first commit

tests/models/llava_next_video/test_modeling_llava_next_video.py

@@ -0,0 +1,483 @@
+# 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 Llava-NeXT-Video model."""
+
+import copy
+import unittest
+
+import numpy as np
+import pytest
+from huggingface_hub import hf_hub_download
+from parameterized import parameterized
+
+from transformers import (
+    AutoProcessor,
+    BitsAndBytesConfig,
+    LlavaNextVideoConfig,
+    LlavaNextVideoForConditionalGeneration,
+    LlavaNextVideoModel,
+    is_torch_available,
+    is_vision_available,
+)
+from transformers.testing_utils import (
+    Expectations,
+    cleanup,
+    require_bitsandbytes,
+    require_torch,
+    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
+
+
+if is_vision_available():
+    from PIL import Image
+
+
+class LlavaNextVideoVisionText2TextModelTester:
+    def __init__(
+        self,
+        parent,
+        ignore_index=-100,
+        image_token_index=0,
+        video_token_index=1,
+        projector_hidden_act="gelu",
+        seq_length=7,
+        vision_feature_select_strategy="default",
+        vision_feature_layer=-1,
+        text_config={
+            "model_type": "llama",
+            "seq_length": 7,
+            "is_training": True,
+            "use_input_mask": True,
+            "use_token_type_ids": False,
+            "use_labels": True,
+            "vocab_size": 99,
+            "hidden_size": 32,
+            "num_hidden_layers": 2,
+            "num_attention_heads": 4,
+            "intermediate_size": 37,
+            "hidden_act": "gelu",
+            "hidden_dropout_prob": 0.1,
+            "attention_probs_dropout_prob": 0.1,
+            "max_position_embeddings": 580,
+            "type_vocab_size": 16,
+            "type_sequence_label_size": 2,
+            "initializer_range": 0.02,
+            "num_labels": 3,
+            "num_choices": 4,
+            "pad_token_id": 3,
+        },
+        is_training=True,
+        vision_config={
+            "image_size": 8,
+            "patch_size": 4,
+            "num_channels": 3,
+            "is_training": True,
+            "hidden_size": 32,
+            "projection_dim": 32,
+            "num_hidden_layers": 2,
+            "num_attention_heads": 4,
+            "intermediate_size": 37,
+            "dropout": 0.1,
+            "attention_dropout": 0.1,
+            "initializer_range": 0.02,
+        },
+    ):
+        self.parent = parent
+        self.ignore_index = ignore_index
+        self.image_token_index = image_token_index
+        self.video_token_index = video_token_index
+        self.projector_hidden_act = projector_hidden_act
+        self.vision_feature_select_strategy = vision_feature_select_strategy
+        self.vision_feature_layer = vision_feature_layer
+        self.text_config = text_config
+        self.vision_config = vision_config
+        self.pad_token_id = text_config["pad_token_id"]
+
+        self.num_hidden_layers = text_config["num_hidden_layers"]
+        self.vocab_size = text_config["vocab_size"]
+        self.hidden_size = text_config["hidden_size"]
+        self.num_attention_heads = text_config["num_attention_heads"]
+        self.is_training = is_training
+
+        self.batch_size = 3
+        self.num_channels = 3
+        self.image_size = 30
+
+        self.image_grid_pinpoints = [[16, 16]]
+        self.num_image_tokens = 24
+        self.num_video_tokens = 8
+        self.seq_length = seq_length + self.num_image_tokens + self.num_video_tokens
+
+    def get_config(self):
+        return LlavaNextVideoConfig(
+            text_config=self.text_config,
+            vision_config=self.vision_config,
+            ignore_index=self.ignore_index,
+            image_token_index=self.image_token_index,
+            video_token_index=self.video_token_index,
+            projector_hidden_act=self.projector_hidden_act,
+            vision_feature_select_strategy=self.vision_feature_select_strategy,
+            vision_feature_layer=self.vision_feature_layer,
+            image_grid_pinpoints=self.image_grid_pinpoints,
+            video_seq_length=self.num_video_tokens,
+            image_seq_length=self.num_image_tokens,
+        )
+
+    def prepare_config_and_inputs(self):
+        pixel_values = floats_tensor(
+            [
+                self.batch_size,
+                5,
+                self.vision_config["num_channels"],
+                self.vision_config["image_size"],
+                self.vision_config["image_size"],
+            ]
+        )
+        pixel_values_videos = floats_tensor(
+            [
+                self.batch_size,
+                8,
+                self.vision_config["num_channels"],
+                self.vision_config["image_size"],
+                self.vision_config["image_size"],
+            ]
+        )
+        config = self.get_config()
+
+        return config, pixel_values, pixel_values_videos
+
+    def prepare_config_and_inputs_for_common(self):
