fix: Pass providers and cuda to multimodal models

.github/workflows/python-tests.yml

@@ -1,10 +1,10 @@
 name: Tests
+run-name: Tests (gpu)
 
 on:
-  pull_request:
+  push:
     branches: [ master, main, gpu ]
-  workflow_dispatch:
-  
+  pull_request:
 
 env:
   CARGO_TERM_COLOR: always
@@ -22,8 +22,6 @@ jobs:
           - '3.13.x'
         os:
           - ubuntu-latest
-          - macos-latest
-          - windows-latest
 
     runs-on: ${{ matrix.os }}
 
@@ -43,4 +41,4 @@ jobs:
 
       - name: Run pytest
         run: |
-          poetry run pytest 
+          poetry run pytest 

README.md

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

fastembed/common/model_description.py

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

fastembed/common/model_management.py

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

fastembed/common/onnx_model.py

@@ -68,7 +68,15 @@ class OnnxModel(Generic[T]):
             if device_id is None:
                 onnx_providers = ["CUDAExecutionProvider"]
             else:
-                onnx_providers = [("CUDAExecutionProvider", {"device_id": device_id})]
+                # kSameAsRequested: Allocates only the requested memory, avoiding over-allocation.
+                # more precise than 'kNextPowerOfTwo', which grows memory aggressively.
+                # source: https://onnxruntime.ai/docs/get-started/with-c.html#features:~:text=Memory%20arena%20shrinkage:
+                onnx_providers = [
+                    (
+                        "CUDAExecutionProvider",
+                        {"device_id": device_id, "arena_extend_strategy": "kSameAsRequested"},
+                    )
+                ]
         else:
             onnx_providers = ["CPUExecutionProvider"]
 
@@ -132,5 +140,7 @@ class EmbeddingWorker(Worker, Generic[T]):
     def start(cls, model_name: str, cache_dir: str, **kwargs: Any) -> "EmbeddingWorker[T]":
         return cls(model_name=model_name, cache_dir=cache_dir, **kwargs)
 
-    def process(self, items: Iterable[tuple[int, Any]]) -> Iterable[tuple[int, Any]]:
+    def process(
+        self, items: Iterable[tuple[int, Any]], **kwargs: Any
+    ) -> Iterable[tuple[int, Any]]:
         raise NotImplementedError("Subclasses must implement this method")

fastembed/common/utils.py

@@ -5,12 +5,12 @@ import tempfile
 import unicodedata
 from pathlib import Path
 from itertools import islice
-from typing import Iterable, Optional, TypeVar
+from typing import Iterable, Optional, TypeVar, Sequence
 
 import numpy as np
 from numpy.typing import NDArray
 
-from fastembed.common.types import NumpyArray
+from fastembed.common.types import NumpyArray, OnnxProvider
 
 T = TypeVar("T")
 
@@ -67,3 +67,18 @@ def get_all_punctuation() -> set[str]:
 
 def remove_non_alphanumeric(text: str) -> str:
     return re.sub(r"[^\w\s]", " ", text, flags=re.UNICODE)
+
+
+def is_cuda_enabled(cuda: bool, providers: Optional[Sequence[OnnxProvider]]) -> bool:
+    """
+    Check if CUDA is enabled based on the `cuda` and `providers` parameters
+    """
+    if cuda:
+        return True
+    if not providers:
+        return False
+    if isinstance(providers, str):
+        return "CUDAExecutionProvider" in providers
+    return isinstance(providers, (list, tuple)) and any(
+        isinstance(p, str) and "CUDAExecutionProvider" in p for p in providers
+    )

fastembed/image/onnx_embedding.py

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

fastembed/image/onnx_image_model.py

@@ -6,13 +6,14 @@ from typing import Any, Iterable, Optional, Sequence, Type, Union
 
 import numpy as np
 from PIL import Image
+import onnxruntime as ort
 
 from fastembed.image.transform.operators import Compose
 from fastembed.common.types import NumpyArray
 from fastembed.common import ImageInput, OnnxProvider
 from fastembed.common.onnx_model import EmbeddingWorker, OnnxModel, OnnxOutputContext, T
 from fastembed.common.preprocessor_utils import load_preprocessor
-from fastembed.common.utils import iter_batch
+from fastembed.common.utils import iter_batch, is_cuda_enabled
 from fastembed.parallel_processor import ParallelWorkerPool
 
 # Holds type of the embedding result
@@ -74,7 +75,21 @@ class OnnxImageModel(OnnxModel[T]):
             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]
+
+        run_options = ort.RunOptions()
+        providers = kwargs.get("providers", None)
+        cuda = kwargs.get("cuda", False)
+        if is_cuda_enabled(cuda, providers):
+            device_id = kwargs.get("device_id", None)
+            device_id = str(device_id if isinstance(device_id, int) else 0)
+            # enables memory arena shrinkage, freeing unused memory after each Run() cycle.
+            # helps prevent excessive memory retention, especially for dynamic workloads.
+            # source: https://onnxruntime.ai/docs/get-started/with-c.html#features:~:text=Memory%20arena%20shrinkage:
+            run_options.add_run_config_entry(
+                "memory.enable_memory_arena_shrinkage", f"gpu:{device_id}"
+            )
+
+        model_output = self.model.run(None, onnx_input, run_options)  # type: ignore[union-attr]
         embeddings = model_output[0].reshape(len(images), -1)
         return OnnxOutputContext(model_output=embeddings)
 
