init

transformers/tests/models/glm4v/test_image_processing_glm4v.py

@@ -0,0 +1,254 @@
+# Copyright 2021 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.testing_utils import require_torch, require_vision
+from transformers.utils import is_torch_available, is_torchvision_available, is_vision_available
+
+from ...test_image_processing_common import ImageProcessingTestMixin, prepare_image_inputs
+
+
+if is_torch_available():
+    import torch
+
+
+if is_vision_available():
+    from PIL import Image
+
+    from transformers import Glm4vImageProcessor
+    from transformers.models.glm4v.image_processing_glm4v import smart_resize
+
+    if is_torchvision_available():
+        from transformers import Glm4vImageProcessorFast
+
+
+class Glm4vImageProcessingTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=7,
+        num_channels=3,
+        min_resolution=30,
+        max_resolution=80,
+        do_resize=True,
+        size=None,
+        do_normalize=True,
+        image_mean=[0.5, 0.5, 0.5],
+        image_std=[0.5, 0.5, 0.5],
+        temporal_patch_size=2,
+        patch_size=14,
+        merge_size=2,
+    ):
+        size = size if size is not None else {"longest_edge": 20, "shortest_edge": 10}
+        self.parent = parent
+        self.batch_size = batch_size
+        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.temporal_patch_size = temporal_patch_size
+        self.patch_size = patch_size
+        self.merge_size = merge_size
+
+    def prepare_image_processor_dict(self):
+        return {
+            "image_mean": self.image_mean,
+            "image_std": self.image_std,
+            "do_normalize": self.do_normalize,
+            "do_resize": self.do_resize,
+            "size": self.size,
+            "temporal_patch_size": self.temporal_patch_size,
+            "patch_size": self.patch_size,
+            "merge_size": self.merge_size,
+        }
+
+    def expected_output_image_shape(self, images):
+        grid_t = 1
+        hidden_dim = self.num_channels * self.temporal_patch_size * self.patch_size * self.patch_size
+        seq_len = 0
+        for image in images:
+            if isinstance(image, list) and isinstance(image[0], Image.Image):
+                image = np.stack([np.array(frame) for frame in image])
+            elif hasattr(image, "shape"):
+                pass
+            else:
+                image = np.array(image)
+            if hasattr(image, "shape") and len(image.shape) >= 3:
+                if isinstance(image, np.ndarray):
+                    if len(image.shape) == 4:
+                        height, width = image.shape[1:3]
+                    elif len(image.shape) == 3:
+                        height, width = image.shape[:2]
+                    else:
+                        height, width = self.min_resolution, self.min_resolution
+                else:
+                    height, width = image.shape[-2:]
+            else:
+                height, width = self.min_resolution, self.min_resolution
+
+            resized_height, resized_width = smart_resize(
+                self.temporal_patch_size,
+                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_image_inputs(self, equal_resolution=False, numpify=False, torchify=False):
+        return 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,
+        )
+
+
+@require_torch
+@require_vision
+class ViTImageProcessingTest(ImageProcessingTestMixin, unittest.TestCase):
+    image_processing_class = Glm4vImageProcessor if is_vision_available() else None
+    fast_image_processing_class = Glm4vImageProcessorFast if is_torchvision_available() else None
+
+    def setUp(self):
+        super().setUp()
+        self.image_processor_tester = Glm4vImageProcessingTester(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_processor_list:
+            image_processing = image_processing_class(**self.image_processor_dict)
+            self.assertTrue(hasattr(image_processing, "image_mean"))
+            self.assertTrue(hasattr(image_processing, "image_std"))
+            self.assertTrue(hasattr(image_processing, "do_normalize"))
+            self.assertTrue(hasattr(image_processing, "do_resize"))
+            self.assertTrue(hasattr(image_processing, "size"))
+
+    def test_image_processor_from_dict_with_kwargs(self):
+        for image_processing_class in self.image_processor_list:
+            image_processor = image_processing_class.from_dict(self.image_processor_dict)
+            self.assertEqual(image_processor.size, {"shortest_edge": 10, "longest_edge": 20})
+
+            image_processor = image_processing_class.from_dict(
+                self.image_processor_dict, size={"shortest_edge": 42, "longest_edge": 42}
