EVS Support (Video tokens pruning) (#22980)

Signed-off-by: Eugene Khvedchenia <ekhvedchenia@nvidia.com>
Signed-off-by: Eugene Khvedchenya <ekhvedchenya@gmail.com>
Co-authored-by: Roger Wang <hey@rogerw.io>

tests/models/multimodal/generation/test_qwen2_5_vl.py

@@ -0,0 +1,132 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import pytest
+
+from vllm.multimodal.video import sample_frames_from_video
+
+from ....conftest import VIDEO_ASSETS
+
+models = ["Qwen/Qwen2.5-VL-3B-Instruct"]
+target_dtype = "bfloat16"
+
+VIDEO_PLACEHOLDER = "<|vision_start|><|video_pad|><|vision_end|>"
+
+
+def qwen2_5_vl_chat_template(*query):
+    return f"<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n{''.join(query)}<|im_end|><|im_start|>assistant\n"  # noqa: E501
+
+
+VIDEO_PROMPTS = VIDEO_ASSETS.prompts({
+    "baby_reading":
+    qwen2_5_vl_chat_template(
+        VIDEO_PLACEHOLDER,
+        "Describe this video with a short sentence ",
+        "(no more than 20 words)",
+    ),
+})
+
+
+@pytest.mark.core_model
+@pytest.mark.parametrize("model", models)
+@pytest.mark.parametrize("video_pruning_rate", [0.0, 0.75])
+@pytest.mark.parametrize("num_frames", [16])
+@pytest.mark.parametrize("dtype", [target_dtype])
+@pytest.mark.parametrize("max_tokens", [128])
+def test_qwen2_5_vl_evs_functionality(vllm_runner, video_assets, model,
+                                      video_pruning_rate: float,
+                                      num_frames: int, dtype: str,
+                                      max_tokens: int) -> None:
+    """Test EVS (Efficient Video Sampling) functionality with different
+    pruning rates.
+    """
+
+    # Sample frames from video assets
+    sampled_vids = [
+        sample_frames_from_video(asset.np_ndarrays, num_frames)
+        for asset in video_assets
+    ]
+
+    prompts = [VIDEO_PROMPTS[0]]
+    videos = [sampled_vids[0]]
+
+    # Initialize model with EVS configuration
+    with vllm_runner(model,
+                     runner="generate",
+                     max_model_len=4000,
+                     max_num_seqs=1,
+                     dtype=dtype,
+                     limit_mm_per_prompt={"video": 1},
+                     tensor_parallel_size=1,
+                     video_pruning_rate=video_pruning_rate) as vllm_model:
+
+        # Generate output - this should not crash
+        outputs = vllm_model.generate_greedy(prompts,
+                                             max_tokens,
+                                             videos=videos)
+
+        # Basic validation that we got a response
+        assert len(outputs) == 1
+        output_ids, output_text = outputs[0]
+
+        # Ensure we got some output
+        assert len(output_ids) > 0
+        assert len(output_text) > 0
+
+        # Ensure the output is a string
+        assert isinstance(output_text, str)
+
+
+@pytest.mark.core_model
+@pytest.mark.parametrize("model", models)
+@pytest.mark.parametrize("video_pruning_rate", [0.0, 0.75])
+@pytest.mark.parametrize("num_frames", [16])
+@pytest.mark.parametrize("dtype", [target_dtype])
+@pytest.mark.parametrize("max_tokens", [128])
+def test_qwen2_5_vl_evs_batched_videos(vllm_runner, video_assets, model,
+                                       video_pruning_rate: float,
+                                       num_frames: int, dtype: str,
+                                       max_tokens: int) -> None:
+    """Test EVS functionality with batched videos.
+
+    This test validates that:
+    1. The model handles batched video inputs correctly with EVS
+    2. Both pruning configurations work with multiple videos
+    3. The model doesn't crash when processing multiple videos simultaneously
+    """
+    # Sample frames from video assets
+    sampled_vids = [
+        sample_frames_from_video(asset.np_ndarrays, num_frames)
+        for asset in video_assets
+    ]
+
+    # Test batched videos
+    prompts = [VIDEO_PROMPTS[0], VIDEO_PROMPTS[0]]
+    videos = [sampled_vids[0],
+              sampled_vids[0]]  # Use same video twice for testing
+
+    # Initialize model with EVS configuration
+    with vllm_runner(model,
+                     runner="generate",
+                     max_model_len=4000,
+                     max_num_seqs=2,
+                     dtype=dtype,
+                     limit_mm_per_prompt={"video": 2},
+                     tensor_parallel_size=1,
+                     video_pruning_rate=video_pruning_rate) as vllm_model:
+
+        # Generate output - this should not crash
+        outputs = vllm_model.generate_greedy(prompts,
+                                             max_tokens,
+                                             videos=videos)
+
+        # Basic validation that we got responses for both videos
+        assert len(outputs) == 2
+
+        for output_ids, output_text in outputs:
+            # Ensure we got some output for each video
+            assert len(output_ids) > 0
+            assert len(output_text) > 0
+
+            # Ensure the output is a string
+            assert isinstance(output_text, str)

vllm/config/model.py

@@ -283,6 +283,7 @@ class ModelConfig:
     mm_encoder_tp_mode: InitVar[Optional[MMEncoderTPMode]] = None
     interleave_mm_strings: InitVar[Optional[bool]] = None
     skip_mm_profiling: InitVar[Optional[bool]] = None
+    video_pruning_rate: InitVar[Optional[float]] = None
 
     def compute_hash(self) -> str:
         """
@@ -311,6 +312,7 @@ class ModelConfig:
         factors.append(self.override_generation_config)
         factors.append(self.rope_scaling)
         factors.append(self.rope_theta)
+        factors.append(self.video_pruning_rate)
 
