[VLM] Merged multi-modal processors for LLaVA-NeXT-Video and LLaVA-OneVision (#11717)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>

tests/models/decoder_only/vision_language/processing/test_llava_next.py

+import pytest
+from PIL import Image
+from transformers import AutoTokenizer
+
+from vllm.inputs import InputProcessingContext
+
+from ....utils import build_model_context
+
+
+# Fixtures lazy import to avoid initializing CUDA during test collection
+@pytest.fixture()
+def processor_for_llava_next():
+    from vllm.model_executor.models.llava_next import (
+        LlavaNextMultiModalProcessor)
+    return LlavaNextMultiModalProcessor
+
+
+# FIXME: image_size [(198, 176), (176, 198)]
+@pytest.mark.parametrize("model_id", ["llava-hf/llava-v1.6-mistral-7b-hf"])
+@pytest.mark.parametrize("image_size", [(1669, 2560), (2560, 1669), (183, 488),
+                                        (488, 183)])
+@pytest.mark.parametrize("num_imgs", [1, 2])
+def test_processor_prompt_replacements(
+    processor_for_llava_next,
+    model_id: str,
+    image_size: tuple[int, int],
+    num_imgs: int,
+):
+    """
+    Ensure LlavaNextMultiModalProcessor handles prompt replacement properly.
+    """
+    ctx = build_model_context(
+        model_name=model_id,
+        tokenizer_name=model_id,
+        mm_processor_kwargs=None,
+        limit_mm_per_prompt={"image": num_imgs},
+    )
+    tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
+    ctx = InputProcessingContext(ctx.model_config, tokenizer)
+
+    # Build the image str / prompt based on the number of images we pass
+    prompt = "<image>" * num_imgs
+    mm_data = {"image": [Image.new("RGB", size=image_size)] * num_imgs}
+
+    # The processor will throw an error if there is a mismatch
+    # in the prompt replacements
+    processor = processor_for_llava_next(ctx)
+    processed_inputs = processor.apply(prompt, mm_data, {})
+
+    image_placeholders = processed_inputs["mm_placeholders"]["image"]
+    assert len(image_placeholders) == num_imgs
+
+    first_placeholder = image_placeholders[0]
+
+    # NOTE: There is a BOS token
+    assert first_placeholder["offset"] == 1
+    assert first_placeholder["length"] == (
+        len(processed_inputs["prompt_token_ids"]) - 1) // num_imgs

tests/models/decoder_only/vision_language/processing/test_llava_onevision.py

+import pytest
+from PIL import Image
+from transformers import AutoTokenizer
+
+from vllm.inputs import InputProcessingContext
+
+from ....utils import build_model_context
+
+
+# Fixtures lazy import to avoid initializing CUDA during test collection
+@pytest.fixture()
+def processor_for_llava_onevision():
+    from vllm.model_executor.models.llava_onevision import (
+        LlavaOnevisionMultiModalProcessor)
+    return LlavaOnevisionMultiModalProcessor
+
+
+@pytest.mark.parametrize("model_id",
+                         ["llava-hf/llava-onevision-qwen2-0.5b-ov-hf"])
+@pytest.mark.parametrize("image_size", [(1669, 2560), (2560, 1669), (183, 488),
+                                        (488, 183), (198, 176), (176, 198)])
+@pytest.mark.parametrize("num_imgs", [1, 2])
+def test_processor_prompt_replacements(
+    processor_for_llava_onevision,
+    model_id: str,
+    image_size: tuple[int, int],
+    num_imgs: int,
+):
+    """
+    Ensure LlavaOnevisionMultiModalProcessor handles prompt replacement
+    properly.
+    """
+    ctx = build_model_context(
+        model_name=model_id,
+        tokenizer_name=model_id,
+        mm_processor_kwargs=None,
+        limit_mm_per_prompt={"image": num_imgs},
+    )
+    tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
+    ctx = InputProcessingContext(ctx.model_config, tokenizer)
+
+    # Build the image str / prompt based on the number of images we pass
+    prompt = "<image>" * num_imgs
+    mm_data = {"image": [Image.new("RGB", size=image_size)] * num_imgs}
+
+    # The processor will throw an error if there is a mismatch
+    # in the prompt replacements
+    processor = processor_for_llava_onevision(ctx)
+    processed_inputs = processor.apply(prompt, mm_data, {})
+
+    image_placeholders = processed_inputs["mm_placeholders"]["image"]
+    assert len(image_placeholders) == num_imgs
+
+    first_placeholder = image_placeholders[0]
+
+    # NOTE: There is a BOS token
+    assert first_placeholder["offset"] == 0
+    assert first_placeholder["length"] == len(
+        processed_inputs["prompt_token_ids"]) // num_imgs

tests/models/decoder_only/vision_language/mm_processor_kwargs/test_phi3v.py → tests/models/decoder_only/vision_language/processing/test_phi3v.py

 """Tests for phi3v's multimodal preprocessing kwargs."""
-from typing import Optional
-
 import pytest
 from transformers import AutoTokenizer
 
@@ -10,8 +8,6 @@ from vllm.model_executor.models.phi3v import _IMAGE_TOKEN_ID
 from .....conftest import _ImageAssets
 from ....utils import build_model_context
 
-models = ["microsoft/Phi-3.5-vision-instruct"]
-
 
 # Wrap lazy imports to avoid initializing CUDA during test collection
 @pytest.fixture()
@@ -20,40 +16,40 @@ def processor_for_phi3v():
     return Phi3VMultiModalProcessor
 