+            )
+            self.assertEqual(image_processor.size, {"shortest_edge": 42, "longest_edge": 42})
+
+    # batch size is flattened
+    def test_call_pil(self):
+        for image_processing_class in self.image_processor_list:
+            # 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=False)
+            for image in image_inputs:
+                self.assertIsInstance(image, Image.Image)
+
+            # Test not batched input
+            encoded_images = image_processing(image_inputs[0], return_tensors="pt").pixel_values
+            expected_output_image_shape = self.image_processor_tester.expected_output_image_shape([image_inputs[0]])
+            self.assertEqual(tuple(encoded_images.shape), expected_output_image_shape)
+
+            # Test batched
+            encoded_images = image_processing(image_inputs, return_tensors="pt").pixel_values
+            expected_output_image_shape = self.image_processor_tester.expected_output_image_shape(image_inputs)
+            self.assertEqual(tuple(encoded_images.shape), expected_output_image_shape)
+
+    def test_call_numpy(self):
+        for image_processing_class in self.image_processor_list:
+            # 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=False, numpify=True)
+            for image in image_inputs:
+                self.assertIsInstance(image, np.ndarray)
+
+            # Test not batched input
+            encoded_images = image_processing(image_inputs[0], return_tensors="pt").pixel_values
+            expected_output_image_shape = self.image_processor_tester.expected_output_image_shape([image_inputs[0]])
+            self.assertEqual(tuple(encoded_images.shape), expected_output_image_shape)
+
+            # Test batched
+            encoded_images = image_processing(image_inputs, return_tensors="pt").pixel_values
+            expected_output_image_shape = self.image_processor_tester.expected_output_image_shape(image_inputs)
+            self.assertEqual(tuple(encoded_images.shape), expected_output_image_shape)
+
+    def test_call_pytorch(self):
+        for image_processing_class in self.image_processor_list:
+            # 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=False, torchify=True)
+
+            for image in image_inputs:
+                self.assertIsInstance(image, torch.Tensor)
+
+            # Test not batched input
+            encoded_images = image_processing(image_inputs[0], return_tensors="pt").pixel_values
+            expected_output_image_shape = self.image_processor_tester.expected_output_image_shape([image_inputs[0]])
+            self.assertEqual(tuple(encoded_images.shape), expected_output_image_shape)
+
+            # Test batched
+            expected_output_image_shape = self.image_processor_tester.expected_output_image_shape(image_inputs)
+            encoded_images = image_processing(image_inputs, return_tensors="pt").pixel_values
+            self.assertEqual(tuple(encoded_images.shape), expected_output_image_shape)
+
+    def test_call_numpy_4_channels(self):
+        for image_processing_class in self.image_processor_list:
+            # Test that can process images which have an arbitrary number of channels
+            # Initialize image_processing
+            image_processor = image_processing_class(**self.image_processor_dict)
+
+            # create random numpy tensors
+            self.image_processor_tester.num_channels = 4
+            image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=False, numpify=True)
+
+            # Test not batched input
+            encoded_images = image_processor(
+                image_inputs[0],
+                return_tensors="pt",
+                input_data_format="channels_last",
+                image_mean=0,
+                image_std=1,
+            ).pixel_values
+            expected_output_image_shape = self.image_processor_tester.expected_output_image_shape([image_inputs[0]])
+            self.assertEqual(tuple(encoded_images.shape), expected_output_image_shape)
+
+            # Test batched
+            encoded_images = image_processor(
+                image_inputs,
+                return_tensors="pt",
+                input_data_format="channels_last",
+                image_mean=0,
+                image_std=1,
+            ).pixel_values
+            expected_output_image_shape = self.image_processor_tester.expected_output_image_shape(image_inputs)
+            self.assertEqual(tuple(encoded_images.shape), expected_output_image_shape)

transformers/tests/models/glm4v/test_modeling_glm4v.py

@@ -0,0 +1,578 @@
+# Copyright 2025 The 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 GLM-4.1V model."""
