first commit

tests/models/ernie4_5_vl_moe/test_image_processing_ernie4_5_vl_moe.py

@@ -0,0 +1,334 @@
+# Copyright 2025 HuggingFace Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import itertools
+import tempfile
+import unittest
+
+import numpy as np
+
+from transformers.image_utils import OPENAI_CLIP_MEAN, OPENAI_CLIP_STD, load_image
+from transformers.models.ernie4_5_vl_moe.image_processing_ernie4_5_vl_moe import smart_resize
+from transformers.testing_utils import require_torch, require_vision
+from transformers.utils import is_torch_available, is_vision_available
+
+from ...test_image_processing_common import ImageProcessingTestMixin, prepare_image_inputs
+from ...test_processing_common import url_to_local_path
+
+
+if is_torch_available():
+    import torch
+
+if is_vision_available():
+    from PIL import Image
+
+
+class Ernie4_5_VLMoeImageProcessorTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=7,
+        num_channels=3,
+        min_resolution=56,
+        max_resolution=1024,
+        size=None,
+        do_resize=True,
+        do_normalize=True,
+        do_convert_rgb=True,
+        image_mean=OPENAI_CLIP_MEAN,
+        image_std=OPENAI_CLIP_STD,
+        patch_size=14,
+        merge_size=2,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.num_channels = num_channels
+        self.min_resolution = min_resolution
+        self.max_resolution = max_resolution
+        if size is None:
+            size = {"shortest_edge": 56 * 56, "longest_edge": 6177 * 28 * 28}
+        self.size = size
+        self.do_resize = do_resize
+        self.do_normalize = do_normalize
+        self.do_convert_rgb = do_convert_rgb
+        self.image_mean = image_mean
+        self.image_std = image_std
+        self.patch_size = patch_size
+        self.merge_size = merge_size
+
+    def prepare_image_processor_dict(self):
+        return {
+            "do_resize": self.do_resize,
+            "image_mean": self.image_mean,
+            "image_std": self.image_std,
+            "size": self.size,
+            "patch_size": self.patch_size,
+            "merge_size": self.merge_size,
+        }
+
+    def prepare_image_inputs(self, equal_resolution=False, numpify=False, torchify=False):
+        images = prepare_image_inputs(
+            batch_size=self.batch_size,
+            num_channels=self.num_channels,
+            min_resolution=self.min_resolution,
+            max_resolution=self.max_resolution,
+            equal_resolution=equal_resolution,
+            numpify=numpify,
+            torchify=torchify,
+        )
+        return [[image] for image in images]
+
+
+@require_torch
+@require_vision
+class Ernie4_5_VLMoeImageProcessingTest(ImageProcessingTestMixin, unittest.TestCase):
+    def setUp(self):
+        super().setUp()
+        self.image_processor_tester = Ernie4_5_VLMoeImageProcessorTester(self)
+
+    @property
+    def image_processor_dict(self):
+        return self.image_processor_tester.prepare_image_processor_dict()
+
+    def test_image_processor_properties(self):
+        for image_processing_class in self.image_processing_classes.values():
+            image_processing = image_processing_class(**self.image_processor_dict)
+            self.assertTrue(hasattr(image_processing, "do_normalize"))
+            self.assertTrue(hasattr(image_processing, "image_mean"))
+            self.assertTrue(hasattr(image_processing, "image_std"))
+            self.assertTrue(hasattr(image_processing, "do_resize"))
+            self.assertTrue(hasattr(image_processing, "size"))
+            self.assertTrue(hasattr(image_processing, "do_convert_rgb"))
+            self.assertTrue(hasattr(image_processing, "patch_size"))
+            self.assertTrue(hasattr(image_processing, "merge_size"))
+
+    def test_image_processor_from_dict_with_kwargs(self):
+        for image_processing_class in self.image_processing_classes.values():
+            image_processor = image_processing_class.from_dict(self.image_processor_dict)
+            self.assertEqual(image_processor.size["shortest_edge"], 56 * 56)
+            self.assertEqual(image_processor.size["longest_edge"], 6177 * 28 * 28)
+
+            image_processor = image_processing_class.from_dict(
+                self.image_processor_dict,
+                size={"shortest_edge": 256 * 256, "longest_edge": 640 * 640},
+            )
+            self.assertEqual(image_processor.size["shortest_edge"], 256 * 256)
+            self.assertEqual(image_processor.size["longest_edge"], 640 * 640)
+
+    def test_select_best_resolution(self):
+        # Test with a final resize resolution
+        best_resolution = smart_resize(561, 278, factor=28)
+        self.assertEqual(best_resolution, (560, 280))
+
+    def test_call_pil(self):
+        for image_processing_class in self.image_processing_classes.values():
+            # Initialize image_processing
+            image_processing = image_processing_class(**self.image_processor_dict)
+            # create random PIL images
+            image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=True)
+            for image in image_inputs:
+                self.assertIsInstance(image[0], Image.Image)
+
+            # Test not batched input
+            process_out = image_processing(image_inputs[0], return_tensors="pt")
+            encoded_images = process_out.pixel_values
+            image_grid_thws = process_out.image_grid_thw
+            expected_output_image_shape = (5476, 588)
+            expected_image_grid_thws = torch.Tensor([[1, 74, 74]])
+            self.assertEqual(tuple(encoded_images.shape), expected_output_image_shape)
+            self.assertTrue((image_grid_thws == expected_image_grid_thws).all())
+
+            # Test batched
+            process_out = image_processing(image_inputs, return_tensors="pt")
+            encoded_images = process_out.pixel_values
+            image_grid_thws = process_out.image_grid_thw
+            expected_output_image_shape = (38332, 588)
+            expected_image_grid_thws = torch.Tensor([[1, 74, 74]] * 7)
+            self.assertEqual(tuple(encoded_images.shape), expected_output_image_shape)
+            self.assertTrue((image_grid_thws == expected_image_grid_thws).all())
+
+    def test_call_numpy(self):
+        for image_processing_class in self.image_processing_classes.values():
+            # Initialize image_processing
+            image_processing = image_processing_class(**self.image_processor_dict)
+            # create random numpy tensors
+            image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=True, numpify=True)
+            for image in image_inputs:
+                self.assertIsInstance(image[0], np.ndarray)
+
+            # Test not batched input
+            process_out = image_processing(image_inputs[0], return_tensors="pt")
+            encoded_images = process_out.pixel_values
+            image_grid_thws = process_out.image_grid_thw
+            expected_output_image_shape = (5476, 588)
+            expected_image_grid_thws = torch.Tensor([[1, 74, 74]])
+            self.assertEqual(tuple(encoded_images.shape), expected_output_image_shape)
+            self.assertTrue((image_grid_thws == expected_image_grid_thws).all())
+
+            # Test batched
+            process_out = image_processing(image_inputs, return_tensors="pt")
+            encoded_images = process_out.pixel_values
+            image_grid_thws = process_out.image_grid_thw
+            expected_output_image_shape = (38332, 588)
+            expected_image_grid_thws = torch.Tensor([[1, 74, 74]] * 7)
+            self.assertEqual(tuple(encoded_images.shape), expected_output_image_shape)
+            self.assertTrue((image_grid_thws == expected_image_grid_thws).all())
+
+    def test_call_pytorch(self):
+        for image_processing_class in self.image_processing_classes.values():
+            # Initialize image_processing
+            image_processing = image_processing_class(**self.image_processor_dict)
+            # create random PyTorch tensors
+            image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=True, torchify=True)
+
+            for image in image_inputs:
+                self.assertIsInstance(image[0], torch.Tensor)
+
+            # Test not batched input
+            process_out = image_processing(image_inputs[0], return_tensors="pt")
+            encoded_images = process_out.pixel_values
+            image_grid_thws = process_out.image_grid_thw
+            expected_output_image_shape = (5476, 588)
+            expected_image_grid_thws = torch.Tensor([[1, 74, 74]])
+            self.assertEqual(tuple(encoded_images.shape), expected_output_image_shape)
+            self.assertTrue((image_grid_thws == expected_image_grid_thws).all())
+
+            # Test batched
+            process_out = image_processing(image_inputs, return_tensors="pt")
+            encoded_images = process_out.pixel_values
+            image_grid_thws = process_out.image_grid_thw
+            expected_output_image_shape = (38332, 588)
+            expected_image_grid_thws = torch.Tensor([[1, 74, 74]] * 7)
+            self.assertEqual(tuple(encoded_images.shape), expected_output_image_shape)
+            self.assertTrue((image_grid_thws == expected_image_grid_thws).all())
+
+    @unittest.skip(reason="Erni4_5_VLImageProcessor doesn't treat 4 channel PIL and numpy consistently yet")
+    def test_call_numpy_4_channels(self):
+        pass
+
+    def test_nested_input(self):
+        for image_processing_class in self.image_processing_classes.values():
+            image_processing = image_processing_class(**self.image_processor_dict)
+            image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=True)
+
+            # Test batched as a list of images
+            process_out = image_processing(image_inputs, return_tensors="pt")
+            encoded_images = process_out.pixel_values
+            image_grid_thws = process_out.image_grid_thw
