Validate processing kwargs with @strict from huggingface_hub

src/transformers/image_processing_utils_fast.py

@@ -50,6 +50,7 @@
     logging,
 )
 from .utils.import_utils import is_rocm_platform
+from .utils.type_validators import TypedDictAdapter
 
 
 if is_vision_available():
@@ -710,6 +711,11 @@ def _validate_preprocess_kwargs(
     def preprocess(self, images: ImageInput, *args, **kwargs: Unpack[ImagesKwargs]) -> BatchFeature:
         # args are not validated, but their order in the `preprocess` and `_preprocess` signatures must be the same
         validate_kwargs(captured_kwargs=kwargs.keys(), valid_processor_keys=self._valid_kwargs_names)
+
+        # Perform type validation on received kwargs
+        type_validator = TypedDictAdapter(self.valid_kwargs)
+        type_validator.validate_fields(**kwargs)
+
         # Set default kwargs from self. This ensures that if a kwarg is not provided
         # by the user, it gets its default value from the instance, or is set to None.
         for kwarg_name in self._valid_kwargs_names:

src/transformers/models/aria/modular_aria.py

@@ -38,7 +38,7 @@
 )
 from ...modeling_flash_attention_utils import FlashAttentionKwargs
 from ...modeling_utils import PreTrainedModel
-from ...processing_utils import MultiModalData, ProcessingKwargs, ProcessorMixin, Unpack
+from ...processing_utils import ImagesKwargs, MultiModalData, ProcessingKwargs, ProcessorMixin, Unpack
 from ...tokenization_utils import PreTokenizedInput, TextInput
 from ...utils import TensorType, TransformersKwargs, auto_docstring, can_return_tuple, logging
 from ..auto import CONFIG_MAPPING, AutoConfig, AutoTokenizer
@@ -904,7 +904,15 @@ def get_number_of_image_patches(self, height: int, width: int, images_kwargs=Non
         return num_patches
 
 
+class AriaImagesKwargs(ImagesKwargs, total=False):
+    split_image: Optional[bool]
+    max_image_size: Optional[int]
+    min_image_size: Optional[int]
+
+
 class AriaProcessorKwargs(ProcessingKwargs, total=False):
+    images_kwargs: AriaImagesKwargs
+
     _defaults = {
         "text_kwargs": {
             "padding": False,

src/transformers/models/aria/processing_aria.py

@@ -24,13 +24,21 @@
 
 from ...image_processing_utils import BatchFeature
 from ...image_utils import ImageInput
-from ...processing_utils import MultiModalData, ProcessingKwargs, ProcessorMixin, Unpack
+from ...processing_utils import ImagesKwargs, MultiModalData, ProcessingKwargs, ProcessorMixin, Unpack
 from ...tokenization_utils import PreTokenizedInput, TextInput
 from ...utils import TensorType
 from ..auto import AutoTokenizer
 
 
+class AriaImagesKwargs(ImagesKwargs, total=False):
+    split_image: Optional[bool]
+    max_image_size: Optional[int]
+    min_image_size: Optional[int]
+
+
 class AriaProcessorKwargs(ProcessingKwargs, total=False):
+    images_kwargs: AriaImagesKwargs
+
     _defaults = {
         "text_kwargs": {
             "padding": False,

src/transformers/models/deepseek_vl/image_processing_deepseek_vl.py

@@ -56,7 +56,7 @@ class DeepseekVLImageProcessorKwargs(ImagesKwargs):
         falls below this value after resizing.
     """
 
-    min_size: int
+    min_size: Optional[int]
 
 
 class DeepseekVLImageProcessor(BaseImageProcessor):

src/transformers/models/deepseek_vl_hybrid/image_processing_deepseek_vl_hybrid.py

@@ -69,11 +69,11 @@ class DeepseekVLHybridImageProcessorKwargs(ImagesKwargs):
         number of channels in the image. Can be overridden by the `high_res_image_std` parameter in the `preprocess` method.
     """
 