 
-@pytest.mark.parametrize("model", models)
+@pytest.mark.parametrize("model_id", ["microsoft/Phi-3.5-vision-instruct"])
+# yapf: disable
 @pytest.mark.parametrize(
-    "num_crops,expected_toks_per_img",
+    ("mm_processor_kwargs", "expected_toks_per_img"),
     [
-        (4, 757),
-        (16, 1921),
+        ({"num_crops": 4}, 757),
+        ({"num_crops": 16}, 1921),
         # the default num_crops of phi-3.5-vision is 4
-        (None, 757),
+        ({}, 757),
     ])
+# yapf: enable
 @pytest.mark.parametrize("num_imgs", [1, 2])
-def test_processor_override(processor_for_phi3v, image_assets: _ImageAssets,
-                            model: str, num_crops: Optional[int],
-                            expected_toks_per_img: int, num_imgs: int):
+def test_processor_override(
+    processor_for_phi3v,
+    image_assets: _ImageAssets,
+    model_id: str,
+    mm_processor_kwargs: dict[str, int],
+    expected_toks_per_img: int,
+    num_imgs: int,
+):
     """Ensure input_processor_for_phi3v handles num_crops properly."""
-    # Same as the previous test - don't initialize mm_processor_kwargs
-    # in this test and assume that the kwargs will be correctly expanded by
-    # the partial when calling the custom input processor.
     ctx = build_model_context(
-        model_name=model,
-        tokenizer_name=model,
+        model_name=model_id,
+        tokenizer_name=model_id,
         trust_remote_code=True,
         limit_mm_per_prompt={"image": num_imgs},
     )
-    tokenizer = AutoTokenizer.from_pretrained(model, trust_remote_code=True)
+    tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
     ctx = InputProcessingContext(ctx.model_config, tokenizer)
+
     # Build the image str / prompt based on the number of images we pass
     img_str = "".join([f"<|image_{idx}|>\n" for idx in range(1, num_imgs + 1)])
     prompt = f"<|user|>\n{img_str}<|end|>\n<|assistant|>\n"
-    images = [image_assets[0].pil_image] * num_imgs
-
-    mm_data = {"image": images}
-    mm_processor_kwargs = {}
-    if num_crops is not None:
-        mm_processor_kwargs = {"num_crops": num_crops}
+    mm_data = {"image": [image_assets[0].pil_image] * num_imgs}
 
     processor = processor_for_phi3v(ctx)
     processed_inputs = processor.apply(prompt, mm_data, mm_processor_kwargs)

tests/models/decoder_only/vision_language/mm_processor_kwargs/test_qwen2_vl.py → tests/models/decoder_only/vision_language/processing/test_qwen2_vl.py

-from typing import Any, Dict, Tuple
-
 import pytest
 from transformers import AutoTokenizer
 
@@ -8,56 +6,45 @@ from vllm.inputs import InputProcessingContext
 from .....conftest import _ImageAssets
 from ....utils import build_model_context
 
-MODEL = "Qwen/Qwen2-VL-2B-Instruct"
-MIN_PIXELS = "min_pixels"
-MAX_PIXELS = "max_pixels"
-
 
 # Fixtures lazy import to avoid initializing CUDA during test collection
-# NOTE: Qwen2VL supports multiple input modalities, so it registers multiple
-# input mappers.
 @pytest.fixture()
 def processor_for_qwen2_vl():
     from vllm.model_executor.models.qwen2_vl import Qwen2VLMultiModalProcessor
     return Qwen2VLMultiModalProcessor
 
 
+@pytest.mark.parametrize("model_id", ["Qwen/Qwen2-VL-2B-Instruct"])
+# yapf: disable
 @pytest.mark.parametrize(
-    "mm_processor_kwargs, expected_toks_per_img, expected_pixels_shape", [
+    ("mm_processor_kwargs", "expected_toks_per_img", "expected_pixels_shape"), [
         ({}, 1426, (5704, 1176)),
-        ({
-            MIN_PIXELS: 64**2,
-            MAX_PIXELS: 512**2
-        }, 330, (1320, 1176)),
+        ({"min_pixels": 64**2, "max_pixels": 512**2}, 330, (1320, 1176)),
     ])
-@pytest.mark.parametrize("model", [MODEL])
+# yapf: enable
 @pytest.mark.parametrize("num_imgs", [1, 2])
 def test_processor_override(
     processor_for_qwen2_vl,
     image_assets: _ImageAssets,
-    model: str,
-    mm_processor_kwargs: Dict[str, Any],
+    model_id: str,
+    mm_processor_kwargs: dict[str, object],
     expected_toks_per_img: int,
-    expected_pixels_shape: Tuple[int, int],
+    expected_pixels_shape: tuple[int, int],
     num_imgs: int,
 ):
     """Ensure Qwen2VLMultiModalProcessor handles min/max pixels properly."""
-    # Same as the previous test - don't initialize mm_processor_kwargs
-    # in this test and assume that the kwargs will be correctly expanded by
-    # the partial when calling the custom input processor.
     ctx = build_model_context(
-        model_name=model,
-        tokenizer_name=model,
+        model_name=model_id,
+        tokenizer_name=model_id,
         mm_processor_kwargs=None,
         limit_mm_per_prompt={"image": num_imgs},
     )
-    tokenizer = AutoTokenizer.from_pretrained(model, trust_remote_code=True)
+    tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
     ctx = InputProcessingContext(ctx.model_config, tokenizer)
+
     # Build the image str / prompt based on the number of images we pass
     prompt = "<|vision_start|><|image_pad|><|vision_end|>" * num_imgs
-    images = [image_assets[0].pil_image] * num_imgs
-
-    mm_data = {"image": images}
+    mm_data = {"image": [image_assets[0].pil_image] * num_imgs}
 
     processor = processor_for_qwen2_vl(ctx)
     processed_inputs = processor.apply(prompt, mm_data, mm_processor_kwargs)

tests/models/decoder_only/vision_language/test_models.py

@@ -274,10 +274,8 @@ VLM_TEST_SETTINGS = {
             ),
             limit_mm_per_prompt={"image": 4},
         )],
-        # Llava-next tests fixed sizes & the default size factors
-        image_sizes=[((1669, 2560), (2560, 1669), (183, 488), (488, 183))],
     ),
-    "llava_one_vision": VLMTestInfo(
+    "llava_onevision": VLMTestInfo(
         models=["llava-hf/llava-onevision-qwen2-0.5b-ov-hf"],
         test_type=VLMTestType.CUSTOM_INPUTS,
         prompt_formatter=lambda vid_prompt: f"<|im_start|>user\n{vid_prompt}<|im_end|>\n<|im_start|>assistant\n",   # noqa: E501
@@ -288,8 +286,6 @@ VLM_TEST_SETTINGS = {
         ),
         auto_cls=AutoModelForVision2Seq,
         vllm_output_post_proc=model_utils.llava_onevision_vllm_to_hf_output,
-        # Llava-one-vision tests fixed sizes & the default size factors
-        image_sizes=[((1669, 2560), (2560, 1669), (183, 488), (488, 183))],
         custom_test_opts=[CustomTestOptions(
             inputs=custom_inputs.multi_video_multi_aspect_ratio_inputs(
                 formatter=lambda vid_prompt: f"<|im_start|>user\n{vid_prompt}<|im_end|>\n<|im_start|>assistant\n",   # noqa: E501
@@ -306,7 +302,6 @@ VLM_TEST_SETTINGS = {
         max_model_len=4096,
         auto_cls=AutoModelForVision2Seq,
         vllm_output_post_proc=model_utils.llava_video_vllm_to_hf_output,
-        image_sizes=[((1669, 2560), (2560, 1669), (183, 488), (488, 183))],
     ),
     "mantis": VLMTestInfo(
         models=["TIGER-Lab/Mantis-8B-siglip-llama3"],
@@ -431,7 +426,7 @@ VLM_TEST_SETTINGS = {
             ) for inp in custom_inputs.different_patch_input_cases_internvl()
         ],
     ),
-    "llava_one_vision-multiple-images": VLMTestInfo(
+    "llava_onevision-multiple-images": VLMTestInfo(
         models=["llava-hf/llava-onevision-qwen2-0.5b-ov-hf"],
         test_type=VLMTestType.CUSTOM_INPUTS,
         max_model_len=16384,

