding
 from fastembed.text.text_embedding import TextEmbedding
-from tests.utils import delete_model_cache
+from tests.utils import delete_model_cache, should_test_model
 
 CANONICAL_VECTOR_VALUES = {
     "BAAI/bge-small-en": np.array([-0.0232, -0.0255, 0.0174, -0.0639, -0.0006]),
@@ -52,7 +55,7 @@ CANONICAL_VECTOR_VALUES = {
         [0.0802303, 0.3700881, -4.3053818, 0.4431803, -0.271572]
     ),
     "thenlper/gte-large": np.array(
-        [-0.01920587, 0.00113156, -0.00708992, -0.00632304, -0.04025577]
+        [-0.00986551, -0.00018734, 0.00605892, -0.03289612, -0.0387564],
     ),
     "mixedbread-ai/mxbai-embed-large-v1": np.array(
         [0.02295546, 0.03196154, 0.016512, -0.04031524, -0.0219634]
@@ -67,86 +70,194 @@ CANONICAL_VECTOR_VALUES = {
     "Qdrant/clip-ViT-B-32-text": np.array([0.0083, 0.0103, -0.0138, 0.0199, -0.0069]),
     "thenlper/gte-base": np.array([0.0038, 0.0355, 0.0181, 0.0092, 0.0654]),
     "jinaai/jina-clip-v1": np.array([-0.0862, -0.0101, -0.0056, 0.0375, -0.0472]),
+    "google/embeddinggemma-300m": np.array(
+        [-0.08181356, 0.0214127, 0.05120273, -0.03690156, -0.0254504]
+    ),
+    "Qwen/Qwen3-Embedding-0.6B": np.array(
+        [-0.01476084, 0.01723184, -0.01195498, -0.07275258, 0.00281229]
+    ),
+    "Qwen/Qwen3-Embedding-0.6B-Q": np.array(
+        [-0.01599521, 0.01676456, -0.01195119, -0.07132675, 0.00346729]
+    ),
+    "google/siglip2-base-patch16-224": np.array(
+        [-0.01181389, 0.00737596, 0.01118064, 0.0103095, 0.3451049]
+    ),
+    "minishlab/potion-base-8M": np.array(
+        [-0.03432461, -0.08020256, -0.14396408, 0.08480079, 0.01958815]
+    ),
+    "minishlab/potion-retrieval-32M": np.array(
+        [0.019733, -0.01530093, -0.08678473, 0.0229059, 0.04700558]
+    ),
+    "minishlab/potion-multilingual-128M": np.array(
+        [0.02366836, 0.02973341, 0.05140258, -0.00745248, -0.06740689]
+    ),
+}
+
+QWEN3_INSTRUCT_PREFIX = (
+    "Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery:"
+)
+
+DOC_PREFIXES = {
+    "google/embeddinggemma-300m": "title: none | text: ",
+}
+QUERY_PREFIXES = {
+    "google/embeddinggemma-300m": "task: search result | query: ",
+    "Qwen/Qwen3-Embedding-0.6B": QWEN3_INSTRUCT_PREFIX,
+    "Qwen/Qwen3-Embedding-0.6B-Q": QWEN3_INSTRUCT_PREFIX,
+}
+CANONICAL_QUERY_VECTOR_VALUES = {
+    "google/embeddinggemma-300m": np.array(
+        [-0.22990295, 0.03311195, 0.04290345, -0.03558498, -0.01399477]
+    ),
+    "Qwen/Qwen3-Embedding-0.6B": np.array(
+        [-0.01908712, 0.01635596, -0.00356586, -0.03947155, -0.01387356]
+    ),
+    "Qwen/Qwen3-Embedding-0.6B-Q": np.array(
+        [-0.02221339, 0.01932909, -0.00361797, -0.03888897, -0.01362813]
+    ),
 }
 
 MULTI_TASK_MODELS = ["jinaai/jina-embeddings-v3"]
 
+_MODELS_TO_CACHE = ("BAAI/bge-small-en-v1.5",)
+MODELS_TO_CACHE = tuple([x.lower() for x in _MODELS_TO_CACHE])
 
