first commit

tests/models/qwen2_vl/test_image_processing_qwen2_vl.py

@@ -0,0 +1,362 @@
+# Copyright 2024 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 io
+import itertools
+import json
+import tempfile
+import unittest
+
+import httpx
+import numpy as np
+
+from transformers.image_utils import OPENAI_CLIP_MEAN, OPENAI_CLIP_STD
+from transformers.models.qwen2_vl.image_processing_qwen2_vl 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, prepare_video_inputs
+
+
+if is_torch_available():
+    import torch
+
+if is_vision_available():
+    from PIL import Image
+
+
+class Qwen2VLImageProcessingTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=7,
+        num_channels=3,
+        num_frames=10,
+        min_resolution=56,
+        max_resolution=1024,
+        min_pixels=56 * 56,
+        max_pixels=28 * 28 * 1280,
+        do_normalize=True,
+        image_mean=OPENAI_CLIP_MEAN,
+        image_std=OPENAI_CLIP_STD,
+        do_resize=True,
+        patch_size=14,
+        temporal_patch_size=2,
+        merge_size=2,
+        do_convert_rgb=True,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.min_resolution = min_resolution
+        self.max_resolution = max_resolution
+        self.num_channels = num_channels
+        self.num_frames = num_frames
+        self.image_mean = OPENAI_CLIP_MEAN
+        self.image_std = OPENAI_CLIP_STD
+        self.min_pixels = min_pixels
+        self.max_pixels = max_pixels
+        self.patch_size = patch_size
+        self.temporal_patch_size = temporal_patch_size
+        self.merge_size = merge_size
+        self.do_resize = do_resize
+        self.do_normalize = do_normalize
+        self.image_mean = image_mean
+        self.image_std = image_std
+        self.do_convert_rgb = do_convert_rgb
+
+    def prepare_image_processor_dict(self):
+        return {
+            "do_resize": self.do_resize,
+            "image_mean": self.image_mean,
+            "image_std": self.image_std,
+            "min_pixels": self.min_pixels,
+            "max_pixels": self.max_pixels,
+            "patch_size": self.patch_size,
+            "temporal_patch_size": self.temporal_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]
+
+    def prepare_video_inputs(self, equal_resolution=False, numpify=False, torchify=False):
+        return prepare_video_inputs(
+            batch_size=self.batch_size,
+            num_channels=self.num_channels,
+            num_frames=self.num_frames,
+            min_resolution=self.min_resolution,
+            max_resolution=self.max_resolution,
+            equal_resolution=equal_resolution,
+            numpify=numpify,
+            torchify=torchify,
+        )
+
+
+@require_torch
+@require_vision
+class Qwen2VLImageProcessingTest(ImageProcessingTestMixin, unittest.TestCase):
+    def setUp(self):
+        super().setUp()
+        self.image_processor_tester = Qwen2VLImageProcessingTester(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, "do_convert_rgb"))
+            self.assertTrue(hasattr(image_processing, "patch_size"))
+            self.assertTrue(hasattr(image_processing, "temporal_patch_size"))
+            self.assertTrue(hasattr(image_processing, "merge_size"))
+
+    def test_image_processor_to_json_string(self):
+        for image_processing_class in self.image_processing_classes.values():
+            image_processor = image_processing_class(**self.image_processor_dict)
+            obj = json.loads(image_processor.to_json_string())
+            for key, value in self.image_processor_dict.items():
+                if key not in ["min_pixels", "max_pixels"]:
+                    self.assertEqual(obj[key], value)
+
+    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 = (4900, 1176)
+            expected_image_grid_thws = torch.Tensor([[1, 70, 70]])
+            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 = (34300, 1176)
+            expected_image_grid_thws = torch.Tensor([[1, 70, 70]] * 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 = (4900, 1176)
+            expected_image_grid_thws = torch.Tensor([[1, 70, 70]])
+            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 = (34300, 1176)
+            expected_image_grid_thws = torch.Tensor([[1, 70, 70]] * 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 = (4900, 1176)
+            expected_image_grid_thws = torch.Tensor([[1, 70, 70]])
+            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 = (34300, 1176)
+            expected_image_grid_thws = torch.Tensor([[1, 70, 70]] * 7)
+            self.assertEqual(tuple(encoded_images.shape), expected_output_image_shape)
+            self.assertTrue((image_grid_thws == expected_image_grid_thws).all())
+
+    @unittest.skip(reason="Qwen2VLImageProcessor 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 = (34300, 1176)
+            expected_image_grid_thws = torch.Tensor([[1, 70, 70]] * 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 = (34300, 1176)
+            expected_image_grid_thws = torch.Tensor([[1, 70, 70]] * 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, max_pixels=56 * 56, min_pixels=28 * 28
+                )
+
+            image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=True)
+            process_out = image_processor_loaded(image_inputs, return_tensors="pt")
+            expected_output_video_shape = [112, 1176]
+            self.assertListEqual(list(process_out.pixel_values.shape), expected_output_video_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["min_pixels"] = min(a_pixels, b_pixels)
+                image_processor_dict["max_pixels"] = 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):
+        if len(self.image_processing_classes) < 2:
+            self.skipTest(reason="Skipping backends equivalence test as there are less than 2 backends")
+
+        dummy_image = Image.open(
+            io.BytesIO(
+                httpx.get("http://images.cocodataset.org/val2017/000000039769.jpg", follow_redirects=True).content
+            )
+        )
+
+        # Create processors for each backend
+        encodings = {}
+        for backend_name, image_processing_class in self.image_processing_classes.items():
+            image_processor = image_processing_class(**self.image_processor_dict)
+            encodings[backend_name] = image_processor(dummy_image, return_tensors="pt")
+
+        # Compare all backends to the first one (reference backend)
+        backend_names = list(encodings.keys())
+        reference_backend = backend_names[0]
+        reference_encoding = encodings[reference_backend]
+        for backend_name in backend_names[1:]:
+            self._assert_tensors_equivalence(reference_encoding.pixel_values, encodings[backend_name].pixel_values)
+            self.assertEqual(reference_encoding.image_grid_thw.dtype, encodings[backend_name].image_grid_thw.dtype)
+            self._assert_tensors_equivalence(
+                reference_encoding.image_grid_thw.float(), encodings[backend_name].image_grid_thw.float()
+            )
+
+    @require_vision
+    @require_torch
+    def test_backends_equivalence_batched(self):
+        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)
+
+        # Create processors for each backend
+        encodings = {}
+        for backend_name, image_processing_class in self.image_processing_classes.items():
+            image_processor = image_processing_class(**self.image_processor_dict)
+            encodings[backend_name] = image_processor(dummy_images, return_tensors="pt")
+
+        # Compare all backends to the first one (reference backend)
+        backend_names = list(encodings.keys())
+        reference_backend = backend_names[0]
+        reference_encoding = encodings[reference_backend]
+        for backend_name in backend_names[1:]:
+            self._assert_tensors_equivalence(reference_encoding.pixel_values, encodings[backend_name].pixel_values)
+            self.assertEqual(reference_encoding.image_grid_thw.dtype, encodings[backend_name].image_grid_thw.dtype)
+            self._assert_tensors_equivalence(
+                reference_encoding.image_grid_thw.float(), encodings[backend_name].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/qwen2_vl/test_modeling_qwen2_vl.py

@@ -0,0 +1,721 @@
+# Copyright 2024 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 Qwen2-VL model."""