+
+import copy
+import unittest
+
+from transformers import (
+    AutoProcessor,
+    Glm4vConfig,
+    Glm4vForConditionalGeneration,
+    Glm4vModel,
+    is_torch_available,
+)
+from transformers.testing_utils import (
+    cleanup,
+    require_flash_attn,
+    require_torch,
+    require_torch_gpu,
+    slow,
+    torch_device,
+)
+
+from ...generation.test_utils import GenerationTesterMixin
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_common import (
+    ModelTesterMixin,
+    floats_tensor,
+    ids_tensor,
+)
+
+
+if is_torch_available():
+    import torch
+
+
+class Glm4vVisionText2TextModelTester:
+    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={
+            "vocab_size": 99,
+            "hidden_size": 16,
+            "intermediate_size": 22,
+            "num_hidden_layers": 2,
+            "num_attention_heads": 2,
+            "num_key_value_heads": 1,
+            "output_channels": 64,
+            "hidden_act": "silu",
+            "max_position_embeddings": 512,
+            "rope_scaling": {"type": "default", "mrope_section": [2, 1, 1]},
+            "rope_theta": 10000,
+            "tie_word_embeddings": True,
+            "bos_token_id": 0,
+            "eos_token_id": 0,
+            "pad_token_id": 0,
+        },
+        vision_config={
+            "depth": 2,
+            "hidden_act": "silu",
+            "hidden_size": 48,
+            "out_hidden_size": 16,
+            "intermediate_size": 22,
+            "patch_size": 14,
+            "spatial_merge_size": 1,
+            "temporal_patch_size": 2,
+        },
+    ):
+        self.parent = parent
+        self.ignore_index = ignore_index
+        self.bos_token_id = text_config["bos_token_id"]
+        self.eos_token_id = text_config["eos_token_id"]
+        self.pad_token_id = text_config["pad_token_id"]
+        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.text_config = text_config
+        self.vision_config = vision_config
+        self.batch_size = batch_size
+        self.num_channels = num_channels
+        self.image_size = image_size
+        self.is_training = is_training
+        self.hidden_size = text_config["hidden_size"]
+        self.num_hidden_layers = text_config["num_hidden_layers"]
+        self.num_attention_heads = text_config["num_attention_heads"]
+        self.vocab_size = text_config["vocab_size"]
+        self.num_image_tokens = 64
+        self.seq_length = seq_length + self.num_image_tokens
+
+    def get_config(self):
+        return Glm4vConfig(
+            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
+        temporal_patch_size = config.vision_config.temporal_patch_size
+        pixel_values = floats_tensor(
+            [
+                self.batch_size * (self.image_size**2) // (patch_size**2),
+                self.num_channels * (patch_size**2) * temporal_patch_size,
+            ]
+        )
+
+        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
+
+        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,
+        }
+        return config, inputs_dict
+
+
+@require_torch
+class Glm4vModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):
+    all_model_classes = (Glm4vModel, Glm4vForConditionalGeneration) if is_torch_available() else ()
+    test_pruning = False
+    test_head_masking = False
+    test_torchscript = False
+    model_split_percents = [0.7, 0.9]  # model too big to split at 0.5
+    _is_composite = True
+
+    def setUp(self):
+        self.model_tester = Glm4vVisionText2TextModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=Glm4vConfig, has_text_modality=False)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    # GLM4V has images shaped as (bs*patch_len, dim) so we can't slice to batches in generate
+    def prepare_config_and_inputs_for_generate(self, batch_size=2):
+        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 to mask attention heads
+            "head_mask",
+            "decoder_head_mask",
+            "cross_attn_head_mask",
+            # 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",
+            # model-specific exceptions should overload/overwrite this function
+        ]
+
+        # 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
+
+    @unittest.skip(reason="No available kernels - not supported")
+    def test_sdpa_can_dispatch_on_flash(self):
+        pass
+
+    @unittest.skip(reason="Size mismatch")
+    def test_multi_gpu_data_parallel_forward(self):
+        pass
+
+    @unittest.skip("Error with compilation")
+    def test_generate_from_inputs_embeds_with_static_cache(self):
+        pass
+
+    def test_inputs_embeds(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 = copy.deepcopy(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"]
+
+            wte = model.get_input_embeddings()
+            inputs["inputs_embeds"] = wte(input_ids)
