transformers 5.0.0__py3-none-any.whl → 5.0.0rc0__py3-none-any.whl

transformers/models/llava_next_video/modeling_llava_next_video.py

@@ -4,6 +4,7 @@
 #             the file from the modular. If any change should be done, please apply the change to the
 #                          modular_llava_next_video.py file directly. One of our CI enforces this.
 #                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+# coding=utf-8
 # Copyright 2024 HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -20,6 +21,7 @@
 
 import math
 from dataclasses import dataclass
+from typing import Optional, Union
 
 import numpy as np
 import torch
@@ -31,11 +33,10 @@ from ...cache_utils import Cache
 from ...generation import GenerationMixin
 from ...image_processing_utils import select_best_resolution
 from ...modeling_flash_attention_utils import FlashAttentionKwargs
-from ...modeling_outputs import BaseModelOutputWithPast, BaseModelOutputWithPooling, ModelOutput
+from ...modeling_outputs import BaseModelOutputWithPast, ModelOutput
 from ...modeling_utils import PreTrainedModel
 from ...processing_utils import Unpack
-from ...utils import TransformersKwargs, auto_docstring, logging, torch_compilable_check
-from ...utils.generic import check_model_inputs
+from ...utils import TransformersKwargs, auto_docstring, can_return_tuple, logging
 from ..auto import AutoModel
 from .configuration_llava_next_video import LlavaNextVideoConfig
 
@@ -64,9 +65,9 @@ class LlavaNextVideoModelOutputWithPast(BaseModelOutputWithPast):
         video_hidden_states of the model produced by the vision encoder and after projecting the last hidden state.
     """
 
-    image_hidden_states: torch.FloatTensor | None = None
+    image_hidden_states: Optional[torch.FloatTensor] = None
 
-    video_hidden_states: torch.FloatTensor | None = None
+    video_hidden_states: Optional[torch.FloatTensor] = None
 
 
 @dataclass
@@ -94,14 +95,14 @@ class LlavaNextVideoCausalLMOutputWithPast(ModelOutput):
         video_hidden_states of the model produced by the vision encoder and after projecting the last hidden state.
     """
 
-    loss: torch.FloatTensor | None = None
-    logits: torch.FloatTensor | None = None
-    past_key_values: Cache | None = None
-    hidden_states: tuple[torch.FloatTensor] | None = None
-    attentions: tuple[torch.FloatTensor] | None = None
-    image_hidden_states: torch.FloatTensor | None = None
+    loss: Optional[torch.FloatTensor] = None
+    logits: Optional[torch.FloatTensor] = None
+    past_key_values: Optional[Cache] = None
+    hidden_states: Optional[tuple[torch.FloatTensor]] = None
+    attentions: Optional[tuple[torch.FloatTensor]] = None
+    image_hidden_states: Optional[torch.FloatTensor] = None
 
-    video_hidden_states: torch.FloatTensor | None = None
+    video_hidden_states: Optional[torch.FloatTensor] = None
 
 
 class LlavaNextVideoPooler(nn.Module):
@@ -320,6 +321,7 @@ class LlavaNextVideoModel(LlavaNextVideoPreTrainedModel):
 
         self.vocab_size = config.text_config.vocab_size
         self.language_model = AutoModel.from_config(config.text_config)
+        self.pad_token_id = self.config.pad_token_id if self.config.pad_token_id is not None else -1
         self.vision_resampler = LlavaNextVideoPooler(config)
         self.post_init()
 
@@ -396,32 +398,41 @@ class LlavaNextVideoModel(LlavaNextVideoPreTrainedModel):
         feature_lens = torch.tensor(feature_lens, dtype=torch.long, device=image_features[0].device)
         return new_image_features, feature_lens
 