+
+import copy
+import gc
+import tempfile
+import unittest
+
+import pytest
+import requests
+
+from transformers import (
+    AutoProcessor,
+    Qwen2VLConfig,
+    Qwen2VLForConditionalGeneration,
+    Qwen2VLModel,
+    is_torch_available,
+    is_vision_available,
+)
+from transformers.testing_utils import (
+    Expectations,
+    backend_empty_cache,
+    require_flash_attn,
+    require_torch,
+    require_torch_accelerator,
+    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,
+)
+from ...test_pipeline_mixin import PipelineTesterMixin
+
+
+if is_torch_available():
+    import torch
+
+
+if is_vision_available():
+    from PIL import Image
+
+
+class Qwen2VLVisionText2TextModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=3,
+        seq_length=7,
+        num_channels=3,
+        ignore_index=-100,
+        image_size=14,
+        text_config={
+            "bos_token_id": 0,
+            "eos_token_id": 1,
+            "pad_token_id": 2,
+            "hidden_act": "silu",
+            "hidden_size": 32,
+            "vocab_size": 99,
+            "intermediate_size": 37,
+            "max_position_embeddings": 512,
+            "max_window_layers": 3,
+            "num_attention_heads": 4,
+            "num_hidden_layers": 2,
+            "num_key_value_heads": 2,
+            "rope_theta": 10000,
+            "tie_word_embeddings": True,
+            "rope_parameters": {"type": "mrope", "mrope_section": [2, 1, 1]},
+        },
+        vision_start_token_id=3,
+        image_token_id=4,
+        video_token_id=5,
+        is_training=True,
+        vision_config={
+            "depth": 2,
+            "embed_dim": 32,
+            "hidden_act": "quick_gelu",
+            "hidden_size": 32,
+            "mlp_ratio": 4,
+            "num_heads": 4,
+            "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.num_hidden_layers = text_config["num_hidden_layers"]
+        self.num_attention_heads = text_config["num_attention_heads"]
+        self.hidden_size = text_config["hidden_size"]
+        self.vision_start_token_id = vision_start_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.vocab_size = text_config["vocab_size"]
+        self.num_image_tokens = 32
+        self.seq_length = seq_length + self.num_image_tokens
+
+    def get_config(self):
+        return Qwen2VLConfig(
+            text_config=self.text_config,
+            vision_config=self.vision_config,
+            vision_start_token_id=self.vision_start_token_id,
+            image_token_id=self.image_token_id,
+            video_token_id=self.video_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[:, -1] = self.pad_token_id
+        attention_mask[:, -1] = 0
+        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.vision_start_token_id] = self.pad_token_id
+        input_ids[:, self.num_image_tokens] = self.image_token_id
+        input_ids[:, self.num_image_tokens - 1] = self.vision_start_token_id
+
+        mm_token_type_ids = torch.zeros_like(input_ids)
+        mm_token_type_ids[:, self.num_image_tokens] = 1
+
+        inputs_dict = {
+            "pixel_values": pixel_values,
+            "image_grid_thw": torch.tensor([[1, 1, 1]] * 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 Qwen2VLModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase):
+    """
+    Model tester for `Qwen2VLForConditionalGeneration`.
+    """
+
+    all_model_classes = (
+        (
+            Qwen2VLModel,
+            Qwen2VLForConditionalGeneration,
+        )
+        if is_torch_available()
+        else ()
+    )
+    pipeline_model_mapping = {
+        "image-text-to-text": Qwen2VLForConditionalGeneration,
+        "any-to-any": Qwen2VLForConditionalGeneration,
+    }
+    _is_composite = True
+
+    def setUp(self):
+        self.model_tester = Qwen2VLVisionText2TextModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=Qwen2VLConfig, has_text_modality=False)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_mismatching_num_image_tokens(self):
+        """
+        Tests that VLMs through an error with explicit message saying what is wrong
+        when number of images don't match number of image tokens in the text.
+        Also we need to test multi-image cases when one prompt has multiple image tokens.
+        """
+        config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        for model_class in self.all_model_classes:
+            model = model_class(config).to(torch_device)
+            model.eval()
+            curr_input_dict = copy.deepcopy(input_dict)
+            _ = model(**curr_input_dict)  # successful forward with no modifications
+
+            # remove one image but leave the image token in text
+            patch_size = config.vision_config.patch_size
+            one_img_length = (self.model_tester.image_size**2) // (patch_size**2)
+            curr_input_dict["pixel_values"] = curr_input_dict["pixel_values"][-one_img_length:, ...]
+            curr_input_dict["image_grid_thw"] = curr_input_dict["image_grid_thw"][-1:, ...]
+            with self.assertRaisesRegex(ValueError, "Image features and image tokens do not match"):
+                _ = model(**curr_input_dict)
+
+            model.base_model.rope_deltas = None
+            # simulate multi-image case by concatenating inputs where each has exactly one image/image-token
+            input_ids = curr_input_dict["input_ids"][:1]
+            mm_token_type_ids = curr_input_dict["mm_token_type_ids"][:1]
+            pixel_values = curr_input_dict["pixel_values"][:one_img_length]
+            image_grid_thw = curr_input_dict["image_grid_thw"][:1]
+            input_ids = torch.cat([input_ids, input_ids], dim=0)
+            mm_token_type_ids = torch.cat([mm_token_type_ids, mm_token_type_ids], dim=0)
+            with self.assertRaisesRegex(ValueError, "Image features and image tokens do not match"):
+                _ = model(
+                    input_ids=input_ids,
+                    pixel_values=pixel_values,
+                    image_grid_thw=image_grid_thw,
+                    mm_token_type_ids=mm_token_type_ids,
+                )
+
+            model.base_model.rope_deltas = None
+            # two images and two image tokens don't raise an error
+            pixel_values = torch.cat([pixel_values, pixel_values], dim=0)
+            image_grid_thw = torch.cat([image_grid_thw, image_grid_thw], dim=0)
+            _ = model(
+                input_ids=input_ids,
+                pixel_values=pixel_values,
+                image_grid_thw=image_grid_thw,
+                mm_token_type_ids=mm_token_type_ids,
+            )
+
+    def test_forward_with_rope_deltas_cached(self):
+        """
+        Tests that Qwen2-VL computes new rope deltas every forward pass with new set of inputs.
