Add InternVL (2.5 MPO) (#35968)

* initial commit

* add convert internvl

* add first end-to-end working internvl

* nit prompt and image proc

* add working chat template

* add conversion llama-based models

* add tests

* pass all tests

* fix isort

* fix modular after main merge

* add video processing for internvl

* add support for interlaced images and videos

* Remove processing and config from modular, add more tests

* add llama model tests

* Modify processor for compatibility with refactored got ocr image processor

* add comments in processor

* Add docs and nits

* change video processing to use custom sample_indices_fn

* rebase and fix tests

* add processor tests

* Add changes Raushan review

* Use the new attention interface for the vision model

* nits

* add support for custom video_load_backend

* remove mention to InternVLTokenizer

* refactor vision model to simplify logic

* refactor processor for better readibility

* fix copies

* fix require av processor test

* refactor internVL vision

* Update processor and fix processing tests

* fix docstring

* update convert_weights for internvl3

* change image processor to fast by default

* remove do_center_crop=True in convert_weights

* force use_cache to True

* push_to_hub before reloading

* fix internVLVision for larger models

* update convert weight for qk norm

* fix convert_weights

* fix eos_token_id in convert

* update docs and integration tests

* make modifs after review

* fix wrong k_norm and reduce modular

* change image_token_index to image_token_id

* change checkpoint to OpenGVLab org

* last nits

* explicitely del self.num_key_value_groups

* add extra special tokens

tests/generation/test_utils.py

@@ -130,6 +130,7 @@ VLM_CLASS_NAMES = [
     "gemma3",
     "mistral3",
     "chameleon",
+    "internvl",
     "qwen2_5_omni",
 ]
 