tests/models/decoder_only/vision_language/test_qwen2_vl.py

@@ -427,130 +427,3 @@ def test_qwen2_vl_video_embeddings_input(vllm_runner, video_assets, model,
         mm_limit=1,
         tensor_parallel_size=1,
     )
-
-
-def run_chunked_prefill_test(
-    vllm_runner: Type[VllmRunner],
-    inputs: List[Tuple[List[str], PromptImageInput, PromptVideoInput]],
-    model: str,
-    *,
-    dtype: str,
-    max_tokens: int,
-    num_logprobs: int,
-    mm_limit: int,
-    tensor_parallel_size: int,
-    distributed_executor_backend: Optional[str] = None,
-):
-    """Compare inference result between
-    chunked prefill disabled and chunked prefill enabled
-    """
-
-    # NOTE:
-    # max_model_len should be greater than image_feature_size
-    with vllm_runner(model,
-                     task="generate",
-                     max_model_len=4000,
-                     max_num_seqs=4,
-                     dtype=dtype,
-                     limit_mm_per_prompt={
-                         "image": mm_limit,
-                         "video": mm_limit
-                     },
-                     tensor_parallel_size=tensor_parallel_size,
-                     distributed_executor_backend=distributed_executor_backend
-                     ) as vllm_model:
-
-        outputs_per_case = [
-            vllm_model.generate_greedy_logprobs(prompts,
-                                                max_tokens,
-                                                num_logprobs=num_logprobs,
-                                                images=images or None,
-                                                videos=videos or None)
-            for prompts, images, videos in inputs
-        ]
-
-    with vllm_runner(
-            model,
-            task="generate",
-            max_model_len=4000,
-            max_num_seqs=4,
-            dtype=dtype,
-            limit_mm_per_prompt={
-                "image": mm_limit,
-                "video": mm_limit
-            },
-            tensor_parallel_size=tensor_parallel_size,
-            distributed_executor_backend=distributed_executor_backend,
-            enable_chunked_prefill=True,
-            # should be small enough to ensure prefilling is chunked
-            max_num_batched_tokens=32,
-            mm_processor_kwargs={
-                "max_pixels": 16 * 28 * 28,
-            }) as vllm_model_chunked:
-        outputs_per_case_chunked = [
-            vllm_model_chunked.generate_greedy_logprobs(
-                prompts,
-                max_tokens,
-                num_logprobs=num_logprobs,
-                images=images or None,
-                videos=videos or None) for prompts, images, videos in inputs
-        ]
-
-    for outputs, \
-        outputs_chunked \
-        in zip(outputs_per_case,
-            outputs_per_case_chunked):
-        check_logprobs_close(
-            outputs_0_lst=outputs,
-            outputs_1_lst=outputs_chunked,
-            name_0="non_chunked",
-            name_1="chunked",
-        )
-
-
-@pytest.mark.core_model
-@pytest.mark.parametrize("model", models)
-@pytest.mark.parametrize("dtype", [target_dtype])
-@pytest.mark.parametrize("max_tokens", [1])
-@pytest.mark.parametrize("num_logprobs", [10])
-def test_qwen2_vl_mrope_chunked_prefill(vllm_runner, example_prompts,
-                                        model: str, dtype: str,
-                                        max_tokens: int,
-                                        num_logprobs: int) -> None:
-    """
-    Test Qwen2-VL's chunked prefill with M-RoPE
-    """
-    prompts = [
-        qwen2_vl_chat_template(IMAGE_PLACEHOLDER, prompt)
-        for prompt in example_prompts[:1]
-    ]
-
-    # 1. Qwen2-VL's M-RoPE works only when there are some multi-modal inputs,
-    #    so an image is included in the inputs
-    # 2. however, Qwen2-VL currently won't work properly
-    #    when chunked prefill is enabled and there are some multi-modal inputs,
-    #    here use a hacky way: provide a **zero-length** image to make it happy
-    #
-    # and finally we achieved:
-    # (1) chunked_prefill enabled; (2) M-RoPE works; to continue our tests
-    zero_len_image = {
-        "image_embeds": torch.empty((0, MODEL_HIDDEN_SIZE)),
-        "image_grid_thw": torch.tensor([[0, 0, 0]])
-    }
-    images = [zero_len_image] * len(prompts)
-
-    inputs_per_case: List[Tuple[List[str], PromptImageInput,
-                                PromptVideoInput]] = [
-                                    (prompts, images, []),
-                                ]
-
-    run_chunked_prefill_test(
-        vllm_runner,
-        inputs_per_case,
-        model,
-        dtype=dtype,
-        max_tokens=max_tokens,
-        num_logprobs=num_logprobs,
-        mm_limit=1,
-        tensor_parallel_size=1,
-    )

tests/multimodal/test_processing.py

@@ -11,8 +11,8 @@ from vllm.config import ModelConfig
 from vllm.inputs import InputProcessingContext
 from vllm.multimodal import MULTIMODAL_REGISTRY
 from vllm.multimodal.processing import (ProcessingCache, PromptReplacement,
-                                        _PlaceholderInfo, find_text_matches,
-                                        find_token_matches, iter_placeholders,
+                                        _PlaceholderInfo, find_mm_placeholders,
+                                        find_text_matches, find_token_matches,
                                         iter_token_matches,
                                         replace_text_matches,
                                         replace_token_matches)
@@ -314,21 +314,27 @@ def test_find_replace_text(
     # Should not be used since there is nothing to convert to text
     mock_tokenizer = cast(AnyTokenizer, object())
 
-    prompt_repls = [
-        PromptReplacement(key, target, repl_by_key[key]).bind(mock_tokenizer)
-        for key, target in target_by_key.items()
+    mm_prompt_repls = {
+        key: [
+            PromptReplacement(key, target,
+                              repl_by_key[key]).bind(mock_tokenizer)
         ]
-    matches = find_text_matches(prompt, prompt_repls)
+        for key, target in target_by_key.items()
+    }
+    mm_matches = {
+        key: find_text_matches(prompt, prompt_repls)
+        for key, prompt_repls in mm_prompt_repls.items()
+    }
 
     result = replace_text_matches(
         prompt,
-        matches,
+        mm_matches,
         {key: mm_count
          for key in repl_by_key},
     )
 
