[Model] EVS support for nano_nemotron_vl (#26269)

Signed-off-by: Tomer Asida <57313761+tomeras91@users.noreply.github.com>
Signed-off-by: tomeras91 <57313761+tomeras91@users.noreply.github.com>
Signed-off-by: Eugene Khvedchenia <ekhvedchenia@nvidia.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Co-authored-by: Eugene Khvedchenia <ekhvedchenia@nvidia.com>

vllm/model_executor/models/nano_nemotron_vl.py

@@ -30,6 +30,7 @@ from vllm.model_executor.models.interfaces import (
     IsHybrid,
     MultiModalEmbeddings,
     SupportsMultiModal,
+    SupportsMultiModalPruning,
 )
 from vllm.model_executor.models.internvl import (
     calculate_internvl_targets,
@@ -44,6 +45,10 @@ from vllm.model_executor.models.utils import (
     maybe_prefix,
 )
 from vllm.multimodal import MULTIMODAL_REGISTRY
+from vllm.multimodal.evs import (
+    compute_retained_tokens_count,
+    compute_retention_mask,
+)
 from vllm.multimodal.inputs import (
     MultiModalDataDict,
     MultiModalFieldConfig,
@@ -62,13 +67,20 @@ from vllm.multimodal.processing import (
     PromptReplacement,
     PromptUpdate,
     PromptUpdateDetails,
+    _seq2tokens,
 )
 from vllm.multimodal.profiling import BaseDummyInputsBuilder
 from vllm.sequence import IntermediateTensors
 from vllm.transformers_utils.configs.radio import RadioConfig
-from vllm.transformers_utils.tokenizer import AnyTokenizer
+from vllm.transformers_utils.tokenizer import (
+    AnyTokenizer,
+    cached_tokenizer_from_config,
+    encode_tokens,
+)
 from vllm.utils.tensor_schema import TensorSchema, TensorShape
 
+from .utils import _merge_multimodal_embeddings
+
 # Configure PIL to handle large images without warnings
 # This prevents DecompressionBombWarning for legitimate large images
 Image.MAX_IMAGE_PIXELS = None  # Disable the limit entirely
@@ -382,6 +394,7 @@ class NanoNemotronVLProcessor(BaseNanoNemotronVLProcessor):
         max_dynamic_patch: Optional[int] = None,
         dynamic_image_size: Optional[bool] = None,
         video_token: Optional[str] = None,
+        video_pruning_rate: Optional[float] = None,
     ) -> None:
         super().__init__(
             config=config,
@@ -392,6 +405,7 @@ class NanoNemotronVLProcessor(BaseNanoNemotronVLProcessor):
         )
         # add extra video token for video processing
         self.video_token = video_token
+        self.video_pruning_rate = video_pruning_rate
 
     @property
     def supports_video(self) -> bool:
@@ -446,12 +460,38 @@ class NanoNemotronVLProcessor(BaseNanoNemotronVLProcessor):
                 ),
             }
 
-            for pixel_values in pixel_values_lst_video:
+            image_size: int = self.config.force_image_size
-                num_patches = pixel_values.shape[0]
+            patch_size: int = self.config.patch_size
+            downsample_ratio = self.config.downsample_ratio
+            tokens_in_single_frame = int(
+                (image_size * image_size // patch_size**2) * (downsample_ratio**2)
+            )
+
+            for pixel_values in pixel_values_lst_video:
+                num_frames = pixel_values.shape[0]
+
+                if (
+                    self.video_pruning_rate is not None
+                    and self.video_pruning_rate > 0.0
+                ):
+                    # Start of EVS-specific code
+                    num_tokens = compute_retained_tokens_count(
+                        tokens_per_frame=tokens_in_single_frame,
+                        num_frames=num_frames,
+                        q=self.video_pruning_rate,
+                    )
+
+                    # Here we just need placeholders that won't actually be replaced -
+                    # we just need to make sure the total number of tokens is correct
+                    # assign all tokens to the first frame
+                    tokens_per_frame = [num_tokens] + [0] * (num_frames - 1)
+
+                    # End of EVS-specific code
+                else:
+                    tokens_per_frame = [tokens_in_single_frame] * num_frames
+
+                video_repl = self.get_video_repl(tokens_per_frame, self.video_token)
 
-                video_repl = self.get_video_repl(
-                    self.num_image_token, num_patches, self.video_token
-                )
                 text = [t.replace("<video>", video_repl.full, 1) for t in text]
         return text, video_inputs
 
@@ -501,20 +541,40 @@ class NanoNemotronVLProcessor(BaseNanoNemotronVLProcessor):
 
         return PromptUpdateDetails.select_text(repl_full, IMG_CONTEXT)
 