tests/models/internvl/test_modeling_internvl.py

@@ -0,0 +1,894 @@
+# coding=utf-8
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Testing suite for the PyTorch InternVL model."""
+
+import unittest
+from io import BytesIO
+
+import requests
+
+from transformers import (
+    AutoProcessor,
+    BitsAndBytesConfig,
+    InternVLConfig,
+    is_torch_available,
+    is_vision_available,
+)
+from transformers.testing_utils import (
+    cleanup,
+    require_av,
+    require_bitsandbytes,
+    require_torch,
+    require_torch_gpu,
+    slow,
+    torch_device,
+)
+
+from ...generation.test_utils import GenerationTesterMixin
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_common import ModelTesterMixin, _config_zero_init, floats_tensor, ids_tensor
+from ...test_pipeline_mixin import PipelineTesterMixin
+
+
+if is_torch_available():
+    import torch
+
+    from transformers import (
+        InternVLForConditionalGeneration,
+    )
+
+
+if is_vision_available():
+    from PIL import Image
+
+
+class InternVLVisionText2TextModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=3,
+        seq_length=7,
+        image_seq_length=64,
+        vision_feature_layer=-1,
+        ignore_index=-100,
+        bos_token_id=0,
+        eos_token_id=0,
+        pad_token_id=0,
+        image_token_id=1,
+        num_channels=3,
+        image_size=64,
+        model_type="internvl",
+        is_training=True,
+        text_config={
+            "model_type": "qwen2",
+            "vocab_size": 99,
+            "hidden_size": 128,
+            "intermediate_size": 37,
+            "num_hidden_layers": 4,
+            "num_attention_heads": 4,
+            "num_key_value_heads": 2,
+            "output_channels": 64,
+            "hidden_act": "silu",
+            "max_position_embeddings": 512,
+            "rope_theta": 10000,
+            "mlp_ratio": 4,
+            "tie_word_embeddings": True,
+            "bos_token_id": 0,
+            "eos_token_id": 0,
+            "pad_token_id": 0,
+        },
+        vision_config={
+            "hidden_size": 32,
+            "num_hidden_layers": 2,
+            "num_attention_heads": 4,
+            "intermediate_size": 128,
+            "image_size": 64,
+            "patch_size": 4,
+            "num_channels": 3,
+            "hidden_act": "quick_gelu",
+            "use_absolute_position_embeddings": True,
+        },
+    ):
+        self.parent = parent
+        self.ignore_index = ignore_index
+        self.bos_token_id = bos_token_id
+        self.eos_token_id = eos_token_id
+        self.pad_token_id = pad_token_id
+        self.image_token_id = image_token_id
+        self.model_type = model_type
+        self.text_config = text_config
+        self.vision_config = vision_config
+        self.batch_size = batch_size
+        self.vision_feature_layer = vision_feature_layer
+        self.is_training = is_training
+        self.image_seq_length = image_seq_length
+        self.num_channels = num_channels
+        self.image_size = image_size
+        self.seq_length = seq_length + image_seq_length
+
+        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"]
+
+    def get_config(self):
+        return InternVLConfig(
+            text_config=self.text_config,
+            vision_config=self.vision_config,
+            model_type=self.model_type,
+            bos_token_id=self.bos_token_id,
+            eos_token_id=self.eos_token_id,
+            pad_token_id=self.pad_token_id,
+            image_token_id=self.image_token_id,
+            image_seq_length=self.image_seq_length,
+            vision_feature_layer=self.vision_feature_layer,
+        )
+
+    def prepare_config_and_inputs(self):
+        config = self.get_config()
+        pixel_values = floats_tensor([self.batch_size, self.num_channels, self.image_size, self.image_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
+        input_ids[input_ids == self.image_token_id] = self.pad_token_id
+        input_ids[:, : self.image_seq_length] = self.image_token_id
+
+        inputs_dict = {
+            "pixel_values": pixel_values,
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+        }
+        return config, inputs_dict
+
+    def create_and_check_model_fp16_forward(self, config, input_ids, pixel_values, attention_mask):
+        model = InternVLForConditionalGeneration(config=config)
+        model.to(torch_device)
+        model.half()
+        model.eval()
+        logits = model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            pixel_values=pixel_values.to(torch.bfloat16),
+            return_dict=True,
+        )["logits"]
+        self.parent.assertFalse(torch.isnan(logits).any().item())
+
+    def create_and_check_model_fp16_autocast_forward(self, config, input_ids, pixel_values, attention_mask):
+        config.torch_dtype = torch.float16
+        model = InternVLForConditionalGeneration(config=config)
+        model.to(torch_device)
+        model.eval()
+        with torch.autocast(device_type="cuda", dtype=torch.float16):
+            logits = model(
+                input_ids=input_ids,
+                attention_mask=attention_mask,
+                pixel_values=pixel_values.to(torch.bfloat16),
+                return_dict=True,
+            )["logits"]
+        self.parent.assertFalse(torch.isnan(logits).any().item())
+
+
+@require_torch
+class InternVLModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase):
+    all_model_classes = (InternVLForConditionalGeneration,) if is_torch_available() else ()
+    all_generative_model_classes = (InternVLForConditionalGeneration,) if is_torch_available() else ()
+    pipeline_model_mapping = (
+        {
+            "image-text-to-text": InternVLForConditionalGeneration,
+        }
+        if is_torch_available()
+        else {}
+    )
+    test_headmasking = False
+    test_pruning = False
+
+    def setUp(self):
+        self.model_tester = InternVLVisionText2TextModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=InternVLConfig, has_text_modality=False)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_initialization(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        configs_no_init = _config_zero_init(config)
+        for model_class in self.all_model_classes:
+            model = model_class(config=configs_no_init)
+            for name, param in model.named_parameters():
+                if param.requires_grad:
+                    self.assertIn(
+                        ((param.data.mean() * 1e9).round() / 1e9).item(),
+                        [0.0, 1.0],