     # Only displayed on error
-    print("matches:", matches)
+    print("mm_matches:", mm_matches)
     print("result:", result)
 
     # Manually constructed results
@@ -380,21 +386,27 @@ def test_find_replace_tokens(
     # Should not be used since there is nothing to convert to tokens
     mock_tokenizer = cast(AnyTokenizer, object())
 
-    prompt_repls = [
-        PromptReplacement(key, target, repl_by_key[key]).bind(mock_tokenizer)
-        for key, target in target_by_key.items()
+    mm_prompt_repls = {
+        key: [
+            PromptReplacement(key, target,
+                              repl_by_key[key]).bind(mock_tokenizer)
         ]
-    matches = find_token_matches(prompt, prompt_repls)
+        for key, target in target_by_key.items()
+    }
+    mm_matches = {
+        key: find_token_matches(prompt, prompt_repls)
+        for key, prompt_repls in mm_prompt_repls.items()
+    }
 
     result = replace_token_matches(
         prompt,
-        matches,
+        mm_matches,
         {key: mm_count
          for key in repl_by_key},
     )
 
     # Only displayed on error
-    print("matches:", matches)
+    print("mm_matches:", mm_matches)
     print("result:", result)
 
     # Manually constructed results
@@ -417,58 +429,76 @@ def test_find_replace_tokens(
     [
         (
             [1, 9833, 28747, 32000, 9833, 28747, 32000, 32000, 918],
-            [
+            {
+                "pattern_1": [
                     _PlaceholderInfo(
                         modality="pattern_1",
+                        item_idx=0,
                         start_idx=6,
                         replacement=[32000, 32000],
                     ),
                 ],
+            }
+
         ),
         (
             [1, 32000, 32000, 9833, 28747, 32000, 32000, 1550, 918, 1550],
-            [
+            {
+                "pattern_1": [
                     _PlaceholderInfo(
                         modality="pattern_1",
+                        item_idx=0,
                         start_idx=1,
                         replacement=[32000, 32000],
                     ),
                     _PlaceholderInfo(
                         modality="pattern_1",
+                        item_idx=1,
                         start_idx=5,
                         replacement=[32000, 32000],
                     ),
+                ],
+                "pattern_3": [
                     _PlaceholderInfo(
                         modality="pattern_3",
+                        item_idx=0,
                         start_idx=7,
                         replacement=[1550, 918, 1550],
                     ),
                 ],
+            }
         ),
         (
             [1, 32000, 32000, 32000, 32000, 32000, 1550, 918, 1550],
-            [
+            {
+                "pattern_1": [
                     _PlaceholderInfo(
                         modality="pattern_1",
+                        item_idx=0,
                         start_idx=1,
                         replacement=[32000, 32000],
                     ),
                     _PlaceholderInfo(
                         modality="pattern_1",
+                        item_idx=1,
                         start_idx=3,
                         replacement=[32000, 32000],
                     ),
+                ],
+                "pattern_3": [
                     _PlaceholderInfo(
                         modality="pattern_3",
+                        item_idx=0,
                         start_idx=6,
                         replacement=[1550, 918, 1550],
                     ),
                 ],
+            }
         ),
     ]
 )
 # yapf: enable
-def test_iter_placeholders(
+def test_find_mm_placeholders(
     repl_by_key,
     prompt,
     expected,
@@ -476,19 +506,18 @@ def test_iter_placeholders(
     # Should not be used since there is nothing to convert to tokens
     mock_tokenizer = cast(AnyTokenizer, object())
 
-    prompt_repls = [
-        PromptReplacement(key, [], repl).bind(mock_tokenizer)
+    mm_prompt_repls = {
+        key: [PromptReplacement(key, [], repl).bind(mock_tokenizer)]
         for key, repl in repl_by_key.items()
-    ]
+    }
 
-    result = list(
-        iter_placeholders(
-            prompt_repls,
+    result = find_mm_placeholders(
+        mm_prompt_repls,
         prompt,
         # Effectively match all occurrences in the prompt
         {key: 3
          for key in repl_by_key},
-        ))
+    )
 
     # Only displayed on error
     print("result:", result)
@@ -694,7 +723,10 @@ def _test_processing_cache_correctness(
         }
 
         mm_counts = {k: len(vs) for k, vs in mm_data.items()}
-        prompt = baseline_processor._get_dummy_mm_inputs(mm_counts).prompt_text
+        prompt = baseline_processor._get_dummy_processor_inputs(
+            model_config.max_model_len,
+            mm_counts,
+        ).prompt_text
 
         # Drop unnecessary keys and test single -> multi conversion
         if rng.rand() < simplify_rate:
@@ -728,6 +760,8 @@ def _test_processing_cache_correctness(
     ("adept/fuyu-8b", {"image": False}),
     ("llava-hf/llava-1.5-7b-hf", {"image": True}),
     ("llava-hf/llava-v1.6-mistral-7b-hf", {"image": True}),
+    ("llava-hf/LLaVA-NeXT-Video-7B-hf", {"video": False}),
+    ("llava-hf/llava-onevision-qwen2-0.5b-ov-hf", {"image": True, "video": True}),  # noqa: E501
     ("TIGER-Lab/Mantis-8B-siglip-llama3", {"image": True}),
     ("mistral-community/pixtral-12b", {"image": True}),
     ("Qwen/Qwen2-VL-2B-Instruct", {"image": True, "video": True}),

vllm/model_executor/models/aria.py

@@ -456,7 +456,7 @@ class AriaMultiModalProcessor(BaseMultiModalProcessor):
         hf_config = self.ctx.get_hf_config()
         return max(hf_config.projector_patch_to_query_dict.values())
 
-    def get_mm_max_tokens_per_item(self) -> Mapping[str, int]:
+    def get_mm_max_tokens_per_item(self, seq_len: int) -> Mapping[str, int]:
         return {"image": self._get_num_image_tokens()}
 
     def _get_mm_fields_config(
@@ -488,8 +488,9 @@ class AriaMultiModalProcessor(BaseMultiModalProcessor):
             )
         ]
 
-    def _get_dummy_mm_inputs(
+    def _get_dummy_processor_inputs(
         self,
+        seq_len: int,
         mm_counts: Mapping[str, int],
     ) -> ProcessorInputs:
         hf_config = self.ctx.get_hf_config()

vllm/model_executor/models/blip2.py

@@ -405,7 +405,7 @@ class Blip2MultiModalProcessor(BaseMultiModalProcessor):
         hf_config = self.ctx.get_hf_config(Blip2Config)
         return hf_config.num_query_tokens
 