+        config, pixel_values, pixel_values_videos = self.prepare_config_and_inputs()
+        input_ids = ids_tensor([self.batch_size, self.seq_length], config.text_config.vocab_size - 2) + 2
+        attention_mask = torch.ones(input_ids.shape, dtype=torch.long).to(torch_device)
+
+        input_ids[input_ids == config.image_token_index] = self.pad_token_id
+        input_ids[input_ids == config.video_token_index] = self.pad_token_id
+        input_ids[:, : self.num_image_tokens] = config.image_token_index
+        input_ids[:, self.num_image_tokens : self.num_video_tokens + self.num_image_tokens] = config.video_token_index
+
+        inputs_dict = {
+            "pixel_values": pixel_values,
+            "pixel_values_videos": pixel_values_videos,
+            "image_sizes": torch.tensor(
+                [[self.vision_config["image_size"], self.vision_config["image_size"]]] * self.batch_size
+            ),
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+        }
+        return config, inputs_dict
+
+
+@require_torch
+class LlavaNextVideoForConditionalGenerationModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):
+    """
+    Model tester for `LlavaNextVideoForConditionalGeneration`.
+    """
+
+    all_model_classes = (
+        (
+            LlavaNextVideoModel,
+            LlavaNextVideoForConditionalGeneration,
+        )
+        if is_torch_available()
+        else ()
+    )
+    # LlavaNextVideo merges batch_size and num_patches in the first output dimension
+    skip_test_image_features_output_shape = True
+    # LlavaNextVideo merges batch_size and num_frames in the first output dimension
+    skip_test_video_features_output_shape = True
+
+    _is_composite = True
+    test_torch_exportable = False
+
+    def setUp(self):
+        self.model_tester = LlavaNextVideoVisionText2TextModelTester(self)
+        common_properties = ["image_token_index", "video_token_index", "vision_feature_layer", "image_seq_length"]
+        self.config_tester = ConfigTester(
+            self, config_class=LlavaNextVideoConfig, has_text_modality=False, common_properties=common_properties
+        )
+
+    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 prompr 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)  # in=place modifications further
+            _ = model(**curr_input_dict)  # successful forward with no modifications
+
+            # remove one image but leave the image token in text
+            curr_input_dict["pixel_values"] = curr_input_dict["pixel_values"][-1:, ...]
+            curr_input_dict["image_sizes"] = curr_input_dict["image_sizes"][-1:, ...]
+            with self.assertRaisesRegex(ValueError, "Image features and image tokens do not match"):
+                _ = model(**curr_input_dict)
+
+            # simulate multi-image case by concatenating inputs where each has exactly one image/image-token
+            input_ids = curr_input_dict["input_ids"][:1]
+            pixel_values = curr_input_dict["pixel_values"][:1]
+            image_sizes = curr_input_dict["image_sizes"][:1]
+            input_ids = torch.cat([input_ids, input_ids], dim=0)
+
+            # one image and two image tokens raise an error
+            with self.assertRaisesRegex(ValueError, "Image features and image tokens do not match"):
+                _ = model(input_ids=input_ids, pixel_values=pixel_values, image_sizes=image_sizes)
+
+            # two images and two image tokens don't raise an error
+            pixel_values = torch.cat([pixel_values, pixel_values], dim=0)
+            image_sizes = torch.cat([image_sizes, image_sizes], dim=0)
+            _ = model(input_ids=input_ids, pixel_values=pixel_values, image_sizes=image_sizes)
+
+    def test_odd_sized_image(self):
+        # prepare model configuration
+        config = self.model_tester.get_config()
+
+        # prepare input
+        num_image_tokens = 24
+        pixel_values = floats_tensor([1, 5, 3, config.vision_config.image_size, config.vision_config.image_size])
+        input_ids = ids_tensor([1, 64], config.text_config.vocab_size - 2) + 2
+        input_ids[:, :num_image_tokens] = config.image_token_index
+        attention_mask = torch.ones(input_ids.shape, dtype=torch.long).to(torch_device)
+        inputs_dict = {
+            "pixel_values": pixel_values,
+            "image_sizes": torch.tensor([[13, 16]]),  # odd-sized image
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+        }
+
+        # forward with odd-sized image input
+        for model_class in self.all_model_classes:
+            model = model_class(config).to(torch_device)
+            model(**inputs_dict)
+
+    @parameterized.expand(
+        [
+            (-1,),
+            ([-1],),
+            ([-1, -2],),
+        ],
+    )
+    def test_vision_feature_layers(self, vision_feature_layer):
+        """
+        Test that we can use either one vision feature layer, or a list of
+        vision feature layers.