+    @classmethod
     def get_video_repl(
-        self,
+        cls,
-        feature_size: int,
+        tokens_per_frame: list[int],
-        num_patches: Optional[int] = None,
         video_context_token: str = IMG_CONTEXT,
     ) -> PromptUpdateDetails[str]:
-        repl_features = video_context_token * self.num_image_token
+        """
-        repl_features_with_sep = IMG_START + repl_features + IMG_END
+        Build prompt replacement for a video.
-        # num_patches is equal to num_frames
+        The replacement returned is not actually used to replace the placeholder
+        tokens - it's just used to make sure we allocate the correct number
+        of tokens.
+        Actual replacement is done in get_multimodal_embeddings of
+        NemotronH_Nano_VL_V2
+        (specifically in _process_video_input -> _create_final_video_embeddings).
+        There, we create the final embeddings with text embeddings for indicator tokens
+        and video embeddings for video tokens.
+        This is a single function that handles all cases - non EVS, EVS dummy, EVS real.
+        The differentiation is done via tokens_per_frame parameter.
+        - non EVS case - constant value same value across all frames
+        - EVS dummy - Doesn't matter how tokens are distributed between frames - just
+                        make sure the total number of tokens is correct.
+        - EVS real (called from get_real_video_repl_for_evs) - different value per frame
+        Args:
+            tokens_per_frame (list[int]): number of tokens per frame
+            video_context_token (str): the token to use for the video context
+        """
         repl_full = "".join(
-            [f"Frame{i + 1}: {repl_features_with_sep}" for i in range(num_patches)]
+            [
+                f"Frame{i + 1}: {IMG_START}{video_context_token * num_tokens}{IMG_END}"
+                for i, num_tokens in enumerate(tokens_per_frame)
+            ]
         )
 
-        return PromptUpdateDetails.select_text(repl_full, video_context_token)
+        return PromptUpdateDetails.from_seq(repl_full)
 
 
 class BaseNanoNemotronVLProcessingInfo(BaseProcessingInfo):
@@ -605,6 +665,9 @@ class NanoNemotronVLProcessingInfo(BaseNanoNemotronVLProcessingInfo):
     def get_video_token(self) -> Optional[str]:
         return IMG_CONTEXT
 
+    def get_video_pruning_rate(self) -> Optional[float]:
+        return self.ctx.get_mm_config().video_pruning_rate
+
     def get_num_frames_with_most_features(
         self,
         seq_len: int,
@@ -628,6 +691,7 @@ class NanoNemotronVLProcessingInfo(BaseNanoNemotronVLProcessingInfo):
             config=self.get_hf_config(),
             tokenizer=self.get_tokenizer(),
             video_token=self.get_video_token(),
+            video_pruning_rate=self.get_video_pruning_rate(),
             **kwargs,
         )
 
@@ -805,8 +869,26 @@ class NanoNemotronVLMultiModalProcessor(
             if num_patches is not None:
                 assert isinstance(num_patches, int)
 
+            video_pruning_rate = self.info.ctx.get_mm_config().video_pruning_rate
+            if video_pruning_rate is not None and video_pruning_rate > 0.0:
+                # Start of EVS-specific code
+                num_tokens = compute_retained_tokens_count(
+                    tokens_per_frame=feature_size,
+                    num_frames=num_patches,
+                    q=video_pruning_rate,
+                )
+                # Here we just need placeholders that won't actually be replaced -
+                # we just need to make sure the total number of tokens is correct
+                # assign all tokens to the first frame
+                tokens_per_frame = [num_tokens] + [0] * (num_patches - 1)
+
+                # End of EVS-specific code
+            else:
+                tokens_per_frame = [feature_size] * num_patches
+
             return hf_processor.get_video_repl(
-                feature_size, num_patches, video_context_token=hf_processor.video_token
+                tokens_per_frame,
+                video_context_token=hf_processor.video_token,
             )
 
         if self.info.supports_video:
@@ -901,7 +983,9 @@ class NanoNemotronVLDummyInputsBuilder(
     info=NanoNemotronVLProcessingInfo,
     dummy_inputs=NanoNemotronVLDummyInputsBuilder,
 )
-class NemotronH_Nano_VL_V2(nn.Module, HasInnerState, IsHybrid, SupportsMultiModal):
+class NemotronH_Nano_VL_V2(
+    nn.Module, HasInnerState, IsHybrid, SupportsMultiModal, SupportsMultiModalPruning
+):
     @classmethod
     def get_placeholder_str(cls, modality: str, i: int) -> Optional[str]:
         if modality.startswith("image"):
@@ -913,7 +997,7 @@ class NemotronH_Nano_VL_V2(nn.Module, HasInnerState, IsHybrid, SupportsMultiModa
     def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
         super().__init__()
         config = vllm_config.model_config.hf_config
-
+        multimodal_config = vllm_config.model_config.multimodal_config
         image_size = config.force_image_size
         patch_size = config.patch_size
         self.patch_size = patch_size
@@ -924,7 +1008,7 @@ class NemotronH_Nano_VL_V2(nn.Module, HasInnerState, IsHybrid, SupportsMultiModa
         self.downsample_ratio = config.downsample_ratio
         self.ps_version = config.ps_version
         self.image_tag_type = config.image_tag_type
-
+        self.video_pruning_rate = multimodal_config.video_pruning_rate
         self.language_model = init_vllm_registered_model(
             vllm_config=vllm_config,
             hf_config=config.text_config,
@@ -957,6 +1041,7 @@ class NemotronH_Nano_VL_V2(nn.Module, HasInnerState, IsHybrid, SupportsMultiModa
         self.img_context_token_id = None
         self.video_context_token_id = None
         self.config = config
+        self.model_config = vllm_config.model_config
 