+            expected_output_image_shape = (38332, 588)
+            expected_image_grid_thws = torch.Tensor([[1, 74, 74]] * 7)
+            self.assertEqual(tuple(encoded_images.shape), expected_output_image_shape)
+            self.assertTrue((image_grid_thws == expected_image_grid_thws).all())
+
+            # Test batched as a nested list of images, where each sublist is one batch
+            image_inputs_nested = image_inputs[:3] + image_inputs[3:]
+            process_out = image_processing(image_inputs_nested, return_tensors="pt")
+            encoded_images_nested = process_out.pixel_values
+            image_grid_thws_nested = process_out.image_grid_thw
+            expected_output_image_shape = (38332, 588)
+            expected_image_grid_thws = torch.Tensor([[1, 74, 74]] * 7)
+            self.assertEqual(tuple(encoded_images_nested.shape), expected_output_image_shape)
+            self.assertTrue((image_grid_thws == expected_image_grid_thws).all())
+
+            # Image processor should return same pixel values, independently of ipnut format
+            self.assertTrue((encoded_images_nested == encoded_images).all())
+            self.assertTrue((image_grid_thws_nested == expected_image_grid_thws).all())
+
+    def test_custom_image_size(self):
+        for image_processing_class in self.image_processing_classes.values():
+            image_processing = image_processing_class(**self.image_processor_dict)
+            with tempfile.TemporaryDirectory() as tmpdirname:
+                image_processing.save_pretrained(tmpdirname)
+                image_processor_loaded = image_processing_class.from_pretrained(
+                    tmpdirname, size={"shortest_edge": 28 * 28, "longest_edge": 56 * 56}
+                )
+
+            image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=True)
+            process_out = image_processor_loaded(image_inputs, return_tensors="pt")
+            expected_output_image_shape = [112, 588]
+            self.assertListEqual(list(process_out.pixel_values.shape), expected_output_image_shape)
+
+    def test_custom_pixels(self):
+        pixel_choices = frozenset(itertools.product((100, 150, 200, 20000), (100, 150, 200, 20000)))
+        for image_processing_class in self.image_processing_classes.values():
+            image_processor_dict = self.image_processor_dict.copy()
+            for a_pixels, b_pixels in pixel_choices:
+                image_processor_dict["size"] = {
+                    "shortest_edge": min(a_pixels, b_pixels),
+                    "longest_edge": max(a_pixels, b_pixels),
+                }
+                image_processor = image_processing_class(**image_processor_dict)
+                image_inputs = self.image_processor_tester.prepare_image_inputs()
+                # Just checking that it doesn't raise an error
+                image_processor(image_inputs, return_tensors="pt")
+
+    @require_vision
+    @require_torch
+    def test_backends_equivalence(self):
+        """Override base test to also compare image_grid_thw."""
+        if len(self.image_processing_classes) < 2:
+            self.skipTest(reason="Skipping backends equivalence test as there are less than 2 backends")
+
+        dummy_image = load_image(url_to_local_path("http://images.cocodataset.org/val2017/000000039769.jpg"))
+
+        image_processor_pil = self.image_processing_classes["pil"](**self.image_processor_dict)
+        image_processor_torchvision = self.image_processing_classes["torchvision"](**self.image_processor_dict)
+
+        encoding_pil = image_processor_pil(dummy_image, return_tensors="pt")
+        encoding_torchvision = image_processor_torchvision(dummy_image, return_tensors="pt")
+
+        self._assert_tensors_equivalence(encoding_pil.pixel_values, encoding_torchvision.pixel_values)
+        self.assertEqual(encoding_pil.image_grid_thw.dtype, encoding_torchvision.image_grid_thw.dtype)
+        self._assert_tensors_equivalence(
+            encoding_pil.image_grid_thw.float(), encoding_torchvision.image_grid_thw.float()
+        )
+
+    @require_vision
+    @require_torch
+    def test_backends_equivalence_batched(self):
+        """Override base test to also compare image_grid_thw."""
+        if len(self.image_processing_classes) < 2:
+            self.skipTest(reason="Skipping backends equivalence test as there are less than 2 backends")
+
+        if hasattr(self.image_processor_tester, "do_center_crop") and self.image_processor_tester.do_center_crop:
+            self.skipTest(
+                reason="Skipping as do_center_crop is True and center_crop functions are not equivalent for fast and slow processors"
+            )
+
+        dummy_images = self.image_processor_tester.prepare_image_inputs(equal_resolution=False, torchify=True)
+        image_processor_pil = self.image_processing_classes["pil"](**self.image_processor_dict)
+        image_processor_torchvision = self.image_processing_classes["torchvision"](**self.image_processor_dict)
+
+        encoding_pil = image_processor_pil(dummy_images, return_tensors="pt")
+        encoding_torchvision = image_processor_torchvision(dummy_images, return_tensors="pt")
+
+        self._assert_tensors_equivalence(encoding_pil.pixel_values, encoding_torchvision.pixel_values)
+        self.assertEqual(encoding_pil.image_grid_thw.dtype, encoding_torchvision.image_grid_thw.dtype)
+        self._assert_tensors_equivalence(
+            encoding_pil.image_grid_thw.float(), encoding_torchvision.image_grid_thw.float()
+        )
+
+    def test_get_num_patches_without_images(self):
+        for image_processing_class in self.image_processing_classes.values():
+            image_processing = image_processing_class(**self.image_processor_dict)
+            num_patches = image_processing.get_number_of_image_patches(height=100, width=100, images_kwargs={})
+            self.assertEqual(num_patches, 64)
+
+            num_patches = image_processing.get_number_of_image_patches(height=200, width=50, images_kwargs={})
+            self.assertEqual(num_patches, 56)
+
+            num_patches = image_processing.get_number_of_image_patches(
+                height=100, width=100, images_kwargs={"patch_size": 28}
+            )
+            self.assertEqual(num_patches, 16)

tests/models/ernie4_5_vl_moe/test_modeling_ernie4_5_vl_moe.py

@@ -0,0 +1,784 @@
+# Copyright 2025 Baidu and HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Testing suite for the PyTorch Ernie 4.5 VL model."""
+
+import unittest
+
+from transformers import (
+    AutoModelForImageTextToText,
+    AutoProcessor,
+    Ernie4_5_VLMoeConfig,
+    Ernie4_5_VLMoeForConditionalGeneration,
+    Ernie4_5_VLMoeModel,
+    is_torch_available,
+    is_vision_available,
+)
+from transformers.testing_utils import (
+    Expectations,
+    cleanup,
+    require_deterministic_for_xpu,
+    require_torch,
+    require_torch_large_accelerator,
+    slow,
+    torch_device,
+)
+from transformers.utils import is_cv2_available
+
+from ...generation.test_utils import GenerationTesterMixin
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_common import (
+    ModelTesterMixin,
+    floats_tensor,
+    ids_tensor,
+)
+
+
+if is_cv2_available():
+    pass
+
+if is_torch_available():
+    import torch
+
+if is_vision_available():
+    pass
+
+
+class Ernie4_5_VLMoeVisionText2TextModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=3,
+        seq_length=7,
+        num_channels=3,
+        ignore_index=-100,
+        image_size=112,
+        video_start_token_id=3,
+        video_end_token_id=4,
+        image_start_token_id=5,
+        image_end_token_id=6,
+        image_token_id=7,
+        video_token_id=8,
+        is_training=True,
+        text_config=None,
+        vision_config=None,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.seq_length = seq_length
+        self.num_channels = num_channels
+        self.ignore_index = ignore_index
+        self.image_size = image_size
+        self.bos_token_id = 0
+        self.eos_token_id = 0
+        self.pad_token_id = 0
+        self.video_start_token_id = video_start_token_id
+        self.video_end_token_id = video_end_token_id
+        self.image_start_token_id = image_start_token_id
+        self.image_end_token_id = image_end_token_id
+        self.image_token_id = image_token_id
+        self.video_token_id = video_token_id
+        self.is_training = is_training
+
+        self.text_config = text_config
+        if text_config is None:
+            self.text_config = {
+                "vocab_size": 99,
+                "hidden_size": 16,
+                "intermediate_size": 32,
+                "num_hidden_layers": 2,
+                "num_attention_heads": 2,
+                "num_key_value_heads": 1,
+                "hidden_act": "silu",
+                "max_position_embeddings": 512,
+                "tie_word_embeddings": True,
+                "rope_parameters": {"type": "default", "rope_theta": 500_000.0, "mrope_section": [1, 1, 2]},
+                "mlp_layer_types": ["dense", "sparse"],
+                "moe_intermediate_size": [32, 32],
+                "moe_k": 2,
+                "moe_num_experts": 8,
+                "moe_num_shared_experts": 2,
+                "moe_norm_min": 1e-12,
+            }
+
+        self.vision_config = vision_config
+        if vision_config is None:
+            self.vision_config = {
+                "depth": 2,
+                "hidden_size": 32,
+                "hidden_act": "silu",
+                "intermediate_size": 32,
+                "num_heads": 2,
+                "spatial_merge_size": 1,
+            }
+
+        self.hidden_size = self.text_config["hidden_size"]
+        self.num_hidden_layers = self.text_config["num_hidden_layers"]
+        self.num_attention_heads = self.text_config["num_attention_heads"]
+        self.vocab_size = self.text_config["vocab_size"]
+
+        self.num_image_tokens = 64