+            with torch.no_grad():
+                model(**inputs)[0]
+
+    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)
+
+
+@require_torch
+class Glm4vIntegrationTest(unittest.TestCase):
+    def setUp(self):
+        cleanup(torch_device, gc_collect=True)
+
+        self.processor = AutoProcessor.from_pretrained("THUDM/GLM-4.1V-9B-Thinking")
+        self.message = [
+            {
+                "role": "user",
+                "content": [
+                    {
+                        "type": "image",
+                        "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/pipeline-cat-chonk.jpeg",
+                    },
+                    {"type": "text", "text": "What kind of dog is this?"},
+                ],
+            }
+        ]
+        self.message2 = [
+            {
+                "role": "user",
+                "content": [
+                    {
+                        "type": "image",
+                        "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/coco_sample.png",
+                    },
+                    {"type": "text", "text": "What kind of dog is this?"},
+                ],
+            }
+        ]
+
+    def tearDown(self):
+        cleanup(torch_device, gc_collect=True)
+
+    @slow
+    def test_small_model_integration_test(self):
+        model = Glm4vForConditionalGeneration.from_pretrained(
+            "THUDM/GLM-4.1V-9B-Thinking", dtype="auto", device_map="auto"
+        )
+
+        inputs = self.processor.apply_chat_template(
+            self.message, tokenize=True, add_generation_prompt=True, return_dict=True, return_tensors="pt"
+        )
+        expected_input_ids = [151331, 151333, 151336, 198, 151339, 151343, 151343, 151343, 151343, 151343, 151343, 151343, 151343, 151343, 151343, 151343, 151343]  # 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 = "\nWhat kind of dog is this?\n<think>Got it, let's look at the image. The animal in the picture doesn't look like a dog; it's actually a cat. Specifically"
+        self.assertEqual(
+            self.processor.decode(output[0], skip_special_tokens=True),
+            EXPECTED_DECODED_TEXT,
+        )
+
+    @slow
+    def test_small_model_integration_test_batch(self):
+        model = Glm4vForConditionalGeneration.from_pretrained(
+            "THUDM/GLM-4.1V-9B-Thinking", dtype="auto", device_map="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 = [
+            "\nWhat kind of dog is this?\n<think>Got it, let's look at the image. The animal in the picture doesn't look like a dog; it's actually a cat. Specifically",
+            "\nWhat kind of dog is this?\n<think>Got it, let's look at the image. The animal in the picture has a stocky body, thick fur, and a face that's"
+        ]  # fmt: skip
+        self.assertEqual(
+            self.processor.batch_decode(output, skip_special_tokens=True),
+            EXPECTED_DECODED_TEXT,
+        )
+
+    @slow
+    def test_small_model_integration_test_with_video(self):
+        processor = AutoProcessor.from_pretrained("THUDM/GLM-4.1V-9B-Thinking", max_image_size={"longest_edge": 50176})
+        model = Glm4vForConditionalGeneration.from_pretrained(
+            "THUDM/GLM-4.1V-9B-Thinking", dtype=torch.float16, device_map="auto"
+        )
+        questions = ["Describe this video."]
+        video_urls = ["https://huggingface.co/datasets/hf-internal-testing/fixtures_videos/resolve/main/tennis.mp4"]
+        messages = [
+            [
+                {
+                    "role": "user",
+                    "content": [
+                        {
+                            "type": "video",
+                            "video": video_url,
+                        },
+                        {"type": "text", "text": question},
+                    ],
+                }
+            ]
+            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 = ["\n012345Describe this video.\n<think>Got it, let's analyze the video. First, the scene is an indoor tennis court. There are two players: one in a white shirt"]  # fmt: skip
+
+        self.assertEqual(
+            processor.batch_decode(output, skip_special_tokens=True),
+            EXPECTED_DECODED_TEXT,
+        )
+
+    @slow
+    def test_small_model_integration_test_expand(self):
+        model = Glm4vForConditionalGeneration.from_pretrained(
+            "THUDM/GLM-4.1V-9B-Thinking", dtype="auto", device_map="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 = [
+            "\nWhat kind of dog is this?\n<think>Got it, let's look at the image. The animal in the picture doesn't look like a dog; it's actually a cat. Specifically",
+            "\nWhat kind of dog is this?\n<think>Got it, let's look at the image. The animal in the picture doesn't look like a dog; it's actually a cat, specifically"
+        ]  # fmt: skip
+        self.assertEqual(
