diff --git a/docs/models/supported_models.md b/docs/models/supported_models.md
index 37d27234aa..98f9cb5b65 100644
--- a/docs/models/supported_models.md
+++ b/docs/models/supported_models.md
@@ -555,7 +555,7 @@ These models primarily accept the [`LLM.generate`](./generative_models.md#llmgen
| `Llama_Nemotron_Nano_VL` | Llama Nemotron Nano VL | T + IE+ | `nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1` | ✅︎ | ✅︎ |
| `LlavaForConditionalGeneration` | LLaVA-1.5, Pixtral (HF Transformers) | T + IE+ | `llava-hf/llava-1.5-7b-hf`, `mistral-community/pixtral-12b`, etc. | ✅︎ | ✅︎ |
| `LlavaNextForConditionalGeneration` | LLaVA-NeXT, Granite Vision | T + IE+ | `llava-hf/llava-v1.6-mistral-7b-hf`, `llava-hf/llava-v1.6-vicuna-7b-hf`, `ibm-granite/granite-vision-3.3-2b`, etc. | | ✅︎ |
-| `LlavaNextVideoForConditionalGeneration` | LLaVA-NeXT-Video | T + V | `llava-hf/LLaVA-NeXT-Video-7B-hf`, etc. | | ✅︎ |
+| `LlavaNextVideoForConditionalGeneration` | LLaVA-NeXT-Video | T + V | `llava-hf/LLaVA-NeXT-Video-7B-hf`, etc. | ✅︎ | ✅︎ |
| `LlavaOnevision2ForConditionalGeneration` | LLaVA-OneVision-2 | T + I+ + V+ | `lmms-lab-encoder/LLaVA-OneVision-2-8B-Instruct` | | |
| `LlavaOnevisionForConditionalGeneration` | LLaVA-Onevision | T + I+ + V+ | `llava-hf/llava-onevision-qwen2-7b-ov-hf`, `llava-hf/llava-onevision-qwen2-0.5b-ov-hf`, etc. | | ✅︎ |
| `MiDashengLMModel` | MiDashengLM | T + A+ | `mispeech/midashenglm-7b` | | ✅︎ |
diff --git a/tests/models/multimodal/processing/test_llava_next_video.py b/tests/models/multimodal/processing/test_llava_next_video.py
new file mode 100644
index 0000000000..2dcbb7d4b5
--- /dev/null
+++ b/tests/models/multimodal/processing/test_llava_next_video.py
@@ -0,0 +1,88 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import math
+
+import pytest
+
+from vllm.model_executor.models.llava_next_video import (
+ LlavaNextVideoForConditionalGeneration,
+)
+from vllm.model_executor.models.vision import get_vision_encoder_info
+
+from ...utils import build_model_context
+
+
+class _StubModel:
+ """Carries only the two attributes the token helpers read from
+ `self`, so the real methods can be exercised without constructing
+ the full `nn.Module` (vision tower, language model, etc.)."""
+
+ patch_grid_length: int
+ pooled_grid_length: int
+
+
+get_num_mm_encoder_tokens = (
+ LlavaNextVideoForConditionalGeneration.get_num_mm_encoder_tokens
+)
+get_num_mm_connector_tokens = (
+ LlavaNextVideoForConditionalGeneration.get_num_mm_connector_tokens
+)
+
+
+@pytest.mark.parametrize("model_id", ["llava-hf/LLaVA-NeXT-Video-7B-hf"])
+def test_num_mm_tokens_match_real_config(model_id):
+ """The stored grid lengths must match what `__init__` derives from
+ the real HF config, and the two helpers must invert each other's
+ frame-level scaling exactly."""
