Add video modality for InstrucBLIP (#30182)

* squash in single commit

* add docs

* dummy obj

* more changes in diff converter

* tiny fix

* make docs happy

* skip test

* repo consistency tests

* update docstring

* style

* fix tests

* change diff imports

* [run-slow] instructblipvideo

* [run-slow] instructblipvideo

* fix tests and remove logit check

* [run-slow] instructblipvideo

docs/source/en/_toctree.yml

@@ -776,6 +776,8 @@
         title: Idefics2
       - local: model_doc/instructblip
         title: InstructBLIP
+      - local: model_doc/instructblipvideo
+        title: InstructBlipVideo
       - local: model_doc/kosmos-2
         title: KOSMOS-2
       - local: model_doc/layoutlm

docs/source/en/index.md

@@ -165,6 +165,7 @@ Flax), PyTorch, and/or TensorFlow.
 |                      [ImageGPT](model_doc/imagegpt)                      |       ✅        |         ❌         |      ❌      |
 |                      [Informer](model_doc/informer)                      |       ✅        |         ❌         |      ❌      |
 |                  [InstructBLIP](model_doc/instructblip)                  |       ✅        |         ❌         |      ❌      |
+|             [InstructBlipVideo](model_doc/instructblipvideo)             |       ✅        |         ❌         |      ❌      |
 |                         [Jamba](model_doc/jamba)                         |       ✅        |         ❌         |      ❌      |
 |                        [JetMoe](model_doc/jetmoe)                        |       ✅        |         ❌         |      ❌      |
 |                       [Jukebox](model_doc/jukebox)                       |       ✅        |         ❌         |      ❌      |

docs/source/en/model_doc/instructblip.md

@@ -50,6 +50,7 @@ InstructBLIP uses the same architecture as [BLIP-2](blip2) with a tiny but impor
 
 [[autodoc]] InstructBlipProcessor
 
+
 ## InstructBlipVisionModel
 
 [[autodoc]] InstructBlipVisionModel

docs/source/en/model_doc/instructblipvideo.md

+<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# InstructBlipVideo
+
+## Overview
+
+## Overview
+
+The InstructBLIPVideo is an extension of the models proposed in [InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning](https://arxiv.org/abs/2305.06500) by Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, Steven Hoi.
+InstructBLIPVideo uses the same architecture as [InstructBLIP](instructblip) and works with the same checkpoints as [InstructBLIP](instructblip). The only difference is the ability to process videos.
+
+The abstract from the paper is the following:
+
+*General-purpose language models that can solve various language-domain tasks have emerged driven by the pre-training and instruction-tuning pipeline. However, building general-purpose vision-language models is challenging due to the increased task discrepancy introduced by the additional visual input. Although vision-language pre-training has been widely studied, vision-language instruction tuning remains relatively less explored. In this paper, we conduct a systematic and comprehensive study on vision-language instruction tuning based on the pre-trained BLIP-2 models. We gather a wide variety of 26 publicly available datasets, transform them into instruction tuning format and categorize them into two clusters for held-in instruction tuning and held-out zero-shot evaluation. Additionally, we introduce instruction-aware visual feature extraction, a crucial method that enables the model to extract informative features tailored to the given instruction. The resulting InstructBLIP models achieve state-of-the-art zero-shot performance across all 13 held-out datasets, substantially outperforming BLIP-2 and the larger Flamingo. Our models also lead to state-of-the-art performance when finetuned on individual downstream tasks (e.g., 90.7% accuracy on ScienceQA IMG). Furthermore, we qualitatively demonstrate the advantages of InstructBLIP over concurrent multimodal models.*
+
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/model_doc/instructblip_architecture.jpg"
+alt="drawing" width="600"/>
+
+<small> InstructBLIPVideo architecture. Taken from the <a href="https://arxiv.org/abs/2305.06500">original paper.</a> </small>
+
+This model was contributed by [RaushanTurganbay](https://huggingface.co/RaushanTurganbay).
+The original code can be found [here](https://github.com/salesforce/LAVIS/tree/main/projects/instructblip).
+
+## Usage tips
+
+- The model was trained by sampling 4 frames per video, so it's recommended to sample 4 frames
+
+## InstructBlipVideoConfig
+
+[[autodoc]] InstructBlipVideoConfig
+    - from_vision_qformer_text_configs
+
+## InstructBlipVideoVisionConfig
+
+[[autodoc]] InstructBlipVideoVisionConfig
+
+## InstructBlipVideoQFormerConfig
+
+[[autodoc]] InstructBlipVideoQFormerConfig
+
+## InstructBlipVideoProcessor
+
+[[autodoc]] InstructBlipVideoProcessor
+
+## InstructBlipVideoImageProcessor
+
+[[autodoc]] InstructBlipVideoImageProcessor
+    - preprocess
+
+## InstructBlipVideoVisionModel
+
+[[autodoc]] InstructBlipVideoVisionModel
+    - forward
+
+## InstructBlipVideoQFormerModel
+
+[[autodoc]] InstructBlipVideoQFormerModel
+    - forward
+
+## InstructBlipVideoForConditionalGeneration
+
+[[autodoc]] InstructBlipVideoForConditionalGeneration
+    - forward
+    - generate
\ No newline at end of file

src/transformers/__init__.py

@@ -473,6 +473,12 @@ _import_structure = {
         "InstructBlipQFormerConfig",
         "InstructBlipVisionConfig",
     ],
+    "models.instructblipvideo": [
+        "InstructBlipVideoConfig",
+        "InstructBlipVideoProcessor",
+        "InstructBlipVideoQFormerConfig",
+        "InstructBlipVideoVisionConfig",
+    ],
     "models.jamba": ["JambaConfig"],
     "models.jetmoe": ["JetMoeConfig"],
     "models.kosmos2": [
@@ -1137,6 +1143,7 @@ else:
     _import_structure["models.idefics"].extend(["IdeficsImageProcessor"])
     _import_structure["models.idefics2"].extend(["Idefics2ImageProcessor"])
     _import_structure["models.imagegpt"].extend(["ImageGPTFeatureExtractor", "ImageGPTImageProcessor"])
+    _import_structure["models.instructblipvideo"].extend(["InstructBlipVideoImageProcessor"])
     _import_structure["models.layoutlmv2"].extend(["LayoutLMv2FeatureExtractor", "LayoutLMv2ImageProcessor"])
     _import_structure["models.layoutlmv3"].extend(["LayoutLMv3FeatureExtractor", "LayoutLMv3ImageProcessor"])
     _import_structure["models.levit"].extend(["LevitFeatureExtractor", "LevitImageProcessor"])
@@ -2318,6 +2325,14 @@ else:
             "InstructBlipVisionModel",
         ]
     )
+    _import_structure["models.instructblipvideo"].extend(
+        [
+            "InstructBlipVideoForConditionalGeneration",
+            "InstructBlipVideoPreTrainedModel",
+            "InstructBlipVideoQFormerModel",
+            "InstructBlipVideoVisionModel",
+        ]
+    )
     _import_structure["models.jamba"].extend(
         [
             "JambaForCausalLM",
@@ -5079,6 +5094,12 @@ if TYPE_CHECKING:
         InstructBlipQFormerConfig,
         InstructBlipVisionConfig,
     )
+    from .models.instructblipvideo import (
+        InstructBlipVideoConfig,
+        InstructBlipVideoProcessor,
+        InstructBlipVideoQFormerConfig,
+        InstructBlipVideoVisionConfig,
+    )
     from .models.jamba import JambaConfig
     from .models.jetmoe import JetMoeConfig
     from .models.kosmos2 import (
@@ -5772,6 +5793,7 @@ if TYPE_CHECKING:
         from .models.idefics import IdeficsImageProcessor
         from .models.idefics2 import Idefics2ImageProcessor
         from .models.imagegpt import ImageGPTFeatureExtractor, ImageGPTImageProcessor
+        from .models.instructblipvideo import InstructBlipVideoImageProcessor
         from .models.layoutlmv2 import (
             LayoutLMv2FeatureExtractor,
             LayoutLMv2ImageProcessor,
@@ -6771,6 +6793,12 @@ if TYPE_CHECKING:
             InstructBlipQFormerModel,
             InstructBlipVisionModel,
         )
+        from .models.instructblipvideo import (
+            InstructBlipVideoForConditionalGeneration,
+            InstructBlipVideoPreTrainedModel,
+            InstructBlipVideoQFormerModel,
+            InstructBlipVideoVisionModel,
+        )
         from .models.jamba import (
             JambaForCausalLM,
             JambaForSequenceClassification,

src/transformers/models/__init__.py

@@ -111,6 +111,7 @@ from . import (
     imagegpt,
     informer,
     instructblip,
+    instructblipvideo,
     jamba,
     jetmoe,
     kosmos2,

src/transformers/models/auto/configuration_auto.py

@@ -128,6 +128,7 @@ CONFIG_MAPPING_NAMES = OrderedDict(
         ("imagegpt", "ImageGPTConfig"),
         ("informer", "InformerConfig"),
         ("instructblip", "InstructBlipConfig"),
+        ("instructblipvideo", "InstructBlipVideoConfig"),
         ("jamba", "JambaConfig"),
         ("jetmoe", "JetMoeConfig"),
         ("jukebox", "JukeboxConfig"),
@@ -404,6 +405,7 @@ MODEL_NAMES_MAPPING = OrderedDict(
         ("imagegpt", "ImageGPT"),
         ("informer", "Informer"),
         ("instructblip", "InstructBLIP"),
+        ("instructblipvideo", "InstructBlipVideo"),
         ("jamba", "Jamba"),
         ("jetmoe", "JetMoe"),
         ("jukebox", "Jukebox"),

src/transformers/models/auto/image_processing_auto.py

@@ -89,6 +89,7 @@ else:
             ("idefics2", ("Idefics2ImageProcessor",)),
             ("imagegpt", ("ImageGPTImageProcessor",)),
             ("instructblip", ("BlipImageProcessor",)),
+            ("instructblipvideo", ("InstructBlipVideoImageProcessor",)),
             ("kosmos-2", ("CLIPImageProcessor",)),
             ("layoutlmv2", ("LayoutLMv2ImageProcessor",)),
             ("layoutlmv3", ("LayoutLMv3ImageProcessor",)),
@@ -156,7 +157,6 @@ for model_type, image_processors in IMAGE_PROCESSOR_MAPPING_NAMES.items():
 
     IMAGE_PROCESSOR_MAPPING_NAMES[model_type] = (slow_image_processor_class, fast_image_processor_class)
 
-
 IMAGE_PROCESSOR_MAPPING = _LazyAutoMapping(CONFIG_MAPPING_NAMES, IMAGE_PROCESSOR_MAPPING_NAMES)
 