         # hf_config can control how the model looks!
         try:
@@ -338,17 +340,19 @@ class ModelConfig:
         return hashlib.sha256(str(factors).encode()).hexdigest()
 
     def __post_init__(
-            self,
-            # Multimodal config init vars
-            limit_mm_per_prompt: Optional[dict[str, int]],
-            media_io_kwargs: Optional[dict[str, dict[str, Any]]],
-            mm_processor_kwargs: Optional[dict[str, Any]],
-            mm_processor_cache_gb: Optional[float],
-            mm_processor_cache_type: Optional[MMCacheType],
-            mm_shm_cache_max_object_size_mb: Optional[int],
-            mm_encoder_tp_mode: Optional[MMEncoderTPMode],
-            interleave_mm_strings: Optional[bool],
-            skip_mm_profiling: Optional[bool]) -> None:
+        self,
+        # Multimodal config init vars
+        limit_mm_per_prompt: Optional[dict[str, int]],
+        media_io_kwargs: Optional[dict[str, dict[str, Any]]],
+        mm_processor_kwargs: Optional[dict[str, Any]],
+        mm_processor_cache_gb: Optional[float],
+        mm_processor_cache_type: Optional[MMCacheType],
+        mm_shm_cache_max_object_size_mb: Optional[int],
+        mm_encoder_tp_mode: Optional[MMEncoderTPMode],
+        interleave_mm_strings: Optional[bool],
+        skip_mm_profiling: Optional[bool],
+        video_pruning_rate: Optional[float],
+    ) -> None:
         # Set the default seed to 0 in V1.
         # NOTE(woosuk): In V0, we set the default seed to None because the
         # driver worker shares the same process as the user process, and thus
@@ -612,6 +616,7 @@ class ModelConfig:
                 mm_encoder_tp_mode=mm_encoder_tp_mode,
                 interleave_mm_strings=interleave_mm_strings,
                 skip_mm_profiling=skip_mm_profiling,
+                video_pruning_rate=video_pruning_rate,
             )
 
             mm_config_kwargs = {

vllm/config/multimodal.py

@@ -78,6 +78,11 @@ class MultiModalConfig:
     This reduces engine startup time but shifts the responsibility to users for
     estimating the peak memory usage of the activation of multimodal encoder and
     embedding cache."""
+    video_pruning_rate: Optional[float] = None
+    """Sets pruning rate for video pruning via Efficient Video Sampling.
+    Value sits in range [0;1) and determines fraction of media tokens
+    from each video to be pruned.
+    """
 
     def compute_hash(self) -> str:
         """
@@ -118,3 +123,7 @@ class MultiModalConfig:
         """
         kwargs = self.mm_processor_kwargs or {}
         return kwargs | dict(inference_kwargs)
+
+    def is_multimodal_pruning_enabled(self):
+        return (self.video_pruning_rate is not None
+                and self.video_pruning_rate > 0)

vllm/engine/arg_utils.py

@@ -391,6 +391,7 @@ class EngineArgs:
     mm_encoder_tp_mode: MMEncoderTPMode = MultiModalConfig.mm_encoder_tp_mode
     io_processor_plugin: Optional[str] = None
     skip_mm_profiling: bool = MultiModalConfig.skip_mm_profiling
+    video_pruning_rate: float = MultiModalConfig.video_pruning_rate
     # LoRA fields
     enable_lora: bool = False
     enable_lora_bias: bool = LoRAConfig.bias_enabled
@@ -813,6 +814,9 @@ class EngineArgs:
         multimodal_group.add_argument("--skip-mm-profiling",
                                       **multimodal_kwargs["skip_mm_profiling"])
 
+        multimodal_group.add_argument(
+            "--video-pruning-rate", **multimodal_kwargs["video_pruning_rate"])
+
         # LoRA related configs
         lora_kwargs = get_kwargs(LoRAConfig)
         lora_group = parser.add_argument_group(
@@ -1032,6 +1036,7 @@ class EngineArgs:
             model_impl=self.model_impl,
             override_attention_dtype=self.override_attention_dtype,
             logits_processors=self.logits_processors,
+            video_pruning_rate=self.video_pruning_rate,
             io_processor_plugin=self.io_processor_plugin,
         )
 

vllm/model_executor/models/interfaces.py

@@ -115,6 +115,42 @@ class SupportsMultiModal(Protocol):
         ...
 
 
+@runtime_checkable
+class SupportsMultiModalPruning(Protocol):
+    """The interface required for models that support returning both input
+    embeddings and positions. Model may require custom positions for dynamic
+    pruning of multimodal embeddings.
+    """
+    supports_multimodal_pruning: ClassVar[Literal[True]] = True
+
+    def recompute_mrope_positions(
+            self, input_ids: list[int],
+            multimodal_embeddings: MultiModalEmbeddings,
+            mrope_positions: torch.LongTensor, num_computed_tokens: int
+    ) -> tuple[MultiModalEmbeddings, Tensor, int]:
+        """
+        Update part of input mrope positions (starting with
+        num_computed_tokens index). Original mrope_positions are computed
+        for unpruned sequence and becomes incorrect once pruning occurs,
+        so once we prune media tokens we should reflect this in the
+        mrope_positions before we feed it to LLM.
+
+        Args:
+            input_ids: (N,) All input tokens of the prompt containing
+                entire sequence.
+            multimodal_embeddings: Tuple of multimodal embeddings that
+                fits into the prefill chunk that is being processed.
+            mrope_positions: Existing mrope positions (3, N) for entire
+                sequence
+            num_computed_tokens: A number of computed tokens so far.
+
+        Returns:
+            Tuple of (multimodal_embeddings, mrope_positions,
+                mrope_position_delta).
+        """
+        ...
+
+
 @overload
 def supports_multimodal(
         model: type[object]) -> TypeIs[type[SupportsMultiModal]]:
@@ -142,6 +178,25 @@ def supports_multimodal_encoder_tp_data(
     return getattr(model, "supports_encoder_tp_data", False)
 