+        Rope deltas are cached when we generate and re-used for decoding phase, byt are not reset
+        automatically after generation ends. See https://github.com/huggingface/transformers/pull/36013 for more
+        """
+        config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        for model_class in self.all_generative_model_classes:
+            model = model_class(config).to(torch_device)
+
+            # Generate and make sure rope_deltas are not `None`
+            self.assertTrue(model.model.rope_deltas is None)
+            generation_output = model.generate(
+                **input_dict, max_new_tokens=4, return_dict_in_generate=True, output_logits=True
+            )
+            self.assertTrue(model.model.rope_deltas is not None)
+
+            # Now if we try to do forward pass, we should get new rope logits, because cache is not passed
+            forward_output = model(**input_dict)
+            torch.testing.assert_close(
+                generation_output.logits[0], forward_output.logits[:, -1, :], rtol=1e-4, atol=1e-4
+            )
+
+            # Same happens if we call `generate` API instead of `forward`
+            generation_output_second = model.generate(
+                **input_dict, max_new_tokens=10, return_dict_in_generate=True, output_logits=True
+            )
+            torch.testing.assert_close(
+                generation_output.logits[0], generation_output_second.logits[0], rtol=1e-4, atol=1e-4
+            )
+
+    def test_vision_position_ids(self):
+        """
+        Tests that vision position ids are built correctly for images and for videos.
+        See https://github.com/huggingface/transformers/pull/45400
+        """
+        config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        model = Qwen2VLModel(config).to(torch_device)
+        batch_size = input_dict["input_ids"].shape[0]
+
+        # Test most simple case when num_image_tokens == 1. Position ids will be sunsequent and text-like
+        position_ids = model.get_rope_index(
+            input_dict["input_ids"], input_dict["mm_token_type_ids"], input_dict["image_grid_thw"]
+        )[0]
+        expected_positions = torch.arange(39)[None, None, :].repeat(3, batch_size, 1)
+        self.assertListEqual(list(position_ids.shape), [3, batch_size, 39])
+        self.assertListEqual(position_ids.tolist(), expected_positions.tolist())
+
+        # Each image encodes to more than 1 token (i.e. 4 height and 3 width patches = 12 tokens)
+        image_token_id = config.image_token_id
+        pad_token_id = config.text_config.pad_token_id
+        input_ids = torch.tensor([[pad_token_id] + [image_token_id] * 12 + [pad_token_id]], device=torch_device)
+        mm_token_type_ids = torch.tensor([[0] + [1] * 12 + [0]], device=torch_device)
+        image_grid_thw = torch.tensor([[1, 4, 3]], device=torch_device)
+        position_ids = model.get_rope_index(input_ids, mm_token_type_ids, image_grid_thw)[0]
+        expected_positions = torch.tensor(
+            [
+                [[0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 5]],
+                [[0, 1, 1, 1, 2, 2, 2, 3, 3, 3, 4, 4, 4, 5]],
+                [[0, 1, 2, 3, 1, 2, 3, 1, 2, 3, 1, 2, 3, 5]],
+            ]
+        )
+
+        self.assertListEqual(list(position_ids.shape), [3, 1, 14])
+        self.assertListEqual(position_ids.tolist(), expected_positions.tolist())
+
+        # Check video position ids with 2 frames, and 4 height, 3 width patches (= 12 * 2 tokens)
+        video_token_id = config.video_token_id
+        input_ids = torch.tensor([[pad_token_id] + [video_token_id] * 24 + [pad_token_id]], device=torch_device)
+        mm_token_type_ids = torch.tensor([[0] + [2] * 24 + [0]], device=torch_device)
+        video_grid_thw = torch.tensor([[2, 4, 3]], device=torch_device)
+        position_ids = model.get_rope_index(input_ids, mm_token_type_ids, video_grid_thw=video_grid_thw)[0]
+        expected_positions = torch.tensor(
+            [
+                [[0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 5]],
+                [[0, 1, 1, 1, 2, 2, 2, 3, 3, 3, 4, 4, 4, 1, 1, 1, 2, 2, 2, 3, 3, 3, 4, 4, 4, 5]],
+                [[0, 1, 2, 3, 1, 2, 3, 1, 2, 3, 1, 2, 3, 1, 2, 3, 1, 2, 3, 1, 2, 3, 1, 2, 3, 5]],
+            ]
+        )
+
+        self.assertListEqual(list(position_ids.shape), [3, 1, 26])
+        self.assertListEqual(position_ids.tolist(), expected_positions.tolist())
+
+    def attention_mask_padding_matches_padding_free_with_position_ids(
+        self, attn_implementation: str, fa_kwargs: bool = False
+    ):
+        max_new_tokens = 30
+        for model_class in self.all_generative_model_classes:
+            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+            dummy_input = inputs_dict[model_class.main_input_name]
+            if dummy_input.dtype in [torch.float32, torch.float16]:
+                dummy_input = dummy_input.to(torch.bfloat16)
+
+            # make sure that all models have enough positions for generation
+            if hasattr(config, "max_position_embeddings"):
+                config.max_position_embeddings = max_new_tokens + dummy_input.shape[1] + 1
+
+            model = model_class(config)
+
+            with tempfile.TemporaryDirectory() as tmpdirname:
+                model.save_pretrained(tmpdirname)
+
+                if 0 in inputs_dict["attention_mask"][:, -1]:
+                    inputs_dict["attention_mask"] = inputs_dict["attention_mask"].flip(1)
+                dummy_attention_mask = inputs_dict["attention_mask"]
+                inputs_dict["input_ids"][~dummy_attention_mask.bool()] = config.get_text_config().pad_token_id
+
+                model = (
+                    model_class.from_pretrained(
+                        tmpdirname,
+                        dtype=torch.bfloat16,
+                        attn_implementation=attn_implementation,
+                    )
+                    .to(torch_device)
+                    .eval()
+                )
+
+                # flatten