+        """
+        config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        config.vision_feature_layer = vision_feature_layer
+
+        num_feature_layers = 1 if isinstance(vision_feature_layer, int) else len(vision_feature_layer)
+        hidden_size = config.vision_config.hidden_size
+        expected_features = hidden_size * num_feature_layers
+
+        for model_class in self.all_model_classes:
+            model = model_class(config).to(torch_device)
+            # We should have the right number of input features,
+            # and should be able to run a forward pass without exploding
+            base_model = getattr(model, "model", model)
+            assert base_model.multi_modal_projector.linear_1.in_features == expected_features
+            model(**input_dict)
+
+    @pytest.mark.xfail(reason="This architecture seems to not compute gradients for some layer.")
+    def test_training_gradient_checkpointing(self):
+        super().test_training_gradient_checkpointing()
+
+    @pytest.mark.xfail(reason="This architecture seems to not compute gradients for some layer.")
+    def test_training_gradient_checkpointing_use_reentrant_false(self):
+        super().test_training_gradient_checkpointing_use_reentrant_false()
+
+    @pytest.mark.xfail(reason="This architecture seems to not compute gradients for some layer.")
+    def test_training_gradient_checkpointing_use_reentrant_true(self):
+        super().test_training_gradient_checkpointing_use_reentrant_true()
+
+    @unittest.skip("FlashAttention only support fp16 and bf16 data type")
+    def test_flash_attn_2_fp32_ln(self):
+        pass
+
+    @unittest.skip(
+        "VLMs need lots of steps to prepare images/mask correctly to get pad-free inputs. Can be tested as part of LLM test"
+    )
+    def test_flash_attention_2_padding_matches_padding_free_with_position_ids(self):
+        pass
+
+    def _video_features_prepare_config_and_inputs(self):
+        """
+        Helper method to extract only image-related inputs from the full set of inputs, for testing `get_image_features`.
+
+        Despite using `pixel_values_videos` in forward, LlavaNextVideo's `get_video_features` method
+        instead uses `pixel_values` as input, so we need to override the inputs accordingly.
+        """
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        inputs_dict = {"pixel_values": inputs_dict["pixel_values_videos"]}
+        return config, inputs_dict
+
+
+@require_torch
+class LlavaNextVideoForConditionalGenerationIntegrationTest(unittest.TestCase):
+    def setUp(self):
+        self.processor = AutoProcessor.from_pretrained("llava-hf/LLaVA-NeXT-Video-7B-hf")
+        image_file = hf_hub_download(
+            repo_id="raushan-testing-hf/images_test", filename="llava_v1_5_radar.jpg", repo_type="dataset"
+        )
+        video_file = hf_hub_download(
+            repo_id="raushan-testing-hf/videos-test", filename="video_demo.npy", repo_type="dataset"
+        )
+        self.image = Image.open(image_file)
+        self.video = np.load(video_file)
+        self.prompt_image = "USER: <image>\nWhat is shown in this image? ASSISTANT:"
+        self.prompt_video = "USER: <video>\nWhy is this video funny? ASSISTANT:"
+
+    def tearDown(self):
+        cleanup(torch_device, gc_collect=True)
+
+    @slow
+    @require_bitsandbytes
+    def test_small_model_integration_test(self):
+        model = LlavaNextVideoForConditionalGeneration.from_pretrained(
+            "llava-hf/LLaVA-NeXT-Video-7B-hf",
+            quantization_config=BitsAndBytesConfig(load_in_4bit=True),
+            cache_dir="./",
+        )
+
+        inputs = self.processor(text=self.prompt_video, videos=self.video, return_tensors="pt")
+        # verify single forward pass
+        inputs = inputs.to(torch_device)
+        with torch.no_grad():
+            output = model(**inputs)
+
+        # verify generation
+        output = model.generate(**inputs, do_sample=False, max_new_tokens=40)
+        expected_decoded_text = Expectations(
+            {
+                ("cuda", None): "USER: \nWhy is this video funny? ASSISTANT: The humor in this video comes from the unexpected and somewhat comical situation of a young child reading a book while another child is attempting to read the same book. The child who is reading the book seems",