 
+@overload
+def supports_multimodal_pruning(
+        model: type[object]) -> TypeIs[type[SupportsMultiModalPruning]]:
+    ...
+
+
+@overload
+def supports_multimodal_pruning(
+        model: object) -> TypeIs[SupportsMultiModalPruning]:
+    ...
+
+
+def supports_multimodal_pruning(
+    model: Union[type[object], object],
+) -> Union[TypeIs[type[SupportsMultiModalPruning]],
+           TypeIs[SupportsMultiModalPruning]]:
+    return getattr(model, "supports_multimodal_pruning", False)
+
+
 @runtime_checkable
 class SupportsScoreTemplate(Protocol):
     """The interface required for all models that support score template."""

vllm/model_executor/models/qwen2_5_vl.py

@@ -25,9 +25,9 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """Inference-only Qwen2.5-VL model compatible with HuggingFace weights."""
-from collections.abc import Iterable, Mapping
+from collections.abc import Iterable, Mapping, Sequence
 from functools import lru_cache, partial
-from typing import Annotated, Callable, Literal, Optional, Union
+from typing import Annotated, Any, Callable, Literal, Optional, Union
 
 import torch
 import torch.nn as nn
@@ -58,7 +58,13 @@ from vllm.model_executor.layers.quantization.gptq_marlin import (
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 from vllm.model_executor.models.module_mapping import MultiModelKeys
 from vllm.multimodal import MULTIMODAL_REGISTRY
-from vllm.multimodal.inputs import MultiModalFieldConfig
+from vllm.multimodal.evs import (compute_mrope_for_media,
+                                 compute_retained_tokens_count,
+                                 compute_retention_mask,
+                                 recompute_mrope_positions)
+from vllm.multimodal.inputs import MultiModalFieldConfig, MultiModalKwargs
+from vllm.multimodal.parse import MultiModalDataItems
+from vllm.multimodal.processing import PromptReplacement, PromptUpdate
 from vllm.platforms import _Backend
 from vllm.sequence import IntermediateTensors
 from vllm.transformers_utils.config import uses_mrope
@@ -66,7 +72,8 @@ from vllm.utils import is_pin_memory_available
 from vllm.utils.tensor_schema import TensorSchema, TensorShape
 
 from .interfaces import (MultiModalEmbeddings, SupportsLoRA,
-                         SupportsMultiModal, SupportsPP, SupportsQuant)
+                         SupportsMultiModal, SupportsMultiModalPruning,
+                         SupportsPP, SupportsQuant)
 from .qwen2_vl import Qwen2VLDummyInputsBuilder as Qwen2_5_VLDummyInputsBuilder
 from .qwen2_vl import (Qwen2VLMultiModalProcessor, Qwen2VLProcessingInfo,
                        apply_rotary_pos_emb_vision)
@@ -86,9 +93,9 @@ class Qwen2_5_VLImagePixelInputs(TensorSchema):
         - np: Number of patches
         - ni: Number of images
         - cps: Number of channels * patch_size * patch_size
-    
+
     Historical context:
-        - pixel_values shape: (num_patches, num_channels * patch_size * 
+        - pixel_values shape: (num_patches, num_channels * patch_size *
           patch_size)
         - image_grid_thw shape: (num_images, 3) in (grid_t, grid_h, grid_w)
           formatnum_channels * patch_size * patch_size
@@ -112,7 +119,7 @@ class Qwen2_5_VLImageEmbeddingInputs(TensorSchema):
         - nf: Number of image features
         - hs: Hidden size
         - ni: Number of images
-    
+
     Historical context:
         - image_embeds shape: (num_image_features, hidden_size)
         - num_image_features varies based on the number and resolution of the
@@ -143,11 +150,11 @@ class Qwen2_5_VLVideoPixelInputs(TensorSchema):
     Dimensions:
         - np: Number of patches
         - nv: Number of videos
-        - ctps: Number of channels * temporal_patch_size * patch_size * 
+        - ctps: Number of channels * temporal_patch_size * patch_size *
           patch_size
-    
+
     Historical context:
-        - pixel_values_videos shape: (num_patches, num_channels * 
+        - pixel_values_videos shape: (num_patches, num_channels *
           temporal_patch_size * patch_size * patch_size)
         - video_grid_thw shape: (num_videos, 3) in (grid_t, grid_h, grid_w)
           format
@@ -179,7 +186,7 @@ class Qwen2_5_VLVideoEmbeddingInputs(TensorSchema):
         - nf: Number of video features
         - hs: Hidden size
         - nv: Number of videos
-    
+
     Historical context:
         - video_embeds shape: (num_video_features, hidden_size)
         - num_video_features varies based on the number and resolution of the
@@ -905,6 +912,55 @@ class Qwen2_5_VLMultiModalProcessor(Qwen2VLMultiModalProcessor):
             second_per_grid_ts=MultiModalFieldConfig.batched("video"),
         )
 