+                padfree_inputs_dict = {
+                    "pixel_values": inputs_dict["pixel_values"],
+                    "image_grid_thw": inputs_dict["image_grid_thw"],
+                    "input_ids": inputs_dict["input_ids"][dummy_attention_mask.bool()].unsqueeze(0),
+                }
+
+                # add position_ids
+                vision_position_ids, deltas = model.model.get_rope_index(
+                    input_ids=inputs_dict["input_ids"],
+                    image_grid_thw=inputs_dict["image_grid_thw"],
+                    attention_mask=inputs_dict["attention_mask"],
+                    mm_token_type_ids=inputs_dict["mm_token_type_ids"],
+                )  # [3, bs, padded-seq-len]
+                vision_padfree_positions = vision_position_ids[:, dummy_attention_mask.bool()].view(
+                    3, -1
+                )  # [3, bs*padfree-len]
+                text_padfree_positions = torch.cat(
+                    [torch.arange(length) for length in dummy_attention_mask.sum(1).tolist()]
+                )  # [1, bs*padfree-len]
+                text_padfree_positions = text_padfree_positions.long().unsqueeze(0).to(torch_device)
+                padfree_inputs_dict["position_ids"] = torch.cat([text_padfree_positions, vision_padfree_positions])[
+                    :, None, :
+                ]
+
+                if fa_kwargs:
+                    cu_seq_lens = [0] + dummy_attention_mask.sum(1).tolist()
+                    cu_seq_lens = torch.tensor(cu_seq_lens, device=torch_device)
+                    max_length = cu_seq_lens.diff().max().item()
+                    padfree_inputs_dict.update(
+                        {
+                            "cu_seq_lens_q": cu_seq_lens.cumsum(-1).to(dtype=torch.int32),
+                            "cu_seq_lens_k": cu_seq_lens.cumsum(-1).to(dtype=torch.int32),
+                            "max_length_q": max_length,
+                            "max_length_k": max_length,
+                        }
+                    )
+
+                # We need to do simple forward without cache in roder to trigger packed SDPA/FLEX/EAGER path
+                res_padded = model(**inputs_dict, use_cache=False)
+                res_padfree = model(**padfree_inputs_dict, use_cache=False)
+
+                logits_padded = res_padded.logits[inputs_dict["attention_mask"].bool()]
+                logits_padfree = res_padfree.logits[0]
+
+                # acceptable numerical instability
+                tol = torch.finfo(torch.bfloat16).eps
+                torch.testing.assert_close(logits_padded, logits_padfree, rtol=tol, atol=tol)
+
+    @unittest.skip(reason="Feedforward chunking is not yet supported")
+    def test_feed_forward_chunking(self):
+        pass
+
+    @unittest.skip(reason="CPU offload is not yet supported")
+    def test_cpu_offload(self):
+        pass
+
+    @unittest.skip(reason="Some undefined behavior encountered with test versions of this model. Skip for now.")
+    def test_disk_offload_bin(self):
+        pass
+
+    @unittest.skip(reason="Some undefined behavior encountered with test versions of this model. Skip for now.")
+    def test_disk_offload_safetensors(self):
+        pass
+
+    @unittest.skip(reason="Some undefined behavior encountered with test versions of this model. Skip for now.")
+    def test_model_parallelism(self):
+        pass
+
+    @unittest.skip(reason="Compile not yet supported because in Qwen2VL models")
+    def test_sdpa_can_dispatch_on_flash(self):
+        pass
+
+    @unittest.skip(reason="Got `CUDA error: misaligned address` with PyTorch 2.0.0.")
+    def test_multi_gpu_data_parallel_forward(self):
+        pass
+
+    def test_enable_input_require_grads_with_gradient_checkpointing(self):
+        if not self.model_tester.is_training:
+            self.skipTest(reason="ModelTester not in training mode")
+
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        config.use_cache = False
+        config.return_dict = True
+
+        for model_class in self.all_model_classes:
+            if not model_class.supports_gradient_checkpointing:
+                continue
+
+            model = model_class(config)
+            model.to(torch_device)
+            model.gradient_checkpointing_enable(gradient_checkpointing_kwargs={"use_reentrant": False})
+            model.enable_input_require_grads()
+            model.train()
+
+            for parameter in model.parameters():
+                parameter.requires_grad = False
+
+            vision_module = None
+            if hasattr(model, "visual"):
+                vision_module = model.visual
+            elif hasattr(model, "model") and hasattr(model.model, "visual"):
+                vision_module = model.model.visual
+
+            if vision_module is None:
+                continue
+
+            target_linear = vision_module.blocks[0].attn.qkv
+            target_linear.weight.requires_grad = True
+            if target_linear.bias is not None:
+                target_linear.bias.requires_grad = True
+
+            inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
+            outputs = model(**inputs)
+
+            if hasattr(outputs, "loss") and outputs.loss is not None:
+                loss = outputs.loss
+            else:
+                logits = outputs.logits if hasattr(outputs, "logits") else outputs[0]
+                loss = logits.sum()
+
+            loss.backward()
+
+            self.assertIsNotNone(
+                target_linear.weight.grad,
+                f"qkv weights should receive gradients when enable_input_require_grads is used with gradient checkpointing. Model: {model_class.__name__}",
+            )
+            self.assertGreater(
+                target_linear.weight.grad.abs().sum().item(),
+                0,
+                f"qkv weights should have non-zero gradients when enable_input_require_grads is used with gradient checkpointing. Model: {model_class.__name__}",
+            )
+
+
+@require_torch
+class Qwen2VLIntegrationTest(unittest.TestCase):
+    def setUp(self):