@@ -104,7 +119,9 @@ 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, cuda=cuda, providers=providers)
+                )
         else:
             if parallel == 0:
                 parallel = os.cpu_count()
@@ -129,7 +146,9 @@ class OnnxImageModel(OnnxModel[T]):
 
 
 class ImageEmbeddingWorker(EmbeddingWorker[T]):
-    def process(self, items: Iterable[tuple[int, Any]]) -> Iterable[tuple[int, Any]]:
+    def process(
+        self, items: Iterable[tuple[int, Any]], **kwargs: Any
+    ) -> Iterable[tuple[int, Any]]:
         for idx, batch in items:
-            embeddings = self.model.onnx_embed(batch)
+            embeddings = self.model.onnx_embed(batch, **kwargs)
             yield idx, embeddings

fastembed/late_interaction/colbert.py

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

fastembed/late_interaction_multimodal/colpali.py

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

fastembed/late_interaction_multimodal/onnx_multimodal_model.py

@@ -6,13 +6,14 @@ from typing import Any, Iterable, Optional, Sequence, Type, Union
 
 import numpy as np
 from PIL import Image
+import onnxruntime as ort
 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.utils import iter_batch
+from fastembed.common.utils import iter_batch, is_cuda_enabled
 from fastembed.image.transform.operators import Compose
 from fastembed.parallel_processor import ParallelWorkerPool
 
@@ -103,7 +104,21 @@ class OnnxMultimodalModel(OnnxModel[T]):
             )
 
         onnx_input = self._preprocess_onnx_text_input(onnx_input, **kwargs)
-        model_output = self.model.run(self.ONNX_OUTPUT_NAMES, onnx_input)  # type: ignore[union-attr]
+
+        run_options = ort.RunOptions()
+        providers = kwargs.get("providers", None)
+        cuda = kwargs.get("cuda", False)
+        if is_cuda_enabled(cuda, providers):
+            device_id = kwargs.get("device_id", None)
+            device_id = str(device_id if isinstance(device_id, int) else 0)
+            # enables memory arena shrinkage, freeing unused memory after each Run() cycle.
+            # helps prevent excessive memory retention, especially for dynamic workloads.
+            # source: https://onnxruntime.ai/docs/get-started/with-c.html#features:~:text=Memory%20arena%20shrinkage:
+            run_options.add_run_config_entry(
+                "memory.enable_memory_arena_shrinkage", f"gpu:{device_id}"
+            )
+
+        model_output = self.model.run(self.ONNX_OUTPUT_NAMES, onnx_input, run_options)  # type: ignore[union-attr]
         return OnnxOutputContext(
             model_output=model_output[0],
             attention_mask=onnx_input.get("attention_mask", attention_mask),
@@ -136,7 +151,9 @@ class OnnxMultimodalModel(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_text_output(self.onnx_embed_text(batch))
+                yield from self._post_process_onnx_text_output(
+                    self.onnx_embed_text(batch, cuda=cuda, providers=providers)
+                )
         else:
             if parallel == 0:
                 parallel = os.cpu_count()
@@ -169,7 +186,21 @@ class OnnxMultimodalModel(OnnxModel[T]):
             encoded = np.array(self.processor(image_files))
         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]
+
+        run_options = ort.RunOptions()
+        providers = kwargs.get("providers", None)
+        cuda = kwargs.get("cuda", False)
+        if is_cuda_enabled(cuda, providers):
+            device_id = kwargs.get("device_id", None)
+            device_id = str(device_id if isinstance(device_id, int) else 0)
+            # enables memory arena shrinkage, freeing unused memory after each Run() cycle.
+            # helps prevent excessive memory retention, especially for dynamic workloads.
+            # source: https://onnxruntime.ai/docs/get-started/with-c.html#features:~:text=Memory%20arena%20shrinkage:
+            run_options.add_run_config_entry(
+                "memory.enable_memory_arena_shrinkage", f"gpu:{device_id}"
+            )
+
+        model_output = self.model.run(None, onnx_input, run_options)  # type: ignore[union-attr]
         embeddings = model_output[0].reshape(len(images), -1)
         return OnnxOutputContext(model_output=embeddings)
 
@@ -199,7 +230,9 @@ class OnnxMultimodalModel(OnnxModel[T]):
                 self.load_onnx_model()
 
             for batch in iter_batch(images, batch_size):
-                yield from self._post_process_onnx_image_output(self.onnx_embed_image(batch))
+                yield from self._post_process_onnx_image_output(
+                    self.onnx_embed_image(batch, cuda=cuda, providers=providers)
+                )
         else:
             if parallel == 0:
                 parallel = os.cpu_count()
@@ -241,9 +274,11 @@ class TextEmbeddingWorker(EmbeddingWorker[T]):
     ) -> OnnxMultimodalModel:
         raise NotImplementedError()
 
-    def process(self, items: Iterable[tuple[int, Any]]) -> Iterable[tuple[int, Any]]:
+    def process(
+        self, items: Iterable[tuple[int, Any]], **kwargs: Any
+    ) -> Iterable[tuple[int, Any]]:
         for idx, batch in items:
-            onnx_output = self.model.onnx_embed_text(batch)
+            onnx_output = self.model.onnx_embed_text(batch, **kwargs)
             yield idx, onnx_output
 
 
@@ -265,7 +300,9 @@ class ImageEmbeddingWorker(EmbeddingWorker[T]):
     ) -> OnnxMultimodalModel:
         raise NotImplementedError()
 
-    def process(self, items: Iterable[tuple[int, Any]]) -> Iterable[tuple[int, Any]]:
+    def process(
+        self, items: Iterable[tuple[int, Any]], **kwargs: Any
+    ) -> Iterable[tuple[int, Any]]:
         for idx, batch in items:
-            embeddings = self.model.onnx_embed_image(batch)
+            embeddings = self.model.onnx_embed_image(batch, **kwargs)
             yield idx, embeddings

fastembed/parallel_processor.py

@@ -3,8 +3,6 @@ import os
 from collections import defaultdict
 from copy import deepcopy
 from enum import Enum
-from dataclasses import dataclass
-from multiprocessing import shared_memory, Manager, Lock
 from multiprocessing import Queue, get_context
 from multiprocessing.context import BaseContext
 from multiprocessing.process import BaseProcess
@@ -12,11 +10,6 @@ from multiprocessing.sharedctypes import Synchronized as BaseValue
 from queue import Empty
 from typing import Any, Iterable, Optional, Type
 