-    def get_mm_max_tokens_per_item(self) -> Mapping[str, int]:
+    def get_mm_max_tokens_per_item(self, seq_len: int) -> Mapping[str, int]:
         return {"image": self._get_num_image_tokens()}
 
     def _get_hf_processor(self) -> Blip2Processor:
@@ -457,8 +457,9 @@ class Blip2MultiModalProcessor(BaseMultiModalProcessor):
 
         return result
 
-    def _get_dummy_mm_inputs(
+    def _get_dummy_processor_inputs(
         self,
+        seq_len: int,
         mm_counts: Mapping[str, int],
     ) -> ProcessorInputs:
         hf_config = self.ctx.get_hf_config(Blip2Config)

vllm/model_executor/models/chameleon.py

@@ -57,7 +57,7 @@ class ChameleonMultiModalProcessor(BaseMultiModalProcessor):
         processor = self._get_hf_processor()
         return processor.image_seq_length
 
-    def get_mm_max_tokens_per_item(self) -> Mapping[str, int]:
+    def get_mm_max_tokens_per_item(self, seq_len: int) -> Mapping[str, int]:
         return {"image": self._get_num_image_tokens()}
 
     def _get_hf_processor(self) -> ChameleonProcessor:
@@ -90,8 +90,9 @@ class ChameleonMultiModalProcessor(BaseMultiModalProcessor):
             )
         ]
 
-    def _get_dummy_mm_inputs(
+    def _get_dummy_processor_inputs(
         self,
+        seq_len: int,
         mm_counts: Mapping[str, int],
     ) -> ProcessorInputs:
         config = self.ctx.get_hf_config(ChameleonConfig)

vllm/model_executor/models/clip.py

@@ -164,15 +164,18 @@ class CLIPEncoderInfo(VisionEncoderInfo[CLIPVisionConfig]):
     def get_max_image_tokens(self) -> int:
         return get_max_clip_image_tokens(self.vision_config)
 
-    def get_num_patches(self) -> int:
+    def get_image_size(self) -> int:
+        return self.vision_config.image_size
+
+    def get_patch_size(self) -> int:
+        return self.vision_config.patch_size
+
+    def get_patch_grid_length(self) -> int:
         return get_clip_patch_grid_length(
             image_size=self.vision_config.image_size,
             patch_size=self.vision_config.patch_size,
         )
 
-    def get_image_size(self) -> int:
-        return self.vision_config.image_size
-
 
 # Adapted from https://github.com/huggingface/transformers/blob/v4.39.0/src/transformers/models/clip/modeling_clip.py#L164 # noqa
 class CLIPVisionEmbeddings(nn.Module):

vllm/model_executor/models/fuyu.py

@@ -96,7 +96,7 @@ class FuyuMultiModalProcessor(BaseMultiModalProcessor):
         nrows = math.ceil(image_height / 30)
         return ncols, nrows
 
-    def get_mm_max_tokens_per_item(self) -> Mapping[str, int]:
+    def get_mm_max_tokens_per_item(self, seq_len: int) -> Mapping[str, int]:
         target_width, target_height = self._get_image_target_size()
 
         max_ncols, max_nrows = self._get_image_feature_grid_size(
@@ -208,8 +208,9 @@ class FuyuMultiModalProcessor(BaseMultiModalProcessor):
 
         return result
 
-    def _get_dummy_mm_inputs(
+    def _get_dummy_processor_inputs(
         self,
+        seq_len: int,
         mm_counts: Mapping[str, int],
     ) -> ProcessorInputs:
         target_width, target_height = self._get_image_target_size()

vllm/model_executor/models/llava.py

@@ -25,11 +25,9 @@ from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig,
                                     NestedTensors)
 from vllm.multimodal.parse import (ImageEmbeddingItems, ImageProcessorItems,
                                    ImageSize)
-from vllm.multimodal.processing import (BaseMultiModalProcessor,
-                                        InputProcessingContext,
+from vllm.multimodal.processing import (InputProcessingContext,
                                         MultiModalDataItems, ProcessingCache,
-                                        ProcessorInputs, PromptReplacement,
-                                        full_groupby_modality)
+                                        ProcessorInputs, PromptReplacement)
 from vllm.sequence import IntermediateTensors
 
 from .clip import CLIPVisionModel
@@ -39,7 +37,7 @@ from .pixtral import (PixtralHFVisionModel,
 from .siglip import SiglipVisionModel
 from .utils import (AutoWeightsLoader, flatten_bn, init_vllm_registered_model,
                     maybe_prefix, merge_multimodal_embeddings)
-from .vision import vision_encoder_info
+from .vision import BaseVisionLanguageMultiModalProcessor
 
 
 class LlavaImagePixelInputs(TypedDict):
@@ -100,19 +98,7 @@ class LlavaLikeConfig(Protocol):
     vision_feature_layer: Final[Union[int, List[int]]]
 
 
-class BaseLlavaMultiModalProcessor(BaseMultiModalProcessor):
-
-    def __init__(self,
-                 ctx: InputProcessingContext,
-                 *,
-                 cache: Optional[ProcessingCache] = None,
-                 enable_sanity_checks: bool = True) -> None:
-        super().__init__(ctx,
-                         cache=cache,
-                         enable_sanity_checks=enable_sanity_checks)
-
-        vision_config = self._get_hf_config().vision_config
-        self._vision_encoder_info = vision_encoder_info(vision_config)
+class BaseLlavaMultiModalProcessor(BaseVisionLanguageMultiModalProcessor):
 
     @abstractmethod
     def _get_hf_config(self) -> LlavaLikeConfig:
@@ -121,6 +107,19 @@ class BaseLlavaMultiModalProcessor(BaseMultiModalProcessor):
     def get_supported_mm_limits(self) -> Mapping[str, Optional[int]]:
         return {"image": None}
 
+    def get_mm_max_tokens_per_item(self, seq_len: int) -> Mapping[str, int]:
+        return {"image": self._get_max_image_tokens()}
+
+    def _get_mm_fields_config(
+        self,
+        hf_inputs: BatchFeature,
+        hf_processor_mm_kwargs: Mapping[str, object],
+    ) -> Mapping[str, MultiModalFieldConfig]:
+        return dict(
+            pixel_values=MultiModalFieldConfig.batched("image"),
+            image_embeds=MultiModalFieldConfig.batched("image"),
+        )
+
     def _apply_feature_select_strategy(
         self,
         strategy: str,
@@ -142,19 +141,6 @@ class BaseLlavaMultiModalProcessor(BaseMultiModalProcessor):
             self._vision_encoder_info.get_max_image_tokens(),
         )
 