+ ctx = build_model_context(model_id, limit_mm_per_prompt={"video": 1})
+ config = ctx.model_config.hf_config
+
+ vision_encoder_info = get_vision_encoder_info(config)
+ patch_grid_length = vision_encoder_info.get_patch_grid_length()
+ pooled_grid_length = math.ceil(patch_grid_length / config.spatial_pool_stride)
+
+ stub = _StubModel()
+ stub.patch_grid_length = patch_grid_length
+ stub.pooled_grid_length = pooled_grid_length
+
+ for num_frames in (1, 2, 8, 16, 32):
+ num_video_tokens = num_frames * pooled_grid_length**2
+
+ encoder_tokens = get_num_mm_encoder_tokens(stub, num_video_tokens)
+ assert encoder_tokens == num_frames * patch_grid_length**2
+
+ connector_tokens = get_num_mm_connector_tokens(stub, encoder_tokens)
+ assert connector_tokens == num_video_tokens
+
+
+@pytest.mark.parametrize(
+ ("patch_grid_length", "pooled_grid_length", "num_frames"),
+ [
+ (24, 12, 1), # llava-hf/LLaVA-NeXT-Video-7B-hf: 336 / 14, stride 2
+ (24, 12, 16),
+ (27, 14, 5), # non-power-of-2 pooled grid (ceil rounding)
+ (16, 8, 6),
+ ],
+)
+def test_num_mm_tokens_roundtrip(patch_grid_length, pooled_grid_length, num_frames):
+ stub = _StubModel()
+ stub.patch_grid_length = patch_grid_length
+ stub.pooled_grid_length = pooled_grid_length
+
+ num_video_tokens = num_frames * pooled_grid_length**2
+
+ encoder_tokens = get_num_mm_encoder_tokens(stub, num_video_tokens)
+ assert encoder_tokens == num_frames * patch_grid_length**2
+
+ connector_tokens = get_num_mm_connector_tokens(stub, encoder_tokens)
+ assert connector_tokens == num_video_tokens
+
+
+def test_num_mm_tokens_zero():
+ stub = _StubModel()
+ stub.patch_grid_length = 24
+ stub.pooled_grid_length = 12
+
+ assert get_num_mm_encoder_tokens(stub, 0) == 0
+ assert get_num_mm_connector_tokens(stub, 0) == 0
diff --git a/vllm/model_executor/models/llava_next_video.py b/vllm/model_executor/models/llava_next_video.py
index e256768ef5..bb9c95017d 100644
--- a/vllm/model_executor/models/llava_next_video.py
+++ b/vllm/model_executor/models/llava_next_video.py
@@ -36,8 +36,14 @@ from vllm.sequence import IntermediateTensors
from vllm.utils.collection_utils import is_list_of
from vllm.utils.tensor_schema import TensorSchema, TensorShape
-from .interfaces import MultiModalEmbeddings, SupportsMultiModal, SupportsPP
+from .interfaces import (
+ MultiModalEmbeddings,
+ SupportsLoRA,
+ SupportsMultiModal,
+ SupportsPP,
+)
from .llava import init_vision_tower_for_llava
+from .module_mapping import MultiModelKeys
from .siglip import SiglipVisionModel
from .utils import (
AutoWeightsLoader,
@@ -298,7 +304,14 @@ class LlavaNextMultiModalProjector(nn.Module):
info=LlavaNextVideoProcessingInfo,
dummy_inputs=LlavaNextVideoDummyInputsBuilder,
)
-class LlavaNextVideoForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP):
+class LlavaNextVideoForConditionalGeneration(
+ nn.Module, SupportsLoRA, SupportsMultiModal, SupportsPP
+):
+ packed_modules_mapping = {
+ "qkv_proj": ["q_proj", "k_proj", "v_proj"],
+ "gate_up_proj": ["gate_proj", "up_proj"],
+ }
+
hf_to_vllm_mapper = WeightsMapper(
orig_to_new_prefix={
# mapping for new names in checkpoint saved after transformers v4.52
@@ -327,6 +340,12 @@ class LlavaNextVideoForConditionalGeneration(nn.Module, SupportsMultiModal, Supp
self.config = config
self.multimodal_config = multimodal_config
+ vision_encoder_info = get_vision_encoder_info(config)
+ self.patch_grid_length = vision_encoder_info.get_patch_grid_length()
+ self.pooled_grid_length = math.ceil(
+ self.patch_grid_length / config.spatial_pool_stride
+ )
+
with self._mark_tower_model(vllm_config, "video"):
# Initialize the vision tower only up to the required feature layer
self.vision_tower = init_vision_tower_for_llava(
@@ -460,3 +479,42 @@ class LlavaNextVideoForConditionalGeneration(nn.Module, SupportsMultiModal, Supp
ignore_unexpected_prefixes=["image_newline"],
)
return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)
+
+ def get_mm_mapping(self) -> MultiModelKeys:
+ """
+ Get the module prefix in multimodal models
+ """
+ return MultiModelKeys.from_string_field(
+ language_model="language_model",
+ connector="multi_modal_projector",
+ tower_model="vision_tower",
+ )
+
+ def get_num_mm_encoder_tokens(
+ self,
+ num_video_tokens: int,
+ ) -> int:
+ # Invert the spatial pooling done by `vision_resampler`: each frame
+ # contributes `pooled_grid_length ** 2` tokens after the language
+ # model's placeholder count, but `patch_grid_length ** 2` tokens
+ # when it leaves the vision encoder.
+ pooled_tokens_per_frame = self.pooled_grid_length**2
+ if num_video_tokens <= 0 or pooled_tokens_per_frame <= 0:
+ return 0
+
+ num_frames = num_video_tokens // pooled_tokens_per_frame
+ return num_frames * (self.patch_grid_length**2)
+
+ def get_num_mm_connector_tokens(
+ self,
+ num_vision_tokens: int,
+ ) -> int:
+ # The projector is length-preserving; it runs after
+ # `vision_resampler` has already pooled each frame down to
+ # `pooled_grid_length ** 2` tokens.
+ patch_tokens_per_frame = self.patch_grid_length**2
+ if num_vision_tokens <= 0 or patch_tokens_per_frame <= 0:
+ return 0
+
+ num_frames = num_vision_tokens // patch_tokens_per_frame
+ return num_frames * (self.pooled_grid_length**2)