+        self.seq_length = seq_length + self.num_image_tokens
+
+    def get_config(self):
+        return Ernie4_5_VLMoeConfig(
+            text_config=self.text_config,
+            vision_config=self.vision_config,
+            image_token_id=self.image_token_id,
+            video_token_id=self.video_token_id,
+            video_start_token_id=self.video_start_token_id,
+            video_end_token_id=self.video_end_token_id,
+            image_start_token_id=self.image_start_token_id,
+            image_end_token_id=self.image_end_token_id,
+        )
+
+    def prepare_config_and_inputs(self):
+        config = self.get_config()
+        patch_size = config.vision_config.patch_size
+        pixel_values = floats_tensor(
+            [self.batch_size * (self.image_size**2) // (patch_size**2), self.num_channels * (patch_size**2)]
+        )
+
+        return config, pixel_values
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        config, pixel_values = config_and_inputs
+        input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
+        attention_mask = torch.ones(input_ids.shape, dtype=torch.long, device=torch_device)
+
+        input_ids[input_ids == self.video_token_id] = self.pad_token_id
+        input_ids[input_ids == self.image_token_id] = self.pad_token_id
+        input_ids[input_ids == self.video_start_token_id] = self.pad_token_id
+        input_ids[input_ids == self.image_start_token_id] = self.pad_token_id
+        input_ids[input_ids == self.video_end_token_id] = self.pad_token_id
+        input_ids[input_ids == self.image_end_token_id] = self.pad_token_id
+
+        input_ids[:, 0] = self.image_start_token_id
+        input_ids[:, 1 : 1 + self.num_image_tokens] = self.image_token_id
+        input_ids[:, 1 + self.num_image_tokens] = self.image_end_token_id
+
+        patch_size = config.vision_config.patch_size
+        patches_per_side = self.image_size // patch_size
+
+        mm_token_type_ids = torch.zeros_like(input_ids)
+        mm_token_type_ids[input_ids == self.image_token_id] = 1
+        inputs_dict = {
+            "pixel_values": pixel_values,
+            "image_grid_thw": torch.tensor(
+                [[1, patches_per_side, patches_per_side]] * self.batch_size, device=torch_device
+            ),
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+            "mm_token_type_ids": mm_token_type_ids,
+        }
+        return config, inputs_dict
+
+
+@require_torch
+class Ernie4_5_VLMoeModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):
+    all_model_classes = (
+        (
+            Ernie4_5_VLMoeModel,
+            Ernie4_5_VLMoeForConditionalGeneration,
+        )
+        if is_torch_available()
+        else ()
+    )
+    model_split_percents = [0.7, 0.9]  # model too big to split at 0.5
+    test_all_params_have_gradient = False  # e score correction bias + moe
+    _is_composite = True
+
+    def setUp(self):
+        self.model_tester = Ernie4_5_VLMoeVisionText2TextModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=Ernie4_5_VLMoeConfig, has_text_modality=False)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def prepare_config_and_inputs_for_generate(self, batch_size=2):
+        """
+        Same as in GLM4V, see `tests/models/glm4v/test_modeling_glm4v.py` for reference
+        """
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        # We don't want a few model inputs in our model input dictionary for generation tests
+        input_keys_to_ignore = [
+            # we don't want encoder-decoder models to start from filled decoder ids
+            "decoder_input_ids",
+            "decoder_attention_mask",
+            # we'll set cache use in each test differently
+            "use_cache",
+            # ignore labels if it is in the input dict
+            "labels",
+        ]
+
+        # The diff from the general `prepare_config_and_inputs_for_generate` lies here
+        patch_size = config.vision_config.patch_size
+        filtered_image_length = batch_size * (self.model_tester.image_size**2) // (patch_size**2)
+        filtered_inputs_dict = {
+            k: v[:batch_size, ...] if isinstance(v, torch.Tensor) else v
+            for k, v in inputs_dict.items()
+            if k not in input_keys_to_ignore
+        }
+        filtered_inputs_dict["pixel_values"] = inputs_dict["pixel_values"][:filtered_image_length]
+
+        # It is important set `eos_token_id` to `None` to avoid early stopping (would break for length-based checks)
+        text_gen_config = config.get_text_config(decoder=True)
+        if text_gen_config.eos_token_id is not None and text_gen_config.pad_token_id is None:
+            text_gen_config.pad_token_id = (
+                text_gen_config.eos_token_id
+                if isinstance(text_gen_config.eos_token_id, int)
+                else text_gen_config.eos_token_id[0]
+            )
+        text_gen_config.eos_token_id = None
+        text_gen_config.forced_eos_token_id = None
+
+        return config, filtered_inputs_dict
+
+    def test_inputs_embeds_matches_input_ids(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            model = model_class(config)
+            model.to(torch_device)
+            model.eval()
+
+            inputs = self._prepare_for_class(inputs_dict, model_class)
+            input_ids = inputs["input_ids"]
+            del inputs["input_ids"]
+            del inputs["pixel_values"]
+            del inputs["image_grid_thw"]
+
+            inputs_embeds = model.get_input_embeddings()(input_ids)
+
+            with torch.no_grad():
+                out_ids = model(input_ids=input_ids, **inputs)[0]
+                out_embeds = model(inputs_embeds=inputs_embeds, **inputs)[0]
+            torch.testing.assert_close(out_embeds, out_ids)
+
+    @unittest.skip(reason="Size mismatch")
+    def test_multi_gpu_data_parallel_forward(self):
+        pass
+
+    def _video_features_prepare_config_and_inputs(self):
+        """
+        Helper method to extract only video-related inputs from the full set of inputs, for testing `get_video_features`.
+
+        The superclass method simply calls the model_tester.prepare_config_and_inputs_for_common(),
+        but that method only prepared image inputs, i.e. where the temporal dimension in grid_thw is 1.
+        This override prepares proper video inputs with 12 frames.
+        """
+        config = self.model_tester.get_config()
+        patch_size = config.vision_config.patch_size
+        batch_size = self.model_tester.batch_size
+        image_size = self.model_tester.image_size
+        num_channels = self.model_tester.num_channels
+        num_frames = 12
+        pixel_values_videos = floats_tensor(
+            [num_frames * batch_size * (image_size**2) // (patch_size**2), num_channels * (patch_size**2)]
+        )
+
+        patches_per_side = image_size // patch_size
+        video_grid_thw = torch.tensor(
+            [[num_frames, patches_per_side, patches_per_side]] * batch_size, device=torch_device
+        )
+        inputs_dict = {
+            "pixel_values_videos": pixel_values_videos,
+            "video_grid_thw": video_grid_thw,
+        }
+        return config, inputs_dict
+
+
+@slow
+@require_torch_large_accelerator(memory=64)  # Tested on A100 / torch 2.9.0
+@require_torch
+class Ernie4_5_VLMoeIntegrationTest(unittest.TestCase):
+    model = None
+    model_id = "baidu/ERNIE-4.5-VL-28B-A3B-PT"
+
+    # TODO: remove revision when PR on the hub is merged
+    def setUp(self):
+        cleanup(torch_device, gc_collect=True)
+
+        self.processor = AutoProcessor.from_pretrained(self.model_id, revision="refs/pr/11")
+        self.message = [
+            {
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": "What kind of dog is this?"},
+                    {
+                        "type": "image",
+                        "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/pipeline-cat-chonk.jpeg",
+                    },
+                ],
+            }
+        ]
+        self.message2 = [
+            {
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": "What kind of dog is this?"},
+                    {
+                        "type": "image",
+                        "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/coco_sample.png",
+                    },
+                ],
+            }
+        ]
+
+    def tearDown(self):
+        cleanup(torch_device, gc_collect=True)
+
+    def load_model(self, dtype, attn_implementation="sdpa"):
+        return AutoModelForImageTextToText.from_pretrained(
+            self.model_id,
+            device_map="auto",
+            dtype=dtype,
+            attn_implementation=attn_implementation,
+            experts_implementation="eager",
+            revision="refs/pr/11",
+        )
+
+    def test_small_model_integration_test(self):
+        model = self.load_model("auto")
+        inputs = self.processor.apply_chat_template(
+            self.message, tokenize=True, add_generation_prompt=True, return_dict=True, return_tensors="pt"
+        )
+        expected_input_ids = [100273, 2969, 93963, 1912, 3836, 315, 9159, 357, 501, 94009, 39082, 93919, 4, 93963, 101304, 100295, 100295]  # fmt: skip
+        assert expected_input_ids == inputs.input_ids[0].tolist()[:17]
+
+        expected_pixel_slice = torch.tensor(
+            [
+                [-0.0988, -0.0842, -0.0842],
+                [-0.5660, -0.5514, -0.4200],
+                [-0.0259, -0.0259, -0.0259],
+                [-0.1280, -0.0988, -0.2010],
+                [-0.4638, -0.5806, -0.6974],
+                [-1.2083, -1.2229, -1.2083],
+            ],
+            dtype=torch.float32,
+            device="cpu",
+        )
+        assert torch.allclose(expected_pixel_slice, inputs.pixel_values[:6, :3], atol=3e-3)
+
+        # verify generation
+        inputs = inputs.to(torch_device)
+
+        # This model on the hub has `do_sample=True`.