-    def get_mm_max_tokens_per_item(self) -> Mapping[str, int]:
-        return {"image": self._get_max_image_tokens()}
-
-    def _get_mm_fields_config(
-        self,
-        hf_inputs: BatchFeature,
-        hf_processor_mm_kwargs: Mapping[str, object],
-    ) -> Mapping[str, MultiModalFieldConfig]:
-        return dict(
-            pixel_values=MultiModalFieldConfig.batched("image"),
-            image_embeds=MultiModalFieldConfig.batched("image"),
-        )
-
     def _get_dummy_image_size(self) -> ImageSize:
         image_size = self._vision_encoder_info.get_image_size()
         return ImageSize(image_size, image_size)
@@ -163,8 +149,9 @@ class BaseLlavaMultiModalProcessor(BaseMultiModalProcessor):
     def _get_image_token(self) -> str:
         raise NotImplementedError
 
-    def _get_dummy_mm_inputs(
+    def _get_dummy_processor_inputs(
         self,
+        seq_len: int,
         mm_counts: Mapping[str, int],
     ) -> ProcessorInputs:
         num_images = mm_counts.get("image", 0)
@@ -709,7 +696,7 @@ class MantisMultiModalProcessor(LlavaMultiModalProcessor):
                 "</Image>)",  # 3 tokens
             ])
 
-        mantis_repls = self._bind_prompt_replacements([
+        mantis_mm_repls = self._bind_and_group_repls([
             PromptReplacement(
                 modality="image",
                 target=[image_token_id] * num_image_tokens,
@@ -719,7 +706,7 @@ class MantisMultiModalProcessor(LlavaMultiModalProcessor):
 
         prompt_ids, prompt_text, _ = self._apply_prompt_replacements(
             result["prompt_token_ids"],
-            mantis_repls,
+            mantis_mm_repls,
             mm_item_counts,
         )
 
@@ -728,15 +715,19 @@ class MantisMultiModalProcessor(LlavaMultiModalProcessor):
             hf_processor_mm_kwargs,
             mm_kwargs,
         )
-        orig_repls = self._bind_prompt_replacements(unbound_orig_repls)
+        orig_repls = self._bind_and_group_repls(unbound_orig_repls)
+
+        mm_placeholders = self._find_mm_placeholders(
+            orig_repls,
+            prompt_ids,
+            mm_item_counts,
+        )
 
-        all_placeholders = self._find_placeholders(orig_repls, prompt_ids,
-                                                   mm_item_counts)
-        assert len(all_placeholders) == mm_item_counts.get("image", 0)
+        self._validate_mm_placeholders(mm_placeholders, mm_item_counts)
 
-        mm_placeholders = {
-            modality: [item.to_range() for item in items]
-            for modality, items in full_groupby_modality(all_placeholders)
+        mm_placeholder_ranges = {
+            modality: [item.to_range() for item in placeholders]
+            for modality, placeholders in mm_placeholders.items()
         }
 
         return MultiModalInputsV2(
@@ -744,7 +735,7 @@ class MantisMultiModalProcessor(LlavaMultiModalProcessor):
             prompt=prompt_text,
             prompt_token_ids=prompt_ids,
             mm_kwargs=mm_kwargs,
-            mm_placeholders=mm_placeholders,
+            mm_placeholders=mm_placeholder_ranges,
         )
 
 

vllm/model_executor/models/llava_next.py

@@ -67,9 +67,6 @@ class LlavaNextMultiModalProcessor(LlavaMultiModalProcessor):
     def _get_hf_processor(self) -> LlavaNextProcessor:
         return self.ctx.get_hf_processor(LlavaNextProcessor)
 
-    def _get_image_token(self) -> str:
-        return self._get_hf_processor().image_token
-
     def _get_mm_fields_config(
         self,
         hf_inputs: BatchFeature,
@@ -81,6 +78,9 @@ class LlavaNextMultiModalProcessor(LlavaMultiModalProcessor):
             image_embeds=MultiModalFieldConfig.batched("image"),
         )
 
+    def _get_image_token(self) -> str:
+        return self._get_hf_processor().image_token
+
     def _get_max_image_tokens(self) -> int:
         largest_feature_size, _ = self._get_pinpoint_with_most_features()
         return largest_feature_size
@@ -97,20 +97,20 @@ class LlavaNextMultiModalProcessor(LlavaMultiModalProcessor):
         image_height: int,
     ) -> int:
         hf_config = self._get_hf_config()
+        vision_encoder_info = self._vision_encoder_info
 
         base_feature_size = self._apply_feature_select_strategy(
             hf_config.vision_feature_select_strategy,
-            self._vision_encoder_info.get_num_image_tokens(
+            vision_encoder_info.get_num_image_tokens(
                 image_width=image_width,
                 image_height=image_height,
             ),
         )
-        num_patches = self._vision_encoder_info.get_num_patches()
 
         num_patch_height, num_patch_width = get_anyres_image_grid_shape(
             image_size=(image_height, image_width),
             grid_pinpoints=hf_config.image_grid_pinpoints,
-            patch_size=self._vision_encoder_info.get_image_size(),
+            patch_size=vision_encoder_info.get_image_size(),
         )
 
         (
@@ -119,7 +119,7 @@ class LlavaNextMultiModalProcessor(LlavaMultiModalProcessor):
         ) = self._get_num_unpadded_features(
             original_height=image_height,
             original_width=image_width,
-            npatches=num_patches,
+            npatches=vision_encoder_info.get_patch_grid_length(),
             num_patch_height=num_patch_height,
             num_patch_width=num_patch_width,
         )
@@ -155,6 +155,7 @@ class LlavaNextMultiModalProcessor(LlavaMultiModalProcessor):
 
         unpadded_features = current_height * current_width
         newline_features = current_height
+
         return (unpadded_features, newline_features)
 
     def _get_pinpoint_with_most_features(self) -> tuple[int, ImageSize]:

vllm/model_executor/models/llava_next_video.py

@@ -3,38 +3,32 @@ from functools import cached_property
 from typing import (Iterable, List, Literal, Mapping, Optional, Set, Tuple,
                     TypedDict, Union)
 