+                        msg=f"Parameter {name} of model {model_class} seems not properly initialized",
+                    )
+
+    # overwrite inputs_embeds tests because we need to delete "pixel values" for LVLMs
+    def test_inputs_embeds(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            model = model_class(config)
+            model.to(torch_device)
+            model.eval()
+
+            inputs = self._prepare_for_class(inputs_dict, model_class)
+
+            input_ids = inputs["input_ids"]
+            del inputs["input_ids"]
+            del inputs["pixel_values"]
+
+            wte = model.get_input_embeddings()
+            inputs["inputs_embeds"] = wte(input_ids)
+
+            with torch.no_grad():
+                model(**inputs)
+
+    # overwrite inputs_embeds tests because we need to delete "pixel values" for LVLMs
+    # while some other models require pixel_values to be present
+    def test_inputs_embeds_matches_input_ids(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            model = model_class(config)
+            model.to(torch_device)
+            model.eval()
+
+            inputs = self._prepare_for_class(inputs_dict, model_class)
+            input_ids = inputs["input_ids"]
+            del inputs["input_ids"]
+            del inputs["pixel_values"]
+
+            inputs_embeds = model.get_input_embeddings()(input_ids)
+
+            with torch.no_grad():
+                out_ids = model(input_ids=input_ids, **inputs)[0]
+                out_embeds = model(inputs_embeds=inputs_embeds, **inputs)[0]
+            torch.testing.assert_close(out_embeds, out_ids)
+
+    @unittest.skip(reason="Compile not yet supported because in LLava models")
+    def test_sdpa_can_compile_dynamic(self):
+        pass
+
+    @unittest.skip("FlashAttention only support fp16 and bf16 data type")
+    def test_flash_attn_2_fp32_ln(self):
+        pass
+
+    @unittest.skip("Qwen2 flash attention does not support right padding")
+    def test_flash_attn_2_inference_equivalence_right_padding(self):
+        pass
+
+
+@slow
+@require_torch_gpu
+class InternVLQwen2IntegrationTest(unittest.TestCase):
+    def setUp(self):
+        self.small_model_checkpoint = "OpenGVLab/InternVL3-1B-hf"
+        self.medium_model_checkpoint = "OpenGVLab/InternVL3-2B-hf"
+
+    def tearDown(self):
+        cleanup(torch_device, gc_collect=True)
+
+    def test_qwen2_small_model_integration_generate(self):
+        processor = AutoProcessor.from_pretrained(self.small_model_checkpoint)
+        model = InternVLForConditionalGeneration.from_pretrained(
+            self.small_model_checkpoint, device_map=torch_device, torch_dtype=torch.bfloat16
+        )
+        url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        image = Image.open(requests.get(url, stream=True).raw)
+
+        prompt = "<|im_start|>user\n<image>\nPlease describe the image explicitly.<|im_end|>\n<|im_start|>assistant\n"
+        inputs = processor(images=image, text=prompt, return_tensors="pt").to(torch_device, dtype=torch.bfloat16)
+        with torch.no_grad():
+            generate_ids = model.generate(**inputs, max_new_tokens=20, do_sample=False)
+            decoded_output = processor.decode(
+                generate_ids[0, inputs["input_ids"].shape[1] :], skip_special_tokens=True
+            )
+        expected_output = "The image shows two cats lying on a pink blanket. The cat on the left is a tabby"
+        self.assertEqual(decoded_output, expected_output)
+
+    def test_qwen2_small_model_integration_forward(self):
+        processor = AutoProcessor.from_pretrained(self.small_model_checkpoint)
+        model = InternVLForConditionalGeneration.from_pretrained(
+            self.small_model_checkpoint, device_map=torch_device, torch_dtype=torch.bfloat16
+        )
+        url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        image = Image.open(requests.get(url, stream=True).raw)
+
+        prompt = "<|im_start|>user\n<image>\nPlease describe the image explicitly.<|im_end|>\n<|im_start|>assistant\n"
+        inputs = processor(images=image, text=prompt, return_tensors="pt").to(torch_device, dtype=torch.bfloat16)
+
+        # Forward
+        with torch.inference_mode():
+            output = model(**inputs)
+
+        actual_logits = output.logits[0, -1, :5].cpu()
+        expected_logits = torch.tensor([11.9375, 14.8750, 14.0625, 10.7500, 6.9062], dtype=torch.bfloat16)
+        self.assertTrue(
+            torch.allclose(actual_logits, expected_logits, atol=0.1),
+            f"Actual logits: {actual_logits}"
+            f"\nExpected logits: {expected_logits}"
+            f"\nDifference: {torch.abs(actual_logits - expected_logits)}",
+        )
+
+    def test_qwen2_small_model_integration_generate_text_only(self):
+        processor = AutoProcessor.from_pretrained(self.small_model_checkpoint)
+        model = InternVLForConditionalGeneration.from_pretrained(
+            self.small_model_checkpoint, device_map=torch_device, torch_dtype=torch.bfloat16
+        )
+        prompt = "<|im_start|>user\nWrite a haiku<|im_end|>\n<|im_start|>assistant\n"
+        inputs = processor(text=prompt, return_tensors="pt").to(torch_device, dtype=torch.bfloat16)
+        with torch.no_grad():
+            generate_ids = model.generate(**inputs, max_new_tokens=200, do_sample=False)
+            decoded_output = processor.decode(
+                generate_ids[0, inputs["input_ids"].shape[1] :], skip_special_tokens=True
+            )
+        expected_output = "Whispers of dawn,\nSilent whispers of the night,\nNew day's light begins."
+        self.assertEqual(decoded_output, expected_output)
+
+    def test_qwen2_small_model_integration_generate_chat_template(self):
+        processor = AutoProcessor.from_pretrained(self.small_model_checkpoint)
+        model = InternVLForConditionalGeneration.from_pretrained(
+            self.small_model_checkpoint, device_map=torch_device, torch_dtype=torch.bfloat16
+        )
+        messages = [
+            {
+                "role": "user",
+                "content": [
+                    {"type": "image", "url": "http://images.cocodataset.org/val2017/000000039769.jpg"},
+                    {"type": "text", "text": "Please describe the image explicitly."},
+                ],
+            }
+        ]
+
+        inputs = processor.apply_chat_template(
+            messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt"
+        ).to(torch_device, dtype=torch.bfloat16)
+        with torch.no_grad():
+            generate_ids = model.generate(**inputs, max_new_tokens=20, do_sample=False)
+            decoded_output = processor.decode(