-import numpy as np
-from numpy.typing import NDArray
-
-from fastembed.common.types import NumpyArray
-
 
 # Single item should be processed in less than:
 processing_timeout = 10 * 60  # seconds
@@ -24,13 +17,6 @@ processing_timeout = 10 * 60  # seconds
 max_internal_batch_size = 200
 
 
-@dataclass
-class OnnxOutputContext:
-    model_output: NumpyArray
-    attention_mask: Optional[NDArray[np.int64]] = None
-    input_ids: Optional[NDArray[np.int64]] = None
-
-
 class QueueSignals(str, Enum):
     stop = "stop"
     confirm = "confirm"
@@ -42,58 +28,18 @@ class Worker:
     def start(cls, *args: Any, **kwargs: Any) -> "Worker":
         raise NotImplementedError()
 
-    def process(self, items: Iterable[tuple[int, Any]]) -> Iterable[tuple[int, Any]]:
+    def process(
+        self, items: Iterable[tuple[int, Any]], **kwargs: Any
+    ) -> Iterable[tuple[int, Any]]:
         raise NotImplementedError()
 
 
-class SharedMemoryPool:
-    def __init__(self, lock: Lock):
-        self._lock = lock
-        self._pool: dict[str, tuple[shared_memory.SharedMemory, int, np.dtype]] = {}
-        self._free_buffers: list[str] = []
-
-    def allocate(self, size: int, dtype: np.dtype) -> tuple[shared_memory.SharedMemory, str]:
-        best_match = None
-        best_size = float("inf")
-        for buf_name in self._free_buffers:
-            shm, buf_size, buf_dtype = self._pool[buf_name]
-            # get best match for needed size
-            if buf_size >= size and buf_dtype == dtype and buf_size < best_size:
-                best_match = buf_name
-                best_size = buf_size
-            if best_match:
-                self._free_buffers.remove(best_match)
-                return self._pool[best_match][0], best_match
-            shm = shared_memory.SharedMemory(create=True, size=size)
-            self._pool[shm.name] = (shm, size, dtype)
-            return shm, shm.name
-        # if no match found, create new buffer
-        shm = shared_memory.SharedMemory(create=True, size=size)
-        self._pool[shm.name] = (shm, size, dtype)
-        return shm, shm.name
-
-    def release(self, name: str) -> None:
-        if name in self._pool and name not in self._free_buffers:
-            self._free_buffers.append(name)
-
-    def cleanup(self) -> None:
-        for shm, _, _ in self._pool.values():
-            shm.close()
-            shm.unlink()
-        self._pool.clear()
-        self._free_buffers.clear()
-
-    def __del__(self):
-        self.cleanup()
-
-
 def _worker(
     worker_class: Type[Worker],
     input_queue: Queue,
     output_queue: Queue,
     num_active_workers: BaseValue,
     worker_id: int,
-    shared_pool: SharedMemoryPool,
     kwargs: Optional[dict[str, Any]] = None,
 ) -> None:
     """
@@ -111,6 +57,7 @@ def _worker(
     try:
         worker = worker_class.start(**kwargs)
 
+        # Keep going until you get an item that's None.
         def input_queue_iterable() -> Iterable[Any]:
             while True:
                 item = input_queue.get()
@@ -118,25 +65,8 @@ def _worker(
                     break
                 yield item
 
-        for processed_item in worker.process(input_queue_iterable()):
-            idx, output_context = processed_item
-            output_metadata = {}
-            for field in ["model_output", "attention_mask", "input_ids"]:
-                array = getattr(output_context, field, None)
-                if array is not None:
-                    shm, shm_name = shared_pool.allocate(array.nbytes, array.dtype)
-                    shm_array = np.ndarray(array.shape, dtype=array.dtype, buffer=shm.buf)
-                    np.copyto(shm_array, array)
-                    output_metadata[field] = {
-                        "name": shm_name,
-                        "shape": array.shape,
-                        "dtype": array.dtype.str,
-                    }
-                    shm.close()
-            output_queue.put((idx, output_metadata))
-            for field in output_metadata:  # mark release to reuse
-                shared_pool.release(output_metadata[field]["name"])
-
+        for processed_item in worker.process(input_queue_iterable(), **kwargs):
+            output_queue.put(processed_item)
     except Exception as e:  # pylint: disable=broad-except
         logging.exception(e)
         output_queue.put(QueueSignals.error)
@@ -180,8 +110,6 @@ class ParallelWorkerPool:
         self.device_ids = device_ids
         self.cuda = cuda
         self.num_active_workers: Optional[BaseValue] = None
-        self.manager = Manager()
-        self.shared_pool = SharedMemoryPool(self.manager.Lock())
 
     def start(self, **kwargs: Any) -> None:
         self.input_queue = self.ctx.Queue(self.queue_size)
@@ -207,7 +135,6 @@ class ParallelWorkerPool:
                     self.output_queue,
                     self.num_active_workers,
                     worker_id,
-                    self.shared_pool,
                     worker_kwargs,
                 ),
             )
@@ -253,17 +180,7 @@ class ParallelWorkerPool:
                     if out_item == QueueSignals.error:
                         self.join_or_terminate()
                         raise RuntimeError("Thread unexpectedly terminated")
-
-                    idx, output_metadata = out_item
-                    output_arrays = {}
-                    for field, meta in output_metadata.items():
-                        shm = shared_memory.SharedMemory(name=meta["name"])
-                        array = np.ndarray(
-                            meta["shape"], dtype=meta["dtype"], buffer=shm.buf
-                        ).copy()
-                        output_arrays[field] = array
-                        shm.close()
-                    yield (idx, OnnxOutputContext(**output_arrays))
+                    yield out_item
                     read += 1
 