-def test_embedding() -> None:
+
+@pytest.fixture(scope="module")
+def model_cache():
+    is_ci = os.getenv("CI")
+    cache = {}
+
+    @contextmanager
+    def get_model(model_name: str):
+        lowercase_model_name = model_name.lower()
+        if lowercase_model_name not in cache:
+            cache[lowercase_model_name] = TextEmbedding(lowercase_model_name)
+        yield cache[lowercase_model_name]
+        if lowercase_model_name not in MODELS_TO_CACHE:
+            model_inst = cache.pop(lowercase_model_name)
+            if is_ci:
+                delete_model_cache(model_inst.model._model_dir)
+            del model_inst
+
+    yield get_model
+
+    if is_ci:
+        for name, model in cache.items():
+            delete_model_cache(model.model._model_dir)
+    cache.clear()
+
+
+@pytest.mark.parametrize("model_name", ["BAAI/bge-small-en-v1.5"])
+def test_embedding(model_cache, model_name: str) -> None:
     is_ci = os.getenv("CI")
     is_mac = platform.system() == "Darwin"
+    is_manual = os.getenv("GITHUB_EVENT_NAME") == "workflow_dispatch"
 
     for model_desc in TextEmbedding._list_supported_models():
-        if (
-            (not is_ci and model_desc.size_in_GB > 1)
-            or model_desc.model in MULTI_TASK_MODELS
-            or (is_mac and model_desc.model == "nomic-ai/nomic-embed-text-v1.5-Q")
+        if model_desc.model in MULTI_TASK_MODELS or (
+            is_mac and model_desc.model == "nomic-ai/nomic-embed-text-v1.5-Q"
         ):
             continue
+        if not should_test_model(model_desc, model_name, is_ci, is_manual):
+            continue
 
         dim = model_desc.dim
 
-        model = TextEmbedding(model_name=model_desc.model)
-        docs = ["hello world", "flag embedding"]
-        embeddings = list(model.embed(docs))
+        with model_cache(model_desc.model) as model:
+            docs = ["hello world", "flag embedding"]
+            if model_desc.model in DOC_PREFIXES:
+                docs = [DOC_PREFIXES[model_desc.model] + doc for doc in docs]
+
+            embeddings = list(model.embed(docs))
+            embeddings = np.stack(embeddings, axis=0)
+            assert embeddings.shape == (2, dim)
+
+            canonical_vector = CANONICAL_VECTOR_VALUES[model_desc.model]
+            assert np.allclose(
+                embeddings[0, : canonical_vector.shape[0]], canonical_vector, atol=1e-3
+            ), model_desc.model
+
+
+def test_query_embedding(model_cache) -> None:
+    is_ci = os.getenv("CI")
+    is_mac = platform.system() == "Darwin"
+    is_manual = os.getenv("GITHUB_EVENT_NAME") == "workflow_dispatch"
+
+    for model_desc in TextEmbedding._list_supported_models():
+        if model_desc.model in MULTI_TASK_MODELS or (
+            is_mac and model_desc.model == "nomic-ai/nomic-embed-text-v1.5-Q"
+        ):
+            continue
+
+        if model_desc.model not in CANONICAL_QUERY_VECTOR_VALUES:
+            continue
+
+        if not should_test_model(model_desc, "", is_ci, is_manual):
+            continue
+
+        dim = model_desc.dim
+        with model_cache(model_desc.model) as model:
+            queries = ["hello world", "flag embedding"]
+            if model_desc.model in QUERY_PREFIXES:
+                queries = [QUERY_PREFIXES[model_desc.model] + query for query in queries]
+
+            embeddings = list(model.query_embed(queries))
+            embeddings = np.stack(embeddings, axis=0)
+            assert embeddings.shape == (2, dim)
+
+            canonical_vector = CANONICAL_QUERY_VECTOR_VALUES[model_desc.model]
+            assert np.allclose(
+                embeddings[0, : canonical_vector.shape[0]], canonical_vector, atol=1e-3
+            ), model_desc.model
+
+
+def test_quantized_model_reports_onnxruntime_requirement(monkeypatch) -> None:
+    """Old onnxruntime only implements 4-bit MatMulNBits, the error should say so."""
+    monkeypatch.setattr(
+        OnnxTextEmbedding,
+        "load_onnx_model",
+        lambda self: (_ for _ in ()).throw(RuntimeError("nbits_ == 4 was false")),
+    )
+    model = LastTokenNormalizedEmbedding(
+        "Qwen/Qwen3-Embedding-0.6B-Q",
+        lazy_load=True,
+        specific_model_path="./",  # disable model downloading and loading
+    )
+
+    with pytest.raises(RuntimeError, match="onnxruntime>=1.23"):
+        model.load_onnx_model()
+
+
+@pytest.mark.parametrize("n_dims,model_name", [(384, "BAAI/bge-small-en-v1.5")])
+def test_batch_embedding(model_cache, n_dims: int, model_name: str) -> None:
+    with model_cache(model_name) as model:
+        docs = ["hello world", "flag embedding"] * 100
+        embeddings = list(model.embed(docs, batch_size=10))
         embeddings = np.stack(embeddings, axis=0)
-        assert embeddings.shape == (2, dim)
 