-    @check_model_inputs(tie_last_hidden_states=False)
-    @auto_docstring(
-        custom_intro="Obtains image last hidden states from the vision tower and apply multimodal projection."
-    )
     def get_image_features(
         self,
         pixel_values: torch.FloatTensor,
         image_sizes: torch.Tensor,
-        vision_feature_layer: int | list[int] | None = None,
-        vision_feature_select_strategy: str | None = None,
-        output_hidden_states: bool | None = None,
-        **kwargs: Unpack[TransformersKwargs],
-    ) -> tuple | BaseModelOutputWithPooling:
-        r"""
-        pixel_values (`torch.FloatTensor]` of shape `(batch_size, num_patches, channels, height, width)`)
-            The tensors corresponding to the input images.
-        image_sizes (`torch.Tensor` of shape `(num_images, 2)`)
-            Actual image size of each images (H, W).
-        vision_feature_layer (`Union[int, list[int]]`, *optional*):
-            The index of the layer to select the vision feature. If multiple indices are provided,
-            the vision feature of the corresponding indices will be concatenated to form the
-            vision features.
-        vision_feature_select_strategy (`str`, *optional*):
-            The feature selection strategy used to select the vision feature from the vision backbone.
-            Can be one of `"default"` or `"full"`
+        vision_feature_layer: Optional[Union[int, list[int]]] = None,
+        vision_feature_select_strategy: Optional[str] = None,
+    ):
+        """
+        Obtains image last hidden states from the vision tower and apply multimodal projection.
+
+        Args:
+            pixel_values (`torch.FloatTensor]` of shape `(batch_size, num_patches, channels, height, width)`)
+               The tensors corresponding to the input images.
+            image_sizes (`torch.Tensor` of shape `(num_images, 2)`)
+                Actual image size of each images (H, W).
+            vision_feature_layer (`Union[int, list[int]]`, *optional*):
+                The index of the layer to select the vision feature. If multiple indices are provided,
+                the vision feature of the corresponding indices will be concatenated to form the
+                vision features.
+            vision_feature_select_strategy (`str`, *optional*):
+                The feature selection strategy used to select the vision feature from the vision backbone.
+                Can be one of `"default"` or `"full"`
+        Returns:
+            image_features (list[`torch.Tensor`]): List of image feature tensor, each contains all the visual feature of all patches
+            and are of shape `(num_patches, image_length, embed_dim)`).
         """
+        vision_feature_layer = (
+            vision_feature_layer if vision_feature_layer is not None else self.config.vision_feature_layer
+        )
+        vision_feature_select_strategy = (
+            vision_feature_select_strategy
+            if vision_feature_select_strategy is not None
+            else self.config.vision_feature_select_strategy
+        )
+
         # ! infer image_num_patches from image_sizes
         image_num_patches = [
             image_size_to_num_patches(
@@ -439,18 +450,13 @@ class LlavaNextVideoModel(LlavaNextVideoPreTrainedModel):
             # otherwise has to be stacked from list of (num_patches, num_channels, height, width)
             raise ValueError(f"pixel_values of shape {pixel_values.shape}, expect to be of 4 or 5 dimensions")
 
-        image_outputs = self.vision_tower(
-            pixel_values,
-            output_hidden_states=True,  # Ignore arg on purpose
-            return_dict=True,
-            **kwargs,
-        )
+        image_features = self.vision_tower(pixel_values, output_hidden_states=True)
         # If we have one vision feature layer, return the corresponding hidden states,
         # otherwise, select the hidden states of each feature layer and concatenate them
         if isinstance(vision_feature_layer, int):
-            selected_image_feature = image_outputs.hidden_states[vision_feature_layer]
+            selected_image_feature = image_features.hidden_states[vision_feature_layer]
         else:
-            hs_pool = [image_outputs.hidden_states[layer_idx] for layer_idx in vision_feature_layer]
+            hs_pool = [image_features.hidden_states[layer_idx] for layer_idx in vision_feature_layer]
             selected_image_feature = torch.cat(hs_pool, dim=-1)
 
         if vision_feature_select_strategy == "default":
@@ -464,16 +470,14 @@ class LlavaNextVideoModel(LlavaNextVideoPreTrainedModel):
             vision_feature_select_strategy,
             image_newline=self.image_newline,
         )
-        image_outputs.pooler_output = image_features
-
-        return image_outputs
+        return image_features
 
     def get_placeholder_mask(
         self,
         input_ids: torch.LongTensor,
         inputs_embeds: torch.FloatTensor,
-        image_features: torch.FloatTensor | None = None,
-        video_features: torch.FloatTensor | None = None,
+        image_features: Optional[torch.FloatTensor] = None,
+        video_features: Optional[torch.FloatTensor] = None,
     ):
         """
         Obtains multimodal placeholder mask from `input_ids` or `inputs_embeds`, and checks that the placeholder token count is
@@ -494,42 +498,41 @@ class LlavaNextVideoModel(LlavaNextVideoPreTrainedModel):
 