+    def _get_prompt_updates(
+        self,
+        mm_items: MultiModalDataItems,
+        hf_processor_mm_kwargs: Mapping[str, Any],
+        out_mm_kwargs: MultiModalKwargs,
+    ) -> Sequence[PromptUpdate]:
+        hf_processor = self.info.get_hf_processor(**hf_processor_mm_kwargs)
+        image_processor = self.info.get_image_processor(
+            **hf_processor_mm_kwargs)
+        tokenizer = self.info.get_tokenizer()
+        vocab = tokenizer.get_vocab()
+
+        placeholder = {
+            "image": vocab[hf_processor.image_token],
+            "video": vocab[hf_processor.video_token],
+        }
+
+        merge_length = image_processor.merge_size**2
+
+        def get_replacement_qwen2vl(item_idx: int, modality: str):
+            out_item = out_mm_kwargs[modality][item_idx]
+            grid_thw = out_item[f"{modality}_grid_thw"].data
+            assert isinstance(grid_thw, torch.Tensor)
+
+            num_tokens = int(grid_thw.prod()) // merge_length
+
+            # EVS-specific code
+            video_pruning_rate = self.info.ctx.get_mm_config(
+            ).video_pruning_rate
+            if (modality == "video" and video_pruning_rate is not None
+                    and video_pruning_rate > 0.0):
+                num_tokens = compute_retained_tokens_count(
+                    grid_thw,
+                    image_processor.merge_size,
+                    video_pruning_rate,
+                )
+            # End of EVS-specific code
+
+            return [placeholder[modality]] * num_tokens
+
+        return [
+            PromptReplacement(
+                modality=modality,
+                target=[placeholder[modality]],
+                replacement=partial(get_replacement_qwen2vl,
+                                    modality=modality),
+            ) for modality in ("image", "video")
+        ]
+
 
 @MULTIMODAL_REGISTRY.register_processor(
     Qwen2_5_VLMultiModalProcessor,
@@ -912,7 +968,8 @@ class Qwen2_5_VLMultiModalProcessor(Qwen2VLMultiModalProcessor):
     dummy_inputs=Qwen2_5_VLDummyInputsBuilder)
 class Qwen2_5_VLForConditionalGeneration(nn.Module, SupportsMultiModal,
                                          SupportsLoRA, SupportsPP,
-                                         SupportsQuant):
+                                         SupportsQuant,
+                                         SupportsMultiModalPruning):
 
     packed_modules_mapping = {
         "qkv_proj": ["q_proj", "k_proj", "v_proj"],
@@ -949,6 +1006,9 @@ class Qwen2_5_VLForConditionalGeneration(nn.Module, SupportsMultiModal,
         self.use_data_parallel = multimodal_config.mm_encoder_tp_mode == "data"
         self.config = config
         self.multimodal_config = multimodal_config
+        self.video_pruning_rate = multimodal_config.video_pruning_rate
+        self.is_multimodal_pruning_enabled = (
+            multimodal_config.is_multimodal_pruning_enabled())
 
         if multimodal_config.get_limit_per_prompt("image") or \
             multimodal_config.get_limit_per_prompt("video"):
@@ -1090,6 +1150,36 @@ class Qwen2_5_VLForConditionalGeneration(nn.Module, SupportsMultiModal,
 
         return image_embeds.split(sizes)
 
+    def _postprocess_image_embeds_evs(
+            self, image_embeds_split: tuple[torch.Tensor, ...],
+            image_input: Qwen2_5_VLImageInputs) -> tuple[torch.Tensor, ...]:
+        """
+        Append mrope positions for each for images.
+        This is necessary to recover correct mrope
+        positions after video pruning
+
+        Args:
+            image_embeds_split: Tuple of image embeddings for
+                each image item.
+            image_input: Image input data.
+
+        Returns:
+            Tuple of image embeddings for each image item.
+            Resulting embeddings will have extra 4 channels for
+            computed mrope positions.
+        """
+        merge_size = self.visual.spatial_merge_size
+        grid_thw = image_input["image_grid_thw"]
+        grid_thw_list = grid_thw.tolist()
+        image_embeds_out = []
+        for emb, size in zip(image_embeds_split, grid_thw_list):
+            positions = compute_mrope_for_media(size,
+                                                merge_size).to(emb.device)
+            emb = torch.cat([emb, positions], dim=1)
+            image_embeds_out.append(emb)
+        image_embeds_split = image_embeds_out
+        return tuple(image_embeds_split)
+
     def _process_video_input(
             self,
             video_input: Qwen2_5_VLVideoInputs) -> tuple[torch.Tensor, ...]:
@@ -1119,6 +1209,114 @@ class Qwen2_5_VLForConditionalGeneration(nn.Module, SupportsMultiModal,
 
         return video_embeds.split(sizes)
 