-        canonical_vector = CANONICAL_VECTOR_VALUES[model_desc.model]
-        assert np.allclose(
-            embeddings[0, : canonical_vector.shape[0]], canonical_vector, atol=1e-3
-        ), model_desc["model"]
-        if is_ci:
-            delete_model_cache(model.model._model_dir)
+        assert embeddings.shape == (len(docs), n_dims)
 
 
-@pytest.mark.parametrize(
-    "n_dims,model_name",
-    [(384, "BAAI/bge-small-en-v1.5"), (768, "jinaai/jina-embeddings-v2-base-en")],
-)
-def test_batch_embedding(n_dims: int, model_name: str) -> None:
-    is_ci = os.getenv("CI")
-    model = TextEmbedding(model_name=model_name)
+@pytest.mark.parametrize("n_dims,model_name", [(384, "BAAI/bge-small-en-v1.5")])
+def test_parallel_processing(model_cache, n_dims: int, model_name: str) -> None:
+    with model_cache(model_name) as model:
+        docs = ["hello world", "flag embedding"] * 100
+        embeddings = list(model.embed(docs, batch_size=10, parallel=2))
+        embeddings = np.stack(embeddings, axis=0)
 
-    docs = ["hello world", "flag embedding"] * 100
-    embeddings = list(model.embed(docs, batch_size=10))
-    embeddings = np.stack(embeddings, axis=0)
+        embeddings_2 = list(model.embed(docs, batch_size=10, parallel=None))
+        embeddings_2 = np.stack(embeddings_2, axis=0)
 
-    assert embeddings.shape == (200, n_dims)
-    if is_ci:
-        delete_model_cache(model.model._model_dir)
+        embeddings_3 = list(model.embed(docs, batch_size=10, parallel=0))
+        embeddings_3 = np.stack(embeddings_3, axis=0)
+
+        assert embeddings.shape == (len(docs), n_dims)
+        assert np.allclose(embeddings, embeddings_2, atol=1e-3)
+        assert np.allclose(embeddings, embeddings_3, atol=1e-3)
 
 
-@pytest.mark.parametrize(
-    "n_dims,model_name",
-    [(384, "BAAI/bge-small-en-v1.5"), (768, "jinaai/jina-embeddings-v2-base-en")],
-)
-def test_parallel_processing(n_dims: int, model_name: str) -> None:
-    is_ci = os.getenv("CI")
-    model = TextEmbedding(model_name=model_name)
-
-    docs = ["hello world", "flag embedding"] * 100
-    embeddings = list(model.embed(docs, batch_size=10, parallel=2))
-    embeddings = np.stack(embeddings, axis=0)
-
-    embeddings_2 = list(model.embed(docs, batch_size=10, parallel=None))
-    embeddings_2 = np.stack(embeddings_2, axis=0)
-
-    embeddings_3 = list(model.embed(docs, batch_size=10, parallel=0))
-    embeddings_3 = np.stack(embeddings_3, axis=0)
-
-    assert embeddings.shape == (200, n_dims)
-    assert np.allclose(embeddings, embeddings_2, atol=1e-3)
-    assert np.allclose(embeddings, embeddings_3, atol=1e-3)
-
-    if is_ci:
-        delete_model_cache(model.model._model_dir)
-
-
-@pytest.mark.parametrize(
-    "model_name",
-    ["BAAI/bge-small-en-v1.5"],
-)
+@pytest.mark.parametrize("model_name", ["BAAI/bge-small-en-v1.5"])
 def test_lazy_load(model_name: str) -> None:
     is_ci = os.getenv("CI")
     model = TextEmbedding(model_name=model_name, lazy_load=True)
@@ -163,3 +274,60 @@ def test_lazy_load(model_name: str) -> None:
 