         n_image_tokens = special_image_mask.sum()
         special_image_mask = special_image_mask.unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device)
-        if image_features is not None:
-            torch_compilable_check(
-                inputs_embeds[special_image_mask].numel() == image_features.numel(),
-                f"Image features and image tokens do not match, tokens: {n_image_tokens}, features: {image_features.shape[0]}",
+        if image_features is not None and inputs_embeds[special_image_mask].numel() != image_features.numel():
+            raise ValueError(
+                f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {image_features.shape[0]}"
             )
 
         n_video_tokens = special_video_mask.sum()
         special_video_mask = special_video_mask.unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device)
-        if video_features is not None:
-            torch_compilable_check(
-                inputs_embeds[special_video_mask].numel() == video_features.numel(),
-                f"Video features and video tokens do not match, tokens: {n_video_tokens}, features: {video_features.shape[0]}",
+        if video_features is not None and inputs_embeds[special_video_mask].numel() != video_features.numel():
+            raise ValueError(
+                f"Videos features and image tokens do not match: tokens: {n_video_tokens}, features {video_features.shape[0]}"
             )
+
         return special_image_mask, special_video_mask
 
-    @check_model_inputs(tie_last_hidden_states=False)
+    @can_return_tuple
     @auto_docstring
     def forward(
         self,
-        input_ids: torch.LongTensor | None = None,
-        pixel_values: torch.FloatTensor | None = None,
-        pixel_values_videos: torch.FloatTensor | None = None,
-        image_sizes: torch.LongTensor | None = None,
-        attention_mask: torch.Tensor | None = None,
-        position_ids: torch.LongTensor | None = None,
-        past_key_values: Cache | None = None,
-        inputs_embeds: torch.FloatTensor | None = None,
-        vision_feature_layer: int | list[int] | None = None,
-        vision_feature_select_strategy: str | None = None,
-        use_cache: bool | None = None,
-        output_attentions: bool | None = None,
-        output_hidden_states: bool | None = None,
-        return_dict: bool | None = None,
-        cache_position: torch.LongTensor | None = None,
+        input_ids: Optional[torch.LongTensor] = None,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        pixel_values_videos: Optional[torch.FloatTensor] = None,
+        image_sizes: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        vision_feature_layer: Optional[Union[int, list[int]]] = None,
+        vision_feature_select_strategy: Optional[str] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
         **kwargs: Unpack[FlashAttentionKwargs],
-    ) -> tuple | LlavaNextVideoModelOutputWithPast:
+    ) -> Union[tuple, LlavaNextVideoModelOutputWithPast]:
         r"""
         vision_feature_select_strategy (`str`, *optional*, defaults to `"default"`):
             The feature selection strategy used to select the vision feature from the vision backbone.
@@ -541,6 +544,14 @@ class LlavaNextVideoModel(LlavaNextVideoPreTrainedModel):
             output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
         )
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        self.vision_feature_layer = (
+            vision_feature_layer if vision_feature_layer is not None else self.config.vision_feature_layer
+        )
+        self.vision_feature_select_strategy = (
+            vision_feature_select_strategy
+            if vision_feature_select_strategy is not None
+            else self.config.vision_feature_select_strategy
+        )
 