+        self.processor = AutoProcessor.from_pretrained("Qwen/Qwen2-VL-7B-Instruct")
+        self.messages = [
+            {
+                "role": "user",
+                "content": [
+                    {"type": "image"},
+                    {"type": "text", "text": "What kind of dog is this?"},
+                ],
+            }
+        ]
+        url = "https://qianwen-res.oss-accelerate-overseas.aliyuncs.com/Qwen2-VL/demo_small.jpg"
+        self.image = Image.open(requests.get(url, stream=True).raw)
+
+    def tearDown(self):
+        gc.collect()
+        backend_empty_cache(torch_device)
+
+    @slow
+    def test_small_model_integration_test(self):
+        model = Qwen2VLForConditionalGeneration.from_pretrained(
+            "Qwen/Qwen2-VL-7B-Instruct", dtype="auto", device_map="auto"
+        )
+
+        text = self.processor.apply_chat_template(self.messages, tokenize=False, add_generation_prompt=True)
+        inputs = self.processor(text=[text], images=[self.image], return_tensors="pt")
+
+        expected_input_ids = [151644, 8948, 198, 2610, 525, 264, 10950, 17847, 13, 151645, 198, 151644, 872, 198, 151652, 151655, 151655]  # fmt: skip
+        assert expected_input_ids == inputs.input_ids[0].tolist()[:17]
+
+        expected_pixel_slice = torch.tensor(
+            [
+                [0.8792, 0.8792, 0.9084],
+                [1.1858, 1.1858, 1.2296],
+                [1.2004, 1.2004, 1.2150],
+                [1.4340, 1.4340, 1.4194],
+                [1.3902, 1.4048, 1.4194],
+                [1.5216, 1.5362, 1.5362],
+            ],
+            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)
+
+        output = model.generate(**inputs, max_new_tokens=30)
+        EXPECTED_DECODED_TEXT = "system\nYou are a helpful assistant.\nuser\nWhat kind of dog is this?\nassistant\nThe dog in the picture appears to be a Labrador Retriever. Labradors are known for their friendly and intelligent nature, making them popular choices"
+
+        self.assertEqual(
+            self.processor.decode(output[0], skip_special_tokens=True),
+            EXPECTED_DECODED_TEXT,
+        )
+
+    @slow
+    def test_small_model_integration_test_batch(self):
+        model = Qwen2VLForConditionalGeneration.from_pretrained(
+            "Qwen/Qwen2-VL-7B-Instruct", dtype="auto", device_map="auto"
+        )
+        text = self.processor.apply_chat_template(self.messages, tokenize=False, add_generation_prompt=True)
+        inputs = self.processor(text=[text, text], images=[self.image, self.image], return_tensors="pt").to(
+            torch_device
+        )
+
+        # it should not matter whether two images are the same size or not
+        output = model.generate(**inputs, max_new_tokens=30)
+
+        EXPECTED_DECODED_TEXT = [
+            'system\nYou are a helpful assistant.\nuser\nWhat kind of dog is this?\nassistant\nThe dog in the picture appears to be a Labrador Retriever. Labradors are known for their friendly and intelligent nature, making them popular choices',
+            'system\nYou are a helpful assistant.\nuser\nWhat kind of dog is this?\nassistant\nThe dog in the picture appears to be a Labrador Retriever. Labradors are known for their friendly and intelligent nature, making them popular choices',
+        ]  # fmt: skip
+        self.assertEqual(
+            self.processor.batch_decode(output, skip_special_tokens=True),
+            EXPECTED_DECODED_TEXT,
+        )
+
+    @slow
+    def test_small_model_integration_test_expand(self):
+        model = Qwen2VLForConditionalGeneration.from_pretrained(
+            "Qwen/Qwen2-VL-7B-Instruct", dtype="auto", device_map="auto"
+        )
+        text = self.processor.apply_chat_template(self.messages, tokenize=False, add_generation_prompt=True)
+        inputs = self.processor(text=[text], images=[self.image], return_tensors="pt").to(torch_device)
+
+        output = model.generate(**inputs, max_new_tokens=30, num_return_sequences=3)
+
+        EXPECTED_DECODED_TEXT = [
+            'system\nYou are a helpful assistant.\nuser\nWhat kind of dog is this?\nassistant\nThe dog in the picture appears to be a Labrador Retriever. Labradors are known for their friendly and intelligent nature, making them popular choices',
+            'system\nYou are a helpful assistant.\nuser\nWhat kind of dog is this?\nassistant\nThe dog in the picture appears to be a Labrador Retriever. Labradors are known for their friendly and intelligent nature, making them popular choices',
+            'system\nYou are a helpful assistant.\nuser\nWhat kind of dog is this?\nassistant\nThe dog in the picture appears to be a Labrador Retriever. Labradors are known for their friendly and intelligent nature, making them popular choices',
+        ]  # 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 = Qwen2VLForConditionalGeneration.from_pretrained(
+            "Qwen/Qwen2-VL-7B-Instruct", dtype="auto", device_map="auto"
+        )
+        text = self.processor.apply_chat_template(self.messages, tokenize=False, add_generation_prompt=True)
+        messages2 = [
+            {"role": "system", "content": "You are a helpful assistant."},
+            {"role": "user", "content": "Who are you?"},
+        ]
+        text2 = self.processor.apply_chat_template(messages2, tokenize=False, add_generation_prompt=True)
+        inputs = self.processor(text=[text, text2], images=[self.image], padding=True, return_tensors="pt").to(
+            torch_device
+        )
+
+        # it should not matter whether two images are the same size or not
+        output = model.generate(**inputs, max_new_tokens=30)
+
+        EXPECTED_DECODED_TEXT = [
+            'system\nYou are a helpful assistant.\nuser\nWhat kind of dog is this?\nassistant\nThe dog in the picture appears to be a Labrador Retriever. Labradors are known for their friendly and intelligent nature, making them popular choices',
+            'system\nYou are a helpful assistant.\nuser\nWho are you?\nassistant\nI am a large language model created by Alibaba Cloud. I am called Qwen.'