 

src/transformers/models/auto/modeling_auto.py

@@ -697,6 +697,7 @@ MODEL_FOR_VISION_2_SEQ_MAPPING_NAMES = OrderedDict(
         ("git", "GitForCausalLM"),
         ("idefics2", "Idefics2ForConditionalGeneration"),
         ("instructblip", "InstructBlipForConditionalGeneration"),
+        ("instructblipvideo", "InstructBlipVideoForConditionalGeneration"),
         ("kosmos-2", "Kosmos2ForConditionalGeneration"),
         ("llava", "LlavaForConditionalGeneration"),
         ("llava_next", "LlavaNextForConditionalGeneration"),

src/transformers/models/auto/processing_auto.py

@@ -64,6 +64,7 @@ PROCESSOR_MAPPING_NAMES = OrderedDict(
         ("idefics", "IdeficsProcessor"),
         ("idefics2", "Idefics2Processor"),
         ("instructblip", "InstructBlipProcessor"),
+        ("instructblipvideo", "InstructBlipVideoProcessor"),
         ("kosmos-2", "Kosmos2Processor"),
         ("layoutlmv2", "LayoutLMv2Processor"),
         ("layoutlmv3", "LayoutLMv3Processor"),

src/transformers/models/auto/tokenization_auto.py

@@ -205,6 +205,7 @@ else:
             ("idefics", (None, "LlamaTokenizerFast" if is_tokenizers_available() else None)),
             ("idefics2", ("LlamaTokenizer", "LlamaTokenizerFast" if is_tokenizers_available() else None)),
             ("instructblip", ("GPT2Tokenizer", "GPT2TokenizerFast" if is_tokenizers_available() else None)),
+            ("instructblipvideo", ("GPT2Tokenizer", "GPT2TokenizerFast" if is_tokenizers_available() else None)),
             (
                 "jamba",
                 (

src/transformers/models/blip_2/modeling_blip_2.py

@@ -317,7 +317,7 @@ class Blip2PreTrainedModel(PreTrainedModel):
                 module.bias.data.zero_()
 
         if isinstance(module, Blip2VisionEmbeddings):
-            if hasattr(self.config, "vision_config"):
+            if hasattr(self.config, "vision_config") and not isinstance(self.config, Blip2VisionConfig):
                 factor = self.config.vision_config.initializer_range
             nn.init.trunc_normal_(module.position_embedding, mean=0.0, std=factor)
             nn.init.trunc_normal_(module.class_embedding, mean=0.0, std=factor)

src/transformers/models/instructblip/configuration_instructblip.py

@@ -164,6 +164,8 @@ class InstructBlipQFormerConfig(PretrainedConfig):
             The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
         layer_norm_eps (`float`, *optional*, defaults to 1e-12):
             The epsilon used by the layer normalization layers.
+        pad_token_id (`int`, *optional*, defaults to 0):
+            Token id used for padding sequences.
         position_embedding_type (`str`, *optional*, defaults to `"absolute"`):
             Type of position embedding. Choose one of `"absolute"`, `"relative_key"`, `"relative_key_query"`. For
             positional embeddings use `"absolute"`. For more information on `"relative_key"`, please refer to

src/transformers/models/instructblip/modeling_instructblip.py

@@ -324,7 +324,7 @@ class InstructBlipPreTrainedModel(PreTrainedModel):
                 module.bias.data.zero_()
 
         if isinstance(module, InstructBlipVisionEmbeddings):
-            if hasattr(self.config, "vision_config"):
+            if hasattr(self.config, "vision_config") and not isinstance(self.config, InstructBlipVisionConfig):
                 factor = self.config.vision_config.initializer_range
             nn.init.trunc_normal_(module.position_embedding, mean=0.0, std=factor)
             nn.init.trunc_normal_(module.class_embedding, mean=0.0, std=factor)

src/transformers/models/instructblipvideo/__init__.py

+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import TYPE_CHECKING
+
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_torch_available, is_vision_available
+
+
+_import_structure = {
+    "configuration_instructblipvideo": [
+        "InstructBlipVideoConfig",
+        "InstructBlipVideoQFormerConfig",
+        "InstructBlipVideoVisionConfig",
+    ],
+    "processing_instructblipvideo": ["InstructBlipVideoProcessor"],
+}
+
+
+try:
+    if not is_vision_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["image_processing_instructblipvideo"] = ["InstructBlipVideoImageProcessor"]
+
+
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_instructblipvideo"] = [
+        "InstructBlipVideoQFormerModel",
+        "InstructBlipVideoPreTrainedModel",
+        "InstructBlipVideoForConditionalGeneration",
+        "InstructBlipVideoVisionModel",
+    ]
+
+if TYPE_CHECKING:
+    from .configuration_instructblipvideo import (
+        InstructBlipVideoConfig,
+        InstructBlipVideoQFormerConfig,
+        InstructBlipVideoVisionConfig,
+    )
+    from .processing_instructblipvideo import InstructBlipVideoProcessor
+
+    try:
+        if not is_vision_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .image_processing_instructblipvideo import InstructBlipVideoImageProcessor
+
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_instructblipvideo import (
+            InstructBlipVideoForConditionalGeneration,
+            InstructBlipVideoPreTrainedModel,
+            InstructBlipVideoQFormerModel,
+            InstructBlipVideoVisionModel,
+        )
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)

src/transformers/models/instructblipvideo/configuration_instructblipvideo.py

+#           🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+#               This file was automatically generated from <path_to_diff_file.py>.
+#         Do NOT edit this file manually as any edits will be overwritten by the generation of
+#         the file from the diff. If any change should be done, please apply the change to the
+#                                    diff.py file directly.
+#           🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+# coding=utf-8
+# Copyright 2024 HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+from typing import Union
+
+from ...configuration_utils import PretrainedConfig
+from ...models.auto.modeling_auto import MODEL_FOR_CAUSAL_LM_MAPPING_NAMES
+from ...utils import (
+    logging,
+)
+from ..auto import CONFIG_MAPPING
+
+
+logger = logging.get_logger(__name__)
+
+
+class InstructBlipVideoVisionConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`InstructBlipVideoVisionModel`]. It is used to
+    instantiate a Instructblipvideo vision encoder according to the specified arguments, defining the model architecture.
+    Instantiating a configuration defaults will yield a similar configuration to that of the Instructblipvideo
+    [Salesforce/instruct-blip-flan-t5](https://huggingface.co/Salesforce/instruct-blip-flan-t5) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        hidden_size (`int`, *optional*, defaults to 1408):
+            Dimensionality of the encoder layers and the pooler layer.
+        intermediate_size (`int`, *optional*, defaults to 6144):
+            Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
+        num_hidden_layers (`int`, *optional*, defaults to 39):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 16):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        image_size (`int`, *optional*, defaults to 224):
+            The size (resolution) of each image.
+        patch_size (`int`, *optional*, defaults to 14):
+            The size (resolution) of each patch.
+        hidden_act (`str` or `function`, *optional*, defaults to `"gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"selu"` and `"gelu_new"` ``"gelu"` are supported. to 1e-5): The epsilon used by the layer
+            normalization layers.
+        layer_norm_eps (`float`, *optional*, defaults to 1e-06):
+            The epsilon used by the layer normalization layers.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        initializer_range (`float`, *optional*, defaults to 1e-10):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        qkv_bias (`bool`, *optional*, defaults to `True`):
+            Whether to add a bias to the queries and values in the self-attention layers.
+
+    Example:
+
+    ```python
+    >>> from transformers import InstructBlipVideoVisionConfig, InstructBlipVideoVisionModel
+
+    >>> # Initializing a InstructBlipVideoVisionConfig with Salesforce/instruct-blip-flan-t5 style configuration
+    >>> configuration = InstructBlipVideoVisionConfig()
+
+    >>> # Initializing a InstructBlipVideoVisionModel (with random weights) from the Salesforce/instruct-blip-flan-t5 style configuration
+    >>> model = InstructBlipVideoVisionModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "instructblipvideo_vision_model"
+
+    def __init__(
+        self,
+        hidden_size=1408,
+        intermediate_size=6144,
+        num_hidden_layers=39,
+        num_attention_heads=16,
+        image_size=224,
+        patch_size=14,
+        hidden_act="gelu",
+        layer_norm_eps=1e-6,
+        attention_dropout=0.0,
+        initializer_range=1e-10,
+        qkv_bias=True,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.patch_size = patch_size
+        self.image_size = image_size
+        self.initializer_range = initializer_range
+        self.attention_dropout = attention_dropout
+        self.layer_norm_eps = layer_norm_eps
+        self.hidden_act = hidden_act
+        self.qkv_bias = qkv_bias
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
+        cls._set_token_in_kwargs(kwargs)
+
+        config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
+
+        # get the vision config dict if we are loading from InstructBlipVideoConfig
+        if config_dict.get("model_type") == "instructblipvideo":
+            config_dict = config_dict["vision_config"]
+
+        if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
+            logger.warning(
+                f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
+                f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
+            )
+
+        return cls.from_dict(config_dict, **kwargs)
+
+
+class InstructBlipVideoQFormerConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`InstructBlipVideoQFormerModel`]. It is used to
+    instantiate a Instructblipvideo Querying Transformer (Q-Former) model according to the specified arguments, defining the