         if (input_ids is None) ^ (inputs_embeds is not None):
             raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
@@ -552,10 +563,9 @@ class LlavaNextVideoModel(LlavaNextVideoPreTrainedModel):
             image_features = self.get_image_features(
                 pixel_values,
                 image_sizes,
-                vision_feature_layer=vision_feature_layer,
-                vision_feature_select_strategy=vision_feature_select_strategy,
-                return_dict=True,
-            ).pooler_output
+                vision_feature_layer=self.vision_feature_layer,
+                vision_feature_select_strategy=self.vision_feature_select_strategy,
+            )
             image_features = torch.cat(image_features, dim=0).to(inputs_embeds.device, inputs_embeds.dtype)
             special_image_mask, _ = self.get_placeholder_mask(
                 input_ids, inputs_embeds=inputs_embeds, image_features=image_features
@@ -565,10 +575,9 @@ class LlavaNextVideoModel(LlavaNextVideoPreTrainedModel):
         if pixel_values_videos is not None:
             video_features = self.get_video_features(
                 pixel_values_videos,
-                vision_feature_layer=vision_feature_layer,
-                vision_feature_select_strategy=vision_feature_select_strategy,
-                return_dict=True,
-            ).pooler_output
+                vision_feature_layer=self.vision_feature_layer,
+                vision_feature_select_strategy=self.vision_feature_select_strategy,
+            )
             video_features = [feature.flatten(0, 1) for feature in video_features]
             video_feature_lens = [feature.size(0) for feature in video_features]
             video_features = torch.cat(video_features, dim=0)
@@ -602,44 +611,48 @@ class LlavaNextVideoModel(LlavaNextVideoPreTrainedModel):
             video_hidden_states=video_features if pixel_values_videos is not None else None,
         )
 
-    @check_model_inputs(tie_last_hidden_states=False)
-    @auto_docstring(
-        custom_intro="Obtains video last hidden states from the vision tower and apply multimodal projection."
-    )
     def get_video_features(
         self,
         pixel_values: torch.FloatTensor,
-        vision_feature_layer: int | list[int] | None = None,
-        vision_feature_select_strategy: str | None = None,
-        output_hidden_states: bool | None = None,
-        **kwargs: Unpack[TransformersKwargs],
-    ) -> tuple | BaseModelOutputWithPooling:
-        r"""
-        pixel_values (`torch.FloatTensor]` of shape `(batch_size, num_frames, channels, height, width)`)
-            The tensors corresponding to the input video.
-        vision_feature_layer (`Union[int, list[int]]`, *optional;*):
-            The index of the layer to select the vision feature. If multiple indices are provided,
-            the vision feature of the corresponding indices will be concatenated to form the
-            vision features.
-        vision_feature_select_strategy (`str`, *optional*):
-            The feature selection strategy used to select the vision feature from the vision backbone.
-            Can be one of `"default"` or `"full"`
+        vision_feature_layer: Optional[Union[int, list[int]]] = None,
+        vision_feature_select_strategy: Optional[str] = None,
+    ):
         """
+        Obtains video last hidden states from the vision tower and apply multimodal projection.
+
+        Args:
+            pixel_values (`torch.FloatTensor]` of shape `(batch_size, num_frames, channels, height, width)`)
+               The tensors corresponding to the input video.
+            vision_feature_layer (`Union[int, list[int]]`, *optional;*):
+                The index of the layer to select the vision feature. If multiple indices are provided,
+                the vision feature of the corresponding indices will be concatenated to form the
+                vision features.
+            vision_feature_select_strategy (`str`, *optional*):
+                The feature selection strategy used to select the vision feature from the vision backbone.
+                Can be one of `"default"` or `"full"`
+        Returns:
+            video_features (list[`torch.Tensor`]): List of video feature tensor, each contains all the visual feature of all patches
+            and are of shape `(num_videos, video_length, embed_dim)`).
+        """
+        vision_feature_layer = (
+            vision_feature_layer if vision_feature_layer is not None else self.config.vision_feature_layer
+        )
+        vision_feature_select_strategy = (
+            vision_feature_select_strategy
+            if vision_feature_select_strategy is not None
+            else self.config.vision_feature_select_strategy
+        )
+
         batch_size, frames, channels, height, width = pixel_values.shape
         pixel_values = pixel_values.reshape(batch_size * frames, channels, height, width)
-        video_outputs = self.vision_tower(
-            pixel_values,
-            output_hidden_states=True,  # Ignore arg on purpose
-            return_dict=True,
-            **kwargs,
-        )
+        video_features = self.vision_tower(pixel_values, output_hidden_states=True)
 