+                generate_ids[0, inputs["input_ids"].shape[1] :], skip_special_tokens=True
+            )
+        expected_output = "The image shows two cats lying on a pink blanket. The cat on the left is a tabby"
+        self.assertEqual(decoded_output, expected_output)
+
+    def test_qwen2_small_model_integration_batched_generate(self):
+        processor = AutoProcessor.from_pretrained(self.small_model_checkpoint)
+        model = InternVLForConditionalGeneration.from_pretrained(
+            self.small_model_checkpoint, device_map=torch_device, torch_dtype=torch.bfloat16
+        )
+        # Prepare inputs
+        prompt = [
+            "<|im_start|>user\n<image>\nWrite a haiku for this image<|im_end|>\n<|im_start|>assistant\n",
+            "<|im_start|>user\n<image>\nDescribe this image<|im_end|>\n<|im_start|>assistant\n",
+        ]
+        image1 = Image.open(requests.get("https://llava-vl.github.io/static/images/view.jpg", stream=True).raw)
+        image2 = Image.open(requests.get("https://www.ilankelman.org/stopsigns/australia.jpg", stream=True).raw)
+
+        inputs = processor(text=prompt, images=[[image1], [image2]], padding=True, return_tensors="pt").to(
+            torch_device, dtype=torch.bfloat16
+        )
+
+        output = model.generate(**inputs, do_sample=False, max_new_tokens=25)
+
+        # Check first output
+        decoded_output = processor.decode(output[0], skip_special_tokens=True)
+        expected_output = "user\n\nWrite a haiku for this image\nassistant\nSilky lake,  \nWooden pier,  \nNature's peace."  # fmt: skip
+        self.assertEqual(
+            decoded_output,
+            expected_output,
+            f"Decoded output: {decoded_output}\nExpected output: {expected_output}",
+        )
+        # Check second output
+        decoded_output = processor.decode(output[1], skip_special_tokens=True)
+        expected_output = 'user\n\nDescribe this image\nassistant\nThe image shows a street scene with a traditional Chinese archway, known as a "Chinese Gate" or "Chinese Gate of'  # fmt: skip
+        self.assertEqual(
+            decoded_output,
+            expected_output,
+            f"Decoded output: {decoded_output}\nExpected output: {expected_output}",
+        )
+
+    def test_qwen2_small_model_integration_batched_generate_multi_image(self):
+        processor = AutoProcessor.from_pretrained(self.small_model_checkpoint)
+        model = InternVLForConditionalGeneration.from_pretrained(
+            self.small_model_checkpoint, device_map=torch_device, torch_dtype=torch.bfloat16
+        )
+        # Prepare inputs
+        prompt = [
+            "<|im_start|>user\n<image>\nWrite a haiku for this image<|im_end|>\n<|im_start|>assistant\n",
+            "<|im_start|>user\n<image><image>\nWhat are the differences between these two images?<|im_end|>\n<|im_start|>assistant\n",
+        ]
+        image1 = Image.open(requests.get("https://llava-vl.github.io/static/images/view.jpg", stream=True).raw)
+        image2 = Image.open(
+            BytesIO(
+                requests.get(
+                    "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
+                ).content
+            )
+        )
+        image3 = Image.open(
+            BytesIO(
+                requests.get(
+                    "https://thumbs.dreamstime.com/b/golden-gate-bridge-san-francisco-purple-flowers-california-echium-candicans-36805947.jpg"
+                ).content
+            )
+        )
+
+        inputs = processor(text=prompt, images=[[image1], [image2, image3]], padding=True, return_tensors="pt").to(
+            torch_device, dtype=torch.bfloat16
+        )
+
+        output = model.generate(**inputs, do_sample=False, max_new_tokens=25)
+
+        # Check first output
+        decoded_output = processor.decode(output[0], skip_special_tokens=True)
+        # Batching seems to alter the output slightly, but it is also the case in the original implementation. This seems to be expected: https://github.com/huggingface/transformers/issues/23017#issuecomment-1649630232
+        expected_output = "user\n\nWrite a haiku for this image\nassistant\nSilky lake,  \nWooden pier,  \nNature's peace."  # fmt: skip
+        self.assertEqual(
+            decoded_output,
+            expected_output,
+            f"Decoded output: {decoded_output}\nExpected output: {expected_output}",
+        )
+
+        # Check second output
+        decoded_output = processor.decode(output[1], skip_special_tokens=True)
+        expected_output = 'user\n\nWhat are the differences between these two images?\nassistant\nThe images show the Statue of Liberty and the Golden Gate Bridge from different angles. Here are the differences:\n\n1. **Angle'  # fmt: skip
+        self.assertEqual(
+            decoded_output,
+            expected_output,
+            f"Decoded output: {decoded_output}\nExpected output: {expected_output}",
+        )
+
+    @require_av
+    @require_bitsandbytes
+    def test_qwen2_medium_model_integration_video(self):
+        processor = AutoProcessor.from_pretrained(self.medium_model_checkpoint)
+        quantization_config = BitsAndBytesConfig(load_in_4bit=True)
+        model = InternVLForConditionalGeneration.from_pretrained(
+            self.medium_model_checkpoint, quantization_config=quantization_config
+        )
+        # Prepare inputs
+        messages = [
+            {
+                "role": "user",
+                "content": [
+                    {
+                        "type": "video",
+                        "url": "https://huggingface.co/datasets/hf-internal-testing/fixtures_videos/resolve/main/tennis.mp4",
+                    },
+                    {"type": "text", "text": "What type of shot is the man performing?"},
+                ],
+            }
+        ]
+        inputs = processor.apply_chat_template(
+            messages,
+            add_generation_prompt=True,
+            tokenize=True,
+            return_dict=True,
+            return_tensors="pt",
+        ).to(torch_device, dtype=torch.float16)
+
+        output = model.generate(**inputs, do_sample=False, max_new_tokens=25)
+
+        decoded_output = processor.decode(output[0, inputs["input_ids"].shape[1] :], skip_special_tokens=True)
+        expected_output = 'The man is performing a forehand shot.'  # fmt: skip
+        self.assertEqual(
+            decoded_output,
+            expected_output,
+            f"Decoded output: {decoded_output}\nExpected output: {expected_output}",
+        )
+
+    @require_av
+    def test_qwen2_small_model_integration_interleaved_images_videos(self):