+        torch.manual_seed(42)
+
+        output = model.generate(**inputs, max_new_tokens=30)
+        EXPECTED_DECODED_TEXT = "The animal in the image is a lynx, not a dog. It's a wild cat species known for its distinctive ear tufts and"
+        self.assertEqual(
+            self.processor.decode(output[0][len(inputs["input_ids"][0]) :], skip_special_tokens=True),
+            EXPECTED_DECODED_TEXT,
+        )
+
+    def test_small_model_integration_test_batch(self):
+        model = self.load_model("auto")
+        batch_messages = [self.message] * 2
+        inputs = self.processor.apply_chat_template(
+            batch_messages, tokenize=True, add_generation_prompt=True, return_dict=True, return_tensors="pt"
+        ).to(torch_device)
+
+        # This model on the hub has `do_sample=True`.
+        torch.manual_seed(42)
+
+        # it should not matter whether two images are the same size or not
+        output = model.generate(**inputs, max_new_tokens=30)
+
+        EXPECTED_DECODED_TEXT = [
+            "The animal in the image is a lynx, not a dog. It's a wild cat species known for its distinctive ear tufts and",
+            "The animal in the image is a lynx, not a dog. It's a wild cat species characterized by its distinctive ear tufts,"
+        ]  # fmt: skip
+
+        self.assertEqual(
+            [
+                self.processor.decode(output[0][len(inputs["input_ids"][0]) :], skip_special_tokens=True),
+                self.processor.decode(output[1][len(inputs["input_ids"][1]) :], skip_special_tokens=True),
+            ],
+            EXPECTED_DECODED_TEXT,
+        )
+
+    def test_small_model_integration_test_with_video(self):
+        processor = AutoProcessor.from_pretrained(
+            self.model_id, max_image_size={"longest_edge": 50176}, revision="refs/pr/11"
+        )
+        model = self.load_model(dtype=torch.float16)
+        questions = ["Only use English during your responses. Describe the following video."]
+        video_urls = ["https://huggingface.co/datasets/raushan-testing-hf/videos-test/resolve/main/tiny_video.mp4"]
+        messages = [
+            [
+                {
+                    "role": "user",
+                    "content": [
+                        {"type": "text", "text": question},
+                        {
+                            "type": "video",
+                            "video": video_url,
+                        },
+                    ],
+                }
+            ]
+            for question, video_url in zip(questions, video_urls)
+        ]
+        inputs = processor.apply_chat_template(
+            messages, tokenize=True, add_generation_prompt=True, return_dict=True, return_tensors="pt", padding=True
+        ).to(torch_device)
+
+        # This model on the hub has `do_sample=True`.
+        torch.manual_seed(42)
+
+        output = model.generate(**inputs, max_new_tokens=30)
+        EXPECTED_DECODED_TEXT = 'A black-and-white image shows a person lying on their back on a mat in a dojo. They are dressed in a white judo gi'  # fmt: skip
+
+        self.assertEqual(
+            self.processor.decode(output[0][len(inputs["input_ids"][0]) :], skip_special_tokens=True),
+            EXPECTED_DECODED_TEXT,
+        )
+
+    def test_small_model_integration_test_expand(self):
+        model = self.load_model("auto")
+        inputs = self.processor.apply_chat_template(
+            self.message, tokenize=True, add_generation_prompt=True, return_dict=True, return_tensors="pt"
+        ).to(torch_device)
+
+        # This model on the hub has `do_sample=True`.
+        torch.manual_seed(42)
+
+        output = model.generate(**inputs, max_new_tokens=30, do_sample=False, num_beams=2, num_return_sequences=2)
+
+        EXPECTED_DECODED_TEXT = [
+            'The animal in the image is a lynx, not a dog. It has the distinctive features of a lynx, such as tuft',
+            'The animal in the image is a lynx, not a dog. It has the distinctive features of a lynx, including a short tail'
+        ]  # fmt: skip
+
+        self.assertEqual(
+            [
+                self.processor.decode(output[0][len(inputs["input_ids"][0]) :], skip_special_tokens=True),
+                self.processor.decode(output[1][len(inputs["input_ids"][0]) :], skip_special_tokens=True),
+            ],
+            EXPECTED_DECODED_TEXT,
+        )
+
+    def test_small_model_integration_test_batch_wo_image(self):
+        model = self.load_model("auto")
+        message_wo_image = [
+            {"role": "user", "content": [{"type": "text", "text": "Who are you?"}]},
+        ]
+        batched_messages = [self.message, message_wo_image]
+        inputs = self.processor.apply_chat_template(
+            batched_messages,
+            tokenize=True,
+            add_generation_prompt=True,
+            return_dict=True,
+            return_tensors="pt",
+            padding=True,
+        ).to(torch_device)
+
+        # This model on the hub has `do_sample=True`.
+        torch.manual_seed(42)
+
+        # it should not matter whether two images are the same size or not
+        output = model.generate(**inputs, max_new_tokens=30)
+
+        EXPECTED_DECODED_TEXT = [
+            "The animal in the image is a lynx. It's a medium-sized wild cat characterized by its distinctive facial ruff, short tail",
+            "I am an AI assistant designed to help answer questions, provide information, and assist with tasks. I don't have personal experiences or a physical form"
+        ]  # fmt: skip
+
+        self.assertEqual(
+            [
+                self.processor.decode(output[0][len(inputs["input_ids"][0]) :], skip_special_tokens=True),
+                self.processor.decode(output[1][len(inputs["input_ids"][1]) :], skip_special_tokens=True),
+            ],
+            EXPECTED_DECODED_TEXT,
+        )
+
+    def test_small_model_integration_test_batch_different_resolutions(self):
+        model = self.load_model("auto")
+        batched_messages = [self.message, self.message2]
+        inputs = self.processor.apply_chat_template(
+            batched_messages,
+            tokenize=True,
+            add_generation_prompt=True,
+            return_dict=True,
+            return_tensors="pt",
+            padding=True,
+        ).to(torch_device)
+
+        # This model on the hub has `do_sample=True`.
+        torch.manual_seed(42)
+
+        # it should not matter whether two images are the same size or not
+        output = model.generate(**inputs, max_new_tokens=30)
+
+        EXPECTED_DECODED_TEXT = [
+            'The animal in the image is a lynx, not a dog. It has the distinctive features of a lynx, such as tuft',
+            'there are no dogs here, there are 2 cats',
+        ]  # fmt: skip
+
+        self.assertEqual(
+            [
+                self.processor.decode(output[0][len(inputs["input_ids"][0]) :], skip_special_tokens=True),
+                self.processor.decode(output[1][len(inputs["input_ids"][1]) :], skip_special_tokens=True),
+            ],
+            EXPECTED_DECODED_TEXT,
+        )
+
+
+# Garbage output expected as it is a dummy model to be run on the CI
+@slow
+@require_torch
+class Ernie4_5_VLMoeSmallIntegrationTest(unittest.TestCase):
+    model = None
+    model_id = "hf-internal-testing/Ernie-VL-Moe-Small"
+
+    def setUp(self):
+        cleanup(torch_device, gc_collect=True)
+
+        self.processor = AutoProcessor.from_pretrained(self.model_id)
+        self.message = [
+            {
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": "What kind of dog is this?"},
+                    {
+                        "type": "image",
+                        "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/pipeline-cat-chonk.jpeg",
+                    },
+                ],
+            }
+        ]
+        self.message2 = [
+            {
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": "What kind of dog is this?"},
+                    {
+                        "type": "image",
+                        "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/coco_sample.png",
+                    },
+                ],
+            }
+        ]
+
+    def tearDown(self):
+        cleanup(torch_device, gc_collect=True)
+
+    def load_model(self, dtype, attn_implementation="sdpa"):
+        return AutoModelForImageTextToText.from_pretrained(
+            self.model_id,
+            device_map="auto",
+            dtype=dtype,
+            attn_implementation=attn_implementation,
+            experts_implementation="eager",
+        )
+
+    def test_small_model_integration_test(self):
+        model = self.load_model("auto")
+        inputs = self.processor.apply_chat_template(
+            self.message, tokenize=True, add_generation_prompt=True, return_dict=True, return_tensors="pt"
+        )
+        expected_input_ids = [100273, 2969, 93963, 1912, 3836, 315, 9159, 357, 501, 94009, 39082, 93919, 4, 93963, 101304, 100295, 100295]  # fmt: skip
+        assert expected_input_ids == inputs.input_ids[0].tolist()[:17]
+
+        expected_pixel_slice = torch.tensor(
+            [
+                [-0.0988, -0.0842, -0.0842],
+                [-0.5660, -0.5514, -0.4200],
+                [-0.0259, -0.0259, -0.0259],
+                [-0.1280, -0.0988, -0.2010],
+                [-0.4638, -0.5806, -0.6974],
+                [-1.2083, -1.2229, -1.2083],
+            ],
+            dtype=torch.float32,
+            device="cpu",
+        )
+        assert torch.allclose(expected_pixel_slice, inputs.pixel_values[:6, :3], atol=3e-3)
+
+        # verify generation
+        inputs = inputs.to(torch_device)
+
+        # This model on the hub has `do_sample=True`.