+    def _postprocess_video_embeds_evs(
+            self, video_embeds_split: tuple[torch.Tensor, ...],
+            video_input: Qwen2_5_VLVideoInputs) -> tuple[torch.Tensor, ...]:
+        """
+        Prunes video embeddings via Efficient Video Sampling (EVS)
+        and then appends mrope positions for each retained embeddings
+
+        Args:
+            video_embeds_split: Tuple of video embeddings for each video item.
+            video_input: Video input data.
+
+        Returns:
+            Tuple of video embeddings for each video item.
+            Resulting embeddings will have extra 4 channels for
+            computed mrope positions.
+        """
+        grid_thw = video_input["video_grid_thw"]
+        assert grid_thw.ndim == 2
+        grid_thw_list = grid_thw.tolist()
+        merge_size = self.visual.spatial_merge_size
+
+        # Cast to long to match the original code
+        # https://github.com/huggingface/transformers/blob/41980ce93e775f6c88500c51c8db7946fc6a2add/src/transformers/models/qwen2_5_vl/modular_qwen2_5_vl.py#L491 # noqa
+        second_per_grid_ts = video_input["second_per_grid_ts"].long()
+        tokens_per_second = self.config.vision_config.tokens_per_second
+
+        video_embeds_out = []
+        for emb, size, video_second_per_grid_t in zip(video_embeds_split,
+                                                      grid_thw_list,
+                                                      second_per_grid_ts):
+            # For each video, we compute retention mask using EVS
+            retention_mask = compute_retention_mask(
+                emb,
+                size,
+                spatial_merge_size=self.visual.spatial_merge_size,
+                q=self.video_pruning_rate,
+            )
+            positions = compute_mrope_for_media(
+                size,
+                merge_size,
+                tokens_per_second=tokens_per_second,
+                video_second_per_grid=video_second_per_grid_t.item(),
+            ).to(emb.device)
+
+            emb = emb[retention_mask]
+            positions = positions[retention_mask]
+            emb = torch.cat([emb, positions], dim=1)
+            video_embeds_out.append(emb)
+        return tuple(video_embeds_out)
+
+    def recompute_mrope_positions(
+        self,
+        input_ids: list[int],
+        multimodal_embeddings: tuple[torch.Tensor, ...],
+        mrope_positions: torch.LongTensor,
+        num_computed_tokens: int,
+    ) -> tuple[tuple[torch.Tensor, ...], torch.Tensor, int]:
+        """
+        Update part of input mrope positions (starting with
+        num_computed_tokens index). Original mrope_positions are computed
+        for unpruned sequence and becomes incorrect once pruning occurs,
+        so once we prune media tokens we should reflect this in the
+        mrope_positions before we feed it to LLM.
+
+        Args:
+            input_ids: (N,) All input tokens of the prompt (Containing
+                entire sequence).
+            multimodal_embeddings: Tuple of multimodal embeddings.
+            mrope_positions: Existing mrope positions (3, N) for entire
+                sequence
+            num_computed_tokens: A number of computed tokens so far.
+
+        Returns:
+            Tuple of (multimodal_embeddings, mrope_positions,
+                mrope_position_delta).
+        """
+        image_token_id = self.config.image_token_id
+        video_token_id = self.config.video_token_id
+        vision_start_token_id = self.config.vision_start_token_id
+
+        # Device
+        device = (multimodal_embeddings[0].device
+                  if len(multimodal_embeddings) else mrope_positions.device)
+
+        # Tensors
+        input_ids_t = torch.as_tensor(input_ids,
+                                      device=device,
+                                      dtype=torch.long)
+
+        # fmt: off
+        mm_embeddings_out = [mm[:, :-4] for mm in
+                             multimodal_embeddings]
+        mm_embeddings_pos = [mm[:, -4:].permute(1, 0).long() for mm in
+                             multimodal_embeddings]
+        # fmt: in
+
+        positions, mrope_positions_delta = recompute_mrope_positions(
+            input_ids_t,
+            mm_embeddings_pos,
+            mrope_positions,
+            num_computed_tokens,
+            vision_start_token_id,
+            image_token_id,
+            video_token_id,
+        )
+
+        return tuple(mm_embeddings_out), positions, mrope_positions_delta
+
     def _parse_and_validate_multimodal_inputs(self, **kwargs: object) -> dict:
         mm_input_by_modality = {}
 
@@ -1156,9 +1354,17 @@ class Qwen2_5_VLForConditionalGeneration(nn.Module, SupportsMultiModal,
             multimodal_input = mm_input_by_modality[modality]
             if modality == "image":
                 vision_embeddings = self._process_image_input(multimodal_input)
+                if self.is_multimodal_pruning_enabled:
+                    vision_embeddings = self._postprocess_image_embeds_evs(
+                        vision_embeddings, multimodal_input
+                    )
                 multimodal_embeddings += vision_embeddings
             if modality == "video":
                 video_embeddings = self._process_video_input(multimodal_input)
+                if self.is_multimodal_pruning_enabled:
+                    video_embeddings = self._postprocess_video_embeds_evs(
+                        video_embeddings, multimodal_input
+                    )
                 multimodal_embeddings += video_embeddings
         return multimodal_embeddings
 
@@ -1184,6 +1390,10 @@ class Qwen2_5_VLForConditionalGeneration(nn.Module, SupportsMultiModal,
         inputs_embeds = self.get_input_embeddings(input_ids)
         if image_input is not None:
             image_embeds = self._process_image_input(image_input)
+            if self.is_multimodal_pruning_enabled:
+                image_embeds = self._postprocess_image_embeds_evs(
+                    image_embeds, image_input
+                )
             inputs_embeds = merge_multimodal_embeddings(
                 input_ids,
                 inputs_embeds,
@@ -1193,6 +1403,10 @@ class Qwen2_5_VLForConditionalGeneration(nn.Module, SupportsMultiModal,
 
         if video_input is not None:
             video_embeds = self._process_video_input(video_input)
+            if self.is_multimodal_pruning_enabled:
+                video_embeds = self._postprocess_video_embeds_evs(
+                    video_embeds, video_input
+                )
             inputs_embeds = merge_multimodal_embeddings(
                 input_ids,
                 inputs_embeds,