+        ]  # 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 = Qwen2VLForConditionalGeneration.from_pretrained(
+            "Qwen/Qwen2-VL-7B-Instruct", dtype="auto", device_map="auto"
+        )
+        text = self.processor.apply_chat_template(self.messages, tokenize=False, add_generation_prompt=True)
+        text2 = self.processor.apply_chat_template(self.messages, tokenize=False, add_generation_prompt=True)
+        image2 = self.image.resize((224, 224))
+        inputs = self.processor(text=[text, text2], images=[self.image, image2], padding=True, return_tensors="pt").to(
+            torch_device
+        )
+
+        # it should not matter whether two images are the same size or not
+        output = model.generate(**inputs, max_new_tokens=30)
+        DECODED_TEXT = self.processor.batch_decode(output, skip_special_tokens=True)
+
+        EXPECTED_DECODED_TEXTS = Expectations(
+            {
+                ("xpu", 3): [
+                    'system\nYou are a helpful assistant.\nuser\nWhat kind of dog is this?\nassistant\nThe dog in the picture appears to be a Labrador Retriever. Labradors are known for their friendly and intelligent nature, making them popular choices',
+                    'system\nYou are a helpful assistant.\nuser\nWhat kind of dog is this?\nassistant\nThe dog in the picture appears to be a Labrador Retriever. Labradors are known for their friendly and intelligent nature, making them popular choices',
+                ],
+                ("cuda", None): [
+                    'system\nYou are a helpful assistant.\nuser\nWhat kind of dog is this?\nassistant\nThe dog in the picture appears to be a Labrador Retriever. Labradors are known for their friendly and intelligent nature, making them popular choices',
+                    'system\nYou are a helpful assistant.\nuser\nWhat kind of dog is this?\nassistant\nThe dog in the picture appears to be a Labrador Retriever. Labradors are known for their friendly and intelligent nature, making them popular pets',
+                ],
+                ("cuda", 8): [
+                    'system\nYou are a helpful assistant.\nuser\nWhat kind of dog is this?\nassistant\nThe dog in the picture appears to be a Labrador Retriever. Labradors are known for their friendly and intelligent nature, making them popular choices',
+                    'system\nYou are a helpful assistant.\nuser\nWhat kind of dog is this?\nassistant\nThe dog in the picture appears to be a Labrador Retriever. Labradors are known for their friendly and intelligent nature, making them popular choices'
+                ],
+            }
+        )  # fmt: skip
+        EXPECTED_DECODED_TEXT = EXPECTED_DECODED_TEXTS.get_expectation()
+
+        self.assertEqual(DECODED_TEXT, EXPECTED_DECODED_TEXT)
+
+    @slow
+    @require_flash_attn
+    @require_torch_accelerator
+    @pytest.mark.flash_attn_test
+    def test_small_model_integration_test_batch_flashatt2(self):
+        model = Qwen2VLForConditionalGeneration.from_pretrained(
+            "Qwen/Qwen2-VL-7B-Instruct",
+            dtype=torch.bfloat16,
+            attn_implementation="flash_attention_2",
+            device_map="auto",
+        )
+        text = self.processor.apply_chat_template(self.messages, tokenize=False, add_generation_prompt=True)
+        inputs = self.processor(text=[text, text], images=[self.image, self.image], return_tensors="pt").to(
+            torch_device
+        )
+
+        # it should not matter whether two images are the same size or not
+        output = model.generate(**inputs, max_new_tokens=30)
+
+        EXPECTED_DECODED_TEXT = [
+            "system\nYou are a helpful assistant.\nuser\nWhat kind of dog is this?\nassistant\nThe dog in the picture appears to be a Labrador Retriever. Labradors are known for their friendly and intelligent nature, making them popular choices",
+            "system\nYou are a helpful assistant.\nuser\nWhat kind of dog is this?\nassistant\nThe dog in the picture appears to be a Labrador Retriever. Labradors are known for their friendly and intelligent nature, making them popular choices",
+        ]
+        self.assertEqual(
+            self.processor.batch_decode(output, skip_special_tokens=True),
+            EXPECTED_DECODED_TEXT,
+        )
+
+    @slow
+    @require_flash_attn
+    @require_torch_accelerator
+    @pytest.mark.flash_attn_test
+    def test_small_model_integration_test_batch_wo_image_flashatt2(self):
+        model = Qwen2VLForConditionalGeneration.from_pretrained(
+            "Qwen/Qwen2-VL-7B-Instruct",
+            dtype=torch.bfloat16,
+            attn_implementation="flash_attention_2",
+            device_map="auto",
+        )
+        text = self.processor.apply_chat_template(self.messages, tokenize=False, add_generation_prompt=True)
+        messages2 = [
+            {"role": "system", "content": "You are a helpful assistant."},
+            {"role": "user", "content": "Who are you?"},
+        ]
+        text2 = self.processor.apply_chat_template(messages2, tokenize=False, add_generation_prompt=True)
+        inputs = self.processor(text=[text, text2], images=[self.image], padding=True, return_tensors="pt").to(
+            torch_device
+        )
+
+        # it should not matter whether two images are the same size or not
+        output = model.generate(**inputs, max_new_tokens=30)
+
+        EXPECTED_DECODED_TEXT = [
+            'system\nYou are a helpful assistant.\nuser\nWhat kind of dog is this?\nassistant\nThe dog in the picture appears to be a Labrador Retriever. Labradors are known for their friendly and intelligent nature, making them popular choices',
+            'system\nYou are a helpful assistant.\nuser\nWho are you?\nassistant\nI am a large language model created by Alibaba Cloud. I am called Qwen.'
+        ]  # fmt: skip
+
+        self.assertEqual(
+            self.processor.batch_decode(output, skip_special_tokens=True),
+            EXPECTED_DECODED_TEXT,
+        )

tests/models/qwen2_vl/test_processing_qwen2_vl.py

@@ -0,0 +1,300 @@
+# Copyright 2024 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 unittest
+
+import numpy as np
+
+from transformers.testing_utils import require_av, require_torch, require_torchvision, require_vision
+from transformers.utils import is_torch_available, is_torchvision_available, is_vision_available
+
+from ...test_processing_common import ProcessorTesterMixin, url_to_local_path
+
+
+if is_vision_available():
+    from transformers import Qwen2VLProcessor
+
+    if is_torchvision_available():
+        pass
+
+if is_torch_available():
+    import torch
+
+
+@require_vision
+@require_torch
+@require_torchvision
+class Qwen2VLProcessorTest(ProcessorTesterMixin, unittest.TestCase):
+    processor_class = Qwen2VLProcessor
+    model_id = "Qwen/Qwen2-VL-7B-Instruct"
+
+    @classmethod
+    def _setup_from_pretrained(cls, model_id, **kwargs):
+        return super()._setup_from_pretrained(model_id, patch_size=4, max_pixels=56 * 56, min_pixels=28 * 28, **kwargs)
+
+    @classmethod
+    def _setup_test_attributes(cls, processor):
+        cls.image_token = processor.image_token
+
+    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)
+
+    @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"]))
+        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,
+            num_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)
+        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 * 192
+        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")
+
+        if "video_processor" not in self.processor_class.get_attributes():
+            self.skipTest("Processor doesn't accept videos at input")
+
+        messages = [
+            [
+                {
+                    "role": "user",
+                    "content": [
+                        {
+                            "type": "video",
+                            "url": url_to_local_path(
+                                "https://huggingface.co/datasets/raushan-testing-hf/videos-test/resolve/main/tiny_video.mp4"
+                            ),
+                        },
+                        {"type": "text", "text": "What is shown in this video?"},
+                    ],
+                },
+            ]
+        ]
+
+        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,
+        )
+        self.assertTrue(self.videos_input_name in out_dict_with_video)
+        self.assertEqual(len(out_dict_with_video[self.videos_input_name]), 360)
+
+        # 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]), 360)
+
+        # 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]), 1080)
+
+        # 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,
+        )
+        self.assertTrue(self.videos_input_name in out_dict_with_video)
+        self.assertEqual(len(out_dict_with_video[self.videos_input_name]), 160)
+
+        # 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_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()
+        inputs = processor(text=input_str, images=image_input, return_tensors="pt")
+        self.assertEqual(inputs[self.images_input_name].shape[0], 100)
+        inputs = processor(text=input_str, images=image_input, max_pixels=56 * 56 * 4, return_tensors="pt")
+        self.assertEqual(inputs[self.images_input_name].shape[0], 612)
+
+    def test_special_mm_token_truncation(self):
+        """Tests that special vision tokens do not get truncated when `truncation=True` is set."""