                 self.input_queue.put((idx, item))
@@ -278,18 +195,9 @@ class ParallelWorkerPool:
                 if out_item == QueueSignals.error:
                     self.join_or_terminate()
                     raise RuntimeError("Thread unexpectedly terminated")
-
-                idx, output_metadata = out_item
-                output_arrays = {}
-                for field, meta in output_metadata.items():
-                    shm = shared_memory.SharedMemory(name=meta["name"])
-                    array = np.ndarray(meta["shape"], dtype=meta["dtype"], buffer=shm.buf).copy()
-                    output_arrays[field] = array
-                    shm.close()
-                yield (idx, OnnxOutputContext(**output_arrays))
+                yield out_item
                 read += 1
         finally:
-            self.shared_pool.cleanup()
             assert self.input_queue is not None, "Input queue is None"
             assert self.output_queue is not None, "Output queue is None"
             self.join()
@@ -325,13 +233,11 @@ class ParallelWorkerPool:
             if process.is_alive():
                 process.terminate()
         self.processes.clear()
-        self.shared_pool.cleanup()
 
     def join(self) -> None:
         for process in self.processes:
             process.join()
         self.processes.clear()
-        self.shared_pool.cleanup()
 
     def __del__(self) -> None:
         """
@@ -346,4 +252,3 @@ class ParallelWorkerPool:
         for process in self.processes:
             if process.is_alive():
                 process.terminate()
-        self.shared_pool.cleanup()

fastembed/rerank/cross_encoder/custom_text_cross_encoder.py

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

fastembed/rerank/cross_encoder/onnx_text_model.py

@@ -4,6 +4,7 @@ from pathlib import Path
 from typing import Any, Iterable, Optional, Sequence, Type
 
 import numpy as np
+import onnxruntime as ort
 from tokenizers import Encoding
 
 from fastembed.common.onnx_model import (
@@ -14,7 +15,7 @@ from fastembed.common.onnx_model import (
 )
 from fastembed.common.types import NumpyArray
 from fastembed.common.preprocessor_utils import load_tokenizer
-from fastembed.common.utils import iter_batch
+from fastembed.common.utils import iter_batch, is_cuda_enabled
 from fastembed.parallel_processor import ParallelWorkerPool
 
 
@@ -71,7 +72,21 @@ class OnnxCrossEncoderModel(OnnxModel[float]):
         tokenized_input = self.tokenize(pairs, **kwargs)
         inputs = self._build_onnx_input(tokenized_input)
         onnx_input = self._preprocess_onnx_input(inputs, **kwargs)
-        outputs = self.model.run(self.ONNX_OUTPUT_NAMES, onnx_input)  # type: ignore[union-attr]
+
+        run_options = ort.RunOptions()
+        providers = kwargs.get("providers", None)
+        cuda = kwargs.get("cuda", False)
+        if is_cuda_enabled(cuda, providers):
+            device_id = kwargs.get("device_id", None)
+            device_id = str(device_id if isinstance(device_id, int) else 0)
+            # Enables memory arena shrinkage, freeing unused memory after each Run() cycle.
+            # Helps prevent excessive memory retention, especially for dynamic workloads.
+            # Source: https://onnxruntime.ai/docs/get-started/with-c.html#features:~:text=Memory%20arena%20shrinkage:
+            run_options.add_run_config_entry(
+                "memory.enable_memory_arena_shrinkage", f"gpu:{device_id}"
+            )
+
+        outputs = self.model.run(self.ONNX_OUTPUT_NAMES, onnx_input, run_options)  # type: ignore[union-attr]
         relevant_output = outputs[0]
         scores: NumpyArray = relevant_output[:, 0]
         return OnnxOutputContext(model_output=scores)
@@ -110,7 +125,9 @@ class OnnxCrossEncoderModel(OnnxModel[float]):
             if not hasattr(self, "model") or self.model is None:
                 self.load_onnx_model()
             for batch in iter_batch(pairs, batch_size):
-                yield from self._post_process_onnx_output(self.onnx_embed_pairs(batch, **kwargs))
+                yield from self._post_process_onnx_output(
+                    self.onnx_embed_pairs(batch, cuda=cuda, providers=providers, **kwargs)
+                )
         else:
             if parallel == 0:
                 parallel = os.cpu_count()
@@ -163,7 +180,9 @@ class TextRerankerWorker(EmbeddingWorker[float]):
     ) -> OnnxCrossEncoderModel:
         raise NotImplementedError()
 
-    def process(self, items: Iterable[tuple[int, Any]]) -> Iterable[tuple[int, Any]]:
+    def process(
+        self, items: Iterable[tuple[int, Any]], **kwargs: Any
+    ) -> Iterable[tuple[int, Any]]:
         for idx, batch in items:
-            onnx_output = self.model.onnx_embed_pairs(batch)
+            onnx_output = self.model.onnx_embed_pairs(batch, **kwargs)
             yield idx, onnx_output

fastembed/rerank/cross_encoder/text_cross_encoder.py

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

fastembed/sparse/bm25.py

@@ -344,7 +344,7 @@ class Bm25Worker(Worker):
         return cls(model_name=model_name, cache_dir=cache_dir, **kwargs)
 
     def process(
-        self, items: Iterable[tuple[int, Any]]
+        self, items: Iterable[tuple[int, Any]], **kwargs: Any
     ) -> Iterable[tuple[int, list[SparseEmbedding]]]:
         for idx, batch in items:
             onnx_output = self.model.raw_embed(batch)

fastembed/text/onnx_embedding.py

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

fastembed/text/onnx_text_model.py

@@ -4,13 +4,14 @@ from pathlib import Path
 from typing import Any, Iterable, Optional, Sequence, Type, Union
 
 import numpy as np
+import onnxruntime as ort
 from numpy.typing import NDArray
 from tokenizers import Encoding, Tokenizer
 
 from fastembed.common.types import NumpyArray, OnnxProvider
 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
+from fastembed.common.utils import iter_batch, is_cuda_enabled
 from fastembed.parallel_processor import ParallelWorkerPool
 