+    model architecture. Instantiating a configuration with the defaults will yield a similar configuration to that of
+    the Instructblipvideo [Salesforce/instruct-blip-flan-t5](https://huggingface.co/Salesforce/instruct-blip-flan-t5)
+    architecture. Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs.
+    Read the documentation from [`PretrainedConfig`] for more information.
+
+    Note that [`InstructBlipVideoQFormerModel`] is very similar to [`BertLMHeadModel`] with interleaved cross-attention.
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 30522):
+            Vocabulary size of the Q-Former model. Defines the number of different tokens that can be represented by
+            the `inputs_ids` passed when calling the model.
+        hidden_size (`int`, *optional*, defaults to 768):
+            Dimensionality of the encoder layers and the pooler layer.
+        num_hidden_layers (`int`, *optional*, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 12):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        intermediate_size (`int`, *optional*, defaults to 3072):
+            Dimensionality of the "intermediate" (often named feed-forward) layer in the Transformer encoder.
+        hidden_act (`str` or `Callable`, *optional*, defaults to `"gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"silu"` and `"gelu_new"` are supported.
+        hidden_dropout_prob (`float`, *optional*, defaults to 0.1):
+            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_probs_dropout_prob (`float`, *optional*, defaults to 0.1):
+            The dropout ratio for the attention probabilities.
+        max_position_embeddings (`int`, *optional*, defaults to 512):
+            The maximum sequence length that this model might ever be used with. Typically set this to something large
+            just in case (e.g., 512 or 1024 or 2048).
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        layer_norm_eps (`float`, *optional*, defaults to 1e-12):
+            The epsilon used by the layer normalization layers.
+        pad_token_id (`int`, *optional*, defaults to 0):
+            Token id used for padding sequences.
+        position_embedding_type (`str`, *optional*, defaults to `"absolute"`):
+            Type of position embedding. Choose one of `"absolute"`, `"relative_key"`, `"relative_key_query"`. For
+            positional embeddings use `"absolute"`. For more information on `"relative_key"`, please refer to
+            [Self-Attention with Relative Position Representations (Shaw et al.)](https://arxiv.org/abs/1803.02155).
+            For more information on `"relative_key_query"`, please refer to *Method 4* in [Improve Transformer Models
+            with Better Relative Position Embeddings (Huang et al.)](https://arxiv.org/abs/2009.13658).
+        cross_attention_frequency (`int`, *optional*, defaults to 2):
+            The frequency of adding cross-attention to the Transformer layers.
+        encoder_hidden_size (`int`, *optional*, defaults to 1408):
+            The hidden size of the hidden states for cross-attention.
+
+    Examples:
+
+    ```python
+    >>> from transformers import InstructBlipVideoQFormerConfig, InstructBlipVideoQFormerModel
+
+    >>> # Initializing a Instructblipvideo Salesforce/instruct-blip-flan-t5 style configuration
+    >>> configuration = InstructBlipVideoQFormerConfig()
+
+    >>> # Initializing a model (with random weights) from the Salesforce/instruct-blip-flan-t5 style configuration
+    >>> model = InstructBlipVideoQFormerModel(configuration)
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "instructblipvideo_qformer"
+
+    def __init__(
+        self,
+        vocab_size=30522,
+        hidden_size=768,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        intermediate_size=3072,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        max_position_embeddings=512,
+        initializer_range=0.02,
+        layer_norm_eps=1e-12,
+        pad_token_id=0,
+        position_embedding_type="absolute",
+        cross_attention_frequency=2,
+        encoder_hidden_size=1408,
+        **kwargs,
+    ):
+        super().__init__(pad_token_id=pad_token_id, **kwargs)
+
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.hidden_act = hidden_act
+        self.intermediate_size = intermediate_size
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.max_position_embeddings = max_position_embeddings
+        self.initializer_range = initializer_range
+        self.layer_norm_eps = layer_norm_eps
+        self.position_embedding_type = position_embedding_type
+        self.cross_attention_frequency = cross_attention_frequency
+        self.encoder_hidden_size = encoder_hidden_size
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
+        cls._set_token_in_kwargs(kwargs)
+
+        config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
+
+        # get the qformer config dict if we are loading from InstructBlipVideoConfig
+        if config_dict.get("model_type") == "instructblipvideo":
+            config_dict = config_dict["qformer_config"]
+
+        if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
+            logger.warning(
+                f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
+                f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
+            )
+
+        return cls.from_dict(config_dict, **kwargs)
+
+
+class InstructBlipVideoConfig(PretrainedConfig):
+    r"""
+    [`InstructBlipVideoConfig`] is the configuration class to store the configuration of a
+    [`InstructBlipVideoForConditionalGeneration`]. It is used to instantiate a Instructblipvideo model according to the specified
+    arguments, defining the vision model, Q-Former model and language model configs. Instantiating a configuration with
+    the defaults will yield a similar configuration to that of the Instructblipvideo
+    [Salesforce/instruct-blip-flan-t5](https://huggingface.co/Salesforce/instruct-blip-flan-t5) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        vision_config (`dict`, *optional*):
+            Dictionary of configuration options used to initialize [`InstructBlipVideoVisionConfig`].
+        qformer_config (`dict`, *optional*):
+            Dictionary of configuration options used to initialize [`InstructBlipVideoQFormerConfig`].
+        text_config (`dict`, *optional*):
+            Dictionary of configuration options used to initialize any [`PretrainedConfig`].
+        num_query_tokens (`int`, *optional*, defaults to 32):
+            The number of query tokens passed through the Transformer.
+
+        kwargs (*optional*):
+            Dictionary of keyword arguments.
+
+    Example:
+
+    ```python
+    >>> from transformers import (
+    ...     InstructBlipVideoVisionConfig,
+    ...     InstructBlipVideoQFormerConfig,
+    ...     OPTConfig,
+    ...     InstructBlipVideoConfig,
+    ...     InstructBlipVideoForConditionalGeneration,
+    ... )
+
+    >>> # Initializing a InstructBlipVideoConfig with Salesforce/instruct-blip-flan-t5 style configuration
+    >>> configuration = InstructBlipVideoConfig()
+
+    >>> # Initializing a InstructBlipVideoForConditionalGeneration (with random weights) from the Salesforce/instruct-blip-flan-t5 style configuration
+    >>> model = InstructBlipVideoForConditionalGeneration(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+
+    >>> # We can also initialize a InstructBlipVideoConfig from a InstructBlipVideoVisionConfig, InstructBlipVideoQFormerConfig and any PretrainedConfig
+
+    >>> # Initializing Instructblipvideo vision, Instructblipvideo Q-Former and language model configurations
+    >>> vision_config = InstructBlipVideoVisionConfig()
+    >>> qformer_config = InstructBlipVideoQFormerConfig()
+    >>> text_config = OPTConfig()
+
+    >>> config = InstructBlipVideoConfig.from_text_vision_configs(vision_config, qformer_config, text_config)
+    ```"""
+
+    model_type = "instructblipvideo"
+
+    def __init__(self, vision_config=None, qformer_config=None, text_config=None, num_query_tokens=32, **kwargs):
+        super().__init__(**kwargs)
+
+        if vision_config is None:
+            vision_config = {}
+            logger.info("vision_config is None. initializing the InstructBlipVideoVisionConfig with default values.")
+
+        if qformer_config is None:
+            qformer_config = {}
+            logger.info("qformer_config is None. Initializing the InstructBlipVideoQFormerConfig with default values.")
+
+        if text_config is None:
+            text_config = {}
+            logger.info("text_config is None. Initializing the text config with default values (`OPTConfig`).")
+
+        self.vision_config = InstructBlipVideoVisionConfig(**vision_config)
+        self.qformer_config = InstructBlipVideoQFormerConfig(**qformer_config)
+        text_model_type = text_config["model_type"] if "model_type" in text_config else "opt"
+        self.text_config = CONFIG_MAPPING[text_model_type](**text_config)
+
+        self.tie_word_embeddings = self.text_config.tie_word_embeddings
+        self.is_encoder_decoder = self.text_config.is_encoder_decoder
+
+        self.num_query_tokens = num_query_tokens
+        self.qformer_config.encoder_hidden_size = self.vision_config.hidden_size
+        self.use_decoder_only_language_model = self.text_config.model_type in MODEL_FOR_CAUSAL_LM_MAPPING_NAMES
+        self.initializer_factor = 1.0
+        self.initializer_range = 0.02
+
+    @classmethod
+    def from_vision_qformer_text_configs(
+        cls,
+        vision_config: InstructBlipVideoVisionConfig,
+        qformer_config: InstructBlipVideoQFormerConfig,
+        text_config: PretrainedConfig,
+        **kwargs,
+    ):
+        r"""
+        Instantiate a [`InstructBlipVideoConfig`] (or a derived class) from a Instructblipvideo vision model, Q-Former and
+        language model configurations.
+
+        Returns:
+            [`InstructBlipVideoConfig`]: An instance of a configuration object
+        """
+
+        return cls(
+            vision_config=vision_config.to_dict(),
+            qformer_config=qformer_config.to_dict(),
+            text_config=text_config.to_dict(),
+            **kwargs,
+        )