-import numpy as np
 import torch
 import torch.nn as nn
-from transformers import (CLIPVisionConfig, LlavaNextVideoConfig,
-                          SiglipVisionConfig)
+from transformers import (BatchFeature, LlavaNextVideoConfig,
+                          LlavaNextVideoProcessor)
 
 from vllm.attention import AttentionMetadata
 from vllm.config import VllmConfig
-from vllm.inputs import (INPUT_REGISTRY, DecoderOnlyInputs, DummyData,
-                         InputContext, token_inputs)
 from vllm.model_executor.layers.activation import get_act_fn
 from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler
 from vllm.model_executor.models.clip import CLIPVisionModel
 from vllm.model_executor.sampling_metadata import SamplingMetadata
 from vllm.multimodal import MULTIMODAL_REGISTRY
-from vllm.multimodal.inputs import NestedTensors
-from vllm.multimodal.utils import (cached_get_tokenizer,
-                                   repeat_and_pad_placeholder_tokens)
+from vllm.multimodal.inputs import MultiModalKwargs, NestedTensors
+from vllm.multimodal.parse import (ImageSize, MultiModalDataItems,
+                                   VideoEmbeddingItems, VideoProcessorItems)
+from vllm.multimodal.processing import (MultiModalFieldConfig, ProcessorInputs,
+                                        PromptReplacement)
 from vllm.sequence import IntermediateTensors
 from vllm.utils import is_list_of
 
-from .clip import dummy_image_for_clip, dummy_seq_data_for_clip
 from .interfaces import SupportsMultiModal, SupportsPP
 from .llava import init_vision_tower_for_llava
-from .siglip import (SiglipVisionModel, dummy_image_for_siglip,
-                     dummy_seq_data_for_siglip)
+from .siglip import SiglipVisionModel
 from .utils import (AutoWeightsLoader, init_vllm_registered_model,
                     maybe_prefix, merge_multimodal_embeddings)
-
-# For profile run
-_MAX_FRAMES_PER_VIDEO = 32
-_MAX_NUM_VIDEOS = 1
+from .vision import BaseVisionLanguageMultiModalProcessor
 
 
 class LlavaNextVideoPixelInputs(TypedDict):
@@ -50,143 +44,148 @@ class LlavaNextVideoPixelInputs(TypedDict):
     """
 
 
-def get_llava_next_video_frame_feature_size(
-        hf_config: LlavaNextVideoConfig) -> int:
-    # Support both CLIPVisionConfig and SiglipVisionConfig
-    image_size = hf_config.vision_config.image_size
-    patch_size = hf_config.vision_config.patch_size
-    spatial_pool_stride = hf_config.spatial_pool_stride
-
-    return int((image_size / patch_size / spatial_pool_stride)**2)
+class LlavaNextVideoMultiModalProcessor(BaseVisionLanguageMultiModalProcessor):
 
+    def _get_hf_config(self) -> LlavaNextVideoConfig:
+        return self.ctx.get_hf_config(LlavaNextVideoConfig)
 
-def _get_max_llm_tokens(ctx: InputContext) -> int:
-    """
-    Calculated from the maximum video frames under the context length
-    constraints of the language model.
-    """
-    hf_text_config = ctx.model_config.hf_text_config
-    model_config = ctx.model_config
-    max_tokens = model_config.max_model_len
-    rope_scaling = model_config.rope_scaling
+    def _get_hf_processor(self) -> LlavaNextVideoProcessor:
+        return self.ctx.get_hf_processor(LlavaNextVideoProcessor)
 
-    if rope_scaling:
-        rope_scaling_factor = hf_text_config.rope_scaling["factor"]
-    else:
-        rope_scaling_factor = 1
+    def get_supported_mm_limits(self) -> Mapping[str, Optional[int]]:
+        return {"video": 1}
 
-    max_tokens *= rope_scaling_factor
+    def get_mm_max_tokens_per_item(self, seq_len: int) -> Mapping[str, int]:
+        num_frames = self._get_dummy_num_frames(seq_len)
+        max_video_tokens = self._get_max_video_tokens(num_frames)
 
-    return max_tokens
+        return {"video": max_video_tokens}
 
+    def _get_mm_fields_config(
+        self,
+        hf_inputs: BatchFeature,
+        hf_processor_mm_kwargs: Mapping[str, object],
+    ) -> Mapping[str, MultiModalFieldConfig]:
+        return dict(pixel_values_videos=MultiModalFieldConfig.batched("video"))
 
-def get_max_llava_next_video_tokens(ctx: InputContext) -> int:
-    # Currently set to 32 frames
-    # TODO: max_tokens = _get_max_llm_tokens(ctx)
-    hf_config = ctx.get_hf_config(LlavaNextVideoConfig)
-    tokens_per_frame = get_llava_next_video_frame_feature_size(hf_config)
-    return _MAX_FRAMES_PER_VIDEO * tokens_per_frame
-
+    def _get_num_frame_tokens(
+        self,
+        *,
+        image_width: int,
+        image_height: int,
+    ) -> int:
+        hf_config = self._get_hf_config()
+        spatial_pool_stride = hf_config.spatial_pool_stride
 
-def dummy_data_for_llava_next_video(ctx: InputContext, seq_len: int,
-                                    mm_counts: Mapping[str, int]):
-    hf_config = ctx.get_hf_config(LlavaNextVideoConfig)
-    vision_config = hf_config.vision_config
+        patch_grid_length = self._vision_encoder_info.get_patch_grid_length()
+        pooled_grid_length = math.ceil(patch_grid_length / spatial_pool_stride)
 
-    # TODO: support multiple videos
-    num_videos = mm_counts["video"]
-    if num_videos != _MAX_NUM_VIDEOS:
-        raise NotImplementedError(
-            f"Only {_MAX_NUM_VIDEOS} videos are supported")
+        return pooled_grid_length * pooled_grid_length
 
-    # TODO: support configuring the number of frames
-    frames_per_video = _MAX_FRAMES_PER_VIDEO
-    # num_images = num_videos * frames_per_video
+    def _get_num_video_tokens(
+        self,
+        *,
+        image_width: int,
+        image_height: int,
+        num_frames: int,
+    ) -> int:
+        num_frame_tokens = self._get_num_frame_tokens(
+            image_width=image_width,
+            image_height=image_height,
+        )
 
-    # fills the sequence with as longer video data as possible
-    tokens_per_frame = get_llava_next_video_frame_feature_size(hf_config)
-    video_feature_size = frames_per_video * tokens_per_frame
+        return num_frame_tokens * num_frames
 