vllm/multimodal/evs.py

@@ -0,0 +1,273 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+import typing
+
+import torch
+
+
+def compute_retained_tokens_count(video_size_thw: torch.LongTensor,
+                                  spatial_merge_size: int, q: float) -> int:
+    """
+    Compute the number of retained tokens for a given video.
+    Method ensures that we retain all the tokens from the first frame
+    regardless of the pruning rate.
+
+    Args:
+        video_size_thw: The size of the video in the format of (T, H, W).
+        spatial_merge_size: The size of the spatial merge.
+        q: The pruning rate.
+
+    Returns:
+        The number of retained tokens.
+    """
+    T, H, W = map(int, video_size_thw)
+    min_num_tokens = (H // spatial_merge_size) * (W // spatial_merge_size)
+    evs_num_tokens = int(T * min_num_tokens * (1 - q))
+    return max(min_num_tokens, evs_num_tokens)
+
+
+def compute_retention_mask(
+    video_embeds: torch.Tensor,
+    video_size_thw: torch.LongTensor,
+    spatial_merge_size: int,
+    q: float,
+) -> torch.Tensor:
+    """
+    Computes the retention mask for input video embeddings.
+
+    Args:
+        video_embeds (`torch.Tensor`): The input video embeddings
+            of shape `(T * H * W // spatial_merge_size ^ 2, hidden_size)`
+        video_size_thw (`torch.LongTensor` of shape `(3)`):
+            The temporal, height and width of video.
+        spatial_merge_size: Size reduction for rows & cols dimensions.
+        q: (`float`): Pruning rate factor [0,1)
+
+    Returns:
+        `torch.Tensor`: The retention mask for the video embeddings of
+            `(T * H * W // spatial_merge_size ^ 2)` shape.
+    """
+    T, H, W = video_size_thw
+
+    # Use reshape instead of einops to avoid graph breaks
+    video_embeds = video_embeds.reshape(
+        T,
+        H // spatial_merge_size,
+        W // spatial_merge_size,
+        video_embeds.size(-1),
+    )
+
+    # Core EVS
+    similarity = torch.nn.functional.cosine_similarity(video_embeds[1:, ...],
+                                                       video_embeds[:-1, ...],
+                                                       dim=-1)
+    dissimilarity = 1 - similarity
+
+    # Always ensure we include all tokens from the first frame
+    dissimilarity = torch.cat(
+        [255 * torch.ones_like(video_embeds[:1, :, :, 0]), dissimilarity],
+        dim=0)
+
+    dissimilarity_flat = dissimilarity.view(-1)
+    order = torch.argsort(dissimilarity_flat,
+                          dim=-1,
+                          descending=True,
+                          stable=True)
+    retain_num_tokens = compute_retained_tokens_count(video_size_thw,
+                                                      spatial_merge_size, q)
+    topk_indices = order[:retain_num_tokens]
+
+    retention_mask = torch.zeros_like(dissimilarity_flat, dtype=torch.bool)
+    retention_mask[topk_indices] = True
+    retention_mask = retention_mask.reshape(dissimilarity.size())
+
+    mask = retention_mask.view(-1)  # "T H W -> (T H W)"
+    return mask
+
+
+def compute_mrope_for_media(
+    video_size_thw: torch.LongTensor,
+    spatial_merge_size: int,
+    tokens_per_second: float = 1.0,
+    video_second_per_grid: float = 1.0,
+) -> torch.Tensor:
+    """
+    Computes the mrope for video embeddings based on the grid dimensions.
+    Computed mrope positions match original qwen 2.5 implementation,
+    but positions are built for media being the first element in sequence.
+
+    Args:
+        video_size_thw: Media size (num frames, rows, cols)
+        spatial_merge_size: Size reduction for rows & cols dimensions.
+        tokens_per_second: Number of tokens per second.
+        video_second_per_grid: Number of seconds per video.
+
+    Returns:
+        Tensor of shape `(T * H * W, 4)` where last dimension
+        represents mrope positions [0:3), while the last channel
+        contains value of llm_grid_w repeated for all positions.
+    """
+    llm_grid_t = video_size_thw[0]
+    llm_grid_h = video_size_thw[1] // spatial_merge_size
+    llm_grid_w = video_size_thw[2] // spatial_merge_size
+
+    t_index = ((torch.arange(llm_grid_t).view(-1, 1).expand(
+        -1, llm_grid_h * llm_grid_w).mul(
+            tokens_per_second * video_second_per_grid)).long().flatten())
+    h_index = (torch.arange(llm_grid_h).view(1, -1,
+                                             1).expand(llm_grid_t, -1,
+                                                       llm_grid_w).flatten())
+    w_index = (torch.arange(llm_grid_w).view(1, 1, -1).expand(
+        llm_grid_t, llm_grid_h, -1).flatten())
+    llm_grid_w = (torch.tensor([llm_grid_w
+                                ]).view(1, 1,
+                                        1).expand(llm_grid_t, llm_grid_h,
+                                                  llm_grid_w).flatten())
+
+    positions = torch.stack([t_index, h_index, w_index, llm_grid_w], dim=1)
+    return positions
+
+
+def recompute_mrope_positions(
+    input_ids: torch.LongTensor,
+    multimodal_positions: list[torch.Tensor],
+    mrope_positions: torch.LongTensor,
+    num_computed_tokens: int,
+    vision_start_token_id: int,
+    image_token_id: int,
+    video_token_id: int,
+) -> tuple[torch.LongTensor, int]:
+    """
+    Update part of input mrope positions.
+    Original mrope_positions are computed incorrectly, so once we prune media