src/transformers/models/instructblipvideo/convert_instructblipvideo_original_to_pytorch.py

+# coding=utf-8
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Convert InstructBlipVideo checkpoints from the original repository.
+
+URL: https://github.com/salesforce/LAVIS/tree/main/projects/instructblipvideo
+"""
+
+import argparse
+
+import requests
+import torch
+
+# pip3 install salesforce-lavis
+# I'm actually installing a slightly modified version: pip3 install git+https://github.com/nielsrogge/LAVIS.git@fix_lavis_float32 (there's also the fix_lavis branch)
+# also note: to convert Vicuna checkpoints, we had to include /home/niels/python_projects/checkpoints/FastChat/vicuna-7b in lavis/configs/models/blip2/blip2_instruct_vicuna7b.yaml
+# same for Vicuna-13b
+from lavis.models import load_model_and_preprocess
+from PIL import Image
+
+from transformers import (
+    AutoTokenizer,
+    BlipImageProcessor,
+    InstructBlipProcessor,
+    InstructBlipVideoConfig,
+    InstructBlipVideoForConditionalGeneration,
+    InstructBlipVideoQFormerConfig,
+    InstructBlipVideoVisionConfig,
+    LlamaConfig,
+    LlamaTokenizerFast,
+    T5Config,
+    T5TokenizerFast,
+)
+from transformers.utils.constants import OPENAI_CLIP_MEAN, OPENAI_CLIP_STD
+
+
+def load_demo_image():
+    url = "https://raw.githubusercontent.com/salesforce/LAVIS/main/docs/_static/Confusing-Pictures.jpg"
+    image = Image.open(requests.get(url, stream=True).raw).convert("RGB")
+
+    return image
+
+
+# here we list all keys to be renamed (original name on the left, our name on the right)
+def create_rename_keys(config):
+    rename_keys = []
+    # fmt: off
+
+    # vision encoder
+    rename_keys.append(("visual_encoder.cls_token", "vision_model.embeddings.class_embedding"))
+    rename_keys.append(("visual_encoder.pos_embed", "vision_model.embeddings.position_embedding"))
+    rename_keys.append(("visual_encoder.patch_embed.proj.weight", "vision_model.embeddings.patch_embedding.weight"))
+    rename_keys.append(("visual_encoder.patch_embed.proj.bias", "vision_model.embeddings.patch_embedding.bias"))
+    rename_keys.append(("ln_vision.weight", "vision_model.post_layernorm.weight"))
+    rename_keys.append(("ln_vision.bias", "vision_model.post_layernorm.bias"))
+
+    for i in range(config.vision_config.num_hidden_layers):
+        rename_keys.append((f"visual_encoder.blocks.{i}.norm1.weight", f"vision_model.encoder.layers.{i}.layer_norm1.weight"))
+        rename_keys.append((f"visual_encoder.blocks.{i}.norm1.bias", f"vision_model.encoder.layers.{i}.layer_norm1.bias"))
+        rename_keys.append((f"visual_encoder.blocks.{i}.norm2.weight", f"vision_model.encoder.layers.{i}.layer_norm2.weight"))
+        rename_keys.append((f"visual_encoder.blocks.{i}.norm2.bias", f"vision_model.encoder.layers.{i}.layer_norm2.bias"))
+        rename_keys.append((f"visual_encoder.blocks.{i}.attn.qkv.weight", f"vision_model.encoder.layers.{i}.self_attn.qkv.weight"))
+        rename_keys.append((f"visual_encoder.blocks.{i}.attn.proj.weight", f"vision_model.encoder.layers.{i}.self_attn.projection.weight",))
+        rename_keys.append((f"visual_encoder.blocks.{i}.attn.proj.bias", f"vision_model.encoder.layers.{i}.self_attn.projection.bias"))
+        rename_keys.append((f"visual_encoder.blocks.{i}.mlp.fc1.weight", f"vision_model.encoder.layers.{i}.mlp.fc1.weight"))
+        rename_keys.append((f"visual_encoder.blocks.{i}.mlp.fc1.bias", f"vision_model.encoder.layers.{i}.mlp.fc1.bias"))
+        rename_keys.append((f"visual_encoder.blocks.{i}.mlp.fc2.weight", f"vision_model.encoder.layers.{i}.mlp.fc2.weight"))
+        rename_keys.append((f"visual_encoder.blocks.{i}.mlp.fc2.bias", f"vision_model.encoder.layers.{i}.mlp.fc2.bias"))
+
+    # QFormer
+    rename_keys.append(("Qformer.bert.embeddings.LayerNorm.weight", "qformer.embeddings.layernorm.weight"))
+    rename_keys.append(("Qformer.bert.embeddings.LayerNorm.bias", "qformer.embeddings.layernorm.bias"))
+
+    # fmt: on
+    return rename_keys
+
+
+def rename_key(dct, old, new):
+    val = dct.pop(old)
+    dct[new] = val
+
+
+def read_in_q_v_bias(state_dict, config):
+    for i in range(config.vision_config.num_hidden_layers):
+        # read in original q and v biases
+        q_bias = state_dict.pop(f"visual_encoder.blocks.{i}.attn.q_bias")
+        v_bias = state_dict.pop(f"visual_encoder.blocks.{i}.attn.v_bias")
+
+        # next, set bias in the state dict
+        qkv_bias = torch.cat((q_bias, torch.zeros_like(v_bias, requires_grad=False), v_bias))
+        state_dict[f"vision_model.encoder.layers.{i}.self_attn.qkv.bias"] = qkv_bias
+
+
+def get_blip2_config(model_name):
+    image_size = 364 if "coco" in model_name else 224
+    vision_config = InstructBlipVideoVisionConfig(image_size=image_size).to_dict()
+
+    # make sure the models have proper bos_token_id and eos_token_id set (important for generation)
+    # seems like flan-T5 models don't have bos_token_id properly set?
+    if "t5-xl" in model_name:
+        text_config = T5Config.from_pretrained("google/flan-t5-xl", dense_act_fn="gelu", bos_token_id=1).to_dict()
+    elif "t5-xxl" in model_name:
+        text_config = T5Config.from_pretrained("google/flan-t5-xxl", dense_act_fn="gelu", bos_token_id=1).to_dict()
+    elif "vicuna-7b" in model_name:
+        text_config = LlamaConfig.from_pretrained("decapoda-research/llama-7b-hf", vocab_size=32001).to_dict()
+    elif "vicuna-13b" in model_name:
+        text_config = LlamaConfig.from_pretrained("decapoda-research/llama-13b-hf", vocab_size=32001).to_dict()
+    else:
+        raise ValueError("Model name not supported")
+
+    # the authors add one special "[DEC]" token to the vocab of Q-Former, hence vocab size = 30522 + 1
+    qformer_config = InstructBlipVideoQFormerConfig(vocab_size=30523).to_dict()
+    config = InstructBlipVideoConfig(
+        vision_config=vision_config, text_config=text_config, qformer_config=qformer_config
+    )
+
+    return config, image_size
+
+
+@torch.no_grad()
+def convert_blip2_checkpoint(model_name, pytorch_dump_folder_path=None, push_to_hub=False):
+    """
+    Copy/paste/tweak model's weights to Transformers design.
+    """
+    qformer_tokenizer = AutoTokenizer.from_pretrained("google-bert/bert-base-uncased", truncation_side="left")
+    qformer_tokenizer.add_special_tokens({"bos_token": "[DEC]"})
+
+    if "t5" in model_name:
+        tokenizer = T5TokenizerFast.from_pretrained("google/flan-t5-xl", truncation_side="left")
+    elif "vicuna" in model_name:
+        # the following was used in the original implementation:
+        # tokenizer = LlamaTokenizer.from_pretrained("huggyllama/llama-7b", use_fast=False, truncation_side="left")
+        # tokenizer.add_special_tokens({"pad_token": "[PAD]"})
+        # tokenizer.add_special_tokens({"bos_token": "</s>"})
+        # tokenizer.add_special_tokens({"eos_token": "</s>"})
+        # tokenizer.add_special_tokens({"unk_token": "</s>"})
+        tokenizer = LlamaTokenizerFast.from_pretrained(
+            "huggyllama/llama-7b", truncation_side="left", bos_token="</s>", unk_token="</s>"
+        )
+        tokenizer.add_special_tokens({"pad_token": "[PAD]"})
+
+    config, image_size = get_blip2_config(model_name)
+    hf_model = InstructBlipVideoForConditionalGeneration(config).eval()
+
+    model_name_to_original = {
+        "instructblipvideo-vicuna-7b": ("blip2_vicuna_instruct", "vicuna7b"),
+        "instructblipvideo-vicuna-13b": ("blip2_vicuna_instruct", "vicuna13b"),
+        "instructblipvideo-flan-t5-xl": ("blip2_t5_instruct", "flant5xl"),
+        "instructblipvideo-flan-t5-xxl": ("blip2_t5_instruct", "flant5xxl"),
+    }
+
+    name, type = model_name_to_original[model_name]
+
+    # load original model
+    print("Loading original model...")
+    hf_model_device = "cuda:1" if torch.cuda.is_available() else "cpu"
+    lavis_device = "cuda:2" if torch.cuda.is_available() else "cpu"
+    original_model, vis_processors, _ = load_model_and_preprocess(
+        name=name, model_type=type, is_eval=True, device=lavis_device
+    )
+    original_model.eval()
+    print("Done!")
+
+    # update state dict keys
+    state_dict = original_model.state_dict()
+    rename_keys = create_rename_keys(config)
+    for src, dest in rename_keys:
+        rename_key(state_dict, src, dest)
+
+    # some keys can be renamed efficiently
+    for key, val in state_dict.copy().items():
+        val = state_dict.pop(key)
+        if key.startswith("Qformer.bert"):
+            key = key.replace("Qformer.bert", "qformer")
+        if "attention.self" in key:
+            key = key.replace("self", "attention")
+        if "llm_proj" in key:
+            key = key.replace("llm_proj", "language_projection")
+        if "t5_proj" in key:
+            key = key.replace("t5_proj", "language_projection")
+        if key.startswith("llm_model"):
+            key = key.replace("llm_model", "language_model")
+        if key.startswith("t5"):
+            key = key.replace("t5", "language")
+        state_dict[key] = val
+
+    # read in qv biases
+    read_in_q_v_bias(state_dict, config)
+
+    # note: weights get loaded in torch.float32 by default
+    hf_model.load_state_dict(state_dict, strict=True)
+
+    image = load_demo_image()
+    prompt = "What is unusual about this image?"
+
+    # create processor
+    image_processor = BlipImageProcessor(
+        size={"height": image_size, "width": image_size}, image_mean=OPENAI_CLIP_MEAN, image_std=OPENAI_CLIP_STD
+    )
+    processor = InstructBlipProcessor(
+        image_processor=image_processor,
+        tokenizer=tokenizer,
+        qformer_tokenizer=qformer_tokenizer,
+    )
+    inputs = processor(images=image, text=prompt, return_tensors="pt").to(hf_model_device)
+
+    # make sure processor creates exact same pixel values
+    original_pixel_values = vis_processors["eval"](image).unsqueeze(0).to(lavis_device)
+    pixel_values = inputs.pixel_values
+    assert torch.allclose(original_pixel_values.to(pixel_values.device), pixel_values)
+
+    original_model.to(lavis_device)
+    hf_model.to(hf_model_device)
+    with torch.no_grad():
+        if "vicuna" in model_name:
+            original_logits = original_model({"image": original_pixel_values, "text_input": [prompt]}).logits
+            logits = hf_model(**inputs).logits
+        else:
+            original_logits = original_model(
+                {"image": original_pixel_values, "text_input": [prompt], "text_output": ["\n"]}
+            ).logits
+            label_input_ids = tokenizer("\n", return_tensors="pt").input_ids.to(hf_model_device)
+            labels = label_input_ids.masked_fill(label_input_ids == tokenizer.pad_token_id, -100)
+            logits = hf_model(**inputs, labels=labels).logits
+
+    print("First values of original logits:", original_logits[0, :3, :3])
+    print("First values of HF logits:", logits[0, :3, :3])
+
+    # assert values
+    assert original_logits.shape == logits.shape
+    atol = 1e-4 if "vicuna" in model_name else 1e-5
+    assert torch.allclose(original_logits.to(logits.device), logits, atol=atol)
+    print("Looks ok!")
+
+    print("Generating with original model...")
+    original_outputs = original_model.generate({"image": original_pixel_values, "prompt": prompt}, num_beams=5)
+
+    # important: we need to cast the weights of the HF model to the appropriate type
+    print("Generating with HF model...")
+    outputs = hf_model.generate(
+        **inputs,
+        do_sample=False,
+        num_beams=5,
+        max_length=256,
+        min_length=1,
+        top_p=0.9,
+        repetition_penalty=1.5,
+        length_penalty=1.0,
+        temperature=1,
+    )
+    if "vicuna" in model_name:
+        # convert output id 0 to 2 (eos_token_id)
+        # TODO add this in the generate method?
+        outputs[outputs == 0] = 2
+    print("Original generation:", original_outputs)
+    output_text = processor.batch_decode(outputs, skip_special_tokens=True)
+    output_text = [text.strip() for text in output_text]
+    print("HF generation:", output_text)
+
+    if pytorch_dump_folder_path is not None:
+        processor.save_pretrained(pytorch_dump_folder_path)
+        hf_model.save_pretrained(pytorch_dump_folder_path)
+
+    if push_to_hub:
+        processor.push_to_hub(f"Salesforce/{model_name}")
+        hf_model.push_to_hub(f"Salesforce/{model_name}")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    choices = [
+        "instructblipvideo-vicuna-7b",
+        "instructblipvideo-vicuna-13b",
+        "instructblipvideo-flan-t5-xl",
+        "instructblipvideo-flan-t5-xxl",
+    ]
+    parser.add_argument(
+        "--model_name",
+        default="instructblipvideo-flan-t5-xl",
+        choices=choices,
+        type=str,
+        help="Path to hf config.json of model to convert",
+    )
+    parser.add_argument("--pytorch_dump_folder_path", default=None, type=str, help="Path to the output PyTorch model.")
+    parser.add_argument(
+        "--push_to_hub",
+        action="store_true",
+        help="Whether to push the model and processor to the hub after converting",
+    )
+
+    args = parser.parse_args()
+
+    convert_blip2_checkpoint(args.model_name, args.pytorch_dump_folder_path, args.push_to_hub)