         # If we have one vision feature layer, return the corresponding hidden states,
         # otherwise, select the hidden states of each feature layer and concatenate them
         if isinstance(vision_feature_layer, int):
-            selected_video_features = video_outputs.hidden_states[vision_feature_layer]
+            selected_video_features = video_features.hidden_states[vision_feature_layer]
         else:
-            hs_pool = [video_outputs.hidden_states[layer_idx] for layer_idx in vision_feature_layer]
+            hs_pool = [video_features.hidden_states[layer_idx] for layer_idx in vision_feature_layer]
             selected_video_features = torch.cat(hs_pool, dim=-1)
 
         if vision_feature_select_strategy == "default":
@@ -649,9 +662,7 @@ class LlavaNextVideoModel(LlavaNextVideoPreTrainedModel):
         video_features = self.vision_resampler(selected_video_features)
         video_features = self.multi_modal_projector(video_features)
         video_features = torch.split(video_features, frames, dim=0)
-        video_outputs.pooler_output = video_features
-
-        return video_outputs
+        return video_features
 
 
 @auto_docstring(
@@ -692,59 +703,43 @@ class LlavaNextVideoForConditionalGeneration(LlavaNextVideoPreTrainedModel, Gene
             image_newline=image_newline,
         )
 
-    @auto_docstring
     def get_image_features(
         self,
         pixel_values: torch.FloatTensor,
         image_sizes: torch.Tensor,
-        vision_feature_layer: int | list[int] | None = None,
-        vision_feature_select_strategy: str | None = None,
-        **kwargs: Unpack[TransformersKwargs],
-    ) -> tuple | BaseModelOutputWithPooling:
-        r"""
-        pixel_values (`torch.FloatTensor]` of shape `(batch_size, num_patches, channels, height, width)`)
-            The tensors corresponding to the input images.
-        image_sizes (`torch.Tensor` of shape `(num_images, 2)`)
-            Actual image size of each images (H, W).
-        vision_feature_layer (`Union[int, list[int]]`, *optional*):
-            The index of the layer to select the vision feature. If multiple indices are provided,
-            the vision feature of the corresponding indices will be concatenated to form the
-            vision features.
-        vision_feature_select_strategy (`str`, *optional*):
-            The feature selection strategy used to select the vision feature from the vision backbone.
-            Can be one of `"default"` or `"full"`
-        """
+        vision_feature_layer: Optional[Union[int, list[int]]] = None,
+        vision_feature_select_strategy: Optional[str] = None,
+    ):
         return self.model.get_image_features(
             pixel_values=pixel_values,
             image_sizes=image_sizes,
             vision_feature_layer=vision_feature_layer,
             vision_feature_select_strategy=vision_feature_select_strategy,
-            **kwargs,
         )
 
-    @check_model_inputs(tie_last_hidden_states=False)
+    @can_return_tuple
     @auto_docstring
     def forward(
         self,
-        input_ids: torch.LongTensor | None = None,
-        pixel_values: torch.FloatTensor | None = None,
-        pixel_values_videos: torch.FloatTensor | None = None,
-        image_sizes: torch.LongTensor | None = None,
-        attention_mask: torch.Tensor | None = None,
-        position_ids: torch.LongTensor | None = None,
-        past_key_values: Cache | None = None,
-        inputs_embeds: torch.FloatTensor | None = None,
-        vision_feature_layer: int | list[int] | None = None,
-        vision_feature_select_strategy: str | None = None,
-        labels: torch.LongTensor | None = None,
-        use_cache: bool | None = None,
-        output_attentions: bool | None = None,
-        output_hidden_states: bool | None = None,
-        return_dict: bool | None = None,
-        cache_position: torch.LongTensor | None = None,
-        logits_to_keep: int | torch.Tensor = 0,
+        input_ids: Optional[torch.LongTensor] = None,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        pixel_values_videos: Optional[torch.FloatTensor] = None,
+        image_sizes: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        vision_feature_layer: Optional[Union[int, list[int]]] = None,
+        vision_feature_select_strategy: Optional[str] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        logits_to_keep: Union[int, torch.Tensor] = 0,
         **kwargs: Unpack[TransformersKwargs],
-    ) -> tuple | LlavaNextVideoCausalLMOutputWithPast:
+    ) -> Union[tuple, LlavaNextVideoCausalLMOutputWithPast]:
         r"""
         labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
             Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
@@ -755,8 +750,7 @@ class LlavaNextVideoForConditionalGeneration(LlavaNextVideoPreTrainedModel, Gene
 