+        torch.manual_seed(42)
+
+        output = model.generate(**inputs, max_new_tokens=30)
+        EXPECTED_DECODED_TEXT = "知道了知道了attaatta不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如"
+        self.assertEqual(
+            self.processor.decode(output[0][len(inputs["input_ids"][0]) :], skip_special_tokens=True),
+            EXPECTED_DECODED_TEXT,
+        )
+
+    def test_small_model_integration_test_batch(self):
+        model = self.load_model("auto")
+        batch_messages = [self.message] * 2
+        inputs = self.processor.apply_chat_template(
+            batch_messages, tokenize=True, add_generation_prompt=True, return_dict=True, return_tensors="pt"
+        ).to(torch_device)
+
+        # This model on the hub has `do_sample=True`.
+        torch.manual_seed(42)
+
+        # it should not matter whether two images are the same size or not
+        output = model.generate(**inputs, max_new_tokens=30)
+
+        # fmt: off
+        expectations = Expectations(
+            {
+                ("xpu", None): [
+                    '知道了知道了attaatta不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如',
+                    '填空填空填空填空填空填空填空填空填空填空填空填空填空填空填空填空填空填空填空填空填空填空填空填空填空填空填空填空填空填空',
+                ],
+                (None, None): [
+                    '知道了知道了attaatta不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如',
+                    '不是啊不是啊不是啊不是啊不是啊不是啊不是啊不是啊不是啊不是啊不是啊不是啊不是啊不是啊不是啊不是啊不是啊不是啊不是啊不是啊不是啊不是啊不是啊不是啊不是啊不是啊不是啊不是啊不是啊不是啊',
+                ],
+            }
+        )
+        EXPECTED_DECODED_TEXT = expectations.get_expectation()
+        # fmt: on
+
+        self.assertEqual(
+            [
+                self.processor.decode(output[0][len(inputs["input_ids"][0]) :], skip_special_tokens=True),
+                self.processor.decode(output[1][len(inputs["input_ids"][1]) :], skip_special_tokens=True),
+            ],
+            EXPECTED_DECODED_TEXT,
+        )
+
+    def test_small_model_integration_test_with_video(self):
+        processor = AutoProcessor.from_pretrained(self.model_id, max_image_size={"longest_edge": 50176})
+        model = self.load_model(dtype=torch.float16)
+        questions = ["Only use English during your responses. Describe the following video."]
+        video_urls = ["https://huggingface.co/datasets/raushan-testing-hf/videos-test/resolve/main/tiny_video.mp4"]
+        messages = [
+            [
+                {
+                    "role": "user",
+                    "content": [
+                        {"type": "text", "text": question},
+                        {
+                            "type": "video",
+                            "video": video_url,
+                        },
+                    ],
+                }
+            ]
+            for question, video_url in zip(questions, video_urls)
+        ]
+        inputs = processor.apply_chat_template(
+            messages, tokenize=True, add_generation_prompt=True, return_dict=True, return_tensors="pt", padding=True
+        ).to(torch_device)
+
+        # This model on the hub has `do_sample=True`.
+        torch.manual_seed(42)
+
+        output = model.generate(**inputs, max_new_tokens=30)
+        EXPECTED_DECODED_TEXT = 'uschuschusch载载载载载载载载载载载载载载载载载载载载载载载载载载载'  # fmt: skip
+
+        self.assertEqual(
+            self.processor.decode(output[0][len(inputs["input_ids"][0]) :], skip_special_tokens=True),
+            EXPECTED_DECODED_TEXT,
+        )
+
+    @require_deterministic_for_xpu
+    def test_small_model_integration_test_expand(self):
+        model = self.load_model("auto")
+        inputs = self.processor.apply_chat_template(
+            self.message, tokenize=True, add_generation_prompt=True, return_dict=True, return_tensors="pt"
+        ).to(torch_device)
+
+        # This model on the hub has `do_sample=True`.
+        torch.manual_seed(42)
+
+        output = model.generate(**inputs, max_new_tokens=30, do_sample=False, num_beams=2, num_return_sequences=2)
+
+        EXPECTED_DECODED_TEXT = [
+            '不是啊不是啊不是啊不是啊不是啊不是啊不是啊不是啊不是啊不是啊不是啊不是啊不是啊不是啊不是啊不是啊不是啊错的错的错的错的错的错的错的错的错的错的错的错的错的',
+            '不是啊不是啊不是啊不是啊不是啊不是啊不是啊不是啊不是啊不是啊不是啊不是啊不是啊不是啊不是啊不是啊不是啊错的错的错的错的错的错的错的错的错的错的错的错的就是这样',
+        ]  # fmt: skip
+
+        self.assertEqual(
+            [
+                self.processor.decode(output[0][len(inputs["input_ids"][0]) :], skip_special_tokens=True),
+                self.processor.decode(output[1][len(inputs["input_ids"][0]) :], skip_special_tokens=True),
+            ],
+            EXPECTED_DECODED_TEXT,
+        )
+
+    def test_small_model_integration_test_batch_wo_image(self):
+        model = self.load_model("auto")
+        message_wo_image = [
+            {"role": "user", "content": [{"type": "text", "text": "Who are you?"}]},
+        ]
+        batched_messages = [self.message, message_wo_image]
+        inputs = self.processor.apply_chat_template(
+            batched_messages,
+            tokenize=True,
+            add_generation_prompt=True,
+            return_dict=True,
+            return_tensors="pt",
+            padding=True,
+        ).to(torch_device)
+
+        # This model on the hub has `do_sample=True`.
+        torch.manual_seed(42)
+
+        # it should not matter whether two images are the same size or not
+        output = model.generate(**inputs, max_new_tokens=30)
+
+        EXPECTED_DECODED_TEXT = [
+            '知道了知道了attaatta不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如',
+            '用具柄柄柄柄柄柄柄柄柄柄柄柄柄柄柄柄柄柄柄柄柄柄柄柄柄柄柄柄柄',
+        ]  # fmt: skip
+
+        self.assertEqual(
+            [
+                self.processor.decode(output[0][len(inputs["input_ids"][0]) :], skip_special_tokens=True),
+                self.processor.decode(output[1][len(inputs["input_ids"][1]) :], skip_special_tokens=True),
+            ],
+            EXPECTED_DECODED_TEXT,
+        )
+
+    def test_small_model_integration_test_batch_different_resolutions(self):
+        model = self.load_model("auto")
+        batched_messages = [self.message, self.message2]
+        inputs = self.processor.apply_chat_template(
+            batched_messages,
+            tokenize=True,
+            add_generation_prompt=True,
+            return_dict=True,
+            return_tensors="pt",
+            padding=True,
+        ).to(torch_device)
+
+        # This model on the hub has `do_sample=True`.