+    tokens we should reflect this in the mrope positions for the LLM.
+
+    This method supports chunked prefill approach where
+    multimodal_embeddings are passed to LLM in chunks, so input
+    multimodal_embeddings may contain zero, some or even some part of all
+    multimodal_embeddings for a given prompt.
+
+    Each multimodal_positions has 4 extra channels
+    (First 3 channels corresponds to original 3 mrope positions, last channel
+    is the maximum width of the media repeated). Provided multimodal_positions
+    do not reflect location of media position in sequence - they are computed
+    like the media is in the 0-th position in the sequence.
+
+    Method works as follows: it recomputes mrope_positions starting from the
+    `num_computed_tokens` for `total_len_of_multimodal_embeddings` and then
+    shifts all text tokens that goes after total_len_of_multimodal_embeddings.
+
+    It also handles case when multimodal_embeddings is partial
+    (e.g. one media is split into two prefill stages)
+
+    Args:
+        input_ids: (N,) All input tokens of the prompt (entire sequence).
+        multimodal_positions: List of mrope positsions for each media.
+        mrope_positions: Existing mrope positions (4, N) for entire sequence.
+        num_computed_tokens: A number of computed tokens so far.
+        vision_start_token_id: Token indicating start of vision media.
+        image_token_id: Image token id
+        video_token_id: Video token id
+
+    Returns:
+        Tuple of (mrope_positions, mrope_position_delta).
+    """
+
+    # Tensors
+    positions: torch.LongTensor = typing.cast(
+        torch.LongTensor, mrope_positions.clone())  # (3, N)
+    N = input_ids.numel()
+
+    image_mask = input_ids.eq(image_token_id)
+    video_mask = input_ids.eq(video_token_id)
+    media_mask = image_mask | video_mask
+    text_mask = ~media_mask
+
+    # Early exit: no media in this chunk
+    if len(multimodal_positions) == 0:
+        delta = (int((positions.max().item() + 1) -
+                     N) if positions.numel() else -N)
+        return positions, delta
+
+    total_mm_tokens = torch.count_nonzero(media_mask)
+    seen_mm_tokens = torch.count_nonzero(media_mask[:num_computed_tokens])
+
+    # Early exit: we've updated positions for all media tokens
+    # (and consequently - for all remaining text tokens)
+    if seen_mm_tokens == total_mm_tokens:
+        delta = (int((positions.max().item() + 1) -
+                     N) if positions.numel() else -N)
+        return positions, delta
+
+    vision_start_indices = (input_ids == vision_start_token_id).nonzero(
+        as_tuple=True)[0]
+
+    for mm_pos in multimodal_positions:
+        # Each mm_pos can be a complete embedding for single media
+        # or it can be a part of a single media (due to chunked prefill)
+
+        # Cases to cover
+        # - Current prefill chunk has no vision start indexes at all
+        # - Vision start token appeared in previous prefill round
+        # - Regular case
+        seen_vision_start_indices = vision_start_indices[vision_start_indices <
+                                                         num_computed_tokens]
+
+        if len(seen_vision_start_indices):
+            # If we have encountered some vision start indexes,
+            # then we should check the condition:
+            # | --- prefill 1 ------| ---- prefill 2 ----- |
+            # | TTTTTTTTTSVVVVVVVVVV|VVVVVVTTTTTTTTTTTTTTTT|
+            last_vision_start_token = seen_vision_start_indices[-1]
+            seem_mm_tokens_before_last_vision_start = torch.count_nonzero(
+                media_mask[:last_vision_start_token])
+            in_the_middle_of_media = (
+                seen_mm_tokens > seem_mm_tokens_before_last_vision_start)
+
+            if in_the_middle_of_media:
+                mm_embeddings_seen = (seen_mm_tokens -
+                                      seem_mm_tokens_before_last_vision_start)
+                global_mm_start = last_vision_start_token
+            else:
+                # We have completed previous mm_embedding part and
+                # ready to start a new one
+                next_vision_start_token = vision_start_indices[
+                    vision_start_indices >= num_computed_tokens][0]
+                mm_embeddings_seen = 0
+                global_mm_start = next_vision_start_token
+
+        else:
+            # If there were no vision start indexes so far,
+            # let's find first vision start index
+            next_vision_start_token = vision_start_indices[
+                vision_start_indices >= num_computed_tokens][0]
+
+            mm_embeddings_seen = 0
+            global_mm_start = next_vision_start_token
+
+        # Offset right after vision_start_token
+        base = positions[-1, global_mm_start] + 1
+        local_start = global_mm_start + 1 + mm_embeddings_seen
+        local_end = local_start + mm_pos.shape[1]
+        positions[:, local_start:local_end] = mm_pos[0:3] + base
+
+        # mm_pos[3, 0] is the max width of the media
+        offset = mm_pos[3, 0] + base
+
+        text_pos_sum = torch.cumsum(text_mask[local_end:].long(), dim=0)
+
+        positions[:, local_end:N] = text_pos_sum + offset - 1
+
+        # Include distance to the next vision start token
+        num_computed_tokens += mm_pos.shape[1]
+
+    mrope_positions_delta = (positions.max() + 1 - N).item()
+    return positions, mrope_positions_delta