+                ("xpu", None): "USER: \nWhy is this video funny? ASSISTANT: The humor in this video comes from the unexpected and somewhat comical situation of a young child reading a book while another child is attempting to read the same book. The child who is reading the book seems",
+                ("rocm", (9, 5)): "USER: \nWhy is this video funny? ASSISTANT: The humor in this video comes from the unexpected and adorable behavior of the young child. The child is seen reading a book, but instead of turning the pages like one would typically do, they",
+            }
+        ).get_expectation()  # fmt: off
+
+        decoded_text = self.processor.decode(output[0], skip_special_tokens=True)
+        self.assertEqual(decoded_text, expected_decoded_text)
+
+    @slow
+    @require_bitsandbytes
+    def test_small_model_integration_test_batch(self):
+        model = LlavaNextVideoForConditionalGeneration.from_pretrained(
+            "llava-hf/LLaVA-NeXT-Video-7B-hf",
+            quantization_config=BitsAndBytesConfig(load_in_4bit=True),
+            cache_dir="./",
+        )
+
+        inputs = self.processor(
+            text=[self.prompt_video, self.prompt_video],
+            videos=[self.video, self.video],
+            return_tensors="pt",
+            padding=True,
+        ).to(torch_device)
+
+        output = model.generate(**inputs, do_sample=False, max_new_tokens=20)
+        decoded_text = self.processor.batch_decode(output, skip_special_tokens=True)
+
+        expected_decoded_text = Expectations(
+            {
+                ("xpu", None): "USER: \nWhy is this video funny? ASSISTANT: The humor in this video comes from the unexpected and somewhat comical situation of a young child reading a",
+                ("cuda", None): "USER: \nWhy is this video funny? ASSISTANT: The humor in this video comes from the unexpected and somewhat comical situation of a young child reading a",
+                ("rocm", (9, 5)): "USER: \nWhy is this video funny? ASSISTANT: The humor in this video comes from the unexpected and adorable behavior of the young child. The",
+            }
+        ).get_expectation()  # fmt: off
+        EXPECTED_DECODED_TEXT = [expected_decoded_text, expected_decoded_text]
+
+        self.assertEqual(decoded_text, EXPECTED_DECODED_TEXT)
+
+    @slow
+    @require_bitsandbytes
+    def test_small_model_integration_test_batch_different_vision_types(self):
+        model = LlavaNextVideoForConditionalGeneration.from_pretrained(
+            "llava-hf/LLaVA-NeXT-Video-7B-hf",
+            quantization_config=BitsAndBytesConfig(load_in_4bit=True),
+            cache_dir="./",
+        )
+
+        inputs = self.processor(
+            text=[self.prompt_image, self.prompt_video],
+            images=self.image,
+            videos=self.video,
+            return_tensors="pt",
+            padding=True,
+        ).to(torch_device)
+
+        # check loss when labels are passed
+        inputs["labels"] = inputs["input_ids"].clone()
+        with torch.no_grad():
+            output = model(**inputs)
+        self.assertTrue(output.loss is not None)
+
+        # verify generation
+        output = model.generate(**inputs, do_sample=False, max_new_tokens=50)
+        EXPECTED_DECODED_TEXT = Expectations(
+            {
+                ("xpu", None): 'USER: \nWhat is shown in this image? ASSISTANT: The image appears to be a graphical representation of a machine learning model\'s performance on a task, likely related to natural language processing or text understanding. It shows a scatter plot with two axes, one labeled "BLIP-2"',
+                ("rocm", (9, 5)): "USER: \nWhat is shown in this image? ASSISTANT: The image displays a chart that appears to be a comparison of different models or versions of a machine learning (ML) model, likely a neural network, based on their performance on a task or dataset. The chart is a scatter plot with axes labeled",
+                ("cuda", None): 'USER: \nWhat is shown in this image? ASSISTANT: The image appears to be a graphical representation of a machine learning model\'s performance on a task, likely related to natural language processing or text understanding. It shows a scatter plot with two axes, one labeled "BLIP-2"',
+            }
+        ).get_expectation()  # fmt: off
+
+        decoded_text = self.processor.decode(output[0], skip_special_tokens=True)