src/transformers/models/instructblipvideo/diff_instructblipvideo.py

+# coding=utf-8
+# Copyright 2024 HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+from typing import Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch.nn import CrossEntropyLoss
+
+from transformers.models.instructblip.configuration_instructblip import (
+    InstructBlipConfig,
+    InstructBlipQFormerConfig,
+    InstructBlipVisionConfig,
+)
+from transformers.models.instructblip.modeling_instructblip import (
+    InstructBlipAttention,
+    InstructBlipEncoder,
+    InstructBlipEncoderLayer,
+    InstructBlipForConditionalGeneration,
+    InstructBlipForConditionalGenerationModelOutput,
+    InstructBlipMLP,
+    InstructBlipPreTrainedModel,
+    InstructBlipQFormerAttention,
+    InstructBlipQFormerEmbeddings,
+    InstructBlipQFormerEncoder,
+    InstructBlipQFormerIntermediate,
+    InstructBlipQFormerLayer,
+    InstructBlipQFormerModel,
+    InstructBlipQFormerOutput,
+    InstructBlipQFormerSelfOutput,
+    InstructBlipVisionEmbeddings,
+    InstructBlipVisionModel,
+)
+
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class InstructBlipVideoVisionConfig(InstructBlipVisionConfig):
+    pass
+
+
+class InstructBlipVideoQFormerConfig(InstructBlipQFormerConfig):
+    pass
+
+
+class InstructBlipVideoConfig(InstructBlipConfig):
+    pass
+
+
+@dataclass
+class InstructBlipVideoForConditionalGenerationModelOutput(InstructBlipForConditionalGenerationModelOutput):
+    pass
+
+
+class InstructBlipVideoVisionEmbeddings(InstructBlipVisionEmbeddings):
+    pass
+
+
+class InstructBlipVideoAttention(InstructBlipAttention):
+    pass
+
+
+class InstructBlipVideoMLP(InstructBlipMLP):
+    pass
+
+
+class InstructBlipVideoEncoderLayer(InstructBlipEncoderLayer):
+    pass
+
+
+class InstructBlipVideoPreTrainedModel(InstructBlipPreTrainedModel):
+    pass
+
+
+class InstructBlipVideoEncoder(InstructBlipEncoder):
+    pass
+
+
+class InstructBlipVideoVisionModel(InstructBlipVisionModel):
+    pass
+
+
+class InstructBlipVideoQFormerSelfOutput(InstructBlipQFormerSelfOutput):
+    pass
+
+
+class InstructBlipVideoQFormerAttention(InstructBlipQFormerAttention):
+    pass
+
+
+class InstructBlipVideoQFormerIntermediate(InstructBlipQFormerIntermediate):
+    pass
+
+
+class InstructBlipVideoQFormerOutput(InstructBlipQFormerOutput):
+    pass
+
+
+class InstructBlipVideoQFormerLayer(InstructBlipQFormerLayer):
+    pass
+
+
+class InstructBlipVideoQFormerEncoder(InstructBlipQFormerEncoder):
+    pass
+
+
+class InstructBlipVideoQFormerEmbeddings(InstructBlipQFormerEmbeddings):
+    pass
+
+
+class InstructBlipVideoQFormerModel(InstructBlipQFormerModel):
+    pass
+
+
+class InstructBlipVideoForConditionalGeneration(InstructBlipForConditionalGeneration):
+    def forward(
+        self,
+        pixel_values: torch.FloatTensor,
+        qformer_input_ids: torch.FloatTensor,
+        qformer_attention_mask: Optional[torch.LongTensor] = None,
+        input_ids: Optional[torch.FloatTensor] = None,
+        attention_mask: Optional[torch.LongTensor] = None,
+        decoder_input_ids: Optional[torch.LongTensor] = None,
+        decoder_attention_mask: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        labels: Optional[torch.LongTensor] = None,
+        return_dict: Optional[bool] = None,
+        interpolate_pos_encoding: bool = False,
+    ) -> Union[Tuple, InstructBlipVideoForConditionalGenerationModelOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the language modeling loss. Indices should be in `[-100, 0, ..., config.vocab_size -
+            1]`. All labels set to `-100` are ignored (masked), the loss is only computed for labels in `[0, ...,
+            config.vocab_size]`
+
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from transformers import InstructBlipVideoProcessor, InstructBlipVideoForConditionalGeneration
+        >>> import torch
+        >>> from huggingface_hub import hf_hub_download
+        >>> from av
+
+        >>> def read_video_pyav(container, indices):
+        ...     '''
+        ...     Decode the video with PyAV decoder.
+        ...     Args:
+        ...         container (`av.container.input.InputContainer`): PyAV container.
+        ...         indices (`List[int]`): List of frame indices to decode.
+        ...     Returns:
+        ...         result (np.ndarray): np array of decoded frames of shape (num_frames, height, width, 3).
+        ...     '''
+        ...     frames = []
+        ...     container.seek(0)
+        ...     start_index = indices[0]
+        ...     end_index = indices[-1]
+        ...     for i, frame in enumerate(container.decode(video=0)):
+        ...         if i > end_index:
+        ...             break
+        ...         if i >= start_index and i in indices:
+        ...             frames.append(frame)
+        ...     return np.stack([x.to_ndarray(format="rgb24") for x in frames])
+
+        >>> model = InstructBlipVideoProcessor.from_pretrained("Salesforce/instructblip-vicuna-7b", device_map="auto")
+        >>> processor = InstructBlipVideoForConditionalGeneration.from_pretrained("Salesforce/instructblip-vicuna-7b")
+
+        >>> file_path = hf_hub_download(
+                repo_id="nielsr/video-demo", filename="eating_spaghetti.mp4", repo_type="dataset"
+            )
+        >>> container = av.open(video_path)
+        >>> # sample uniformly 4 frames from the videWhy is this video funny?o
+        >>> total_frames = container.streams.video[0].frames
+        >>> indices = np.arange(0, total_frames, total_frames / 4).astype(int)
+        >>> clip = read_video_pyav(container, indices)
+
+        >>> prompt = "What is happening in the video?"
+        >>> inputs = processor(videos=clip, text=prompt, return_tensors="pt").to(device)
+
+        >>> outputs = model.generate(
+        ...     **inputs,
+        ...     do_sample=False,
+        ...     num_beams=5,
+        ...     max_length=256,
+        ...     repetition_penalty=1.5,
+        ...     length_penalty=1.0,
+        ... )
+        >>> generated_text = processor.batch_decode(outputs, skip_special_tokens=True)[0].strip()
+        >>> print(generated_text)
+        "A person is eating a bowl of pasta, and they are using a fork to eat it. The person is sitting at a table, and the plate of pasta is on the table in front"
+        ```"""
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # step 1: forward the images through the vision encoder,
+        # we process in a batched way, later unbatch it back (video has frames=4 always)
+        batch_size, frames, channel, height, width = pixel_values.shape
+        pixel_values = pixel_values.reshape(batch_size * frames, channel, height, width)
+
+        vision_outputs = self.vision_model(
+            pixel_values=pixel_values,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            interpolate_pos_encoding=interpolate_pos_encoding,
+        )
+        image_embeds = vision_outputs[0]
+
+        # step 2: forward the query tokens through the QFormer, using the image embeddings for cross-attention
+        image_attention_mask = torch.ones(image_embeds.size()[:-1], dtype=torch.long, device=image_embeds.device)
+
+        # difference with BLIP-2 here: we also feed the instruction prompt to the Q-Former
+        query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)
+        query_attention_mask = torch.ones(query_tokens.size()[:-1], dtype=torch.long, device=image_embeds.device)
+
+        if qformer_attention_mask is None:
+            qformer_attention_mask = torch.ones_like(qformer_input_ids)
+
+        qformer_input_ids = qformer_input_ids.repeat_interleave(frames, dim=0)
+        qformer_attention_mask = qformer_attention_mask.repeat_interleave(frames, dim=0)
+        qformer_attention_mask = torch.cat([query_attention_mask, qformer_attention_mask], dim=1)
+        query_outputs = self.qformer(
+            input_ids=qformer_input_ids,
+            attention_mask=qformer_attention_mask,
+            query_embeds=query_tokens,
+            encoder_hidden_states=image_embeds,
+            encoder_attention_mask=image_attention_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        query_output = query_outputs[0][:, : query_tokens.size(1), :]
+
+        # step 3: use the language model, conditioned on the query outputs and the prompt
+        language_model_inputs = self.language_projection(query_output)
+
+        # unbatch inputs back, each video-frame gets `num_query_tokens` seq length
+        language_model_inputs = language_model_inputs.reshape(batch_size, self.config.num_query_tokens * frames, -1)
+        language_model_attention_mask = torch.ones(
+            language_model_inputs.size()[:-1], dtype=torch.long, device=language_model_inputs.device
+        )
+
+        inputs_embeds = self.language_model.get_input_embeddings()(input_ids)
+        inputs_embeds = torch.cat([language_model_inputs, inputs_embeds.to(language_model_inputs.device)], dim=1)
+
+        if attention_mask is None:
+            attention_mask = torch.ones_like(input_ids)
+        attention_mask = torch.cat([language_model_attention_mask.to(attention_mask.device), attention_mask], dim=1)
+
+        if self.config.use_decoder_only_language_model:
+            outputs = self.language_model(
+                inputs_embeds=inputs_embeds,
+                attention_mask=attention_mask,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+            )
+            logits = outputs.logits if return_dict else outputs[0]
+            loss = None
+            # we compute the loss here since we need to take into account the sequence length of the query embeds
+            if labels is not None:
+                labels = labels.to(logits.device)
+                logits = logits[:, -labels.size(1) :, :]
+                # Shift so that tokens < n predict n
+                shift_logits = logits[..., :-1, :].contiguous()
+                shift_labels = labels[..., 1:].contiguous().to(logits.device)
+
+                # Flatten the tokens
+                loss_fct = CrossEntropyLoss(reduction="mean")
+
+                loss = loss_fct(shift_logits.view(-1, self.config.text_config.vocab_size), shift_labels.view(-1))
+        else:
+            outputs = self.language_model(
+                inputs_embeds=inputs_embeds,
+                attention_mask=attention_mask,
+                decoder_input_ids=decoder_input_ids,
+                decoder_attention_mask=decoder_attention_mask,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+                labels=labels,
+            )
+            loss = outputs.loss if return_dict else outputs[0]
+            logits = outputs.logits if return_dict else outputs[1]
+
+        if not return_dict:
+            output = (logits, vision_outputs, query_outputs, outputs)
+            return ((loss,) + output) if loss is not None else output
+
+        return InstructBlipVideoForConditionalGenerationModelOutput(
+            loss=loss,
+            logits=logits,
+            vision_outputs=vision_outputs,
+            qformer_outputs=query_outputs,
+            language_model_outputs=outputs,
+        )
+
+    @torch.no_grad()
+    def generate(
+        self,
+        pixel_values: torch.FloatTensor,
+        qformer_input_ids: Optional[torch.LongTensor] = None,
+        qformer_attention_mask: Optional[torch.LongTensor] = None,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.LongTensor] = None,
+        interpolate_pos_encoding: bool = False,
+        **generate_kwargs,
+    ) -> torch.LongTensor:
+        """
+        Overrides `generate` function to be able to use the model as a conditional generator.
+
+        Args:
+           pixel_values (`torch.FloatTensor` of shape (batch_size, num_channels, height, width) or
+                (batch_size, num_frames, num_channels, height, width)): Input images or videos to be processed.
+            qformer_input_ids (`torch.LongTensor` of shape (batch_size, sequence_length), *optional*):
+                The sequence used as a prompt to be fed to the Q-Former module.
+            qformer_attention_mask (`torch.LongTensor` of shape (batch_size, sequence_length), *optional*):
+                Mask to avoid performing attention on padding token indices.
+            input_ids (`torch.LongTensor` of shape (batch_size, sequence_length), *optional*):
+                The sequence used as a prompt for the generation.
+            attention_mask (`torch.LongTensor` of shape (batch_size, sequence_length), *optional*):
+                Mask to avoid performing attention on padding token indices.
+            interpolate_pos_encoding (`bool`, *optional*, defaults to `False`):
+                Whether to interpolate the positional encoding of the image embeddings.
+
+        Returns:
+            captions (list): A list of strings of length batch_size * num_captions.
+        """
+        if hasattr(self, "hf_device_map"):
+            # preprocess for `accelerate`
+            self._preprocess_accelerate()
+
+        # we process in a batched way, later unbatch it back (video has frames=4)
+        batch_size, frames, channel, height, width = pixel_values.shape
+        pixel_values = pixel_values.reshape(batch_size * frames, channel, height, width)
+
+        image_embeds = self.vision_model(
+            pixel_values,
+            return_dict=True,
+            interpolate_pos_encoding=interpolate_pos_encoding,
+        ).last_hidden_state
+        image_attention_mask = torch.ones(image_embeds.size()[:-1], dtype=torch.long, device=image_embeds.device)
+
+        query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)
+        query_attention_mask = torch.ones(query_tokens.size()[:-1], dtype=torch.long, device=image_embeds.device)
+        if qformer_attention_mask is None:
+            qformer_attention_mask = torch.ones_like(qformer_input_ids)
+
+        qformer_input_ids = qformer_input_ids.repeat_interleave(frames, dim=0)
+        qformer_attention_mask = qformer_attention_mask.repeat_interleave(frames, dim=0)
+        qformer_attention_mask = torch.cat([query_attention_mask, qformer_attention_mask], dim=1)
+        query_outputs = self.qformer(
+            input_ids=qformer_input_ids,
+            attention_mask=qformer_attention_mask,
+            query_embeds=query_tokens,
+            encoder_hidden_states=image_embeds,
+            encoder_attention_mask=image_attention_mask,
+            return_dict=True,
+        )
+        query_output = query_outputs.last_hidden_state[:, : query_tokens.size(1), :]
+
+        language_model_inputs = self.language_projection(query_output)
+
+        # unbatch the embeddings back by moving frames to seq-len
+        language_model_inputs = language_model_inputs.reshape(batch_size, self.config.num_query_tokens * frames, -1)
+        language_attention_mask = torch.ones(