 
@@ -82,7 +83,21 @@ class OnnxTextModel(OnnxModel[T]):
             )
         onnx_input = self._preprocess_onnx_input(onnx_input, **kwargs)
 
-        model_output = self.model.run(self.ONNX_OUTPUT_NAMES, onnx_input)  # type: ignore[union-attr]
+        run_options = ort.RunOptions()
+        providers = kwargs.get("providers", None)
+        cuda = kwargs.get("cuda", False)
+
+        if is_cuda_enabled(cuda, providers):
+            device_id = kwargs.get("device_id", None)
+            device_id = str(device_id if isinstance(device_id, int) else 0)
+            # enables memory arena shrinkage, freeing unused memory after each Run() cycle.
+            # helps prevent excessive memory retention, especially for dynamic workloads.
+            # source: https://onnxruntime.ai/docs/get-started/with-c.html#features:~:text=Memory%20arena%20shrinkage:
+            run_options.add_run_config_entry(
+                "memory.enable_memory_arena_shrinkage", f"gpu:{device_id}"
+            )
+
+        model_output = self.model.run(self.ONNX_OUTPUT_NAMES, onnx_input, run_options)  # type: ignore[union-attr]
         return OnnxOutputContext(
             model_output=model_output[0],
             attention_mask=onnx_input.get("attention_mask", attention_mask),
@@ -115,7 +130,9 @@ 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, cuda=cuda, providers=providers)
+                )
         else:
             if parallel == 0:
                 parallel = os.cpu_count()
@@ -140,7 +157,9 @@ class OnnxTextModel(OnnxModel[T]):
 
 
 class TextEmbeddingWorker(EmbeddingWorker[T]):
-    def process(self, items: Iterable[tuple[int, Any]]) -> Iterable[tuple[int, OnnxOutputContext]]:
+    def process(
+        self, items: Iterable[tuple[int, Any]], **kwargs: Any
+    ) -> Iterable[tuple[int, OnnxOutputContext]]:
         for idx, batch in items:
-            onnx_output = self.model.onnx_embed(batch)
+            onnx_output = self.model.onnx_embed(batch, **kwargs)
             yield idx, onnx_output

fastembed/text/pooled_embedding.py

@@ -82,7 +82,6 @@ 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"],
@@ -116,7 +115,7 @@ class PooledEmbedding(OnnxTextEmbedding):
 
         embeddings = output.model_output
         attn_mask = output.attention_mask
-        return self.mean_pooling(embeddings, attn_mask)
+        return self.mean_pooling(embeddings, attn_mask).astype(np.float32)
 
 
 class PooledEmbeddingWorker(OnnxTextEmbeddingWorker):

fastembed/text/pooled_normalized_embedding.py

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

pyproject.toml

@@ -1,5 +1,5 @@
 [tool.poetry]
-name = "fastembed"
+name = "fastembed-gpu"
 version = "0.6.0"
 description = "Fast, light, accurate library built for retrieval embedding generation"
 authors = ["Qdrant Team <info@qdrant.tech>", "NirantK <nirant.bits@gmail.com>"]
@@ -18,7 +18,7 @@ numpy = [
     { version = ">=2.1.0", python = ">=3.13" },
     { version = ">=1.21,<2.1.0", python = "<3.10" },
 ]
-onnxruntime = [
+onnxruntime-gpu = [
     { 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" },

tests/test_custom_models.py

@@ -3,17 +3,10 @@ import os
 import numpy as np
 import pytest
 
-from fastembed.common.model_description import (
-    PoolingType,
-    ModelSource,
-    DenseModelDescription,
-    BaseModelDescription,
-)
+from fastembed.common.model_description import PoolingType, ModelSource, DenseModelDescription
 from fastembed.common.onnx_model import OnnxOutputContext
 from fastembed.common.utils import normalize, mean_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
 
@@ -21,10 +14,8 @@ from tests.utils import delete_model_cache
 @pytest.fixture(autouse=True)
 def restore_custom_models_fixture():
     CustomTextEmbedding.SUPPORTED_MODELS = []
-    CustomTextCrossEncoder.SUPPORTED_MODELS = []
     yield
     CustomTextEmbedding.SUPPORTED_MODELS = []
-    CustomTextCrossEncoder.SUPPORTED_MODELS = []
 
 
 def test_text_custom_model():
@@ -74,43 +65,6 @@ def test_text_custom_model():
         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_mock_add_custom_models():
     dim = 5
     size_in_gb = 0.1
@@ -137,11 +91,15 @@ def test_mock_add_custom_models():
     expected_output = {
         f"{PoolingType.MEAN.lower()}-normalized": normalize(
             mean_pooling(dummy_token_embedding, dummy_attention_mask)
-        ),
+        ).astype(np.float32),
         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()}-normalized": normalize(dummy_token_embedding[:, 0]).astype(
+            np.float32
+        ),
         f"{PoolingType.CLS.lower()}": dummy_token_embedding[:, 0],
-        f"{PoolingType.DISABLED.lower()}-normalized": normalize(dummy_pooled_embedding),
+        f"{PoolingType.DISABLED.lower()}-normalized": normalize(dummy_pooled_embedding).astype(
+            np.float32
+        ),
         f"{PoolingType.DISABLED.lower()}": dummy_pooled_embedding,
     }
 
@@ -202,28 +160,3 @@ def test_do_not_add_existing_model():
             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

@@ -8,7 +8,7 @@ from PIL import Image
 
 from fastembed import ImageEmbedding
 from tests.config import TEST_MISC_DIR
-from tests.utils import delete_model_cache, should_test_model
+from tests.utils import delete_model_cache
 
 CANONICAL_VECTOR_VALUES = {
     "Qdrant/clip-ViT-B-32-vision": np.array([-0.0098, 0.0128, -0.0274, 0.002, -0.0059]),
@@ -27,13 +27,11 @@ CANONICAL_VECTOR_VALUES = {
 }
 