     def pixel_shuffle(self, x, scale_factor=0.5):
         n, w, h, c = x.size()
@@ -1049,7 +1134,7 @@ class NemotronH_Nano_VL_V2(nn.Module, HasInnerState, IsHybrid, SupportsMultiModa
 
     def _process_image_input(
         self, image_input: NanoNemotronVLImageInputs
-    ) -> torch.Tensor:
+    ) -> tuple[torch.Tensor, ...]:
         if image_input["type"] == "image_embeds":
             return image_input["data"]
 
@@ -1071,6 +1156,109 @@ class NemotronH_Nano_VL_V2(nn.Module, HasInnerState, IsHybrid, SupportsMultiModa
         ]
         return image_embeds.split(image_feature_sizes)
 
+    def _process_video_input(
+        self, video_input: NanoNemotronVLVideoPixelInputs
+    ) -> tuple[torch.Tensor, ...]:
+        """Process video input and create final embeddings with video content
+        and indicator tokens."""
+        # Get video embeddings using the same processing as images
+        video_embeddings = self._process_image_input(video_input)
+
+        final_video_embeddings: tuple[torch.Tensor, ...] = ()
+
+        image_rows = image_cols = self.config.force_image_size
+        downsample_ratio = self.config.downsample_ratio
+        patch_size = self.config.patch_size
+        rows = int(image_rows * downsample_ratio // patch_size)
+        cols = int(image_cols * downsample_ratio // patch_size)
+        video_pruning_rate = self.video_pruning_rate
+
+        # Calculate video feature dimensions (number of frames and
+        # their feature size (AKA tokens per frame))
+        # TODO: Maybe this can be optimized to avoid the loop?
+        for i, single_video_embeddings in enumerate(video_embeddings):
+            num_frames = video_input["num_patches"][i].item()
+            assert single_video_embeddings.shape[0] % num_frames == 0
+
+            if video_pruning_rate is not None and video_pruning_rate > 0.0:
+                # Start of EVS-specific code
+                retention_mask = compute_retention_mask(
+                    single_video_embeddings,
+                    video_size_thw=(num_frames, rows, cols),
+                    spatial_merge_size=1,
+                    q=video_pruning_rate,
+                )
+
+                # apply retention mask
+                single_video_embeddings = single_video_embeddings[retention_mask]
+
+                # calculate the actual number of retained tokens per frame
+                retention_mask_thw = retention_mask.reshape(num_frames, rows, cols)
+                num_tokens_per_frame = (
+                    retention_mask_thw.sum(dim=(1, 2)).long().tolist()
+                )
+                # End of EVS-specific code
+            else:
+                feature_size = single_video_embeddings.shape[0] // num_frames
+                num_tokens_per_frame = [feature_size] * num_frames
+
+            final_video_embeddings += (
+                self._create_final_video_embeddings(
+                    single_video_embeddings,
+                    num_tokens_per_frame,
+                ),
+            )
+
+        return final_video_embeddings
+
+    def _create_final_video_embeddings(
+        self,
+        video_embeddings: torch.Tensor,
+        num_tokens_per_frame: list[int],
+    ) -> torch.Tensor:
+        """Create final embeddings that combine video embeddings with
+        text embeddings of indicator tokens.
+
+        These final embeddings contain:
+        - Actual video embeddings in positions corresponding to video content
+        - Text embeddings for indicator tokens (<img>, </img>, and
+          frame separation text) in their respective positions
+
+        These embeddings will replace the placeholder embeddings to create
+        input_embeds for the LLM.
+        """
+        device = video_embeddings.device
+
+        # Generate video replacement text and convert to token IDs
+        video_repl_text = NanoNemotronVLProcessor.get_video_repl(
+            num_tokens_per_frame,
+            IMG_CONTEXT,
+        ).full
+
+        tokenizer = cached_tokenizer_from_config(self.model_config)
+        repl_token_ids = torch.tensor(
+            _seq2tokens(tokenizer, video_repl_text), device=device
+        )
+
+        # Get embedding token IDs for image context
+        embed_token_ids = torch.tensor(
+            encode_tokens(tokenizer, IMG_CONTEXT), device=device
+        )
+
+        # Create mask for video embedding positions
+        is_video_embed = torch.isin(repl_token_ids, embed_token_ids)
+
+        # Create final video embeddings, merging text embeddings for indicator
+        # tokens with video embeddings
+        text_embeddings = self.get_language_model().get_input_embeddings(repl_token_ids)
+        final_video_embeddings = _merge_multimodal_embeddings(
+            inputs_embeds=text_embeddings,
+            multimodal_embeddings=video_embeddings,
+            is_multimodal=is_video_embed,
+        )
+
+        return final_video_embeddings
+
     def _parse_and_validate_video_input(
         self, **kwargs: object
     ) -> Optional[NanoNemotronVLVideoPixelInputs]:
@@ -1152,7 +1340,7 @@ class NemotronH_Nano_VL_V2(nn.Module, HasInnerState, IsHybrid, SupportsMultiModa
                 multimodal_embeddings += vision_embeddings
             if modality == "videos":
                 video_input = modalities["videos"]
-                video_embeddings = self._process_image_input(video_input)
+                video_embeddings = self._process_video_input(video_input)
                 multimodal_embeddings += video_embeddings
 
