Add optimized PixtralImageProcessorFast

docs/source/en/_config.py

@@ -11,4 +11,4 @@
     "{processor_class}": "FakeProcessorClass",
     "{model_class}": "FakeModelClass",
     "{object_class}": "FakeObjectClass",
-}
+}

docs/source/en/model_doc/pixtral.md

@@ -88,6 +88,11 @@ output = processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up
 [[autodoc]] PixtralImageProcessor
     - preprocess
 
+## PixtralImageProcessorFast
+
+[[autodoc]] PixtralImageProcessorFast
+    - preprocess
+
 ## PixtralProcessor
 
 [[autodoc]] PixtralProcessor

src/transformers/__init__.py

@@ -1258,6 +1258,7 @@
     ]
 else:
     _import_structure["image_processing_utils_fast"] = ["BaseImageProcessorFast"]
+    _import_structure["models.pixtral"].append("PixtralImageProcessorFast")
     _import_structure["models.vit"].append("ViTImageProcessorFast")
 
 # PyTorch-backed objects
@@ -6188,6 +6189,7 @@
         from .utils.dummy_torchvision_objects import *
     else:
         from .image_processing_utils_fast import BaseImageProcessorFast
+        from .models.pixtral import PixtralImageProcessorFast
         from .models.vit import ViTImageProcessorFast
 
     # Modeling

src/transformers/image_utils.py

@@ -24,6 +24,7 @@
 
 from .utils import (
     ExplicitEnum,
+    TensorType,
     is_jax_tensor,
     is_numpy_array,
     is_tf_tensor,
@@ -447,6 +448,44 @@ def validate_preprocess_arguments(
         raise ValueError("`size` and `resample` must be specified if `do_resize` is `True`.")
 
 
+def validate_fast_preprocess_arguments(
+    do_rescale: Optional[bool] = None,
+    rescale_factor: Optional[float] = None,
+    do_normalize: Optional[bool] = None,
+    image_mean: Optional[Union[float, List[float]]] = None,
+    image_std: Optional[Union[float, List[float]]] = None,
+    do_pad: Optional[bool] = None,
+    size_divisibility: Optional[int] = None,
+    do_center_crop: Optional[bool] = None,
+    crop_size: Optional[Dict[str, int]] = None,
+    do_resize: Optional[bool] = None,
+    size: Optional[Dict[str, int]] = None,
+    resample: Optional["PILImageResampling"] = None,
+    return_tensors: Optional[Union[str, TensorType]] = None,
+    data_format: Optional[ChannelDimension] = ChannelDimension.FIRST,
+):
+    """
+    Checks validity of typically used arguments in an `ImageProcessorFast` `preprocess` method.
+    Raises `ValueError` if arguments incompatibility is caught.
+    """
+    validate_preprocess_arguments(
+        do_rescale=do_rescale,
+        rescale_factor=rescale_factor,
+        do_normalize=do_normalize,
+        image_mean=image_mean,
+        image_std=image_std,
+        do_resize=do_resize,
+        size=size,
+        resample=resample,
+    )
+    # Extra checks for ImageProcessorFast
+    if return_tensors != "pt":
+        raise ValueError("Only returning PyTorch tensors is currently supported.")
+
+    if data_format != ChannelDimension.FIRST:
+        raise ValueError("Only channel first data format is currently supported.")
+
+
 # In the future we can add a TF implementation here when we have TF models.
 class ImageFeatureExtractionMixin:
     """

src/transformers/models/auto/image_processing_auto.py

@@ -117,7 +117,7 @@
             ("paligemma", ("SiglipImageProcessor",)),
             ("perceiver", ("PerceiverImageProcessor",)),
             ("pix2struct", ("Pix2StructImageProcessor",)),
-            ("pixtral", ("PixtralImageProcessor",)),
+            ("pixtral", ("PixtralImageProcessor", "PixtralImageProcessorFast")),
             ("poolformer", ("PoolFormerImageProcessor",)),
             ("pvt", ("PvtImageProcessor",)),
             ("pvt_v2", ("PvtImageProcessor",)),

src/transformers/models/pixtral/__init__.py

@@ -13,7 +13,13 @@
 # limitations under the License.
 from typing import TYPE_CHECKING
 
-from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_torch_available, is_vision_available
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_torch_available,
+    is_torchvision_available,
+    is_vision_available,
+)
 
 
 _import_structure = {
@@ -41,6 +47,14 @@
 else:
     _import_structure["image_processing_pixtral"] = ["PixtralImageProcessor"]
 
+try:
+    if not is_torchvision_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["image_processing_pixtral_fast"] = ["PixtralImageProcessorFast"]
+
 
 if TYPE_CHECKING:
     from .configuration_pixtral import PixtralVisionConfig
@@ -65,6 +79,14 @@
     else:
         from .image_processing_pixtral import PixtralImageProcessor
 
+    try:
+        if not is_torchvision_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .image_processing_pixtral_fast import PixtralImageProcessorFast
+
 else:
     import sys