 
-@pytest.mark.parametrize("model_name", ["Qdrant/clip-ViT-B-32-vision"])
-def test_embedding(model_name: str) -> None:
+def test_embedding() -> None:
     is_ci = os.getenv("CI")
-    is_manual = os.getenv("GITHUB_EVENT_NAME") == "workflow_dispatch"
 
     for model_desc in ImageEmbedding._list_supported_models():
-        if not should_test_model(model_desc, model_name, is_ci, is_manual):
+        if not is_ci and model_desc.size_in_GB > 1:
             continue
 
         dim = model_desc.dim
@@ -76,12 +74,8 @@ def test_batch_embedding(n_dims: int, model_name: str) -> None:
 
     embeddings = list(model.embed(images, batch_size=10))
     embeddings = np.stack(embeddings, axis=0)
-    assert np.allclose(embeddings[1], embeddings[2])
-
-    canonical_vector = CANONICAL_VECTOR_VALUES[model_name]
 
     assert embeddings.shape == (len(test_images) * n_images, n_dims)
-    assert np.allclose(embeddings[0, : canonical_vector.shape[0]], canonical_vector, atol=1e-3)
     if is_ci:
         delete_model_cache(model.model._model_dir)
 

tests/test_late_interaction_embeddings.py

@@ -6,7 +6,7 @@ import numpy as np
 from fastembed.late_interaction.late_interaction_text_embedding import (
     LateInteractionTextEmbedding,
 )
-from tests.utils import delete_model_cache, should_test_model
+from tests.utils import delete_model_cache
 
 # vectors are abridged and rounded for brevity
 CANONICAL_COLUMN_VALUES = {
@@ -153,37 +153,31 @@ CANONICAL_QUERY_VALUES = {
 docs = ["Hello World"]
 
 
-@pytest.mark.parametrize("model_name", ["answerdotai/answerai-colbert-small-v1"])
-def test_batch_embedding(model_name: str):
+def test_batch_embedding():
     is_ci = os.getenv("CI")
     docs_to_embed = docs * 10
 
-    model = LateInteractionTextEmbedding(model_name=model_name)
-    result = list(model.embed(docs_to_embed, batch_size=6))
-    expected_result = CANONICAL_COLUMN_VALUES[model_name]
+    for model_name, expected_result in CANONICAL_COLUMN_VALUES.items():
+        print("evaluating", model_name)
+        model = LateInteractionTextEmbedding(model_name=model_name)
+        result = list(model.embed(docs_to_embed, batch_size=6))
 
-    for value in result:
-        token_num, abridged_dim = expected_result.shape
-        assert np.allclose(value[:, :abridged_dim], expected_result, atol=2e-3)
+        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)
+        if is_ci:
+            delete_model_cache(model.model._model_dir)
 
 
-@pytest.mark.parametrize("model_name", ["answerdotai/answerai-colbert-small-v1"])
-def test_single_embedding(model_name: str):
+def test_single_embedding():
     is_ci = os.getenv("CI")
-    is_manual = os.getenv("GITHUB_EVENT_NAME") == "workflow_dispatch"
     docs_to_embed = docs
 
-    for model_desc in LateInteractionTextEmbedding._list_supported_models():
-        if not should_test_model(model_desc, model_name, is_ci, is_manual):
-            continue
-
+    for model_name, expected_result in CANONICAL_COLUMN_VALUES.items():
         print("evaluating", model_name)
         model = LateInteractionTextEmbedding(model_name=model_name)
         result = next(iter(model.embed(docs_to_embed, batch_size=6)))
-        expected_result = CANONICAL_COLUMN_VALUES[model_name]
         token_num, abridged_dim = expected_result.shape
         assert np.allclose(result[:, :abridged_dim], expected_result, atol=2e-3)
 
@@ -191,20 +185,14 @@ def test_single_embedding(model_name: str):
             delete_model_cache(model.model._model_dir)
 
 
-@pytest.mark.parametrize("model_name", ["answerdotai/answerai-colbert-small-v1"])
-def test_single_embedding_query(model_name: str):
+def test_single_embedding_query():
     is_ci = os.getenv("CI")
-    is_manual = os.getenv("GITHUB_EVENT_NAME") == "workflow_dispatch"
     queries_to_embed = docs
 
-    for model_desc in LateInteractionTextEmbedding._list_supported_models():
-        if not should_test_model(model_desc, model_name, is_ci, is_manual):
-            continue
-
+    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)))
-        expected_result = CANONICAL_QUERY_VALUES[model_name]
         token_num, abridged_dim = expected_result.shape
         assert np.allclose(result[:, :abridged_dim], expected_result, atol=2e-3)
 
@@ -212,11 +200,10 @@ def test_single_embedding_query(model_name: str):
             delete_model_cache(model.model._model_dir)
 
 
-@pytest.mark.parametrize("token_dim,model_name", [(96, "answerdotai/answerai-colbert-small-v1")])
-def test_parallel_processing(token_dim: int, model_name: str):
+def test_parallel_processing():
     is_ci = os.getenv("CI")
-    model = LateInteractionTextEmbedding(model_name=model_name)
-
+    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)
@@ -235,7 +222,10 @@ def test_parallel_processing(token_dim: int, model_name: str):
         delete_model_cache(model.model._model_dir)
 
 
-@pytest.mark.parametrize("model_name", ["answerdotai/answerai-colbert-small-v1"])
+@pytest.mark.parametrize(
+    "model_name",
+    ["colbert-ir/colbertv2.0"],
+)
 def test_lazy_load(model_name: str):
     is_ci = os.getenv("CI")
 

tests/test_late_interaction_multimodal.py

@@ -1,6 +1,5 @@
 import os
 
-import pytest
 from PIL import Image
 import numpy as np
 
@@ -46,38 +45,38 @@ images = [
 
 
 def test_batch_embedding():
-    if os.getenv("CI"):
-        pytest.skip("Colpali is too large to test in CI")
+    is_ci = os.getenv("CI")
 