-    min_size: int
-    high_res_size: dict
-    high_res_resample: "PILImageResampling"
-    high_res_image_mean: list[float]
-    high_res_image_std: list[float]
+    min_size: Optional[int]
+    high_res_size: Optional[dict]
+    high_res_resample: Optional[Union["PILImageResampling", int]]
+    high_res_image_mean: Optional[Union[float, list[float], tuple[float, ...]]]
+    high_res_image_std: Optional[Union[float, list[float], tuple[float, ...]]]
 
 
 class DeepseekVLHybridImageProcessor(BaseImageProcessor):

src/transformers/models/deepseek_vl_hybrid/modular_deepseek_vl_hybrid.py

@@ -448,11 +448,11 @@ class DeepseekVLHybridImageProcessorKwargs(ImagesKwargs):
         number of channels in the image. Can be overridden by the `high_res_image_std` parameter in the `preprocess` method.
     """
 
-    min_size: int
-    high_res_size: dict
-    high_res_resample: "PILImageResampling"
-    high_res_image_mean: list[float]
-    high_res_image_std: list[float]
+    min_size: Optional[int]
+    high_res_size: Optional[dict]
+    high_res_resample: Optional[Union["PILImageResampling", int]]
+    high_res_image_mean: Optional[Union[float, list[float], tuple[float, ...]]]
+    high_res_image_std: Optional[Union[float, list[float], tuple[float, ...]]]
 
 
 class DeepseekVLHybridImageProcessor(DeepseekVLImageProcessor):

src/transformers/models/dia/processing_dia.py

@@ -55,7 +55,9 @@ class DiaProcessorKwargs(ProcessingKwargs, total=False):
             "generation": True,
             "sampling_rate": 44100,
         },
-        "common_kwargs": {"return_tensors": "pt"},
+        "common_kwargs": {
+            "return_tensors": "pt",
+        },
     }
 
 

src/transformers/models/efficientnet/image_processing_efficientnet.py

@@ -52,8 +52,8 @@ class EfficientNetImageProcessorKwargs(ImagesKwargs):
         Normalize the image again with the standard deviation only for image classification if set to True.
     """
 
-    rescale_offset: bool
-    include_top: bool
+    rescale_offset: Optional[bool]
+    include_top: Optional[bool]
 
 
 class EfficientNetImageProcessor(BaseImageProcessor):

src/transformers/models/eomt/image_processing_eomt.py

@@ -66,8 +66,8 @@ class EomtImageProcessorKwargs(ImagesKwargs):
         denoted with 0 (background) will be replaced with `ignore_index`.
     """
 
-    do_split_image: bool
-    ignore_index: Optional[int] = None
+    do_split_image: Optional[bool]
+    ignore_index: Optional[int]
 
 
 # Adapted from transformers.models.maskformer.image_processing_maskformer.convert_segmentation_map_to_binary_masks

src/transformers/models/internvl/video_processing_internvl.py

@@ -28,7 +28,7 @@
 
 
 class InternVLVideoProcessorInitKwargs(VideosKwargs):
-    initial_shift: Union[bool, float, int]
+    initial_shift: Optional[Union[bool, float, int]]
 
 
 class InternVLVideoProcessor(BaseVideoProcessor):

src/transformers/models/janus/image_processing_janus.py

@@ -58,7 +58,7 @@ class JanusImageProcessorKwargs(ImagesKwargs):
         falls below this value after resizing.
     """
 
-    min_size: int
+    min_size: Optional[int]
 
 
 class JanusImageProcessor(BaseImageProcessor):

src/transformers/models/kosmos2/processing_kosmos2.py

@@ -33,9 +33,11 @@
     list[list[tuple[float, float, float]]],
 ]
 
+NestedList = list[Union[Optional[int], "NestedList"]]
+
 
 class Kosmos2ImagesKwargs(ImagesKwargs, total=False):
-    bboxes: Optional[list[float]]
+    bboxes: Optional[NestedList]
     num_image_tokens: Optional[int]
     first_image_token_id: Optional[int]
 

src/transformers/models/lfm2_vl/processing_lfm2_vl.py

@@ -21,6 +21,7 @@
     ImagesKwargs,
     ProcessingKwargs,
     ProcessorMixin,
+    TextKwargs,
     Unpack,
 )
 from ...tokenization_utils_base import BatchEncoding, TextInput
@@ -46,8 +47,13 @@ class Lfm2VlImagesKwargs(ImagesKwargs, total=False):
     return_row_col_info: Optional[bool]
 