+        processor = AutoProcessor.from_pretrained(self.small_model_checkpoint)
+        model = InternVLForConditionalGeneration.from_pretrained(
+            self.small_model_checkpoint, torch_dtype=torch.bfloat16, device_map=torch_device
+        )
+        messages = [
+            [
+                {
+                    "role": "user",
+                    "content": [
+                        {
+                            "type": "image",
+                            "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg",
+                        },
+                        {
+                            "type": "image",
+                            "url": "https://thumbs.dreamstime.com/b/golden-gate-bridge-san-francisco-purple-flowers-california-echium-candicans-36805947.jpg",
+                        },
+                        {"type": "text", "text": "What are the differences between these two images?"},
+                    ],
+                },
+            ],
+            [
+                {
+                    "role": "user",
+                    "content": [
+                        {
+                            "type": "video",
+                            "url": "https://huggingface.co/datasets/hf-internal-testing/fixtures_videos/resolve/main/tennis.mp4",
+                        },
+                        {"type": "text", "text": "What type of shot is the man performing?"},
+                    ],
+                },
+            ],
+            [
+                {
+                    "role": "user",
+                    "content": [
+                        {
+                            "type": "image",
+                            "url": "https://llava-vl.github.io/static/images/view.jpg",
+                        },
+                        {"type": "text", "text": "Write a haiku for this image"},
+                    ],
+                }
+            ],
+        ]
+        inputs = processor.apply_chat_template(
+            messages,
+            add_generation_prompt=True,
+            tokenize=True,
+            return_dict=True,
+            return_tensors="pt",
+            padding=True,
+        ).to(torch_device, dtype=torch.bfloat16)
+
+        output = model.generate(**inputs, do_sample=False, max_new_tokens=25)
+
+        decoded_output = processor.decode(output[0], skip_special_tokens=True)
+        # Batching seems to alter the output slightly, but it is also the case in the original implementation. This seems to be expected: https://github.com/huggingface/transformers/issues/23017#issuecomment-1649630232
+        expected_output = 'user\n\n\nWhat are the differences between these two images?\nassistant\nThe images depict two distinct scenes:\n\n1. **Left Image**: This shows the Statue of Liberty on Liberty Island, with the'  # fmt: skip
+        self.assertEqual(
+            decoded_output,
+            expected_output,
+            f"Decoded output: {decoded_output}\nExpected output: {expected_output}",
+        )
+        # Check second output
+        decoded_output = processor.decode(output[1], skip_special_tokens=True)
+        expected_output = 'user\nFrame1: \nFrame2: \nFrame3: \nFrame4: \nFrame5: \nFrame6: \nFrame7: \nFrame8: \nWhat type of shot is the man performing?\nassistant\nA forehand shot'  # fmt: skip
+        self.assertEqual(
+            decoded_output,
+            expected_output,
+            f"Decoded output: {decoded_output}\nExpected output: {expected_output}",
+        )
+
+        # Check third output
+        decoded_output = processor.decode(output[2], skip_special_tokens=True)
+        expected_output = "user\n\nWrite a haiku for this image\nassistant\nSilky lake,  \nWooden pier,  \nNature's peace."  # fmt: skip
+        self.assertEqual(
+            decoded_output,
+            expected_output,
+            f"Decoded output: {decoded_output}\nExpected output: {expected_output}",
+        )
+
+
+@slow
+@require_torch_gpu
+class InternVLLlamaIntegrationTest(unittest.TestCase):
+    def setUp(self):
+        self.small_model_checkpoint = "OpenGVLab/InternVL2_5-2B-MPO-hf"
+        self.medium_model_checkpoint = "OpenGVLab/InternVL2_5-8B-MPO-hf"
+
+    def tearDown(self):
+        cleanup(torch_device, gc_collect=True)
+
+    def test_llama_small_model_integration_generate(self):
+        processor = AutoProcessor.from_pretrained(self.small_model_checkpoint)
+        model = InternVLForConditionalGeneration.from_pretrained(
+            self.small_model_checkpoint, device_map=torch_device, torch_dtype=torch.bfloat16
+        )
+        url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        image = Image.open(requests.get(url, stream=True).raw)
+
+        prompt = "<|im_start|>user\n<image>\nPlease describe the image explicitly.<|im_end|>\n<|im_start|>assistant\n"
+        inputs = processor(images=image, text=prompt, return_tensors="pt").to(torch_device, dtype=torch.bfloat16)
+        with torch.no_grad():
+            generate_ids = model.generate(**inputs, max_new_tokens=20, do_sample=False)
+            decoded_output = processor.decode(
+                generate_ids[0, inputs["input_ids"].shape[1] :], skip_special_tokens=True
+            )
+        expected_output = "The image shows two cats sleeping on a pink couch. They are lying side by side, with their"
+        self.assertEqual(decoded_output, expected_output)
+
+    def test_llama_small_model_integration_forward(self):
+        processor = AutoProcessor.from_pretrained(self.small_model_checkpoint)
+        model = InternVLForConditionalGeneration.from_pretrained(
+            self.small_model_checkpoint, device_map=torch_device, torch_dtype=torch.bfloat16
+        )
+        url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        image = Image.open(requests.get(url, stream=True).raw)
+
+        prompt = "<|im_start|>user\n<image>\nPlease describe the image explicitly.<|im_end|>\n<|im_start|>assistant\n"
+        inputs = processor(images=image, text=prompt, return_tensors="pt").to(torch_device, dtype=torch.bfloat16)
+
+        # Forward
+        with torch.inference_mode():
+            output = model(**inputs)
+
+        actual_logits = output.logits[0, -1, :5].cpu()
+        expected_logits = torch.tensor([-9.8750, -0.4258, 1.4844, -10.3125, -10.3125], dtype=torch.bfloat16)
+        # The original implementation and the transformers implementation do not match exactly, hence the higher tolerance.
+        # The difference is likely due to the different implementations of the attention mechanism (different order of operations)