-    for model_name, expected_result in CANONICAL_IMAGE_VALUES.items():
-        print("evaluating", model_name)
-        model = LateInteractionMultimodalEmbedding(model_name=model_name)
-        result = list(model.embed_image(images, batch_size=2))
+    if not is_ci:
+        for model_name, expected_result in CANONICAL_IMAGE_VALUES.items():
+            print("evaluating", model_name)
+            model = LateInteractionMultimodalEmbedding(model_name=model_name)
+            result = list(model.embed_image(images, batch_size=2))
 
-        for value in result:
-            token_num, abridged_dim = expected_result.shape
-            assert np.allclose(value[:token_num, :abridged_dim], expected_result, atol=2e-3)
+            for value in result:
+                token_num, abridged_dim = expected_result.shape
+                assert np.allclose(value[:token_num, :abridged_dim], expected_result, atol=2e-3)
 
 
 def test_single_embedding():
-    if os.getenv("CI"):
-        pytest.skip("Colpali is too large to test in CI")
-
-    for model_name, expected_result in CANONICAL_IMAGE_VALUES.items():
-        print("evaluating", model_name)
-        model = LateInteractionMultimodalEmbedding(model_name=model_name)
-        result = next(iter(model.embed_image(images, batch_size=6)))
-        token_num, abridged_dim = expected_result.shape
-        assert np.allclose(result[:token_num, :abridged_dim], expected_result, atol=2e-3)
+    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)
+            result = next(iter(model.embed_image(images, batch_size=6)))
+            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():
-    if os.getenv("CI"):
-        pytest.skip("Colpali is too large to test in CI")
+    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)))
-        token_num, abridged_dim = expected_result.shape
-        assert np.allclose(result[:token_num, :abridged_dim], expected_result, atol=2e-3)
+        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)

tests/test_sparse_embeddings.py

@@ -5,10 +5,10 @@ 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, should_test_model
+from tests.utils import delete_model_cache
 
 CANONICAL_COLUMN_VALUES = {
-    "prithivida/Splade_PP_en_v1": {
+    "prithvida/Splade_PP_en_v1": {
         "indices": [
             2040,
             2047,
@@ -49,41 +49,28 @@ CANONICAL_COLUMN_VALUES = {
 docs = ["Hello World"]
 
 
-@pytest.mark.parametrize("model_name", ["prithivida/Splade_PP_en_v1"])
-def test_batch_embedding(model_name: str) -> None:
+def test_batch_embedding() -> None:
     is_ci = os.getenv("CI")
     docs_to_embed = docs * 10
 
-    model = SparseTextEmbedding(model_name=model_name)
-    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 model_name, expected_result in CANONICAL_COLUMN_VALUES.items():
+        model = SparseTextEmbedding(model_name=model_name)
+        result = next(iter(model.embed(docs_to_embed, batch_size=6)))
+        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)
+        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)
 
 
-@pytest.mark.parametrize("model_name", ["prithivida/Splade_PP_en_v1"])
-def test_single_embedding(model_name: str) -> None:
+def test_single_embedding() -> None:
     is_ci = os.getenv("CI")
-    is_manual = os.getenv("GITHUB_EVENT_NAME") == "workflow_dispatch"
-
-    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_name, is_ci, is_manual):
-            continue
-
+    for model_name, expected_result in CANONICAL_COLUMN_VALUES.items():
         model = SparseTextEmbedding(model_name=model_name)
 
         passage_result = next(iter(model.embed(docs, batch_size=6)))
         query_result = next(iter(model.query_embed(docs)))
-        expected_result = CANONICAL_COLUMN_VALUES[model_name]
         for result in [passage_result, query_result]:
             assert result.indices.tolist() == expected_result["indices"]
 
@@ -93,10 +80,9 @@ def test_single_embedding(model_name: str) -> None:
             delete_model_cache(model.model._model_dir)
 
 
-@pytest.mark.parametrize("model_name", ["prithivida/Splade_PP_en_v1"])
-def test_parallel_processing(model_name: str) -> None:
+def test_parallel_processing() -> None:
     is_ci = os.getenv("CI")
-    model = SparseTextEmbedding(model_name=model_name)
+    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))
@@ -186,7 +172,10 @@ 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"])
+@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)

tests/test_text_cross_encoder.py

@@ -4,7 +4,7 @@ import numpy as np
 import pytest
 
 from fastembed.rerank.cross_encoder import TextCrossEncoder
-from tests.utils import delete_model_cache, should_test_model
+from tests.utils import delete_model_cache
 
 CANONICAL_SCORE_VALUES = {
     "Xenova/ms-marco-MiniLM-L-6-v2": np.array([8.500708, -2.541011]),
@@ -15,37 +15,44 @@ 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",
+}
 
-@pytest.mark.parametrize("model_name", ["Xenova/ms-marco-MiniLM-L-6-v2"])
+
+@pytest.mark.parametrize(
+    "model_name",
+    [model_name for model_name in CANONICAL_SCORE_VALUES],
+)
 def test_rerank(model_name: str) -> None:
     is_ci = os.getenv("CI")
-    is_manual = os.getenv("GITHUB_EVENT_NAME") == "workflow_dispatch"
 
-    for model_desc in TextCrossEncoder._list_supported_models():
-        if not should_test_model(model_desc, model_name, is_ci, is_manual):
-            continue
+    model = TextCrossEncoder(model_name=model_name)
 
-        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."]
+    scores = np.array(list(model.rerank(query, documents)))
 
-        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_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)
+    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)
 
 
-@pytest.mark.parametrize("model_name", ["Xenova/ms-marco-MiniLM-L-6-v2"])
+@pytest.mark.parametrize(
+    "model_name",
+    [model_name for model_name in SELECTED_MODELS.values()],
+)
 def test_batch_rerank(model_name: str) -> None:
     is_ci = os.getenv("CI")
 