 
+class Lfm2VlTextKwargs(TextKwargs, total=False):
+    use_image_special_tokens: Optional[bool]
+
+
 class Lfm2VlProcessorKwargs(ProcessingKwargs, total=False):
     images_kwargs: Lfm2VlImagesKwargs
+    text_kwargs: Lfm2VlTextKwargs
 
     _defaults = {
         "images_kwargs": {
@@ -75,8 +81,6 @@ class Lfm2VlProcessor(ProcessorMixin):
             An instance of [`PreTrainedTokenizerBase`]. This should correspond with the model's text model. The tokenizer is a required input.
         chat_template (`str`, *optional*):
             A Jinja template which will be used to convert lists of messages in a chat into a tokenizable string.
-        use_image_special_tokens (`bool`, *optional*, defaults to `True`):
-            Whether to use image special tokens or not when processing.
     """
 
     attributes = ["image_processor", "tokenizer"]
@@ -88,12 +92,10 @@ def __init__(
         image_processor,
         tokenizer,
         chat_template: Optional[str] = None,
-        use_image_special_tokens: Optional[bool] = True,
         **kwargs,
     ):
         self.image_token = tokenizer.image_token
         self.image_token_id = tokenizer.image_token_id
-        self.use_image_special_tokens = use_image_special_tokens
         self.image_start_token = tokenizer.image_start_token
         self.image_end_token = tokenizer.image_end_token
         self.image_thumbnail_token = tokenizer.image_thumbnail

src/transformers/models/llava_onevision/image_processing_llava_onevision_fast.py

@@ -76,8 +76,7 @@ def preprocess(self, images: ImageInput, **kwargs: Unpack[LlavaOnevisionImagePro
             batch_num_images = [1] * len(images)
         else:
             batch_num_images = [1]
-        kwargs["batch_num_images"] = batch_num_images
-        return super().preprocess(images, **kwargs)
+        return super().preprocess(images, batch_num_images, **kwargs)
 
     def _resize_for_patching(
         self,
@@ -202,6 +201,7 @@ def _pad_for_batching(
     def _preprocess(
         self,
         images: list["torch.Tensor"],
+        batch_num_images: list[int],
         do_resize: bool,
         size: SizeDict,
         image_grid_pinpoints: list[list[int]],
@@ -214,7 +214,6 @@ def _preprocess(
         image_mean: Optional[Union[float, list[float]]],
         image_std: Optional[Union[float, list[float]]],
         do_pad: bool,
-        batch_num_images: list[int],
         disable_grouping: Optional[bool],
         return_tensors: Optional[Union[str, TensorType]],
         **kwargs,

src/transformers/models/llava_onevision/modular_llava_onevision.py

@@ -35,7 +35,7 @@
 
 from ...cache_utils import Cache
 from ...image_processing_utils import BatchFeature
-from ...image_processing_utils_fast import group_images_by_shape, reorder_images
+from ...image_processing_utils_fast import BaseImageProcessorFast, group_images_by_shape, reorder_images
 from ...image_utils import (
     OPENAI_CLIP_MEAN,
     OPENAI_CLIP_STD,
@@ -128,12 +128,12 @@ def preprocess(self, images: ImageInput, **kwargs: Unpack[LlavaOnevisionImagePro
             batch_num_images = [1] * len(images)
         else:
             batch_num_images = [1]
-        kwargs["batch_num_images"] = batch_num_images
-        return super().preprocess(images, **kwargs)
+        return BaseImageProcessorFast.preprocess(images, batch_num_images, **kwargs)
 