+        # between the transformers Llama model and the original InternLM model.
+        # The difference has almost no effect on the output tokens, but it does affect the logits a lot more.
+        self.assertTrue(
+            torch.allclose(actual_logits, expected_logits, atol=1),
+            f"Actual logits: {actual_logits}"
+            f"\nExpected logits: {expected_logits}"
+            f"\nDifference: {torch.abs(actual_logits - expected_logits)}",
+        )
+
+    def test_llama_small_model_integration_generate_text_only(self):
+        processor = AutoProcessor.from_pretrained(self.small_model_checkpoint)
+        model = InternVLForConditionalGeneration.from_pretrained(
+            self.small_model_checkpoint, device_map=torch_device, torch_dtype=torch.bfloat16
+        )
+        prompt = "<|im_start|>user\nWrite a haiku<|im_end|>\n<|im_start|>assistant\n"
+        inputs = processor(text=prompt, return_tensors="pt").to(torch_device, dtype=torch.bfloat16)
+        with torch.no_grad():
+            generate_ids = model.generate(**inputs, max_new_tokens=200, do_sample=False)
+            decoded_output = processor.decode(
+                generate_ids[0, inputs["input_ids"].shape[1] :], skip_special_tokens=True
+            )
+        expected_output = "Autumn leaves fall,\nNature's breath, a season's sigh,\nSilent woods awake."
+        self.assertEqual(decoded_output, expected_output)
+
+    def test_llama_small_model_integration_generate_chat_template(self):
+        processor = AutoProcessor.from_pretrained(self.small_model_checkpoint)
+        model = InternVLForConditionalGeneration.from_pretrained(
+            self.small_model_checkpoint, device_map=torch_device, torch_dtype=torch.bfloat16
+        )
+        messages = [
+            {
+                "role": "user",
+                "content": [
+                    {"type": "image", "url": "http://images.cocodataset.org/val2017/000000039769.jpg"},
+                    {"type": "text", "text": "Please describe the image explicitly."},
+                ],
+            }
+        ]
+
+        inputs = processor.apply_chat_template(
+            messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt"
+        ).to(torch_device, dtype=torch.bfloat16)
+        with torch.no_grad():
+            generate_ids = model.generate(**inputs, max_new_tokens=20, do_sample=False)
+            decoded_output = processor.decode(
+                generate_ids[0, inputs["input_ids"].shape[1] :], skip_special_tokens=True
+            )
+        expected_output = "The image shows two cats sleeping on a pink couch. They are lying side by side, with their"
+        self.assertEqual(decoded_output, expected_output)
+
+    def test_llama_small_model_integration_batched_generate(self):
+        processor = AutoProcessor.from_pretrained(self.small_model_checkpoint)
+        model = InternVLForConditionalGeneration.from_pretrained(
+            self.small_model_checkpoint, device_map=torch_device, torch_dtype=torch.bfloat16
+        )
+        # Prepare inputs
+        prompt = [
+            "<|im_start|>user\n<image>\nWrite a haiku for this image<|im_end|>\n<|im_start|>assistant\n",
+            "<|im_start|>user\n<image>\nDescribe this image<|im_end|>\n<|im_start|>assistant\n",
+        ]
+        image1 = Image.open(requests.get("https://llava-vl.github.io/static/images/view.jpg", stream=True).raw)
+        image2 = Image.open(requests.get("https://www.ilankelman.org/stopsigns/australia.jpg", stream=True).raw)
+
+        inputs = processor(text=prompt, images=[[image1], [image2]], padding=True, return_tensors="pt").to(
+            torch_device, dtype=torch.bfloat16
+        )
+
+        output = model.generate(**inputs, do_sample=False, max_new_tokens=25)
+
+        # Check first output
+        decoded_output = processor.decode(output[0], skip_special_tokens=True)
+        expected_output = 'user\n\nWrite a haiku for this image\nassistant\nMajestic snow-capped peaks,\nWooden dock stretches to the sea,\nSilent water mirrors.'  # fmt: skip
+        self.assertEqual(
+            decoded_output,
+            expected_output,
+            f"Decoded output: {decoded_output}\nExpected output: {expected_output}",
+        )
+
+        # Check second output
+        decoded_output = processor.decode(output[1], skip_special_tokens=True)
+        expected_output = 'user\n\nDescribe this image\nassistant\nThe image shows a street scene with a traditional Chinese gate in the background, adorned with red and gold colors and Chinese characters'  # fmt: skip
+        self.assertEqual(
+            decoded_output,
+            expected_output,
+            f"Decoded output: {decoded_output}\nExpected output: {expected_output}",
+        )
+
+    def test_llama_small_model_integration_batched_generate_multi_image(self):
+        processor = AutoProcessor.from_pretrained(self.small_model_checkpoint)
+        model = InternVLForConditionalGeneration.from_pretrained(
+            self.small_model_checkpoint, device_map=torch_device, torch_dtype=torch.bfloat16
+        )
+        # Prepare inputs
+        prompt = [
+            "<|im_start|>user\n<image>\nWrite a haiku for this image<|im_end|>\n<|im_start|>assistant\n",
+            "<|im_start|>user\n<image><image>\nWhat are the difference between these two images?<|im_end|>\n<|im_start|>assistant\n",
+        ]
+        image1 = Image.open(requests.get("https://llava-vl.github.io/static/images/view.jpg", stream=True).raw)
+        image2 = Image.open(
+            BytesIO(
+                requests.get(
+                    "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
+                ).content
+            )
+        )
+        image3 = Image.open(
+            BytesIO(
+                requests.get(
+                    "https://thumbs.dreamstime.com/b/golden-gate-bridge-san-francisco-purple-flowers-california-echium-candicans-36805947.jpg"
+                ).content
+            )
+        )
+
+        inputs = processor(text=prompt, images=[[image1], [image2, image3]], padding=True, return_tensors="pt").to(
+            torch_device, dtype=torch.bfloat16
+        )
+
+        output = model.generate(**inputs, do_sample=False, max_new_tokens=25)
+
+        # Check first output
+        decoded_output = processor.decode(output[0], skip_special_tokens=True)
+        # Batching seems to alter the output slightly, but it is also the case in the original implementation. This seems to be expected: https://github.com/huggingface/transformers/issues/23017#issuecomment-1649630232