+            self.processor.batch_decode(output, skip_special_tokens=True),
+            EXPECTED_DECODED_TEXT,
+        )
+
+    @slow
+    def test_small_model_integration_test_batch_wo_image(self):
+        model = Glm4vForConditionalGeneration.from_pretrained(
+            "THUDM/GLM-4.1V-9B-Thinking", dtype="auto", device_map="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 = [
+            "\nWhat kind of dog is this?\n<think>Got it, let's look at the image. The animal in the picture doesn't look like a dog; it's actually a cat. Specifically",
+            "\nWho are you?\n<think>Got it, let's look at the user's question: \"Who are you?\" This is a common question when someone is just starting a conversation"
+        ]  # fmt: skip
+        self.assertEqual(
+            self.processor.batch_decode(output, skip_special_tokens=True),
+            EXPECTED_DECODED_TEXT,
+        )
+
+    @slow
+    def test_small_model_integration_test_batch_different_resolutions(self):
+        model = Glm4vForConditionalGeneration.from_pretrained(
+            "THUDM/GLM-4.1V-9B-Thinking", dtype="auto", device_map="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 = [
+            "\nWhat kind of dog is this?\n<think>Got it, let's look at the image. The animal in the picture doesn't look like a dog; it's actually a cat. Specifically",
+            "\nWhat kind of dog is this?\n<think>Got it, let's look at the image. Wait, the animals here are cats, not dogs. The question is about a dog, but",
+        ]  # fmt: skip
+        self.assertEqual(
+            self.processor.batch_decode(output, skip_special_tokens=True),
+            EXPECTED_DECODED_TEXT,
+        )
+
+    @slow
+    @require_flash_attn
+    @require_torch_gpu
+    def test_small_model_integration_test_batch_flashatt2(self):
+        model = Glm4vForConditionalGeneration.from_pretrained(
+            "THUDM/GLM-4.1V-9B-Thinking",
+            dtype=torch.bfloat16,
+            attn_implementation="flash_attention_2",
+            device_map="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 = [
+            "\nWhat kind of dog is this?\n<think>Got it, let's look at the image. The animal in the picture doesn't look like a dog. Wait, it's a cat,",
+            "\nWhat kind of dog is this?\n<think>Got it, let's look at the image. Wait, the animals here are cats, not dogs. The question is about a dog, but"
+        ]  # fmt: skip
+        self.assertEqual(
+            self.processor.batch_decode(output, skip_special_tokens=True),
+            EXPECTED_DECODED_TEXT,
+        )
+
+    @slow
+    @require_flash_attn
+    @require_torch_gpu
+    def test_small_model_integration_test_batch_wo_image_flashatt2(self):
+        model = Glm4vForConditionalGeneration.from_pretrained(
+            "THUDM/GLM-4.1V-9B-Thinking",
+            dtype=torch.bfloat16,
+            attn_implementation="flash_attention_2",
+            device_map="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 = [
+            "\nWhat kind of dog is this?\n<think>Got it, let's look at the image. The animal in the picture doesn't look like a dog; it's actually a cat. Specifically",
+            "\nWho are you?\n<think>Got it, let's look at the user's question: \"Who are you?\" This is a common question when someone is just starting a conversation"
+        ]  # fmt: skip
+
+        self.assertEqual(
+            self.processor.batch_decode(output, skip_special_tokens=True),
+            EXPECTED_DECODED_TEXT,
+        )

transformers/tests/models/glm4v/test_processor_glm4v.py

@@ -0,0 +1,279 @@
+# Copyright 2025 The HuggingFace 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
+
+from transformers import AutoProcessor
+from transformers.testing_utils import require_av, require_torch, require_vision
+from transformers.utils import is_torch_available, is_vision_available
+
+from ...test_processing_common import ProcessorTesterMixin, url_to_local_path
+
+
+if is_vision_available():
+    from transformers import Glm4vProcessor
+
+if is_torch_available():
+    import torch
+
+
+@require_vision
+@require_torch
+class Glm4vProcessorTest(ProcessorTesterMixin, unittest.TestCase):
+    processor_class = Glm4vProcessor
+
+    @classmethod
+    def setUpClass(cls):
+        cls.tmpdirname = tempfile.mkdtemp()
+        processor = Glm4vProcessor.from_pretrained(
+            "THUDM/GLM-4.1V-9B-Thinking", patch_size=4, size={"shortest_edge": 12 * 12, "longest_edge": 18 * 18}
+        )
+        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)
+
+    @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.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"]))
+        self.assertEqual(len(out_dict_text["input_ids"]), batch_size)
+        self.assertEqual(len(out_dict_text["attention_mask"]), 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,