+        torch.manual_seed(42)
+
+        # it should not matter whether two images are the same size or not
+        output = model.generate(**inputs, max_new_tokens=30)
+
+        EXPECTED_DECODED_TEXT = [
+            '知道了知道了attaatta不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如不如',
+            '填空填空填空填空填空填空填空填空填空填空填空填空填空填空填空填空填空填空填空填空填空填空填空填空填空填空填空填空填空填空',
+        ]  # fmt: skip
+
+        self.assertEqual(
+            [
+                self.processor.decode(output[0][len(inputs["input_ids"][0]) :], skip_special_tokens=True),
+                self.processor.decode(output[1][len(inputs["input_ids"][1]) :], skip_special_tokens=True),
+            ],
+            EXPECTED_DECODED_TEXT,
+        )

tests/models/ernie4_5_vl_moe/test_processing_ernie4_5_vl_moe.py

@@ -0,0 +1,378 @@
+# Copyright 2025 HuggingFace Inc team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+import shutil
+import tempfile
+import unittest
+
+import numpy as np
+import pytest
+
+from transformers import AutoProcessor, TokenizersBackend
+from transformers.testing_utils import require_av, require_torch, require_torchvision, require_vision
+from transformers.utils import is_torch_available, is_vision_available
+
+from ...test_processing_common import ProcessorTesterMixin
+
+
+if is_vision_available():
+    from transformers import Ernie4_5_VLMoeImageProcessor, Ernie4_5_VLMoeProcessor
+
+if is_torch_available():
+    import torch
+
+
+@require_vision
+@require_torch
+@require_torchvision
+class Ernie4_5_VLMoeProcessorTest(ProcessorTesterMixin, unittest.TestCase):
+    processor_class = Ernie4_5_VLMoeProcessor
+
+    @classmethod
+    def setUpClass(cls):
+        cls.tmpdirname = tempfile.mkdtemp()
+        processor = Ernie4_5_VLMoeProcessor.from_pretrained(
+            "hf-internal-testing/Ernie-VL-Moe-Small",
+            patch_size=4,
+            size={"shortest_edge": 28 * 28, "longest_edge": 56 * 56},
+        )
+        processor.save_pretrained(cls.tmpdirname)
+        cls.image_token = processor.image_token
+
+    def get_tokenizer(self, **kwargs):
+        return AutoProcessor.from_pretrained(self.tmpdirname, **kwargs).tokenizer
+
+    def get_image_processor(self, **kwargs):
+        return AutoProcessor.from_pretrained(self.tmpdirname, **kwargs).image_processor
+
+    def get_video_processor(self, **kwargs):
+        return AutoProcessor.from_pretrained(self.tmpdirname, **kwargs).video_processor
+
+    def get_processor(self, **kwargs):
+        return AutoProcessor.from_pretrained(self.tmpdirname, **kwargs)
+
+    @classmethod
+    def tearDownClass(cls):
+        shutil.rmtree(cls.tmpdirname, ignore_errors=True)
+
+    # Copied from tests.models.llava.test_processing_llava.LlavaProcessorTest.test_get_num_vision_tokens
+    def test_get_num_vision_tokens(self):
+        "Tests general functionality of the helper used internally in vLLM"
+
+        processor = self.get_processor()
+
+        output = processor._get_num_multimodal_tokens(image_sizes=[(100, 100), (300, 100), (500, 30)])
+        self.assertTrue("num_image_tokens" in output)
+        self.assertEqual(len(output["num_image_tokens"]), 3)
+
+        self.assertTrue("num_image_patches" in output)
+        self.assertEqual(len(output["num_image_patches"]), 3)
+
+    def test_save_load_pretrained_default(self):
+        tokenizer = self.get_tokenizer()
+        image_processor = self.get_image_processor()
+        video_processor = self.get_video_processor()
+
+        processor = Ernie4_5_VLMoeProcessor(
+            tokenizer=tokenizer, image_processor=image_processor, video_processor=video_processor
+        )
+        processor.save_pretrained(self.tmpdirname)
+        processor = Ernie4_5_VLMoeProcessor.from_pretrained(self.tmpdirname)
+
+        self.assertEqual(processor.tokenizer.get_vocab(), tokenizer.get_vocab())
+        self.assertEqual(processor.image_processor.to_json_string(), image_processor.to_json_string())
+        self.assertIsInstance(processor.tokenizer, TokenizersBackend)
+        self.assertIsInstance(processor.image_processor, Ernie4_5_VLMoeImageProcessor)
+
+    def test_image_processor(self):
+        image_processor = self.get_image_processor()
+        tokenizer = self.get_tokenizer()
+        video_processor = self.get_video_processor()
+
+        processor = Ernie4_5_VLMoeProcessor(
+            tokenizer=tokenizer, image_processor=image_processor, video_processor=video_processor
+        )
+
+        image_input = self.prepare_image_inputs()
+
+        input_image_proc = image_processor(image_input, return_tensors="pt")
+        input_processor = processor(images=image_input, text="dummy", return_tensors="pt")
+
+        for key in input_image_proc:
+            self.assertAlmostEqual(input_image_proc[key].sum(), input_processor[key].sum(), delta=1e-2)
+
+    def test_processor(self):
+        image_processor = self.get_image_processor()
+        tokenizer = self.get_tokenizer()
+        video_processor = self.get_video_processor()
+
+        processor = Ernie4_5_VLMoeProcessor(
+            tokenizer=tokenizer, image_processor=image_processor, video_processor=video_processor
+        )
+
+        input_str = "lower newer"
+        image_input = self.prepare_image_inputs()
+        inputs = processor(text=input_str, images=image_input)
+
+        self.assertListEqual(
+            list(inputs.keys()),
+            [
+                "input_ids",
+                "attention_mask",
+                "mm_token_type_ids",
+                "moe_mm_token_type_ids",
+                "pixel_values",
+                "image_grid_thw",
+            ],
+        )
+
+        # test if it raises when no input is passed
+        with pytest.raises(ValueError):
+            processor()
+
+        # test if it raises when no text is passed
+        with pytest.raises(TypeError):
+            processor(images=image_input)
+
+    @require_torch
+    @require_av
+    def _test_apply_chat_template(
+        self,
+        modality: str,
+        batch_size: int,
+        return_tensors: str,
+        input_name: str,
+        processor_name: str,
+        input_data: list[str],
+    ):
+        processor = self.get_processor()
+        if processor.chat_template is None:
+            self.skipTest("Processor has no chat template")
+
+        if processor_name not in self.processor_class.get_attributes():
+            self.skipTest(f"{processor_name} attribute not present in {self.processor_class}")
+
+        batch_messages = [
+            [
+                {
+                    "role": "user",
+                    "content": [{"type": "text", "text": "Describe this."}],
+                },
+            ]
+        ] * batch_size
+
+        # Test that jinja can be applied
+        formatted_prompt = processor.apply_chat_template(batch_messages, add_generation_prompt=True, tokenize=False)
+        self.assertEqual(len(formatted_prompt), batch_size)
+
+        # Test that tokenizing with template and directly with `self.tokenizer` gives same output
+        formatted_prompt_tokenized = processor.apply_chat_template(
+            batch_messages, add_generation_prompt=True, tokenize=True, return_tensors=return_tensors
+        )
+        add_special_tokens = True
+        if processor.tokenizer.bos_token is not None and formatted_prompt[0].startswith(processor.tokenizer.bos_token):
+            add_special_tokens = False
+        tok_output = processor.tokenizer(
+            formatted_prompt, return_tensors=return_tensors, add_special_tokens=add_special_tokens
+        )
+        expected_output = tok_output.input_ids
+        self.assertListEqual(expected_output.tolist(), formatted_prompt_tokenized.tolist())
+
+        # Test that kwargs passed to processor's `__call__` are actually used
+        tokenized_prompt_100 = processor.apply_chat_template(
+            batch_messages,
+            add_generation_prompt=True,
+            tokenize=True,
+            padding="max_length",
+            truncation=True,
+            return_tensors=return_tensors,
+            max_length=100,
+        )
+        self.assertEqual(len(tokenized_prompt_100[0]), 100)
+
+        # Test that `return_dict=True` returns text related inputs in the dict
+        out_dict_text = processor.apply_chat_template(
+            batch_messages,
+            add_generation_prompt=True,
+            tokenize=True,
+            return_dict=True,
+            return_tensors=return_tensors,
+        )
+        self.assertTrue(
+            all(
+                key in out_dict_text
+                for key in ["input_ids", "attention_mask", "mm_token_type_ids", "moe_mm_token_type_ids"]
+            )
+        )
+        self.assertEqual(len(out_dict_text["input_ids"]), batch_size)
+        self.assertEqual(len(out_dict_text["attention_mask"]), batch_size)
+        self.assertEqual(len(out_dict_text["mm_token_type_ids"]), batch_size)
+        self.assertEqual(len(out_dict_text["moe_mm_token_type_ids"]), batch_size)
+
+        # Test that with modality URLs and `return_dict=True`, we get modality inputs in the dict
+        for idx, url in enumerate(input_data[:batch_size]):
+            batch_messages[idx][0]["content"] = [batch_messages[idx][0]["content"][0], {"type": modality, "url": url}]
+
+        out_dict = processor.apply_chat_template(
+            batch_messages,
+            add_generation_prompt=True,
+            tokenize=True,
+            return_dict=True,
+            return_tensors=return_tensors,
+            max_frames=2,  # by default no more than 2 frames, otherwise too slow
+        )
+        input_name = getattr(self, input_name)
+        self.assertTrue(input_name in out_dict)
+        self.assertEqual(len(out_dict["input_ids"]), batch_size)
+        self.assertEqual(len(out_dict["attention_mask"]), batch_size)
+        self.assertEqual(len(out_dict["mm_token_type_ids"]), batch_size)
+        self.assertEqual(len(out_dict["moe_mm_token_type_ids"]), batch_size)
+
+        if modality == "video":
+            # qwen pixels don't scale with bs same way as other models, calculate expected video token count based on video_grid_thw
+            expected_video_token_count = 0
+            for thw in out_dict["video_grid_thw"]:
+                expected_video_token_count += thw[0] * thw[1] * thw[2]
+            mm_len = expected_video_token_count
+        else:
+            # Calculate expected image token count based on image_grid_thw
+            expected_image_token_count = 0
+            for thw in out_dict["image_grid_thw"]:
+                expected_image_token_count += thw[0] * thw[1] * thw[2]
+            mm_len = expected_image_token_count
+        self.assertEqual(len(out_dict[input_name]), mm_len)
+
+        return_tensor_to_type = {"pt": torch.Tensor, "np": np.ndarray, None: list}
+        for k in out_dict:
+            self.assertIsInstance(out_dict[k], return_tensor_to_type[return_tensors])
+
+    @require_av
+    def test_apply_chat_template_video_frame_sampling(self):
+        processor = self.get_processor()
+        if processor.chat_template is None:
+            self.skipTest("Processor has no chat template")
+
+        signature = inspect.signature(processor.__call__)
+        if "videos" not in {*signature.parameters.keys()} or (
+            signature.parameters.get("videos") is not None
+            and signature.parameters["videos"].annotation == inspect._empty
+        ):
+            self.skipTest("Processor doesn't accept videos at input")
+
+        messages = [
+            [
+                {
+                    "role": "user",
+                    "content": [
+                        {"type": "video"},
+                        {"type": "text", "text": "What is shown in this video?"},
+                    ],
+                },
+            ]
+        ]
+
+        formatted_prompt = processor.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
+        self.assertEqual(len(formatted_prompt), 1)
+
+        formatted_prompt_tokenized = processor.apply_chat_template(messages, add_generation_prompt=True, tokenize=True)
+        expected_output = processor.tokenizer(formatted_prompt, return_tensors=None).input_ids
+        self.assertListEqual(expected_output, formatted_prompt_tokenized)
+
+        out_dict = processor.apply_chat_template(messages, add_generation_prompt=True, tokenize=True, return_dict=True)
+        self.assertListEqual(
+            list(out_dict.keys()), ["input_ids", "attention_mask", "mm_token_type_ids", "moe_mm_token_type_ids"]
+        )
+
+        # Add video URL for return dict and load with `num_frames` arg
+        messages[0][0]["content"][0] = {
+            "type": "video",
+            "url": "https://huggingface.co/datasets/raushan-testing-hf/videos-test/resolve/main/tiny_video.mp4",
+        }
+        num_frames = 3
+        out_dict_with_video = processor.apply_chat_template(
+            messages,
+            add_generation_prompt=True,
+            tokenize=True,
+            return_dict=True,
+            num_frames=num_frames,
+            min_frames=3,  # default is 16
+        )
+        self.assertTrue(self.videos_input_name in out_dict_with_video)
+        self.assertEqual(len(out_dict_with_video[self.videos_input_name]), 720)
+
+        # Load with `fps` arg
+        fps = 1
+        out_dict_with_video = processor.apply_chat_template(
+            messages,
+            add_generation_prompt=True,
+            tokenize=True,
+            return_dict=True,
+            fps=fps,
+        )
+        self.assertTrue(self.videos_input_name in out_dict_with_video)
+        self.assertEqual(len(out_dict_with_video[self.videos_input_name]), 2160)
+
+        # Load with `fps` and `num_frames` args, should raise an error
+        with self.assertRaises(ValueError):
+            out_dict_with_video = processor.apply_chat_template(
+                messages,
+                add_generation_prompt=True,
+                tokenize=True,
+                return_dict=True,
+                fps=fps,
+                num_frames=num_frames,
+            )
+
+        # Load without any arg should load the whole video
+        out_dict_with_video = processor.apply_chat_template(
+            messages,
+            add_generation_prompt=True,
+            tokenize=True,
+            return_dict=True,
+        )
+        self.assertTrue(self.videos_input_name in out_dict_with_video)
+        self.assertEqual(len(out_dict_with_video[self.videos_input_name]), 2160)
+
+        # Load video as a list of frames (i.e. images). NOTE: each frame should have same size
+        # because we assume they come from one video
+        messages[0][0]["content"][0] = {
+            "type": "video",
+            "url": [
+                "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/australia.jpg",
+                "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/australia.jpg",
+            ],
+        }
+        out_dict_with_video = processor.apply_chat_template(
+            messages,
+            add_generation_prompt=True,
+            tokenize=True,
+            return_dict=True,
+            do_sample_frames=False,
+        )
+        self.assertTrue(self.videos_input_name in out_dict_with_video)
+        self.assertEqual(len(out_dict_with_video[self.videos_input_name]), 320)
+
+    def test_kwargs_overrides_custom_image_processor_kwargs(self):
+        processor = self.get_processor()
+        self.skip_processor_without_typed_kwargs(processor)
+
+        input_str = self.prepare_text_inputs()
+        image_input = self.prepare_image_inputs()
+
+        size = {"shortest_edge": processor.image_processor.size["shortest_edge"], "longest_edge": 56 * 56 * 4}
+        inputs = processor(text=input_str, images=image_input, size=size, return_tensors="pt")
+        self.assertEqual(inputs[self.images_input_name].shape[0], 612)
+        inputs = processor(text=input_str, images=image_input, return_tensors="pt")
+        self.assertEqual(inputs[self.images_input_name].shape[0], 100)

tests/models/ernie4_5_vl_moe/test_video_processing_ernie4_5_vl_moe.py

@@ -0,0 +1,335 @@
+# Copyright 2025 HuggingFace Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import unittest
+
+import numpy as np
+
+from transformers.image_utils import IMAGENET_STANDARD_MEAN, IMAGENET_STANDARD_STD
+from transformers.testing_utils import require_torch, require_vision
+from transformers.utils import is_torch_available, is_torchvision_available, is_vision_available
+
+from ...test_video_processing_common import VideoProcessingTestMixin, prepare_video_inputs
+
+
+if is_torch_available():
+    from PIL import Image
+
+if is_vision_available():
+    if is_torchvision_available():
+        from transformers import Ernie4_5_VLMoeVideoProcessor
+        from transformers.models.ernie4_5_vl_moe.video_processing_ernie4_5_vl_moe import smart_resize
+
+
+class Ernie4_5_VLMoeVideoProcessingTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=5,
+        num_frames=8,
+        num_channels=3,
+        min_resolution=30,
+        max_resolution=80,
+        temporal_patch_size=2,
+        patch_size=14,
+        merge_size=2,
+        do_resize=True,
+        size=None,
+        do_normalize=True,
+        image_mean=IMAGENET_STANDARD_MEAN,
+        image_std=IMAGENET_STANDARD_STD,
+        do_convert_rgb=True,
+        draw_on_frames=False,
+    ):
+        size = size if size is not None else {"longest_edge": 20, "shortest_edge": 10}
+        self.parent = parent
+        self.batch_size = batch_size
+        self.num_frames = num_frames
+        self.num_channels = num_channels
+        self.min_resolution = min_resolution
+        self.max_resolution = max_resolution
+        self.do_resize = do_resize
+        self.size = size
+        self.do_normalize = do_normalize
+        self.image_mean = image_mean
+        self.image_std = image_std
+        self.do_convert_rgb = do_convert_rgb
+        self.temporal_patch_size = temporal_patch_size
+        self.patch_size = patch_size
+        self.merge_size = merge_size
+        self.draw_on_frames = draw_on_frames
+
+    def prepare_video_processor_dict(self):
+        return {
+            "do_resize": self.do_resize,
+            "size": self.size,
+            "do_normalize": self.do_normalize,
+            "image_mean": self.image_mean,
+            "image_std": self.image_std,
+            "do_convert_rgb": self.do_convert_rgb,
+            "do_sample_frames": True,
+            "draw_on_frames": self.draw_on_frames,
+        }
+
+    def prepare_video_metadata(self, videos):
+        video_metadata = []
+        for video in videos:
+            if isinstance(video, list):
+                num_frames = len(video)
+            elif hasattr(video, "shape"):
+                if len(video.shape) == 4:  # (T, H, W, C)
+                    num_frames = video.shape[0]
+                else:
+                    num_frames = 1
+            else:
+                num_frames = self.num_frames
+
+            metadata = {
+                "fps": 2,
+                "duration": num_frames / 2,
+                "total_num_frames": num_frames,
+            }
+            video_metadata.append(metadata)
+        return video_metadata
+
+    def expected_output_video_shape(self, videos):
+        grid_t = self.num_frames
+        hidden_dim = self.num_channels * self.patch_size * self.patch_size
+        seq_len = 0
+        for video in videos:
+            if isinstance(video, list) and isinstance(video[0], Image.Image):
+                video = np.stack([np.array(frame) for frame in video])
+            elif hasattr(video, "shape"):
+                pass
+            else:
+                video = np.array(video)
+
+            if hasattr(video, "shape") and len(video.shape) >= 3:
+                if len(video.shape) == 4:
+                    _, height, width = video.shape[:3]
+                elif len(video.shape) == 3:
+                    height, width = video.shape[:2]
+                else:
+                    height, width = self.num_frames, self.min_resolution, self.min_resolution
+            else:
+                height, width = self.min_resolution, self.min_resolution
+
+            resized_height, resized_width = smart_resize(
+                height,
+                width,
+                factor=self.patch_size * self.merge_size,
+                min_pixels=self.size["shortest_edge"],
+                max_pixels=self.size["longest_edge"],
+            )
+            grid_h, grid_w = resized_height // self.patch_size, resized_width // self.patch_size
+            seq_len += grid_t * grid_h * grid_w
+        return [seq_len, hidden_dim]
+
+    def prepare_video_inputs(self, equal_resolution=False, return_tensors="pil"):