+        expected_output = 'user\n\nWrite a haiku for this image\nassistant\nMajestic snow-capped peaks,\nA wooden path leads to the sea,\nPeaceful, still waters.'  # fmt: skip
+        self.assertEqual(
+            decoded_output,
+            expected_output,
+            f"Decoded output: {decoded_output}\nExpected output: {expected_output}",
+        )
+
+        # Check second output
+        decoded_output = processor.decode(output[1], skip_special_tokens=True)
+        expected_output = 'user\n\nWhat are the difference between these two images?\nassistant\nI apologize for the confusion in my previous response. After closely examining the images again, I can see that there are several differences'  # fmt: skip
+        self.assertEqual(
+            decoded_output,
+            expected_output,
+            f"Decoded output: {decoded_output}\nExpected output: {expected_output}",
+        )
+
+    @require_av
+    @require_bitsandbytes
+    def test_llama_medium_model_integration_video(self):
+        processor = AutoProcessor.from_pretrained(self.medium_model_checkpoint)
+        quantization_config = BitsAndBytesConfig(load_in_4bit=True)
+        model = InternVLForConditionalGeneration.from_pretrained(
+            self.medium_model_checkpoint, quantization_config=quantization_config
+        )
+        # Prepare inputs
+        messages = [
+            {
+                "role": "user",
+                "content": [
+                    {
+                        "type": "video",
+                        "url": "https://huggingface.co/datasets/hf-internal-testing/fixtures_videos/resolve/main/tennis.mp4",
+                    },
+                    {"type": "text", "text": "What type of shot is the man performing?"},
+                ],
+            }
+        ]
+        inputs = processor.apply_chat_template(
+            messages,
+            add_generation_prompt=True,
+            tokenize=True,
+            return_dict=True,
+            return_tensors="pt",
+        ).to(torch_device, dtype=torch.float16)
+
+        output = model.generate(**inputs, do_sample=False, max_new_tokens=25)
+
+        decoded_output = processor.decode(output[0, inputs["input_ids"].shape[1] :], skip_special_tokens=True)
+        expected_output = "The man is performing a forehand shot."
+        self.assertEqual(
+            decoded_output,
+            expected_output,
+            f"Decoded output: {decoded_output}\nExpected output: {expected_output}",
+        )
+
+    @require_av
+    def test_llama_small_model_integration_interleaved_images_videos(self):
+        processor = AutoProcessor.from_pretrained(self.small_model_checkpoint)
+        model = InternVLForConditionalGeneration.from_pretrained(
+            self.small_model_checkpoint, torch_dtype=torch.bfloat16, device_map=torch_device
+        )
+        messages = [
+            [
+                {
+                    "role": "user",
+                    "content": [
+                        {
+                            "type": "image",
+                            "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg",
+                        },
+                        {
+                            "type": "image",
+                            "url": "https://thumbs.dreamstime.com/b/golden-gate-bridge-san-francisco-purple-flowers-california-echium-candicans-36805947.jpg",
+                        },
+                        {"type": "text", "text": "What are the difference between these two images?"},
+                    ],
+                },
+            ],
+            [
+                {
+                    "role": "user",
+                    "content": [
+                        {
+                            "type": "video",
+                            "url": "https://huggingface.co/datasets/hf-internal-testing/fixtures_videos/resolve/main/tennis.mp4",
+                        },
+                        {"type": "text", "text": "What type of shot is the man performing?"},
+                    ],
+                },
+            ],
+            [
+                {
+                    "role": "user",
+                    "content": [
+                        {
+                            "type": "image",
+                            "url": "https://llava-vl.github.io/static/images/view.jpg",
+                        },
+                        {"type": "text", "text": "Write a haiku for this image"},
+                    ],
+                }
+            ],
+        ]
+        inputs = processor.apply_chat_template(
+            messages,
+            add_generation_prompt=True,
+            tokenize=True,
+            return_dict=True,
+            return_tensors="pt",
+            padding=True,
+        ).to(torch_device, dtype=torch.bfloat16)
+
+        output = model.generate(**inputs, do_sample=False, max_new_tokens=25)
+
+        decoded_output = processor.decode(output[0], skip_special_tokens=True)
+        # Batching seems to alter the output slightly, but it is also the case in the original implementation. This seems to be expected: https://github.com/huggingface/transformers/issues/23017#issuecomment-1649630232
+        expected_output = 'user\n\n\nWhat are the difference between these two images?\nassistant\nI apologize for the confusion in my previous response. Upon closer inspection, the differences between the two images are:\n\n1. **'  # fmt: skip
+        self.assertEqual(
+            decoded_output,
+            expected_output,
+            f"Decoded output: {decoded_output}\nExpected output: {expected_output}",
+        )
+
+        # Check second output
+        decoded_output = processor.decode(output[1], skip_special_tokens=True)
+        expected_output = 'user\nFrame1: \nFrame2: \nFrame3: \nFrame4: \nFrame5: \nFrame6: \nFrame7: \nFrame8: \nWhat type of shot is the man performing?\nassistant\nThe man is performing a forehand shot. This is a common shot in tennis where the player swings the racket across their'  # fmt: skip
+        self.assertEqual(
+            decoded_output,
+            expected_output,
+            f"Decoded output: {decoded_output}\nExpected output: {expected_output}",
+        )
+
+        # Check third output
+        decoded_output = processor.decode(output[2], skip_special_tokens=True)
+        expected_output = 'user\n\nWrite a haiku for this image\nassistant\nMajestic snow-capped peaks,\nA wooden path leads to the sea,\nPeaceful, untouched dreams.'  # fmt: skip
+        self.assertEqual(
+            decoded_output,
+            expected_output,
+            f"Decoded output: {decoded_output}\nExpected output: {expected_output}",
+        )