+
+        processor = self.get_processor()
+
+        input_str = self.prepare_text_inputs(batch_size=2, modalities="image")
+        image_input = self.prepare_image_inputs(batch_size=2)
+
+        _ = processor(
+            text=input_str,
+            images=image_input,
+            return_tensors="pt",
+            truncation=None,
+            padding=True,
+        )
+
+        with self.assertRaises(ValueError):
+            _ = processor(
+                text=input_str,
+                images=image_input,
+                return_tensors="pt",
+                truncation=True,
+                padding=True,
+                max_length=20,
+            )

tests/models/qwen2_vl/test_video_processing_qwen2_vl.py

@@ -0,0 +1,395 @@
+# 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 json
+import tempfile
+import unittest
+
+import numpy as np
+
+from transformers.image_utils import OPENAI_CLIP_MEAN, OPENAI_CLIP_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():
+    import torch
+
+if is_vision_available():
+    from PIL import Image
+
+    from transformers.image_utils import get_image_size
+    from transformers.models.qwen2_vl.video_processing_qwen2_vl import smart_resize
+
+    if is_torchvision_available():
+        from transformers import Qwen2VLVideoProcessor
+
+
+class Qwen2VLVideoProcessingTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=5,
+        num_frames=8,
+        num_channels=3,
+        min_resolution=30,
+        max_resolution=80,
+        do_resize=True,
+        size=None,
+        do_normalize=True,
+        image_mean=OPENAI_CLIP_MEAN,
+        image_std=OPENAI_CLIP_STD,
+        do_convert_rgb=True,
+        temporal_patch_size=2,
+        patch_size=14,
+        min_pixels=20 * 20,
+        max_pixels=100 * 100,
+        merge_size=2,
+    ):
+        size = size if size is not None else {"shortest_edge": 400, "longest_edge": 10000}
+        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.min_pixels = min_pixels
+        self.max_pixels = max_pixels
+        self.merge_size = merge_size
+
+    def prepare_video_processor_dict(self):
+        return {
+            "do_resize": self.do_resize,
+            "do_normalize": self.do_normalize,
+            "image_mean": self.image_mean,
+            "image_std": self.image_std,
+            "do_convert_rgb": self.do_convert_rgb,
+            "temporal_patch_size": self.temporal_patch_size,
+            "patch_size": self.patch_size,
+            "min_pixels": self.min_pixels,
+            "max_pixels": self.max_pixels,
+            "merge_size": self.merge_size,
+        }
+
+    @require_vision
+    def expected_output_video_shape(self, videos, num_frames=None):
+        num_frames = num_frames if num_frames is not None else self.num_frames
+        grid_t = 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[0], Image.Image):
+                video = np.stack([np.array(frame) for frame in video])
+            height, width = get_image_size(video)
+            resized_height, resized_width = smart_resize(
+                height,
+                width,
+                factor=self.patch_size * self.merge_size,
+                min_pixels=self.min_pixels,
+                max_pixels=self.max_pixels,
+            )
+            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 Qwen2VLVideoProcessingTest(VideoProcessingTestMixin, unittest.TestCase):
+    fast_video_processing_class = Qwen2VLVideoProcessor if is_torchvision_available() else None
+
+    def setUp(self):
+        super().setUp()
+        self.video_processor_tester = Qwen2VLVideoProcessingTester(self)
+
+    @property
+    def video_processor_dict(self):
+        return self.video_processor_tester.prepare_video_processor_dict()
+
+    def test_video_processor_properties(self):
+        video_processing = self.fast_video_processing_class(**self.video_processor_dict)
+        self.assertTrue(hasattr(video_processing, "do_resize"))
+        self.assertTrue(hasattr(video_processing, "size"))
+        self.assertTrue(hasattr(video_processing, "do_normalize"))
+        self.assertTrue(hasattr(video_processing, "image_mean"))
+        self.assertTrue(hasattr(video_processing, "image_std"))
+        self.assertTrue(hasattr(video_processing, "do_convert_rgb"))
+
+    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, {"shortest_edge": 400, "longest_edge": 10000})
+
+        video_processor = self.fast_video_processing_class.from_dict(
+            self.video_processor_dict, size={"shortest_edge": 100, "longest_edge": 200}
+        )
+        # min_pixels and max_pixels take precedence over size, like in the image processor.