         return multimodal_embeddings

vllm/model_executor/models/qwen2_5_vl.py

@@ -1017,9 +1017,13 @@ class Qwen2_5_VLMultiModalProcessor(Qwen2VLMultiModalProcessor):
                 and video_pruning_rate is not None
                 and video_pruning_rate > 0.0
             ):
+                T, H, W = map(int, grid_thw)
+                tokens_per_frame = (H // image_processor.merge_size) * (
+                    W // image_processor.merge_size
+                )
                 num_tokens = compute_retained_tokens_count(
-                    grid_thw,
+                    tokens_per_frame,
-                    image_processor.merge_size,
+                    T,
                     video_pruning_rate,
                 )
             # End of EVS-specific code

vllm/multimodal/evs.py

@@ -9,12 +9,13 @@
 # license agreement from NVIDIA CORPORATION is strictly prohibited.
 
 import typing
+from typing import Union
 
 import torch
 
 
 def compute_retained_tokens_count(
-    video_size_thw: torch.LongTensor, spatial_merge_size: int, q: float
+    tokens_per_frame: int, num_frames: int, q: float
 ) -> int:
     """
     Compute the number of retained tokens for a given video.
@@ -22,22 +23,22 @@ def compute_retained_tokens_count(
     regardless of the pruning rate.
 
     Args:
-        video_size_thw: The size of the video in the format of (T, H, W).
+        tokens_per_frame: The number of tokens per frame.
-        spatial_merge_size: The size of the spatial merge.
+        num_frames: The total number of frames.
         q: The pruning rate.
 
     Returns:
         The number of retained tokens.
     """
-    T, H, W = map(int, video_size_thw)
+    total_tokens = tokens_per_frame * num_frames
-    min_num_tokens = (H // spatial_merge_size) * (W // spatial_merge_size)
+    evs_num_tokens = int(total_tokens * (1 - q))
-    evs_num_tokens = int(T * min_num_tokens * (1 - q))
+    min_num_tokens = tokens_per_frame
     return max(min_num_tokens, evs_num_tokens)
 
 
 def compute_retention_mask(
     video_embeds: torch.Tensor,
-    video_size_thw: torch.LongTensor,
+    video_size_thw: Union[torch.LongTensor, tuple[int, int, int]],
     spatial_merge_size: int,
     q: float,
 ) -> torch.Tensor:
@@ -56,7 +57,7 @@ def compute_retention_mask(
         `torch.Tensor`: The retention mask for the video embeddings of
             `(T * H * W // spatial_merge_size ^ 2)` shape.
     """
-    T, H, W = video_size_thw
+    T, H, W = map(int, video_size_thw)
 
     # Use reshape instead of einops to avoid graph breaks
     video_embeds = video_embeds.reshape(
@@ -65,7 +66,7 @@ def compute_retention_mask(
         W // spatial_merge_size,
         video_embeds.size(-1),
     )
-
+    tokens_per_frame = (H // spatial_merge_size) * (W // spatial_merge_size)
     # Core EVS
     similarity = torch.nn.functional.cosine_similarity(
         video_embeds[1:, ...], video_embeds[:-1, ...], dim=-1
@@ -80,7 +81,7 @@ def compute_retention_mask(
     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
+        tokens_per_frame=tokens_per_frame, num_frames=T, q=q
     )
     topk_indices = order[:retain_num_tokens]