+            language_model_inputs.size()[:-1], dtype=torch.long, device=language_model_inputs.device
+        )
+
+        if input_ids is None:
+            input_ids = (
+                torch.LongTensor([[self.config.text_config.bos_token_id]])
+                .repeat(batch_size, 1)
+                .to(image_embeds.device)
+            )
+        if attention_mask is None:
+            attention_mask = torch.ones_like(input_ids)
+        attention_mask = torch.cat([language_attention_mask, attention_mask.to(language_attention_mask.device)], dim=1)
+
+        # concatenate query embeddings with prompt embeddings
+        inputs_embeds = self.get_input_embeddings()(input_ids)
+        inputs_embeds = torch.cat([language_model_inputs, inputs_embeds.to(language_model_inputs.device)], dim=1)
+
+        # add image_embeds length to max_length, so that the final max_length in counted only on token embeds
+        # -1 is to account for the prepended BOS after `generate.`
+        if not self.language_model.config.is_encoder_decoder:
+            generate_kwargs["max_length"] = generate_kwargs.get("max_length", 20) + language_model_inputs.shape[1] - 1
+            generate_kwargs["min_length"] = generate_kwargs.get("min_length", 0) + language_model_inputs.shape[1]
+
+        outputs = self.language_model.generate(
+            inputs_embeds=inputs_embeds,
+            attention_mask=attention_mask,
+            **generate_kwargs,
+        )
+
+        # this is a temporary workaround to be consistent with other generation models and
+        # have BOS as the first token, even though under the hood we are calling LM with embeds
+        if not self.language_model.config.is_encoder_decoder:
+            # the InstructBLIP authors used inconsistent tokenizer/model files during training,
+            # with the tokenizer's bos token being set to </s> which has ID=2,
+            # whereas the model's text config has bos token id = 0
+            bos_token_id = (
+                2
+                if self.config.text_config.architectures[0] == "LLaMAForCausalLM"
+                else self.config.text_config.bos_token_id
+            )
+            bos_tokens = torch.LongTensor([[bos_token_id]]).repeat(batch_size, 1).to(image_embeds.device)
+            if not isinstance(outputs, torch.Tensor):
+                outputs.sequences = torch.cat([bos_tokens, outputs.sequences], dim=-1)
+            else:
+                outputs = torch.cat([bos_tokens, outputs], dim=-1)
+
+        return outputs

src/transformers/models/instructblipvideo/image_processing_instructblipvideo.py

+# coding=utf-8
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Image processor class for InstructBLIPVideo. Largely copy of Blip2Processor with addition of a video processing abilities
+"""
+
+from typing import Dict, List, Optional, Union
+
+import numpy as np
+
+from ...image_processing_utils import BaseImageProcessor, BatchFeature, get_size_dict
+from ...image_transforms import convert_to_rgb, resize, to_channel_dimension_format
+from ...image_utils import (
+    OPENAI_CLIP_MEAN,
+    OPENAI_CLIP_STD,
+    ChannelDimension,
+    ImageInput,
+    PILImageResampling,
+    VideoInput,
+    infer_channel_dimension_format,
+    is_scaled_image,
+    is_valid_image,
+    to_numpy_array,
+    valid_images,
+    validate_kwargs,
+    validate_preprocess_arguments,
+)
+from ...utils import TensorType, is_vision_available, logging
+
+
+if is_vision_available():
+    import PIL
+
+
+logger = logging.get_logger(__name__)
+
+
+def make_batched_videos(videos) -> List[VideoInput]:
+    if isinstance(videos, (list, tuple)) and isinstance(videos[0], (list, tuple)) and is_valid_image(videos[0][0]):
+        return videos
+
+    elif isinstance(videos, (list, tuple)) and is_valid_image(videos[0]):
+        if isinstance(videos[0], PIL.Image.Image):
+            return [videos]
+        elif len(videos[0].shape) == 4:
+            return [list(video) for video in videos]
+
+    elif is_valid_image(videos) and len(videos.shape) == 4:
+        return [list(videos)]
+
+    raise ValueError(f"Could not make batched video from {videos}")
+
+
+# Copied from transformers.models.blip.image_processing_blip.BlipImageProcessor with Blip->InstructBlipVideo, BLIP->InstructBLIPVideo
+class InstructBlipVideoImageProcessor(BaseImageProcessor):
+    r"""
+    Constructs a InstructBLIPVideo image processor.
+
+    Args:
+        do_resize (`bool`, *optional*, defaults to `True`):
+            Whether to resize the image's (height, width) dimensions to the specified `size`. Can be overridden by the
+            `do_resize` parameter in the `preprocess` method.
+        size (`dict`, *optional*, defaults to `{"height": 384, "width": 384}`):
+            Size of the output image after resizing. Can be overridden by the `size` parameter in the `preprocess`
+            method.
+        resample (`PILImageResampling`, *optional*, defaults to `Resampling.BICUBIC`):
+            Resampling filter to use if resizing the image. Only has an effect if `do_resize` is set to `True`. Can be
+            overridden by the `resample` parameter in the `preprocess` method.
+        do_rescale (`bool`, *optional*, defaults to `True`):
+            Whether to rescale the image by the specified scale `rescale_factor`. Can be overridden by the
+            `do_rescale` parameter in the `preprocess` method.
+        rescale_factor (`int` or `float`, *optional*, defaults to `1/255`):
+            Scale factor to use if rescaling the image. Only has an effect if `do_rescale` is set to `True`. Can be
+            overridden by the `rescale_factor` parameter in the `preprocess` method.
+        do_normalize (`bool`, *optional*, defaults to `True`):
+            Whether to normalize the image. Can be overridden by the `do_normalize` parameter in the `preprocess`
+            method. Can be overridden by the `do_normalize` parameter in the `preprocess` method.
+        image_mean (`float` or `List[float]`, *optional*, defaults to `IMAGENET_STANDARD_MEAN`):
+            Mean to use if normalizing the image. This is a float or list of floats the length of the number of
+            channels in the image. Can be overridden by the `image_mean` parameter in the `preprocess` method. Can be
+            overridden by the `image_mean` parameter in the `preprocess` method.
+        image_std (`float` or `List[float]`, *optional*, defaults to `IMAGENET_STANDARD_STD`):
+            Standard deviation to use if normalizing the image. This is a float or list of floats the length of the
+            number of channels in the image. Can be overridden by the `image_std` parameter in the `preprocess` method.
+            Can be overridden by the `image_std` parameter in the `preprocess` method.
+        do_convert_rgb (`bool`, *optional*, defaults to `True`):
+            Whether to convert the image to RGB.
+    """
+
+    model_input_names = ["pixel_values"]
+
+    def __init__(
+        self,
+        do_resize: bool = True,
+        size: Dict[str, int] = None,
+        resample: PILImageResampling = PILImageResampling.BICUBIC,
+        do_rescale: bool = True,
+        rescale_factor: Union[int, float] = 1 / 255,
+        do_normalize: bool = True,
+        image_mean: Optional[Union[float, List[float]]] = None,
+        image_std: Optional[Union[float, List[float]]] = None,
+        do_convert_rgb: bool = True,
+        **kwargs,
+    ) -> None:
+        super().__init__(**kwargs)
+        size = size if size is not None else {"height": 384, "width": 384}
+        size = get_size_dict(size, default_to_square=True)
+
+        self.do_resize = do_resize
+        self.size = size
+        self.resample = resample
+        self.do_rescale = do_rescale
+        self.rescale_factor = rescale_factor
+        self.do_normalize = do_normalize
+        self.image_mean = image_mean if image_mean is not None else OPENAI_CLIP_MEAN
+        self.image_std = image_std if image_std is not None else OPENAI_CLIP_STD
+        self.do_convert_rgb = do_convert_rgb
+        self._valid_processor_keys = [
+            "images",
+            "do_resize",
+            "size",
+            "resample",
+            "do_rescale",
+            "rescale_factor",
+            "do_normalize",
+            "image_mean",
+            "image_std",
+            "do_convert_rgb",
+            "return_tensors",
+            "data_format",
+            "input_data_format",
+        ]
+
+    # Copied from transformers.models.vit.image_processing_vit.ViTImageProcessor.resize with PILImageResampling.BILINEAR->PILImageResampling.BICUBIC
+    def resize(
+        self,
+        image: np.ndarray,
+        size: Dict[str, int],
+        resample: PILImageResampling = PILImageResampling.BICUBIC,
+        data_format: Optional[Union[str, ChannelDimension]] = None,
+        input_data_format: Optional[Union[str, ChannelDimension]] = None,
+        **kwargs,
+    ) -> np.ndarray:
+        """
+        Resize an image to `(size["height"], size["width"])`.
+
+        Args:
+            image (`np.ndarray`):
+                Image to resize.
+            size (`Dict[str, int]`):
+                Dictionary in the format `{"height": int, "width": int}` specifying the size of the output image.
+            resample (`PILImageResampling`, *optional*, defaults to `PILImageResampling.BICUBIC`):
+                `PILImageResampling` filter to use when resizing the image e.g. `PILImageResampling.BICUBIC`.
+            data_format (`ChannelDimension` or `str`, *optional*):
+                The channel dimension format for the output image. If unset, the channel dimension format of the input
+                image is used. Can be one of:
+                - `"channels_first"` or `ChannelDimension.FIRST`: image in (num_channels, height, width) format.
+                - `"channels_last"` or `ChannelDimension.LAST`: image in (height, width, num_channels) format.
+                - `"none"` or `ChannelDimension.NONE`: image in (height, width) format.
+            input_data_format (`ChannelDimension` or `str`, *optional*):
+                The channel dimension format for the input image. If unset, the channel dimension format is inferred
+                from the input image. Can be one of:
+                - `"channels_first"` or `ChannelDimension.FIRST`: image in (num_channels, height, width) format.
+                - `"channels_last"` or `ChannelDimension.LAST`: image in (height, width, num_channels) format.
+                - `"none"` or `ChannelDimension.NONE`: image in (height, width) format.
+
+        Returns:
+            `np.ndarray`: The resized image.
+        """
+        size = get_size_dict(size)
+        if "height" not in size or "width" not in size:
+            raise ValueError(f"The `size` dictionary must contain the keys `height` and `width`. Got {size.keys()}")
+
+        output_size = (size["height"], size["width"])
+        return resize(
+            image,
+            size=output_size,
+            resample=resample,
+            data_format=data_format,
+            input_data_format=input_data_format,
+            **kwargs,
+        )
+
+    # Ignore copy
+    def preprocess(
+        self,
+        images: VideoInput = None,
+        do_resize: Optional[bool] = None,
+        size: Optional[Dict[str, int]] = None,
+        resample: PILImageResampling = None,
+        do_rescale: Optional[bool] = None,
+        rescale_factor: Optional[float] = None,
+        do_normalize: Optional[bool] = None,
+        image_mean: Optional[Union[float, List[float]]] = None,
+        image_std: Optional[Union[float, List[float]]] = None,
+        return_tensors: Optional[Union[str, TensorType]] = None,
+        do_convert_rgb: bool = None,
+        data_format: ChannelDimension = ChannelDimension.FIRST,
+        input_data_format: Optional[Union[str, ChannelDimension]] = None,
+        **kwargs,
+    ) -> PIL.Image.Image:
+        """
+        Preprocess a video or batch of images/videos.
+
+        Args:
+            videos (`VideoInput`):
+                Video frames to preprocess. Expects a single or batch of videos as a list of frames with pixel values
+                ranging from 0 to 255. If passing in video with pixel values between 0 and 1, set `do_rescale=False`.
+            do_resize (`bool`, *optional*, defaults to `self.do_resize`):
+                Whether to resize the video.
+            size (`Dict[str, int]`, *optional*, defaults to `self.size`):
+                Controls the size of the video after `resize`. The shortest edge of the image is resized to
+                `size["shortest_edge"]` whilst preserving the aspect ratio. If the longest edge of this resized image
+                is > `int(size["shortest_edge"] * (1333 / 800))`, then the image is resized again to make the longest
+                edge equal to `int(size["shortest_edge"] * (1333 / 800))`.
+            resample (`PILImageResampling`, *optional*, defaults to `self.resample`):
+                Resampling filter to use if resizing the video. Only has an effect if `do_resize` is set to `True`.
+            do_rescale (`bool`, *optional*, defaults to `self.do_rescale`):
+                Whether to rescale the video values between [0 - 1].
+            rescale_factor (`float`, *optional*, defaults to `self.rescale_factor`):
+                Rescale factor to rescale the video by if `do_rescale` is set to `True`.
+            do_normalize (`bool`, *optional*, defaults to `self.do_normalize`):
+                Whether to normalize the video.
+            image_mean (`float` or `List[float]`, *optional*, defaults to `self.image_mean`):
+                Image mean to normalize the video by if `do_normalize` is set to `True`.
+            image_std (`float` or `List[float]`, *optional*, defaults to `self.image_std`):
+                Image standard deviation to normalize the video by if `do_normalize` is set to `True`.
+            do_convert_rgb (`bool`, *optional*, defaults to `self.do_convert_rgb`):
+                Whether to convert the image to RGB.
+            return_tensors (`str` or `TensorType`, *optional*):
+                The type of tensors to return. Can be one of:
+                    - Unset: Return a list of `np.ndarray`.
+                    - `TensorType.TENSORFLOW` or `'tf'`: Return a batch of type `tf.Tensor`.
+                    - `TensorType.PYTORCH` or `'pt'`: Return a batch of type `torch.Tensor`.
+                    - `TensorType.NUMPY` or `'np'`: Return a batch of type `np.ndarray`.
+                    - `TensorType.JAX` or `'jax'`: Return a batch of type `jax.numpy.ndarray`.
+            data_format (`ChannelDimension` or `str`, *optional*, defaults to `ChannelDimension.FIRST`):
+                The channel dimension format for the output image. Can be one of:
+                - `"channels_first"` or `ChannelDimension.FIRST`: image in (num_channels, height, width) format.