+        self.assertEqual(decoded_text, EXPECTED_DECODED_TEXT)
+
+    @slow
+    @require_bitsandbytes
+    def test_small_model_integration_test_batch_matches_single(self):
+        model = LlavaNextVideoForConditionalGeneration.from_pretrained(
+            "llava-hf/LLaVA-NeXT-Video-7B-hf",
+            quantization_config=BitsAndBytesConfig(load_in_4bit=True),
+            cache_dir="./",
+        )
+
+        inputs_batched = self.processor(
+            text=[self.prompt_video, self.prompt_image],
+            images=[self.image],
+            videos=[self.video],
+            return_tensors="pt",
+            padding=True,
+        ).to(torch_device)
+
+        inputs_single = self.processor(text=self.prompt_video, videos=[self.video], return_tensors="pt").to(
+            torch_device
+        )
+
+        # verify generation
+        output_batched = model.generate(**inputs_batched, do_sample=False, max_new_tokens=50)
+        output_single = model.generate(**inputs_single, do_sample=False, max_new_tokens=50)
+        self.assertEqual(
+            self.processor.decode(output_batched[0], skip_special_tokens=True),
+            self.processor.decode(output_single[0], skip_special_tokens=True),
+        )

tests/models/llava_next_video/test_processing_llava_next_video.py

@@ -0,0 +1,109 @@
+# 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 unittest
+
+import torch
+
+from transformers import LlavaNextVideoProcessor
+from transformers.testing_utils import require_vision
+from transformers.utils import is_torchvision_available, is_vision_available
+
+from ...test_processing_common import ProcessorTesterMixin
+
+
+if is_vision_available():
+    if is_torchvision_available():
+        pass
+
+
+@require_vision
+class LlavaNextVideoProcessorTest(ProcessorTesterMixin, unittest.TestCase):
+    processor_class = LlavaNextVideoProcessor
+
+    @classmethod
+    def _setup_tokenizer(cls):
+        tokenizer_class = cls._get_component_class_from_processor("tokenizer")
+        tokenizer = tokenizer_class.from_pretrained("llava-hf/LLaVA-NeXT-Video-7B-hf")
+        tokenizer.add_special_tokens({"additional_special_tokens": ["<image>", "<video>"]})
+        return tokenizer
+
+    @classmethod
+    def _setup_test_attributes(cls, processor):
+        cls.image_token = processor.image_token
+        cls.video_token = processor.video_token
+
+    @classmethod
+    def prepare_processor_dict(cls):
+        return {
+            "chat_template": "{% for message in messages %}{{'<|im_start|>' + message['role'] + ' '}}{# Render all images first #}{% for content in message['content'] | selectattr('type', 'equalto', 'image') %}{{ '<image>' }}{% endfor %}{# Render all video then #}{% for content in message['content'] | selectattr('type', 'equalto', 'video') %}{{ '<video>' }}{% endfor %}{# Render all text next #}{% if message['role'] != 'assistant' %}{% for content in message['content'] | selectattr('type', 'equalto', 'text') %}{{ '\n' + content['text'] }}{% endfor %}{% else %}{% for content in message['content'] | selectattr('type', 'equalto', 'text') %}{% generation %}{{ '\n' + content['text'] }}{% endgeneration %}{% endfor %}{% endif %}{{'<|im_end|>'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% endif %}",
+            "num_additional_image_tokens": 0,
+            "patch_size": 128,
+            "vision_feature_select_strategy": "default",
+        }
+
+    # Copied from tests.models.llava.test_processing_llava.LlavaProcessorTest.test_get_num_vision_tokens
+    def test_get_num_vision_tokens(self):
+        "Tests general functionality of the helper used internally in vLLM"
+
+        processor = self.get_processor()
+
+        output = processor._get_num_multimodal_tokens(image_sizes=[(100, 100), (300, 100), (500, 30)])
+        self.assertTrue("num_image_tokens" in output)
+        self.assertEqual(len(output["num_image_tokens"]), 3)
+
+        self.assertTrue("num_image_patches" in output)
+        self.assertEqual(len(output["num_image_patches"]), 3)
+
+    # Copied from tests.models.llava.test_processing_llava.LlavaProcessorTest.test_chat_template_is_saved
+    def test_chat_template_is_saved(self):
+        processor_loaded = self.processor_class.from_pretrained(self.tmpdirname)
+        processor_dict_loaded = json.loads(processor_loaded.to_json_string())
