Model: ModernVBERT/colmodernvbert (#588)

* Add ColModernVBERT to LateInteractionMultimodalEmbedding registry

* Implement image processing based on Idefics3ImageProcessor logic

* Fix padding support

* Implement ColModernVBERT logic

* Remove TODOs

* Handle empty pixel values with proper image_size

* Add ColModernVBERT tests

* Run pre-commit

* mypy fixes

* mypy fixes

* mypy fixes

* mypy fixes

* Fix typo in the class name

* Add processor_config.json to additional files

* Fix mypy errors

* Refactor onnx_embed_image

* Fix mypy errors

* fix: colmodernvbert tests and query processing

* fix: remove Union references

* fix: fix exit stack, update tests, implement token count

* fix: uncomment colpali in tests

* fix: lowercase models to cache

* fix: fix models to cache

* refactor: move colmodernvbert related onnx embed to its class

---------

Co-authored-by: George Panchuk <george.panchuk@qdrant.tech>

Author: kacperlukawski

fastembed/common/onnx_model.py

@@ -21,6 +21,7 @@ class OnnxOutputContext:
     model_output: NumpyArray
     attention_mask: NDArray[np.int64] | None = None
     input_ids: NDArray[np.int64] | None = None
+    metadata: dict[str, Any] | None = None
 
 
 class OnnxModel(Generic[T]):

fastembed/common/preprocessor_utils.py

@@ -50,9 +50,10 @@ def load_tokenizer(model_dir: Path) -> tuple[Tokenizer, dict[str, int]]:
 
     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"]
-    )
+    if not tokenizer.padding:
+        tokenizer.enable_padding(
+            pad_id=config.get("pad_token_id", 0), pad_token=tokenizer_config["pad_token"]
+        )
 
     for token in tokens_map.values():
         if isinstance(token, str):

fastembed/image/onnx_image_model.py

@@ -76,9 +76,11 @@ def _build_onnx_input(self, encoded: NumpyArray) -> dict[str, NumpyArray]:
         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"

fastembed/image/transform/functional.py

@@ -145,3 +145,77 @@ 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