+                - `"channels_last"` or `ChannelDimension.LAST`: image in (height, width, num_channels) format.
+                - Unset: Use the channel dimension format of the input image.
+            input_data_format (`ChannelDimension` or `str`, *optional*):
+                The channel dimension format for the input image. If unset, the channel dimension format is inferred
+                from the input image. Can be one of:
+                - `"channels_first"` or `ChannelDimension.FIRST`: image in (num_channels, height, width) format.
+                - `"channels_last"` or `ChannelDimension.LAST`: image in (height, width, num_channels) format.
+                - `"none"` or `ChannelDimension.NONE`: image in (height, width) format.
+        """
+        do_resize = do_resize if do_resize is not None else self.do_resize
+        resample = resample if resample is not None else self.resample
+        do_rescale = do_rescale if do_rescale is not None else self.do_rescale
+        rescale_factor = rescale_factor if rescale_factor is not None else self.rescale_factor
+        do_normalize = do_normalize if do_normalize is not None else self.do_normalize
+        image_mean = image_mean if image_mean is not None else self.image_mean
+        image_std = image_std if image_std is not None else self.image_std
+        do_convert_rgb = do_convert_rgb if do_convert_rgb is not None else self.do_convert_rgb
+
+        size = size if size is not None else self.size
+        size = get_size_dict(size, default_to_square=False)
+
+        videos = make_batched_videos(images)
+        validate_kwargs(captured_kwargs=kwargs.keys(), valid_processor_keys=self._valid_processor_keys)
+
+        validate_preprocess_arguments(
+            do_rescale=do_rescale,
+            rescale_factor=rescale_factor,
+            do_normalize=do_normalize,
+            image_mean=image_mean,
+            image_std=image_std,
+            do_resize=do_resize,
+            size=size,
+            resample=resample,
+        )
+
+        if not valid_images(videos):
+            raise ValueError(
+                "Invalid input type. Must be of type PIL.Image.Image, numpy.ndarray, "
+                "torch.Tensor, tf.Tensor or jax.ndarray."
+            )
+
+        pixel_values = [
+            [
+                self._preprocess_image(
+                    image=frame,
+                    do_resize=do_resize,
+                    size=size,
+                    resample=resample,
+                    do_rescale=do_rescale,
+                    rescale_factor=rescale_factor,
+                    do_normalize=do_normalize,
+                    image_mean=image_mean,
+                    image_std=image_std,
+                    do_convert_rgb=do_convert_rgb,
+                    data_format=data_format,
+                    input_data_format=input_data_format,
+                )
+                for frame in video
+            ]
+            for video in videos
+        ]
+
+        encoded_outputs = BatchFeature(data={"pixel_values": pixel_values}, tensor_type=return_tensors)
+        return encoded_outputs
+
+    # Ignore copy
+    def _preprocess_image(
+        self,
+        image: ImageInput = None,
+        do_resize: Optional[bool] = None,
+        size: Optional[Dict[str, int]] = None,
+        resample: PILImageResampling = None,
+        do_rescale: Optional[bool] = None,
+        rescale_factor: Optional[float] = None,
+        do_normalize: Optional[bool] = None,
+        image_mean: Optional[Union[float, List[float]]] = None,
+        image_std: Optional[Union[float, List[float]]] = None,
+        do_convert_rgb: bool = None,
+        data_format: ChannelDimension = ChannelDimension.FIRST,
+        input_data_format: Optional[Union[str, ChannelDimension]] = None,
+    ) -> np.ndarray:
+        # PIL RGBA images are converted to RGB
+        if do_convert_rgb:
+            image = convert_to_rgb(image)
+
+        # All transformations expect numpy arrays.
+        image = to_numpy_array(image)
+
+        if is_scaled_image(image) and do_rescale:
+            logger.warning_once(
+                "It looks like you are trying to rescale already rescaled video frames. If the input"
+                " images have pixel values between 0 and 1, set `do_rescale=False` to avoid rescaling them again."
+            )
+
+        if input_data_format is None:
+            # We assume that all images have the same channel dimension format.
+            input_data_format = infer_channel_dimension_format(image)
+
+        if do_resize:
+            image = self.resize(image=image, size=size, resample=resample, input_data_format=input_data_format)
+
+        if do_rescale:
+            image = self.rescale(image=image, scale=rescale_factor, input_data_format=input_data_format)
+
+        if do_normalize:
+            image = self.normalize(image=image, mean=image_mean, std=image_std, input_data_format=input_data_format)
+
+        image = to_channel_dimension_format(image, data_format, input_channel_dim=input_data_format)
+
+        return image

src/transformers/models/instructblipvideo/modeling_instructblipvideo.py

+#           🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+#               This file was automatically generated from <path_to_diff_file.py>.
+#         Do NOT edit this file manually as any edits will be overwritten by the generation of
+#         the file from the diff. If any change should be done, please apply the change to the
+#                                    diff.py file directly.
+#           🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+# coding=utf-8
+# Copyright 2024 HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import math
+from dataclasses import dataclass
+from typing import Any, Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import CrossEntropyLoss
+
+from ...activations import ACT2FN
+from ...modeling_outputs import (
+    BaseModelOutput,
+    BaseModelOutputWithPastAndCrossAttentions,
+    BaseModelOutputWithPooling,
+    BaseModelOutputWithPoolingAndCrossAttentions,
+)
+from ...modeling_utils import PreTrainedModel
+from ...pytorch_utils import apply_chunking_to_forward, find_pruneable_heads_and_indices, prune_linear_layer
+from ...utils import (
+    ModelOutput,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    logging,
+    replace_return_docstrings,
+)
+from ..auto import AutoModelForCausalLM, AutoModelForSeq2SeqLM
+from .configuration_instructblipvideo import (
+    InstructBlipVideoConfig,
+    InstructBlipVideoQFormerConfig,
+    InstructBlipVideoVisionConfig,
+)
+
+
+logger = logging.get_logger(__name__)
+
+
+@dataclass
+# Copied from transformers.models.blip_2.modeling_blip_2.Blip2ForConditionalGenerationModelOutput with Blip2->InstructBlipVideo
+class InstructBlipVideoForConditionalGenerationModelOutput(ModelOutput):
+    """
+    Class defining the outputs of [`InstructBlipVideoForConditionalGeneration`].
+
+    Args:
+        loss (`torch.FloatTensor`, *optional*, returned when `labels` is provided, `torch.FloatTensor` of shape `(1,)`):
+            Language modeling loss from the language model.
+        logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
+            Prediction scores of the language modeling head of the language model.
+        vision_outputs (`BaseModelOutputWithPooling`):
+            Outputs of the vision encoder.
+        qformer_outputs (`BaseModelOutputWithPoolingAndCrossAttentions`):
+            Outputs of the Q-Former (Querying Transformer).
+        language_model_outputs (`CausalLMOutputWithPast` or `Seq2SeqLMOutput`):
+            Outputs of the language model.
+    """
+
+    loss: Optional[Tuple[torch.FloatTensor]] = None
+    logits: Optional[Tuple[torch.FloatTensor]] = None
+    vision_outputs: Optional[torch.FloatTensor] = None
+    qformer_outputs: Optional[Tuple[torch.FloatTensor]] = None
+    language_model_outputs: Optional[Tuple[torch.FloatTensor]] = None
+
+    def to_tuple(self) -> Tuple[Any]:
+        return tuple(
+            self[k]
+            if k not in ["vision_outputs", "qformer_outputs", "language_model_outputs"]
+            else getattr(self, k).to_tuple()
+            for k in self.keys()
+        )
+
+
+# Copied from transformers.models.blip.modeling_blip.BlipVisionEmbeddings with Blip->InstructBlipVideo
+class InstructBlipVideoVisionEmbeddings(nn.Module):
+    def __init__(self, config: InstructBlipVideoVisionConfig):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.image_size = config.image_size
+        self.patch_size = config.patch_size
+
+        self.class_embedding = nn.Parameter(torch.randn(1, 1, self.embed_dim))
+
+        self.patch_embedding = nn.Conv2d(
+            in_channels=3, out_channels=self.embed_dim, kernel_size=self.patch_size, stride=self.patch_size
+        )
+
+        self.num_patches = (self.image_size // self.patch_size) ** 2
+        self.num_positions = self.num_patches + 1
+
+        self.position_embedding = nn.Parameter(torch.randn(1, self.num_positions, self.embed_dim))
+
+    def interpolate_pos_encoding(self, embeddings: torch.Tensor, height: int, width: int) -> torch.Tensor:
+        """
+        This method allows to interpolate the pre-trained position encodings, to be able to use the model on higher
+        resolution images.
+
+        Source:
+        https://github.com/facebookresearch/dino/blob/de9ee3df6cf39fac952ab558447af1fa1365362a/vision_transformer.py#L174
+        """
+        num_patches = embeddings.shape[1] - 1
+        num_positions = self.position_embedding.shape[1] - 1
+
+        if num_patches == num_positions and height == width:
+            return self.position_embedding
+
+        class_pos_embed = self.position_embedding[:, 0, :]
+        patch_pos_embed = self.position_embedding[:, 1:, :]
+        dim = embeddings.shape[-1]
+        h0 = height // self.config.patch_size
+        w0 = width // self.config.patch_size
+        # we add a small number to avoid floating point error in the interpolation
+        # see discussion at https://github.com/facebookresearch/dino/issues/8
+        h0, w0 = h0 + 0.1, w0 + 0.1
+        patch_pos_embed = patch_pos_embed.reshape(1, int(math.sqrt(num_positions)), int(math.sqrt(num_positions)), dim)
+        patch_pos_embed = patch_pos_embed.permute(0, 3, 1, 2)
+        patch_pos_embed = nn.functional.interpolate(
+            patch_pos_embed,
+            scale_factor=(h0 / math.sqrt(num_positions), w0 / math.sqrt(num_positions)),
+            mode="bicubic",
+            align_corners=False,
+        )
+        patch_pos_embed = patch_pos_embed.permute(0, 2, 3, 1).view(1, -1, dim)
+        return torch.cat((class_pos_embed.unsqueeze(0), patch_pos_embed), dim=1)
+
+    def forward(self, pixel_values: torch.FloatTensor, interpolate_pos_encoding: bool = False) -> torch.Tensor:
+        batch_size, _, height, width = pixel_values.shape
+        target_dtype = self.patch_embedding.weight.dtype
+        patch_embeds = self.patch_embedding(pixel_values.to(dtype=target_dtype))  # shape = [*, width, grid, grid]
+        patch_embeds = patch_embeds.flatten(2).transpose(1, 2)
+        class_embeds = self.class_embedding.expand(batch_size, 1, -1).to(target_dtype)
+        embeddings = torch.cat([class_embeds, patch_embeds], dim=1)
+        if interpolate_pos_encoding:
+            position_embedding = self.interpolate_pos_encoding(embeddings, height, width)
+        else:
+            position_embedding = self.position_embedding
+        embeddings = embeddings + position_embedding[:, : embeddings.size(1), :].to(target_dtype)
+        return embeddings
+
+
+# Copied from transformers.models.blip_2.modeling_blip_2.Blip2Attention with Blip2->InstructBlipVideo
+class InstructBlipVideoAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.embed_dim // self.num_heads
+        if self.head_dim * self.num_heads != self.embed_dim:
+            raise ValueError(
+                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:"
+                f" {self.num_heads})."
+            )
+        self.scale = self.head_dim**-0.5
+        self.dropout = nn.Dropout(config.attention_dropout)
+
+        # small tweak here compared to CLIP, no bias here
+        self.qkv = nn.Linear(self.embed_dim, 3 * self.embed_dim, bias=False)
+
+        if config.qkv_bias:
+            q_bias = nn.Parameter(torch.zeros(self.embed_dim))
+            v_bias = nn.Parameter(torch.zeros(self.embed_dim))
+        else:
+            q_bias = None
+            v_bias = None
+
+        if q_bias is not None:
+            qkv_bias = torch.cat((q_bias, torch.zeros_like(v_bias, requires_grad=False), v_bias))
+            self.qkv.bias = nn.Parameter(qkv_bias)
+
+        self.projection = nn.Linear(self.embed_dim, self.embed_dim)
+
+    def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
+        return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        head_mask: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        """Input shape: Batch x Time x Channel"""
+
+        bsz, tgt_len, embed_dim = hidden_states.size()
+
+        mixed_qkv = self.qkv(hidden_states)
+
+        mixed_qkv = mixed_qkv.reshape(bsz, tgt_len, 3, self.num_heads, embed_dim // self.num_heads).permute(
+            2, 0, 3, 1, 4
+        )
+        query_states, key_states, value_states = mixed_qkv[0], mixed_qkv[1], mixed_qkv[2]
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        attention_scores = torch.matmul(query_states, key_states.transpose(-1, -2))
+
+        attention_scores = attention_scores * self.scale
+
+        # Normalize the attention scores to probabilities.
+        attention_probs = nn.functional.softmax(attention_scores, dim=-1)
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        attention_probs = self.dropout(attention_probs)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attention_probs = attention_probs * head_mask
+
+        context_layer = torch.matmul(attention_probs, value_states).permute(0, 2, 1, 3)
+
+        new_context_layer_shape = context_layer.size()[:-2] + (self.embed_dim,)
+        context_layer = context_layer.reshape(new_context_layer_shape)
+
+        output = self.projection(context_layer)
+
+        outputs = (output, attention_probs) if output_attentions else (output, None)
+
+        return outputs
+
+
+# Copied from transformers.models.blip.modeling_blip.BlipMLP
+class InstructBlipVideoMLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.activation_fn = ACT2FN[config.hidden_act]
+        self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
+        self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.fc1(hidden_states)
+        hidden_states = self.activation_fn(hidden_states)
+        hidden_states = self.fc2(hidden_states)
+        return hidden_states
+
+