         ```python
         >>> from PIL import Image
-        >>> import httpx
-        >>> from io import BytesIO
+        >>> import requests
         >>> import av
         >>> from transformers import AutoProcessor, LlavaNextVideoForConditionalGeneration
 
@@ -796,8 +790,7 @@ class LlavaNextVideoForConditionalGeneration(LlavaNextVideoPreTrainedModel, Gene
         >>> # load an image to generate from an image
         >>> prompt = "USER:<image>\nWhat is shown in this image? ASSISTANT:"
         >>> url = "https://www.ilankelman.org/stopsigns/australia.jpg"
-        >>> with httpx.stream("GET", url) as response:
-        ...     image = Image.open(BytesIO(response.read()))
+        >>> image = Image.open(requests.get(url, stream=True).raw)
         >>> inputs_image = processor(text=prompt, images=image, return_tensors="pt").to(model.device)
 
         >>> # Generate from video
@@ -815,6 +808,14 @@ class LlavaNextVideoForConditionalGeneration(LlavaNextVideoPreTrainedModel, Gene
             output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
         )
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        vision_feature_layer = (
+            vision_feature_layer if vision_feature_layer is not None else self.config.vision_feature_layer
+        )
+        vision_feature_select_strategy = (
+            vision_feature_select_strategy
+            if vision_feature_select_strategy is not None
+            else self.config.vision_feature_select_strategy
+        )
 
         outputs = self.model(
             input_ids=input_ids,
@@ -867,7 +868,6 @@ class LlavaNextVideoForConditionalGeneration(LlavaNextVideoPreTrainedModel, Gene
         attention_mask=None,
         cache_position=None,
         logits_to_keep=None,
-        is_first_iteration=False,
         **kwargs,
     ):
         # Overwritten -- extra custom processing
@@ -879,15 +879,12 @@ class LlavaNextVideoForConditionalGeneration(LlavaNextVideoPreTrainedModel, Gene
             attention_mask=attention_mask,
             cache_position=cache_position,
             logits_to_keep=logits_to_keep,
-            is_first_iteration=is_first_iteration,
             **kwargs,
         )
 
-        # Pixel values are used only in the first iteration if available
-        # In subsquent iterations, they are already merged with text and cached
-        # NOTE: first iteration doesn't have to be prefill, it can be the first
-        # iteration with a question and cached system prompt (continue generate from cache)
-        if is_first_iteration or not kwargs.get("use_cache", True):
+        # If we're in cached decoding stage, pixel values should be None because input ids do not contain special image token anymore
+        # Otherwise we need pixel values to be passed to model
+        if cache_position[0] == 0:
             model_inputs["pixel_values"] = pixel_values
             model_inputs["pixel_values_videos"] = pixel_values_videos
             model_inputs["image_sizes"] = image_sizes
@@ -949,30 +946,16 @@ class LlavaNextVideoForConditionalGeneration(LlavaNextVideoPreTrainedModel, Gene
 
         return causal_mask
 
-    @auto_docstring
     def get_video_features(
         self,
         pixel_values: torch.FloatTensor,
-        vision_feature_layer: int | list[int] | None = None,
-        vision_feature_select_strategy: str | None = None,
-        **kwargs: Unpack[TransformersKwargs],
-    ) -> tuple | BaseModelOutputWithPooling:
-        r"""
-        pixel_values (`torch.FloatTensor]` of shape `(batch_size, num_frames, channels, height, width)`)
-            The tensors corresponding to the input video.
-        vision_feature_layer (`Union[int, list[int]]`, *optional;*):
-            The index of the layer to select the vision feature. If multiple indices are provided,
-            the vision feature of the corresponding indices will be concatenated to form the
-            vision features.
-        vision_feature_select_strategy (`str`, *optional*):
-            The feature selection strategy used to select the vision feature from the vision backbone.
-            Can be one of `"default"` or `"full"`
-        """
+        vision_feature_layer: Optional[Union[int, list[int]]] = None,
+        vision_feature_select_strategy: Optional[str] = None,
+    ):
         return self.model.get_video_features(
             pixel_values=pixel_values,
             vision_feature_layer=vision_feature_layer,
             vision_feature_select_strategy=vision_feature_select_strategy,
-            **kwargs,
         )