vllm/v1/worker/gpu_model_runner.py

@@ -40,11 +40,15 @@ from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase
 from vllm.model_executor.layers.mamba.abstract import MambaBase
 from vllm.model_executor.layers.rotary_embedding import MRotaryEmbedding
 from vllm.model_executor.model_loader import TensorizerLoader, get_model_loader
+# yapf conflicts with isort for this block
+# yapf: disable
 from vllm.model_executor.models.interfaces import (SupportsMultiModal,
                                                    is_mixture_of_experts,
                                                    supports_eagle3,
                                                    supports_mrope,
+                                                   supports_multimodal_pruning,
                                                    supports_transcription)
+# yapf: enable
 from vllm.model_executor.models.interfaces_base import (
     VllmModelForPooling, is_pooling_model, is_text_generation_model)
 from vllm.multimodal import MULTIMODAL_REGISTRY
@@ -206,7 +210,8 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
         self.enable_prompt_embeds = model_config.enable_prompt_embeds
         self.is_multimodal_raw_input_only_model = (
             model_config.is_multimodal_raw_input_only_model)
-
+        # This will be overridden in load_model()
+        self.is_multimodal_pruning_enabled = False
         self.max_model_len = model_config.max_model_len
         self.dcp_world_size = self.parallel_config.decode_context_parallel_size
         self.max_num_tokens = scheduler_config.max_num_batched_tokens
@@ -1530,29 +1535,47 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
         # encoder outputs.
         model = cast(SupportsMultiModal, self.model)
         encoder_outputs = []
-        for _, num_items, mm_kwargs_group in group_mm_kwargs_by_modality(
+        for modality, num_items, mm_kwargs_group in group_mm_kwargs_by_modality(
                 mm_kwargs,
                 device=self.device,
                 pin_memory=self.pin_memory,
                 merge_by_field_config=model.merge_by_field_config,
         ):
-            # Run the encoder.
-            # `curr_group_outputs` is either of the following:
-            # 1. A tensor of shape (num_items, feature_size, hidden_size)
-            # in case feature_size is fixed across all multimodal items.
-            # 2. A list or tuple (length: num_items) of tensors, each of shape
-            # (feature_size, hidden_size) in case the feature size is dynamic
-            # depending on the input multimodal items.
-            curr_group_outputs = model.get_multimodal_embeddings(
-                **mm_kwargs_group)
+            # (ekhvedchenia): Temporary hack to limit peak memory usage when
+            # processing multimodal data.This solves the issue with scheduler
+            # putting too many video samples into a single batch. Scheduler
+            # uses pruned vision tokens count to compare it versus compute
+            # budget which is incorrect (Either input media size or non-pruned
+            # output vision tokens count should be considered)
+            curr_group_outputs = []
+
+            if self.is_multimodal_pruning_enabled and modality == "video":
+                micro_batch_size = 1
+                for i in range(0, num_items, micro_batch_size):
+                    micro_batch_mm_inputs = dict(
+                        (k, v[i:i + micro_batch_size])
+                        for k, v in mm_kwargs_group.items())
+
+                    micro_batch_outputs = model.get_multimodal_embeddings(
+                        **micro_batch_mm_inputs)
+
+                    curr_group_outputs.extend(micro_batch_outputs)
+            else:
+                # Run the encoder.
+                # `curr_group_outputs` is either of the following:
+                # 1. A tensor of shape (num_items, feature_size, hidden_size)
+                # in case feature_size is fixed across all multimodal items.
+                # 2. A list or tuple (length: num_items) of tensors,
+                # each of shape (feature_size, hidden_size) in case the feature
+                # size is dynamic depending on the input multimodal items.
+                curr_group_outputs = model.get_multimodal_embeddings(
+                    **mm_kwargs_group)
 
             sanity_check_mm_encoder_outputs(
                 curr_group_outputs,
                 expected_num_items=num_items,
             )
-
-            for output in curr_group_outputs:
-                encoder_outputs.append(output)
+            encoder_outputs.extend(curr_group_outputs)
 
         # Cache the encoder outputs by mm_hash
         for (mm_hash, pos_info), output in zip(mm_hashes_pos, encoder_outputs):
@@ -1566,8 +1589,11 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
         scheduler_output: "SchedulerOutput",
         shift_computed_tokens: int = 0,
     ) -> list[torch.Tensor]:
+        should_sync_mrope_positions = False
         mm_embeds: list[torch.Tensor] = []
         for req_id in self.input_batch.req_ids:
+            mm_embeds_req: list[torch.Tensor] = []
+
             num_scheduled_tokens = scheduler_output.num_scheduled_tokens[
                 req_id]
             req_state = self.requests[req_id]
@@ -1609,7 +1635,28 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
                     encoder_output[start_idx:end_idx],
                     is_embed=is_embed,
                 )
-                mm_embeds.append(mm_embeds_item)
+                mm_embeds_req.append(mm_embeds_item)
+
+            if self.is_multimodal_pruning_enabled and self.uses_mrope:
+                should_sync_mrope_positions = True
+                mm_embeds_req, new_mrope_positions, new_delta = (
+                    self.model.recompute_mrope_positions(
+                        input_ids=req_state.prompt_token_ids,
+                        multimodal_embeddings=mm_embeds_req,
+                        mrope_positions=req_state.mrope_positions,
+                        num_computed_tokens=req_state.num_computed_tokens,
+                    ))
+                assert req_state.mrope_positions is not None
+                req_state.mrope_positions.copy_(new_mrope_positions)
+                req_state.mrope_position_delta = new_delta
+
+            mm_embeds.extend(mm_embeds_req)
+
+        if should_sync_mrope_positions:
+            self._calc_mrope_positions(scheduler_output)
+            self.mrope_positions.copy_to_gpu(
+                scheduler_output.total_num_scheduled_tokens)
+
         return mm_embeds
 
     def _extract_encoder_inputs(
@@ -2589,6 +2636,10 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
                     time_after_load - time_before_load)
         prepare_communication_buffer_for_model(self.model)
 
+        self.is_multimodal_pruning_enabled = (supports_multimodal_pruning(
+            self.model) and self.model_config.multimodal_config.
+                                              is_multimodal_pruning_enabled())
+
         if is_mixture_of_experts(
                 self.model) and self.parallel_config.enable_eplb:
             logger.info("EPLB is enabled for model %s.",
@@ -2843,7 +2894,7 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
         Args:
             num_tokens: Number of tokens to run the dummy forward pass.
             cudagraph_runtime_mode: used to control the behavior.
-                - if not set will determine the cudagraph mode based on using 
+                - if not set will determine the cudagraph mode based on using
                     the self.cudagraph_dispatcher.
                 - CUDAGraphMode.NONE: No cudagraph, for warm up and profile run
                 - CUDAGraphMode.PIECEWISE: Piecewise cudagraph.