+        videos = prepare_video_inputs(
+            batch_size=self.batch_size,
+            num_frames=self.num_frames,
+            num_channels=self.num_channels,
+            min_resolution=self.min_resolution,
+            max_resolution=self.max_resolution,
+            equal_resolution=equal_resolution,
+            return_tensors=return_tensors,
+        )
+        return videos
+
+
+@require_torch
+@require_vision
+class Ernie4_5_VLMoeVideoProcessingTest(VideoProcessingTestMixin, unittest.TestCase):
+    fast_video_processing_class = Ernie4_5_VLMoeVideoProcessor if is_torchvision_available() else None
+    input_name = "pixel_values_videos"
+
+    def setUp(self):
+        super().setUp()
+        self.video_processor_tester = Ernie4_5_VLMoeVideoProcessingTester(self)
+
+    @property
+    def video_processor_dict(self):
+        return self.video_processor_tester.prepare_video_processor_dict()
+
+    def test_video_processor_from_dict_with_kwargs(self):
+        video_processor = self.fast_video_processing_class.from_dict(self.video_processor_dict)
+        self.assertEqual(video_processor.size, {"longest_edge": 20, "shortest_edge": 10})
+
+        video_processor = self.fast_video_processing_class.from_dict(
+            self.video_processor_dict, size={"longest_edge": 42, "shortest_edge": 42}
+        )
+        self.assertEqual(video_processor.size, {"longest_edge": 42, "shortest_edge": 42})
+
+    def test_call_pil(self):
+        for video_processing_class in self.video_processor_list:
+            video_processing = video_processing_class(**self.video_processor_dict)
+            video_inputs = self.video_processor_tester.prepare_video_inputs(
+                equal_resolution=False, return_tensors="pil"
+            )
+
+            for video in video_inputs:
+                self.assertIsInstance(video[0], Image.Image)
+
+            video_metadata = self.video_processor_tester.prepare_video_metadata(video_inputs)
+            encoded_videos = video_processing(
+                video_inputs[0], video_metadata=[video_metadata[0]], return_tensors="pt"
+            )[self.input_name]
+            expected_output_video_shape = self.video_processor_tester.expected_output_video_shape([video_inputs[0]])
+            self.assertEqual(list(encoded_videos.shape), expected_output_video_shape)
+            encoded_videos = video_processing(video_inputs, video_metadata=video_metadata, return_tensors="pt")[
+                self.input_name
+            ]
+            expected_output_video_shape = self.video_processor_tester.expected_output_video_shape(video_inputs)
+            self.assertEqual(list(encoded_videos.shape), expected_output_video_shape)
+
+    def test_call_numpy(self):
+        for video_processing_class in self.video_processor_list:
+            video_processing = video_processing_class(**self.video_processor_dict)
+            video_inputs = self.video_processor_tester.prepare_video_inputs(
+                equal_resolution=False, return_tensors="np"
+            )
+
+            video_metadata = self.video_processor_tester.prepare_video_metadata(video_inputs)
+            encoded_videos = video_processing(
+                video_inputs[0], video_metadata=[video_metadata[0]], return_tensors="pt"
+            )[self.input_name]
+            expected_output_video_shape = self.video_processor_tester.expected_output_video_shape([video_inputs[0]])
+            self.assertEqual(list(encoded_videos.shape), expected_output_video_shape)
+
+            encoded_videos = video_processing(video_inputs, video_metadata=video_metadata, return_tensors="pt")[
+                self.input_name
+            ]
+            expected_output_video_shape = self.video_processor_tester.expected_output_video_shape(video_inputs)
+            self.assertEqual(list(encoded_videos.shape), expected_output_video_shape)
+
+    def test_call_pytorch(self):
+        for video_processing_class in self.video_processor_list:
+            video_processing = video_processing_class(**self.video_processor_dict)
+            video_inputs = self.video_processor_tester.prepare_video_inputs(
+                equal_resolution=False, return_tensors="pt"
+            )
+            video_metadata = self.video_processor_tester.prepare_video_metadata(video_inputs)
+            encoded_videos = video_processing(
+                video_inputs[0], video_metadata=[video_metadata[0]], return_tensors="pt"
+            )[self.input_name]
+            expected_output_video_shape = self.video_processor_tester.expected_output_video_shape([video_inputs[0]])
+            self.assertEqual(list(encoded_videos.shape), expected_output_video_shape)
+            encoded_videos = video_processing(video_inputs, video_metadata=video_metadata, return_tensors="pt")[
+                self.input_name
+            ]
+            expected_output_video_shape = self.video_processor_tester.expected_output_video_shape(video_inputs)
+            self.assertEqual(list(encoded_videos.shape), expected_output_video_shape)
+
+    @unittest.skip("Skip for now, the test needs adjustment for Ernie 4.5 VL")
+    def test_call_numpy_4_channels(self):
+        for video_processing_class in self.video_processor_list:
+            # Test that can process videos which have an arbitrary number of channels
+            # Initialize video_processing
+            video_processor = video_processing_class(**self.video_processor_dict)
+
+            # create random numpy tensors
+            self.video_processor_tester.num_channels = 4
+            video_inputs = self.video_processor_tester.prepare_video_inputs(
+                equal_resolution=False, return_tensors="np"
+            )
+
+            # Test not batched input
+            encoded_videos = video_processor(
+                video_inputs[0],
+                return_tensors="pt",
+                input_data_format="channels_last",
+                image_mean=(0.0, 0.0, 0.0, 0.0),
+                image_std=(1.0, 1.0, 1.0, 1.0),
+            )[self.input_name]
+            expected_output_video_shape = self.video_processor_tester.expected_output_video_shape([video_inputs[0]])
+            self.assertEqual(list(encoded_videos.shape), expected_output_video_shape)
+
+            # Test batched
+            encoded_videos = video_processor(
+                video_inputs,
+                return_tensors="pt",
+                input_data_format="channels_last",
+                image_mean=(0.0, 0.0, 0.0, 0.0),
+                image_std=(1.0, 1.0, 1.0, 1.0),
+            )[self.input_name]
+            expected_output_video_shape = self.video_processor_tester.expected_output_video_shape(video_inputs)
+            self.assertEqual(list(encoded_videos.shape), expected_output_video_shape)
+
+    def test_nested_input(self):
+        """Tests that the processor can work with nested list where each video is a list of arrays"""
+        for video_processing_class in self.video_processor_list:
+            video_processing = video_processing_class(**self.video_processor_dict)
+            video_inputs = self.video_processor_tester.prepare_video_inputs(
+                equal_resolution=False, return_tensors="np"
+            )
+
+            video_inputs_nested = [list(video) for video in video_inputs]
+            video_metadata = self.video_processor_tester.prepare_video_metadata(video_inputs)
+
+            # Test not batched input
+            encoded_videos = video_processing(
+                video_inputs_nested[0], video_metadata=[video_metadata[0]], return_tensors="pt"
+            )[self.input_name]
+            expected_output_video_shape = self.video_processor_tester.expected_output_video_shape([video_inputs[0]])
+            self.assertEqual(list(encoded_videos.shape), expected_output_video_shape)
+
+            # Test batched
+            encoded_videos = video_processing(video_inputs_nested, video_metadata=video_metadata, return_tensors="pt")[
+                self.input_name
+            ]
+            expected_output_video_shape = self.video_processor_tester.expected_output_video_shape(video_inputs)
+            self.assertEqual(list(encoded_videos.shape), expected_output_video_shape)
+
+    def test_call_sample_frames(self):
+        for video_processing_class in self.video_processor_list:
+            video_processor_dict = self.video_processor_dict.copy()
+            video_processing = video_processing_class(**video_processor_dict)
+
+            prev_num_frames = self.video_processor_tester.num_frames
+            self.video_processor_tester.num_frames = 8
+            prev_min_resolution = getattr(self.video_processor_tester, "min_resolution", None)
+            prev_max_resolution = getattr(self.video_processor_tester, "max_resolution", None)
+            self.video_processor_tester.min_resolution = 56
+            self.video_processor_tester.max_resolution = 112
+
+            video_inputs = self.video_processor_tester.prepare_video_inputs(
+                equal_resolution=False,
+                return_tensors="torch",
+            )
+
+            metadata = [[{"total_num_frames": 8, "fps": 4}]]
+            batched_metadata = metadata * len(video_inputs)
+
+            encoded_videos = video_processing(video_inputs[0], return_tensors="pt", video_metadata=metadata)[
+                self.input_name
+            ]
+            encoded_videos_batched = video_processing(
+                video_inputs, return_tensors="pt", video_metadata=batched_metadata
+            )[self.input_name]
+
+            self.assertIsNotNone(encoded_videos)
+            self.assertIsNotNone(encoded_videos_batched)
+            self.assertEqual(len(encoded_videos.shape), 2)
+            self.assertEqual(len(encoded_videos_batched.shape), 2)
+
+            # error out when sampled frames would go over total number of frames
+            with self.assertRaises(ValueError):
+                video_processing(video_inputs[0], num_frames=10, return_tensors="pt")[self.input_name]
+
+            self.video_processor_tester.num_frames = prev_num_frames
+            if prev_min_resolution is not None:
+                self.video_processor_tester.min_resolution = prev_min_resolution
+            if prev_max_resolution is not None:
+                self.video_processor_tester.max_resolution = prev_max_resolution