+        self.assertEqual(video_processor.size, {"shortest_edge": 400, "longest_edge": 10000})
+
+        processor_dict = self.video_processor_dict.copy()
+        processor_dict.pop("min_pixels")
+        processor_dict.pop("max_pixels")
+        video_processor = self.fast_video_processing_class.from_dict(
+            processor_dict, size={"shortest_edge": 100, "longest_edge": 200}
+        )
+        self.assertEqual(video_processor.size, {"shortest_edge": 100, "longest_edge": 200})
+
+    def test_video_processor_to_json_string(self):
+        for video_processing_class in self.video_processor_list:
+            video_processor = video_processing_class(**self.video_processor_dict)
+            obj = json.loads(video_processor.to_json_string())
+            for key, value in self.video_processor_dict.items():
+                if key not in ["min_pixels", "max_pixels"]:
+                    self.assertEqual(obj[key], value)
+
+    def test_call_pil(self):
+        for video_processing_class in self.video_processor_list:
+            # Initialize video_processing
+            video_processing = video_processing_class(**self.video_processor_dict)
+            video_inputs = self.video_processor_tester.prepare_video_inputs(
+                equal_resolution=False, return_tensors="pil"
+            )
+
+            # Each video is a list of PIL Images
+            for video in video_inputs:
+                self.assertIsInstance(video[0], Image.Image)
+
+            # Test not batched input
+            encoded_videos = video_processing(video_inputs[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, 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:
+            # Initialize video_processing
+            video_processing = video_processing_class(**self.video_processor_dict)
+            # create random numpy tensors
+            video_inputs = self.video_processor_tester.prepare_video_inputs(
+                equal_resolution=False, return_tensors="np"
+            )
+            for video in video_inputs:
+                self.assertIsInstance(video, np.ndarray)
+
+            # Test not batched input
+            encoded_videos = video_processing(video_inputs[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, 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:
+            # Initialize video_processing
+            video_processing = video_processing_class(**self.video_processor_dict)
+            # create random PyTorch tensors
+            video_inputs = self.video_processor_tester.prepare_video_inputs(
+                equal_resolution=False, return_tensors="torch"
+            )
+
+            for video in video_inputs:
+                self.assertIsInstance(video, torch.Tensor)
+
+            # Test not batched input
+            encoded_videos = video_processing(video_inputs[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
+            expected_output_video_shape = self.video_processor_tester.expected_output_video_shape(video_inputs)
+            encoded_videos = video_processing(video_inputs, return_tensors="pt")[self.input_name]
+            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"
+            )
+
+            # Test not batched input
+            video_inputs_nested = [list(video) for video in video_inputs]
+            encoded_videos = video_processing(video_inputs_nested[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
+            expected_output_video_shape = self.video_processor_tester.expected_output_video_shape(video_inputs)
+            encoded_videos = video_processing(video_inputs_nested, return_tensors="pt")[self.input_name]
+            self.assertEqual(list(encoded_videos.shape), expected_output_video_shape)
+
+    @unittest.skip("Skip for now, the test needs adjustment fo Qwen2VL")
+    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_call_sample_frames(self):
+        for video_processing_class in self.video_processor_list:
+            video_processing = video_processing_class(**self.video_processor_dict)
+
+            prev_num_frames = self.video_processor_tester.num_frames
+            self.video_processor_tester.num_frames = 8
+            video_inputs = self.video_processor_tester.prepare_video_inputs(
+                equal_resolution=False,
+                return_tensors="torch",
+            )
+
+            # Force set sampling to False. No sampling is expected even when `num_frames` exists
+            video_processing.do_sample_frames = False
+
+            encoded_videos = video_processing(video_inputs[0], return_tensors="pt", num_frames=3)[self.input_name]
+            encoded_videos_batched = video_processing(video_inputs, return_tensors="pt", num_frames=3)[self.input_name]
+            expected_output_video_shape = self.video_processor_tester.expected_output_video_shape([video_inputs[0]])
+            expected_output_video_shape_batched = self.video_processor_tester.expected_output_video_shape(video_inputs)
+            self.assertListEqual(list(encoded_videos.shape), expected_output_video_shape)
+            self.assertListEqual(list(encoded_videos_batched.shape), expected_output_video_shape_batched)
+
+            # Set sampling to True. Video frames should be sampled with `num_frames` in the output
+            video_processing.do_sample_frames = True
+
+            encoded_videos = video_processing(video_inputs[0], return_tensors="pt", num_frames=4)[self.input_name]
+            encoded_videos_batched = video_processing(video_inputs, return_tensors="pt", num_frames=4)[self.input_name]
+            expected_output_video_shape = self.video_processor_tester.expected_output_video_shape(
+                [video_inputs[0]], num_frames=4
+            )
+            expected_output_video_shape_batched = self.video_processor_tester.expected_output_video_shape(
+                video_inputs, num_frames=4
+            )
+            self.assertListEqual(list(encoded_videos.shape), expected_output_video_shape)
+            self.assertListEqual(list(encoded_videos_batched.shape), expected_output_video_shape_batched)
+
+            metadata = [[{"duration": 2.0, "total_num_frames": 8, "fps": 4}]]
+            batched_metadata = metadata * len(video_inputs)
+            encoded_videos = video_processing(video_inputs[0], return_tensors="pt", fps=3, video_metadata=metadata)[
+                self.input_name
+            ]
+            encoded_videos_batched = video_processing(
+                video_inputs, return_tensors="pt", fps=3, video_metadata=batched_metadata
+            )[self.input_name]
+            expected_output_video_shape = self.video_processor_tester.expected_output_video_shape(
+                [video_inputs[0]], num_frames=6
+            )
+            expected_output_video_shape_batched = self.video_processor_tester.expected_output_video_shape(
+                video_inputs, num_frames=6
+            )
+            self.assertListEqual(list(encoded_videos.shape), expected_output_video_shape)
+            self.assertListEqual(list(encoded_videos_batched.shape), expected_output_video_shape_batched)
+
+            # We should raise error when asked to sample more frames than there are in input video
+            with self.assertRaises(ValueError):
+                encoded_videos = video_processing(video_inputs[0], return_tensors="pt", num_frames=10)[self.input_name]
+                encoded_videos_batched = video_processing(video_inputs, return_tensors="pt", num_frames=10)[
+                    self.input_name
+                ]
+
+            # Assign back the actual num frames in tester
+            self.video_processor_tester.num_frames = prev_num_frames
+
+    def test_num_frames_equal_temporal_patch_size_plus_two(self):
+        for video_processing_class in self.video_processor_list:
+            video_processor_dict = self.video_processor_dict.copy()
+            video_processor_dict["size"] = {"longest_edge": 5 * 28 * 28, "shortest_edge": 28 * 28}
+            video_processor_dict["do_sample_frames"] = False
+            temporal_patch_size = 3
+            video_processor_dict["temporal_patch_size"] = temporal_patch_size
+            video_processing = video_processing_class(**video_processor_dict)
+
+            n, w, h = 5, 28, 28
+            video_inputs = [(np.random.randint(0, 256, (h, w, 3), dtype=np.uint8)) for _ in range(n)]
+
+            video_processed = video_processing(video_inputs, return_tensors="pt")
+            encoded_videos = video_processed[self.input_name]
+            self.assertEqual(list(encoded_videos.shape), [8, temporal_patch_size * 3 * 14 * 14])
+
+            video_grid_thw = video_processed["video_grid_thw"]
+            self.assertEqual(video_grid_thw.tolist(), [[2, 2, 2]])
+
+    def test_bc_min_max_pixels(self):
+        for video_processing_class in self.video_processor_list:
+            video_processing = video_processing_class(**self.video_processor_dict)
+            with tempfile.TemporaryDirectory() as tmpdirname:
+                video_processing.save_pretrained(tmpdirname)
+                video_processing_loaded = video_processing_class.from_pretrained(
+                    tmpdirname, max_pixels=56 * 56, min_pixels=28 * 28
+                )
+
+            video_inputs = self.video_processor_tester.prepare_video_inputs(
+                equal_resolution=True,
+                return_tensors="torch",
+            )
+            processed = video_processing_loaded(video_inputs, return_tensors="pt")
+            expected_output_video_shape = [320, 1176]
+            self.assertListEqual(list(processed.pixel_values_videos.shape), expected_output_video_shape)