-    if isinstance(vision_config, CLIPVisionConfig):
-        seq_data, ranges = dummy_seq_data_for_clip(
-            vision_config,
-            seq_len,
-            num_videos,
-            image_token_id=hf_config.video_token_index,
-            image_feature_size_override=video_feature_size,
-            mm_key="video",
-        )
+    def _get_max_video_tokens(self, num_frames: int) -> int:
+        return self._get_num_video_tokens(image_width=999999,
+                                          image_height=999999,
+                                          num_frames=num_frames)
 
-        pil_frame = dummy_image_for_clip(vision_config, num_images=1)
-        np_frame = np.array(pil_frame["image"])
-        mm_data_per_video = np.repeat([np_frame], frames_per_video, axis=0)
-        mm_data = {"video": mm_data_per_video}
-        return DummyData(seq_data, mm_data, ranges)
-    elif isinstance(vision_config, SiglipVisionConfig):
-        seq_data, ranges = dummy_seq_data_for_siglip(
-            vision_config,
-            seq_len,
-            num_videos,
-            image_token_id=hf_config.video_token_index,
-            image_feature_size_override=video_feature_size,
-            mm_key="video",
-        )
+    def _get_max_video_frames(self, max_tokens: int) -> int:
+        num_frames = 0
 
-        pil_frame = dummy_image_for_siglip(vision_config, num_images=1)
-        np_frame = np.array(pil_frame["image"])
-        mm_data_per_video = np.repeat([np_frame], frames_per_video, axis=0)
-        mm_data = {"video": mm_data_per_video}
-        return DummyData(seq_data, mm_data, ranges)
+        while True:
+            next_num_frames = num_frames + 1
 
-    msg = f"Unsupported vision config: {type(vision_config)}"
-    raise NotImplementedError(msg)
+            if self._get_max_video_tokens(next_num_frames) > max_tokens:
+                break
 
+            num_frames = next_num_frames
 
-def input_processor_for_llava_next_video(ctx: InputContext,
-                                         inputs: DecoderOnlyInputs):
-    multi_modal_data = inputs.get("multi_modal_data")
-    if multi_modal_data is None or "video" not in multi_modal_data:
-        return inputs
+        return num_frames
 
-    if "multi_modal_placeholders" in inputs and "video" in inputs[
-            "multi_modal_placeholders"]:
-        # The inputs already have placeholders.
-        return inputs
+    def _get_dummy_num_frames(self, seq_len: int) -> int:
+        mm_config = self.ctx.get_mm_config()
+        max_videos = mm_config.limit_per_prompt.get("video", 1)
 
-    video_data = multi_modal_data["video"]
+        max_total_frames = self._get_max_video_frames(seq_len)
 
-    model_config = ctx.model_config
-    hf_config = ctx.get_hf_config(LlavaNextVideoConfig)
-    vision_config = hf_config.vision_config
+        return max(max_total_frames // max(max_videos, 1), 1)
 
-    if isinstance(video_data, np.ndarray):
-        # Supports both CLIP and Siglip
-        num_frames = video_data.shape[0]
-        frame_feature_size = \
-            get_llava_next_video_frame_feature_size(hf_config)
-        video_feature_size = num_frames * frame_feature_size
+    def _get_dummy_image_size(self) -> ImageSize:
+        image_size = self._vision_encoder_info.get_image_size()
+        return ImageSize(image_size, image_size)
 
-        tokenizer = cached_get_tokenizer(model_config.tokenizer)
+    def _get_video_token(self) -> str:
+        return self._get_hf_processor().video_token
 
-        new_prompt, new_token_ids, ranges = repeat_and_pad_placeholder_tokens(
-            tokenizer,
-            inputs.get("prompt"),
-            inputs["prompt_token_ids"],
-            placeholder_token_id=hf_config.video_token_index,
-            repeat_count=video_feature_size,
+    def _get_prompt_replacements(
+        self,
+        mm_items: MultiModalDataItems,
+        hf_processor_mm_kwargs: Mapping[str, object],
+        out_mm_kwargs: MultiModalKwargs,
+    ) -> list[PromptReplacement]:
+        hf_config = self._get_hf_config()
+        video_token_id = hf_config.video_token_index
+
+        def get_replacement(item_idx: int):
+            videos = mm_items.get_items(
+                "video", (VideoEmbeddingItems, VideoProcessorItems))
+
+            if isinstance(videos, VideoEmbeddingItems):
+                num_video_tokens = videos.get_feature_size(item_idx)
+            else:
+                image_size = videos.get_frame_size(item_idx)
+                num_video_tokens = self._get_num_video_tokens(
+                    image_width=image_size.width,
+                    image_height=image_size.height,
+                    num_frames=videos.get_num_frames(item_idx),
                 )
 
-        return token_inputs(prompt_token_ids=new_token_ids,
-                            prompt=new_prompt,
-                            multi_modal_data=multi_modal_data,
-                            multi_modal_placeholders={"video": ranges})
+            return [video_token_id] * num_video_tokens
 
-    elif is_list_of(video_data, np.ndarray):
-        raise NotImplementedError(
-            "Processing multiple videos is not supported")
+        return [
+            PromptReplacement(
+                modality="video",
+                target=[video_token_id],
+                replacement=get_replacement,
+            ),
+        ]
 
-    msg = f"Unsupported vision config: {type(vision_config)}"
-    raise NotImplementedError(msg)
+    def _get_dummy_processor_inputs(
+        self,
+        seq_len: int,
+        mm_counts: Mapping[str, int],
+    ) -> ProcessorInputs:
+        num_videos = mm_counts.get("video", 0)
+
+        video_token = self._get_video_token()
+        target_width, target_height = self._get_dummy_image_size()
+
+        mm_data = {
+            "video":
+            self._get_dummy_videos(
+                width=target_width,
+                height=target_height,
+                num_frames=self._get_dummy_num_frames(seq_len),
+                num_videos=num_videos,
+            )
+        }
+
+        return ProcessorInputs(
+            prompt_text=video_token * num_videos,
+            mm_data=mm_data,
+        )
 
 
 # adopted from transformers modeling_llava_next_video.py
@@ -246,11 +245,7 @@ class LlavaNextMultiModalProjector(nn.Module):
         return hidden_states
 
 
-@MULTIMODAL_REGISTRY.register_input_mapper("video")
-@MULTIMODAL_REGISTRY.register_max_multimodal_tokens(
-    "video", get_max_llava_next_video_tokens)
-@INPUT_REGISTRY.register_dummy_data(dummy_data_for_llava_next_video)
-@INPUT_REGISTRY.register_input_processor(input_processor_for_llava_next_video)
+@MULTIMODAL_REGISTRY.register_processor(LlavaNextVideoMultiModalProcessor)
 class LlavaNextVideoForConditionalGeneration(nn.Module, SupportsMultiModal,
                                              SupportsPP):