     def _preprocess(
         self,
         images: list["torch.Tensor"],
+        batch_num_images: list[int],
         do_resize: bool,
         size: SizeDict,
         image_grid_pinpoints: list[list[int]],
@@ -146,7 +146,6 @@ def _preprocess(
         image_mean: Optional[Union[float, list[float]]],
         image_std: Optional[Union[float, list[float]]],
         do_pad: bool,
-        batch_num_images: list[int],
         disable_grouping: Optional[bool],
         return_tensors: Optional[Union[str, TensorType]],
         **kwargs,

src/transformers/models/mllama/processing_mllama.py

@@ -258,9 +258,7 @@ def __call__(
             tokenizer_init_kwargs=self.tokenizer.init_kwargs,
             **kwargs,
         )
-
         return_tensors = output_kwargs["text_kwargs"].pop("return_tensors", None)
-        images_kwargs = output_kwargs["images_kwargs"]
 
         data = {}
         if text is not None:
@@ -306,7 +304,7 @@ def __call__(
                     )
 
         if images is not None:
-            image_features = self.image_processor(images, **images_kwargs)
+            image_features = self.image_processor(images, **output_kwargs["images_kwargs"])
             num_tiles = image_features.pop("num_tiles")
             data.update(image_features)
 

src/transformers/models/perception_lm/image_processing_perception_lm_fast.py

@@ -68,7 +68,7 @@ class PerceptionLMImageProcessorFast(BaseImageProcessorFast):
     do_rescale = True
     do_normalize = True
     do_convert_rgb = True
-    vision_input_type = "thumb+tail"
+    vision_input_type = "thumb+tile"
     tile_size = 448
     max_num_tiles = 36
     size = {"width": 448, "height": 448}  # for backward compatibility in tests

src/transformers/models/pixtral/image_processing_pixtral.py

@@ -52,11 +52,11 @@
 
 class PixtralImageProcessorKwargs(ImagesKwargs):
     """
-    patch_size (`dict[str, int]` *optional*, defaults to `{"height": 16, "width": 16}`):
+    patch_size (`Union[dict[str, int], int]` *optional*, defaults to `{"height": 16, "width": 16}`):
         Size of the patches in the model, used to calculate the output image size. Can be overridden by `patch_size` in the `preprocess` method.
     """
 
-    patch_size: Optional[dict[str, int]]
+    patch_size: Optional[Union[dict[str, int], int]]
 
 
 # Adapted from function in image_transforms.py to ensure any transparent pixels are converted to white.

src/transformers/models/qwen2_5_omni/processing_qwen2_5_omni.py

@@ -42,7 +42,7 @@ class Qwen2_5_OmniVideosKwargs(VideosKwargs):
     max_frames: Optional[int]
     use_audio_in_video: Optional[bool]
     seconds_per_chunk: Optional[float]
-    position_id_per_seconds: Optional[int]
+    position_id_per_seconds: Optional[Union[int, float]]
 
 
 class Qwen2_5OmniProcessorKwargs(ProcessingKwargs, total=False):

src/transformers/models/qwen3_omni_moe/processing_qwen3_omni_moe.py

@@ -20,7 +20,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import re
-from typing import Optional
+from typing import Optional, Union
 
 import numpy as np
 
@@ -44,7 +44,7 @@ class Qwen3OmniMoeVideosKwargs(VideosKwargs):
     max_frames: Optional[int]
     use_audio_in_video: Optional[bool]
     seconds_per_chunk: Optional[float]
-    position_id_per_seconds: Optional[int]
+    position_id_per_seconds: Optional[Union[int, float]]
 
 
 class Qwen3OmniMoeProcessorKwargs(ProcessingKwargs, total=False):

src/transformers/models/udop/processing_udop.py

@@ -31,7 +31,7 @@
 
 class UdopTextKwargs(TextKwargs, total=False):
     word_labels: Optional[Union[list[int], list[list[int]]]]
-    boxes: Union[list[list[int]], list[list[list[int]]]]
+    boxes: Optional[Union[list[list[int]], list[list[list[int]]]]]
 
 
 class UdopProcessorKwargs(ProcessingKwargs, total=False):