@@ -71,7 +78,10 @@ def test_batch_rerank(model_name: str) -> None:
         delete_model_cache(model.model._model_dir)
 
 
-@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_lazy_load(model_name: str) -> None:
     is_ci = os.getenv("CI")
     model = TextCrossEncoder(model_name=model_name, lazy_load=True)
@@ -85,7 +95,10 @@ def test_lazy_load(model_name: str) -> None:
         delete_model_cache(model.model._model_dir)
 
 
-@pytest.mark.parametrize("model_name", ["Xenova/ms-marco-MiniLM-L-6-v2"])
+@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")
 

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 JinaEmbeddingV3, Task
+from fastembed.text.multitask_embedding import Task
 from tests.utils import delete_model_cache
 
 
@@ -60,43 +60,52 @@ CANONICAL_VECTOR_VALUES = {
 docs = ["Hello World", "Follow the white rabbit."]
 
 
-@pytest.mark.parametrize("dim,model_name", [(1024, "jinaai/jina-embeddings-v3")])
-def test_batch_embedding(dim: int, model_name: str):
+def test_batch_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")
-
     docs_to_embed = docs * 10
     default_task = Task.RETRIEVAL_PASSAGE
 
-    model = TextEmbedding(model_name=model_name)
+    for model_desc in TextEmbedding._list_supported_models():
+        if not is_ci and model_desc.size_in_GB > 1:
+            continue
 
-    embeddings = list(model.embed(documents=docs_to_embed, batch_size=6))
-    embeddings = np.stack(embeddings, axis=0)
+        model_name = model_desc.model
+        dim = model_desc.dim
 
-    assert embeddings.shape == (len(docs_to_embed), dim)
+        if model_name not in CANONICAL_VECTOR_VALUES.keys():
+            continue
 
-    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
+        model = TextEmbedding(model_name=model_name)
 
-    if is_ci:
-        delete_model_cache(model.model._model_dir)
+        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)
 
 
 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 JinaEmbeddingV3._list_supported_models():
-        # todo: once we add more models, we should not test models >1GB size locally
+    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
         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]:
@@ -118,17 +127,18 @@ def test_single_embedding():
 
 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 JinaEmbeddingV3._list_supported_models():
-        # todo: once we add more models, we should not test models >1GB size locally
+    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
         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}")
@@ -149,18 +159,18 @@ 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 JinaEmbeddingV3._list_supported_models():
-        # todo: once we add more models, we should not test models >1GB size locally
+    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
         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}")
@@ -179,15 +189,14 @@ def test_single_embedding_passage():
             delete_model_cache(model.model._model_dir)
 
 
-@pytest.mark.parametrize("dim,model_name", [(1024, "jinaai/jina-embeddings-v3")])
-def test_parallel_processing(dim: int, model_name: str):
+def test_parallel_processing():
     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
@@ -209,14 +218,14 @@ def test_parallel_processing(dim: int, model_name: str):
 
 def test_task_assignment():
     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")
+    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
+        if model_name not in CANONICAL_VECTOR_VALUES.keys():
+            continue
 
         model = TextEmbedding(model_name=model_name)
 
@@ -228,14 +237,12 @@ def test_task_assignment():
             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

@@ -5,7 +5,7 @@ import numpy as np
 import pytest
 
 from fastembed.text.text_embedding import TextEmbedding
-from tests.utils import delete_model_cache, should_test_model
+from tests.utils import delete_model_cache
 
 CANONICAL_VECTOR_VALUES = {
     "BAAI/bge-small-en": np.array([-0.0232, -0.0255, 0.0174, -0.0639, -0.0006]),
@@ -72,19 +72,17 @@ CANONICAL_VECTOR_VALUES = {
 MULTI_TASK_MODELS = ["jinaai/jina-embeddings-v3"]
 
 
-@pytest.mark.parametrize("model_name", ["BAAI/bge-small-en-v1.5"])
-def test_embedding(model_name: str) -> None:
+def test_embedding() -> 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"
+        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")
         ):
             continue
-        if not should_test_model(model_desc, model_name, is_ci, is_manual):
-            continue
 
         dim = model_desc.dim
 
@@ -97,12 +95,15 @@ def test_embedding(model_name: str) -> None:
         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
+        ), model_desc["model"]
         if is_ci:
             delete_model_cache(model.model._model_dir)
 
 
-@pytest.mark.parametrize("n_dims,model_name", [(384, "BAAI/bge-small-en-v1.5")])
+@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)
@@ -111,12 +112,15 @@ def test_batch_embedding(n_dims: int, model_name: str) -> None:
     embeddings = list(model.embed(docs, batch_size=10))
     embeddings = np.stack(embeddings, axis=0)
 
-    assert embeddings.shape == (len(docs), n_dims)
+    assert embeddings.shape == (200, n_dims)
     if is_ci:
         delete_model_cache(model.model._model_dir)
 
 
-@pytest.mark.parametrize("n_dims,model_name", [(384, "BAAI/bge-small-en-v1.5")])
+@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)
@@ -131,7 +135,7 @@ def test_parallel_processing(n_dims: int, model_name: str) -> None:
     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 embeddings.shape == (200, n_dims)
     assert np.allclose(embeddings, embeddings_2, atol=1e-3)
     assert np.allclose(embeddings, embeddings_3, atol=1e-3)
 
@@ -139,7 +143,10 @@ def test_parallel_processing(n_dims: int, model_name: str) -> None:
         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)

tests/utils.py

@@ -3,9 +3,7 @@ import traceback
 
 from pathlib import Path
 from types import TracebackType
-from typing import Union, Callable, Any, Type, Optional
-
-from fastembed.common.model_description import BaseModelDescription
+from typing import Union, Callable, Any, Type
 
 
 def delete_model_cache(model_dir: Union[str, Path]) -> None:
@@ -37,31 +35,3 @@ 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: Optional[str],
-    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 scenarious.
-    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