+        # chat templates aren't serialized to json in processors
+        self.assertFalse("chat_template" in processor_dict_loaded)
+
+        # they have to be saved as separate file and loaded back from that file
+        # so we check if the same template is loaded
+        processor_dict = self.prepare_processor_dict()
+        self.assertTrue(processor_loaded.chat_template == processor_dict.get("chat_template", None))
+
+    def test_image_token_filling(self):
+        processor = self.processor_class.from_pretrained(self.tmpdirname)
+        processor.patch_size = 14
+        processor.vision_feature_select_strategy = "default"
+        processor.image_processor.crop_size = {"height": 336, "width": 336}
+        processor.image_processor.size = {"shortest_edge": 336}
+        processor.image_processor.image_grid_pinpoints = [[672, 336]]
+        # Important to check with non square image
+        image = torch.randint(0, 2, (3, 503, 316))
+        expected_image_tokens = 1525
+        image_token_index = processor.image_token_id
+
+        messages = [
+            {
+                "role": "user",
+                "content": [
+                    {"type": "image"},
+                    {"type": "text", "text": "What is shown in this image?"},
+                ],
+            },
+        ]
+        inputs = processor(
+            text=[processor.apply_chat_template(messages)],
+            images=[image],
+            return_tensors="pt",
+        )
+        image_tokens = (inputs["input_ids"] == image_token_index).sum().item()
+        self.assertEqual(expected_image_tokens, image_tokens)

tests/models/llava_next_video/test_video_processing_llava_next_video.py

@@ -0,0 +1,123 @@
+# 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
+
+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_torchvision_available, is_vision_available
+
+from ...test_video_processing_common import VideoProcessingTestMixin, prepare_video_inputs
+
+
+if is_vision_available():
+    if is_torchvision_available():
+        from transformers import LlavaNextVideoVideoProcessor
+
+
+class LlavaNextVideoProcessingTester:
+    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_center_crop=True,
+        crop_size=None,
+        do_normalize=True,
+        image_mean=OPENAI_CLIP_MEAN,
+        image_std=OPENAI_CLIP_STD,
+        do_convert_rgb=True,
+    ):
+        size = size if size is not None else {"height": 20, "width": 20}
+        crop_size = crop_size if crop_size is not None else {"height": 18, "width": 18}
+        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_center_crop = do_center_crop
+        self.crop_size = crop_size
+        self.do_normalize = do_normalize
+        self.image_mean = image_mean
+        self.image_std = image_std
+        self.do_convert_rgb = do_convert_rgb
+
+    def prepare_video_processor_dict(self):
+        return {
+            "do_resize": self.do_resize,
+            "size": self.size,
+            "do_center_crop": self.do_center_crop,
+            "crop_size": self.crop_size,
+            "do_normalize": self.do_normalize,
+            "image_mean": self.image_mean,
+            "image_std": self.image_std,
+            "do_convert_rgb": self.do_convert_rgb,
+        }
+
+    def expected_output_video_shape(self, images):
+        return self.num_frames, self.num_channels, self.crop_size["height"], self.crop_size["width"]
+
+    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 LlavaNextVideoProcessingTest(VideoProcessingTestMixin, unittest.TestCase):
+    fast_video_processing_class = LlavaNextVideoVideoProcessor if is_torchvision_available() else None
+
+    def setUp(self):
+        super().setUp()
+        self.video_processor_tester = LlavaNextVideoProcessingTester(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_center_crop"))
+        self.assertTrue(hasattr(video_processing, "center_crop"))
+        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, {"height": 20, "width": 20})
+        self.assertEqual(video_processor.crop_size, {"height": 18, "width": 18})
+
+        video_processor = self.fast_video_processing_class.from_dict(self.video_processor_dict, size=42, crop_size=84)
+        self.assertEqual(video_processor.size, {"shortest_edge": 42})
+        self.assertEqual(video_processor.crop_size, {"height": 84, "width": 84})