+            fps=2
+            if isinstance(input_data[0], str)
+            else None,  # by default no more than 2 frames per second, otherwise too slow
+            do_sample_frames=bool(isinstance(input_data[0], str)),  # don't sample frames if decoded video is used
+        )
+        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)
+
+        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:
+            mm_len = batch_size * 4
+        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"])
+
+        # Add video URL for return dict and load with `num_frames` arg
+        messages[0][0]["content"][0] = {
+            "type": "video",
+            "url": url_to_local_path(
+                "https://huggingface.co/datasets/raushan-testing-hf/videos-test/resolve/main/tiny_video.mp4"
+            ),
+        }
+
+        # Load with `video_fps` arg
+        video_fps = 10
+        out_dict_with_video = processor.apply_chat_template(
+            messages,
+            add_generation_prompt=True,
+            tokenize=True,
+            return_dict=True,
+            video_fps=video_fps,
+        )
+        self.assertTrue(self.videos_input_name in out_dict_with_video)
+        self.assertEqual(len(out_dict_with_video[self.videos_input_name]), 8)
+
+        # Load the whole video
+        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]), 24)
+
+        # 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": [
+                url_to_local_path(
+                    "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/australia.jpg"
+                ),
+                url_to_local_path(
+                    "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]), 4)
+
+        # When the inputs are frame URLs/paths we expect that those are already
+        # sampled and will raise an error is asked to sample again.
+        with self.assertRaisesRegex(
+            ValueError, "Sampling frames from a list of images is not supported! Set `do_sample_frames=False`"
+        ):
+            out_dict_with_video = processor.apply_chat_template(
+                messages,
+                add_generation_prompt=True,
+                tokenize=True,
+                return_dict=True,
+                do_sample_frames=True,
+            )
+
+    def test_model_input_names(self):
+        processor = self.get_processor()
+
+        text = self.prepare_text_inputs(modalities=["image", "video"])
+        image_input = self.prepare_image_inputs()
+        video_inputs = self.prepare_video_inputs()
+        inputs_dict = {"text": text, "images": image_input, "videos": video_inputs}
+        inputs = processor(**inputs_dict, return_tensors="pt", do_sample_frames=False)
+
+        self.assertSetEqual(set(inputs.keys()), set(processor.model_input_names))

transformers/tests/models/glm4v/test_video_processing_glm4v.py

@@ -0,0 +1,334 @@
+# coding=utf-8
+# 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 Glm4vVideoProcessor
+        from transformers.models.glm4v.video_processing_glm4v import smart_resize
+
+
+class Glm4vVideoProcessingTester:
+    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,
+    ):
+        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
+
+    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,
+        }
+
+    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 // self.temporal_patch_size
+        hidden_dim = self.num_channels * self.temporal_patch_size * 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:
+                    t, height, width = video.shape[:3]
+                elif len(video.shape) == 3:
+                    height, width = video.shape[:2]
+                    t = 1
+                else:
+                    t, height, width = self.num_frames, self.min_resolution, self.min_resolution
+            else:
+                t, height, width = self.num_frames, self.min_resolution, self.min_resolution
+
+            resized_height, resized_width = smart_resize(
+                t,
+                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 Glm4vVideoProcessingTest(VideoProcessingTestMixin, unittest.TestCase):
+    fast_video_processing_class = Glm4vVideoProcessor if is_torchvision_available() else None
+    input_name = "pixel_values_videos"
+
+    def setUp(self):
+        super().setUp()
+        self.video_processor_tester = Glm4vVideoProcessingTester(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 GLM-4.1V")
+    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,
+                image_std=1,
+            )[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,
+                image_std=1,
+            )[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)
+
+            with self.assertRaises(ValueError):
+                video_processing(video_inputs[0], 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