tests/models/internvl/test_processor_internvl.py

@@ -0,0 +1,327 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+import shutil
+import tempfile
+import unittest
+
+from huggingface_hub import hf_hub_download
+
+from transformers import AutoProcessor, AutoTokenizer, InternVLProcessor
+from transformers.testing_utils import require_av, require_torch, require_vision
+from transformers.utils import is_torch_available, is_vision_available
+
+from ...test_processing_common import ProcessorTesterMixin
+
+
+if is_torch_available():
+    import torch
+
+
+if is_vision_available():
+    from transformers import GotOcr2ImageProcessor
+
+
+@require_vision
+class InternVLProcessorTest(ProcessorTesterMixin, unittest.TestCase):
+    processor_class = InternVLProcessor
+    videos_input_name = "pixel_values"
+
+    @classmethod
+    def setUpClass(cls):
+        cls.tmpdirname = tempfile.mkdtemp()
+
+        image_processor = GotOcr2ImageProcessor(
+            do_resize=True,
+            size={"height": 20, "width": 20},
+            max_patches=2,
+            do_rescale=True,
+            rescale_factor=1 / 255,
+            do_normalize=True,
+            do_center_crop=True,
+            image_mean=[0.485, 0.456, 0.406],
+            image_std=[0.229, 0.224, 0.225],
+            do_convert_rgb=True,
+        )
+        tokenizer = AutoTokenizer.from_pretrained("OpenGVLab/InternVL3-1B-hf", padding_side="left")
+        processor_kwargs = cls.prepare_processor_dict()
+        processor = InternVLProcessor.from_pretrained(
+            "OpenGVLab/InternVL3-1B-hf",
+            image_processor=image_processor,
+            tokenizer=tokenizer,
+            **processor_kwargs,
+        )
+        processor.save_pretrained(cls.tmpdirname)
+        cls.image_token = processor.fake_image_token
+
+    @staticmethod
+    def prepare_processor_dict():
+        return {"image_seq_length": 10}
+
+    def get_tokenizer(self, **kwargs):
+        return AutoProcessor.from_pretrained(self.tmpdirname, **kwargs).tokenizer
+
+    def get_image_processor(self, **kwargs):
+        return AutoProcessor.from_pretrained(self.tmpdirname, **kwargs).image_processor
+
+    def get_processor(self, **kwargs):
+        return AutoProcessor.from_pretrained(self.tmpdirname, **kwargs)
+
+    @classmethod
+    def tearDownClass(cls):
+        shutil.rmtree(cls.tmpdirname, ignore_errors=True)
+
+    @require_av
+    @require_torch
+    def test_process_interleaved_images_videos(self):
+        processor = self.get_processor()
+
+        messages = [
+            [
+                {
+                    "role": "user",
+                    "content": [
+                        {
+                            "type": "image",
+                            "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg",
+                        },
+                        {
+                            "type": "image",
+                            "url": "https://thumbs.dreamstime.com/b/golden-gate-bridge-san-francisco-purple-flowers-california-echium-candicans-36805947.jpg",
+                        },
+                        {"type": "text", "text": "What are the differences between these two images?"},
+                    ],
+                },
+            ],
+            [
+                {
+                    "role": "user",
+                    "content": [
+                        {
+                            "type": "video",
+                            "url": "https://huggingface.co/datasets/hf-internal-testing/fixtures_videos/resolve/main/tennis.mp4",
+                        },
+                        {"type": "text", "text": "What type of shot is the man performing?"},
+                    ],
+                },
+            ],
+            [
+                {
+                    "role": "user",
+                    "content": [
+                        {
+                            "type": "image",
+                            "url": "https://llava-vl.github.io/static/images/view.jpg",
+                        },
+                        {"type": "text", "text": "Write a haiku for this image"},
+                    ],
+                }
+            ],
+        ]
+
+        inputs_batched = processor.apply_chat_template(
+            messages,
+            add_generation_prompt=True,
+            tokenize=True,
+            return_dict=True,
+            return_tensors="pt",
+            padding=True,
+        )
+
+        # Process non batched inputs to check if the pixel_values and input_ids are reconstructed in the correct order when batched together
+        images_patches_index = 0
+        for i, message in enumerate(messages):
+            inputs = processor.apply_chat_template(
+                message,
+                add_generation_prompt=True,
+                tokenize=True,
+                return_dict=True,
+                return_tensors="pt",
+                padding=True,
+            )
+            # We slice with [-inputs["input_ids"].shape[1] :] as the input_ids are left padded
+            torch.testing.assert_close(
+                inputs["input_ids"][0], inputs_batched["input_ids"][i][-inputs["input_ids"].shape[1] :]
+            )
+            torch.testing.assert_close(
+                inputs["pixel_values"],
+                inputs_batched["pixel_values"][
+                    images_patches_index : images_patches_index + inputs["pixel_values"].shape[0]
+                ],
+            )
+            images_patches_index += inputs["pixel_values"].shape[0]
+
+    # Override video chat_template tests as InternVLProcessor returns flattened video features
+    @require_av
+    def test_apply_chat_template_video_special_processing(self):
+        """
+        Tests that models can use their own preprocessing to preprocess conversations.
+        """
+        processor = self.get_processor()
+        if processor.chat_template is None:
+            self.skipTest("Processor has no chat template")
+
+        signature = inspect.signature(processor.__call__)
+        if "videos" not in {*signature.parameters.keys()} or (
+            signature.parameters.get("videos") is not None
+            and signature.parameters["videos"].annotation == inspect._empty
+        ):
+            self.skipTest("Processor doesn't accept videos at input")
+
+        video_file_path = hf_hub_download(
+            repo_id="raushan-testing-hf/videos-test", filename="sample_demo_1.mp4", repo_type="dataset"
+        )
+        messages = [
+            [
+                {
+                    "role": "user",
+                    "content": [
+                        {"type": "video", "path": video_file_path},
+                        {"type": "text", "text": "What is shown in this video?"},
+                    ],
+                },
+            ]
+        ]
+
+        def _process_messages_for_chat_template(
+            conversation,
+            batch_images,
+            batch_videos,
+            batch_video_metadata,
+            **chat_template_kwargs,
+        ):
+            # Let us just always return a dummy prompt
+            new_msg = [
+                [
+                    {
+                        "role": "user",
+                        "content": [
+                            {"type": "video"},  # no need to use path, video is loaded already by this moment
+                            {"type": "text", "text": "Dummy prompt for preprocess testing"},
+                        ],
+                    },
+                ]
+            ]
+            return new_msg
+
+        processor._process_messages_for_chat_template = _process_messages_for_chat_template
+        out_dict_with_video = processor.apply_chat_template(
+            messages,
+            add_generation_prompt=True,
+            tokenize=True,
+            return_dict=True,
+            return_tensors="np",
+        )
+        self.assertTrue(self.videos_input_name in out_dict_with_video)
+
+        # Check with `in` because we don't know how each template formats the prompt with BOS/EOS/etc
+        formatted_text = processor.batch_decode(out_dict_with_video["input_ids"], skip_special_tokens=True)[0]
+        self.assertTrue("Dummy prompt for preprocess testing" in formatted_text)
+        # Difference with common tests, InternVLProcessor returns flattened video features, and uses 8 frames by default
+        self.assertEqual(len(out_dict_with_video[self.videos_input_name]), 8)
+
+    def test_apply_chat_template_video_frame_sampling(self):
+        processor = self.get_processor()
+
+        if processor.chat_template is None:
+            self.skipTest("Processor has no chat template")
+
+        signature = inspect.signature(processor.__call__)
+        if "videos" not in {*signature.parameters.keys()} or (
+            signature.parameters.get("videos") is not None
+            and signature.parameters["videos"].annotation == inspect._empty
+        ):
+            self.skipTest("Processor doesn't accept videos at input")
+
+        messages = [
+            [
+                {
+                    "role": "user",
+                    "content": [
+                        {
+                            "type": "video",
+                            "url": "https://test-videos.co.uk/vids/bigbuckbunny/mp4/h264/720/Big_Buck_Bunny_720_10s_10MB.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,
+            return_tensors="np",
+        )
+        self.assertTrue(self.videos_input_name in out_dict_with_video)
+        self.assertEqual(len(out_dict_with_video[self.videos_input_name]), num_frames)
+
+        # Load with `video_fps` arg
+        video_fps = 1
+        out_dict_with_video = processor.apply_chat_template(
+            messages,
+            add_generation_prompt=True,
+            tokenize=True,
+            return_dict=True,
+            video_fps=video_fps,
+            num_frames=None,  # force to use default num_frames
+            return_tensors="np",
+        )
+        self.assertTrue(self.videos_input_name in out_dict_with_video)
+        self.assertEqual(len(out_dict_with_video[self.videos_input_name]), video_fps * 10)
+
+        # Load with `video_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,
+                video_fps=video_fps,
+                num_frames=num_frames,
+            )
+
+        # Load without any arg should use the default loading method
+        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)
+        # Difference with common tests, InternVLProcessor returns flattened video features, and uses 8 frames by default
+        self.assertEqual(len(out_dict_with_video[self.videos_input_name]), 8)
+
+        # Load video as a list of frames (i.e. images). NOTE: each frame should have same size
+        # because we assume they come from one video
+        messages[0][0]["content"][0] = {
+            "type": "video",
+            "url": [
+                "https://www.ilankelman.org/stopsigns/australia.jpg",
+                "https://www.ilankelman.org/stopsigns/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]), 2)