     if is_ci:
         delete_model_cache(model.model._model_dir)
+
+
+def test_get_embedding_size() -> None:
+    assert TextEmbedding.get_embedding_size("sentence-transformers/all-MiniLM-L6-v2") == 384
+    assert TextEmbedding.get_embedding_size("sentence-transformers/all-minilm-l6-v2") == 384
+
+
+def test_embedding_size() -> None:
+    is_ci = os.getenv("CI")
+    model_name = "sentence-transformers/all-MiniLM-L6-v2"
+    model = TextEmbedding(model_name=model_name, lazy_load=True)
+    assert model.embedding_size == 384
+
+    model_name = "sentence-transformers/all-minilm-l6-v2"
+    model = TextEmbedding(model_name=model_name, lazy_load=True)
+    assert model.embedding_size == 384
+
+    if is_ci:
+        delete_model_cache(model.model._model_dir)
+
+
+@pytest.mark.parametrize("model_name", ["sentence-transformers/all-MiniLM-L6-v2"])
+def test_session_options(model_cache, model_name) -> None:
+    with model_cache(model_name) as default_model:
+        default_session_options = default_model.model.model.get_session_options()
+        assert default_session_options.enable_cpu_mem_arena is True
+        model = TextEmbedding(model_name=model_name, enable_cpu_mem_arena=False)
+        session_options = model.model.model.get_session_options()
+        assert session_options.enable_cpu_mem_arena is False
+
+
+@pytest.mark.parametrize("model_name", ["sentence-transformers/all-MiniLM-L6-v2"])
+def test_token_count(model_cache, model_name) -> None:
+    with model_cache(model_name) as model:
+        documents = [
+            "Name me a couple of cities were the capitals of Germany?",
+            "Berlin is the current capital of Germany, Bonn is a former capital of Germany.",
+        ]
+        first_doc_token_count = model.token_count(documents[0])
+        second_doc_token_count = model.token_count(documents[1])
+        doc_token_count = model.token_count(documents)
+        assert first_doc_token_count + second_doc_token_count == doc_token_count
+        assert doc_token_count == model.token_count(documents, batch_size=1)
+
+
+@pytest.mark.parametrize(
+    "model_name,dim",
+    [("sentence-transformers/all-MiniLM-L6-v2", 384), ("thenlper/gte-base", 768)],
+)
+def test_mixed_length_batch_with_fixed_padding(model_cache, model_name: str, dim: int) -> None:
+    # both models serialize a fixed padding length of 128 in tokenizer.json; gte-base truncates
+    # at 512, so a document longer than 128 makes the batch ragged unless the padding is relaxed
+    with model_cache(model_name) as model:
+        assert model.model.tokenizer.padding["length"] is None
+
+        embeddings = np.stack(list(model.embed(["hello world", "retrieval " * 200])), axis=0)
+        assert embeddings.shape == (2, dim)

tests/utils.py

@@ -3,10 +3,12 @@ import traceback
 
 from pathlib import Path
 from types import TracebackType
-from typing import Union, Callable, Any, Type
+from typing import Callable, Any, Type
+
+from fastembed.common.model_description import BaseModelDescription
 
 
-def delete_model_cache(model_dir: Union[str, Path]) -> None:
+def delete_model_cache(model_dir: str | Path) -> None:
     """Delete the model cache directory.
 
     If a model was downloaded from the HuggingFace model hub, then _model_dir is the dir to snapshots, removing
@@ -35,3 +37,31 @@ def delete_model_cache(model_dir: Union[str, Path]) -> None:
     if model_dir.exists():
         # todo: PermissionDenied is raised on blobs removal in Windows, with blobs > 2GB
         shutil.rmtree(model_dir, onerror=on_error)
+
+
+def should_test_model(
+    model_desc: BaseModelDescription,
+    autotest_model_name: str,
+    is_ci: str | None,
+    is_manual: bool,
+):
+    """Determine if a model should be tested based on environment
+
+    Tests can be run either in ci or locally.
+    Testing all models each time in ci is too long.
+    The testing scheme in ci and on a local machine are different, therefore, there are 3 possible scenarios.
+    1) Run lightweight tests in ci:
+        - test only one model that has been manually chosen as a representative for a certain class family
+    2) Run heavyweight (manual) tests in ci:
+        - test all models
+        Running tests in ci each time is too expensive, however, it's fine to run it one time with a manual dispatch
+    3) Run tests locally:
+        - test all models, which are not too heavy, since network speed might be a bottleneck
+
+    """
+    if not is_ci:
+        if model_desc.size_in_GB > 1:
+            return False
+    elif not is_manual and model_desc.model != autotest_model_name:
+        return False
+    return True