+# Copied from transformers.models.blip.modeling_blip.BlipEncoderLayer with Blip->InstructBlipVideo
+class InstructBlipVideoEncoderLayer(nn.Module):
+    def __init__(self, config: InstructBlipVideoConfig):
+        super().__init__()
+        self.embed_dim = config.hidden_size
+        self.self_attn = InstructBlipVideoAttention(config)
+        self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
+        self.mlp = InstructBlipVideoMLP(config)
+        self.layer_norm2 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.FloatTensor]:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+            attention_mask (`torch.FloatTensor`): attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+                `(config.encoder_attention_heads,)`.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+        """
+        residual = hidden_states
+
+        hidden_states = self.layer_norm1(hidden_states)
+        hidden_states, attn_weights = self.self_attn(
+            hidden_states=hidden_states,
+            head_mask=attention_mask,
+            output_attentions=output_attentions,
+        )
+        hidden_states = hidden_states + residual
+        residual = hidden_states
+        hidden_states = self.layer_norm2(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+
+        hidden_states = hidden_states + residual
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (attn_weights,)
+
+        return outputs
+
+
+class InstructBlipVideoPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = InstructBlipVideoConfig
+    base_model_prefix = "blip"
+    supports_gradient_checkpointing = True
+    _no_split_modules = [
+        "InstructBlipVideoQFormerEmbeddings",
+        "InstructBlipVideoAttention",
+        "InstructBlipVideoQFormerMultiHeadAttention",
+        "InstructBlipVideoQFormerSelfOutput",
+    ]
+    _keep_in_fp32_modules = []
+
+    # Copied from transformers.models.blip_2.modeling_blip_2.Blip2PreTrainedModel._init_weights with Blip2->InstructBlipVideo
+    def _init_weights(self, module):
+        """Initialize the weights"""
+        factor = self.config.initializer_range
+        if isinstance(module, nn.Conv2d) or isinstance(module, nn.Embedding) or isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=factor)
+            if hasattr(module, "bias") and module.bias is not None:
+                module.bias.data.zero_()
+
+        if isinstance(module, InstructBlipVideoVisionEmbeddings):
+            if hasattr(self.config, "vision_config") and not isinstance(self.config, InstructBlipVideoVisionConfig):
+                factor = self.config.vision_config.initializer_range
+            nn.init.trunc_normal_(module.position_embedding, mean=0.0, std=factor)
+            nn.init.trunc_normal_(module.class_embedding, mean=0.0, std=factor)
+
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+        elif isinstance(module, nn.Linear) and module.bias is not None:
+            module.bias.data.zero_()
+
+
+INSTRUCTBLIPVIDEO_VISION_INPUTS_DOCSTRING = r"""
+    Args:
+        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            Pixel values. Pixel values can be obtained using [`InstructBlipVideoProcessor`]. See
+            [`InstructBlipVideoProcessor.__call__`] for details.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+        interpolate_pos_encoding (`bool`, *optional*, defaults to `False`):
+            Whether to interpolate the pre-trained position encodings.
+"""
+
+
+# Copied from transformers.models.blip.modeling_blip.BlipEncoder with Blip->InstructBlipVideo
+class InstructBlipVideoEncoder(nn.Module):
+    """
+    Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a
+    [`InstructBlipVideoEncoderLayer`].
+
+    Args:
+        config (`InstructBlipVideoConfig`):
+            The corresponding vision configuration for the `InstructBlipVideoEncoder`.
+    """
+
+    def __init__(self, config: InstructBlipVideoConfig):
+        super().__init__()
+        self.config = config
+        self.layers = nn.ModuleList([InstructBlipVideoEncoderLayer(config) for _ in range(config.num_hidden_layers)])
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        inputs_embeds,
+        attention_mask: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutput]:
+        r"""
+        Args:
+            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
+                Embedded representation of the inputs. Should be float, not int tokens.
+            attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+                - 1 for tokens that are **not masked**,
+                - 0 for tokens that are **masked**.
+
+                [What are attention masks?](../glossary#attention-mask)
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            output_hidden_states (`bool`, *optional*):
+                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
+                for more detail.
+            return_dict (`bool`, *optional*):
+                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            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
+
+        encoder_states = () if output_hidden_states else None
+        all_attentions = () if output_attentions else None
+
+        hidden_states = inputs_embeds
+        for idx, encoder_layer in enumerate(self.layers):
+            if output_hidden_states:
+                encoder_states = encoder_states + (hidden_states,)
+            if self.gradient_checkpointing and self.training:
+                layer_outputs = self._gradient_checkpointing_func(
+                    encoder_layer.__call__,
+                    hidden_states,
+                    attention_mask,
+                    output_attentions,
+                )
+            else:
+                layer_outputs = encoder_layer(
+                    hidden_states,
+                    attention_mask,
+                    output_attentions=output_attentions,
+                )
+
+            hidden_states = layer_outputs[0]
+
+            if output_attentions:
+                all_attentions = all_attentions + (layer_outputs[1],)
+
+        if output_hidden_states:
+            encoder_states = encoder_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None)
+        return BaseModelOutput(
+            last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions
+        )
+
+
+INSTRUCTBLIPVIDEO_START_DOCSTRING = r"""
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+
+    Parameters:
+        config ([`InstructBlipVideoConfig`]): Model configuration class with all the parameters of the model.
+            Initializing with a config file does not load the weights associated with the model, only the
+            configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+INSTRUCTBLIPVIDEO_INPUTS_DOCSTRING = r"""
+    Args:
+        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            Pixel values. Pixel values can be obtained using [`InstructBlipVideoProcessor`]. See
+            [`InstructBlipVideoProcessor.__call__`] for details.
+
+        qformer_input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of input sequence tokens in the vocabulary of the Q-Former. Input tokens can optionally be provided
+            to serve as text prompt, which the Q-Former model will encode.
+
+            Indices can be obtained using [`InstructBlipVideoProcessor`]. See [`InstructBlipVideoProcessor.__call__`] for
+            details.
+
+            [What are input IDs?](../glossary#input-ids)
+
+        qformer_attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of input sequence tokens in the vocabulary of the language model. Input tokens can optionally be
+            provided to serve as text prompt, which the language model can continue.
+
+            Indices can be obtained using [`InstructBlipVideoProcessor`]. See [`InstructBlipVideoProcessor.__call__`] for
+            details.
+
+            [What are input IDs?](../glossary#input-ids)
+
+        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+
+        decoder_input_ids (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*):
+            Indices of decoder input sequence tokens in the vocabulary of the language model. Only relevant in case an
+            encoder-decoder language model (like T5) is used.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details. [What are decoder input IDs?](../glossary#decoder-input-ids)
+
+        decoder_attention_mask (`torch.BoolTensor` of shape `(batch_size, target_sequence_length)`, *optional*):
+            Default behavior: generate a tensor that ignores pad tokens in `decoder_input_ids`. Causal mask will also
+            be used by default.
+
+            Only relevant in case an encoder-decoder language model (like T5) is used.
+
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+        interpolate_pos_encoding (`bool`, *optional*, defaults to `False`):
+            Whether to interpolate the pre-trained position encodings.
+"""
+
+
+# Copied from transformers.models.blip.modeling_blip.BlipVisionModel with Blip->InstructBlipVideo, BLIP->INSTRUCTBLIPVIDEO
+class InstructBlipVideoVisionModel(InstructBlipVideoPreTrainedModel):
+    main_input_name = "pixel_values"
+    config_class = InstructBlipVideoVisionConfig
+
+    def __init__(self, config: InstructBlipVideoVisionConfig):
+        super().__init__(config)
+        self.config = config
+        embed_dim = config.hidden_size
+
+        self.embeddings = InstructBlipVideoVisionEmbeddings(config)
+        self.encoder = InstructBlipVideoEncoder(config)
+        self.post_layernorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
+
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(INSTRUCTBLIPVIDEO_VISION_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=InstructBlipVideoVisionConfig)
+    def forward(
+        self,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        interpolate_pos_encoding: bool = False,
+    ) -> Union[Tuple, BaseModelOutputWithPooling]:
+        r"""
+        Returns:
+
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            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
+
+        if pixel_values is None:
+            raise ValueError("You have to specify pixel_values")
+
+        hidden_states = self.embeddings(pixel_values, interpolate_pos_encoding=interpolate_pos_encoding)
+
+        encoder_outputs = self.encoder(
+            inputs_embeds=hidden_states,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        last_hidden_state = encoder_outputs[0]
+        last_hidden_state = self.post_layernorm(last_hidden_state)
+
+        pooled_output = last_hidden_state[:, 0, :]
+        pooled_output = self.post_layernorm(pooled_output)
+
+        if not return_dict:
+            return (last_hidden_state, pooled_output) + encoder_outputs[1:]
+
+        return BaseModelOutputWithPooling(
+            last_hidden_state=last_hidden_state,
+            pooler_output=pooled_output,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )
+
+    def get_input_embeddings(self):
+        return self.embeddings
+
+
+class InstructBlipVideoQFormerMultiHeadAttention(nn.Module):
+    def __init__(self, config, is_cross_attention=False):
+        super().__init__()
+        self.config = config
+        if config.hidden_size % config.num_attention_heads != 0 and not hasattr(config, "embedding_size"):
+            raise ValueError(
+                "The hidden size (%d) is not a multiple of the number of attention heads (%d)"
+                % (config.hidden_size, config.num_attention_heads)
+            )
+
+        self.num_attention_heads = config.num_attention_heads
+        self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
+        self.all_head_size = self.num_attention_heads * self.attention_head_size
+
+        self.query = nn.Linear(config.hidden_size, self.all_head_size)
+        if is_cross_attention:
+            self.key = nn.Linear(config.encoder_hidden_size, self.all_head_size)
+            self.value = nn.Linear(config.encoder_hidden_size, self.all_head_size)
+        else:
+            self.key = nn.Linear(config.hidden_size, self.all_head_size)
+            self.value = nn.Linear(config.hidden_size, self.all_head_size)
+
+        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
+        self.position_embedding_type = getattr(config, "position_embedding_type", "absolute")
+        if self.position_embedding_type == "relative_key" or self.position_embedding_type == "relative_key_query":
+            self.max_position_embeddings = config.max_position_embeddings
+            self.distance_embedding = nn.Embedding(2 * config.max_position_embeddings - 1, self.attention_head_size)
+        self.save_attention = False
+
+    def save_attn_gradients(self, attn_gradients):
+        self.attn_gradients = attn_gradients
+
+    def get_attn_gradients(self):
+        return self.attn_gradients
+
+    def save_attention_map(self, attention_map):
+        self.attention_map = attention_map
+
+    def get_attention_map(self):
+        return self.attention_map
+
+    def transpose_for_scores(self, x):
+        new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
+        x = x.view(*new_x_shape)
+        return x.permute(0, 2, 1, 3)
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_value=None,
+        output_attentions=False,
+    ):
+        # If this is instantiated as a cross-attention module, the keys
+        # and values come from an encoder; the attention mask needs to be
+        # such that the encoder's padding tokens are not attended to.
+        is_cross_attention = encoder_hidden_states is not None
+
+        if is_cross_attention:
+            key_layer = self.transpose_for_scores(self.key(encoder_hidden_states))
+            value_layer = self.transpose_for_scores(self.value(encoder_hidden_states))
+            attention_mask = encoder_attention_mask
+        elif past_key_value is not None:
+            key_layer = self.transpose_for_scores(self.key(hidden_states))
+            value_layer = self.transpose_for_scores(self.value(hidden_states))
+            key_layer = torch.cat([past_key_value[0], key_layer], dim=2)
+            value_layer = torch.cat([past_key_value[1], value_layer], dim=2)
+        else:
+            key_layer = self.transpose_for_scores(self.key(hidden_states))
+            value_layer = self.transpose_for_scores(self.value(hidden_states))
+
+        mixed_query_layer = self.query(hidden_states)
+
+        query_layer = self.transpose_for_scores(mixed_query_layer)
+
+        past_key_value = (key_layer, value_layer)
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
+
+        if self.position_embedding_type == "relative_key" or self.position_embedding_type == "relative_key_query":
+            seq_length = hidden_states.size()[1]
+            position_ids_l = torch.arange(seq_length, dtype=torch.long, device=hidden_states.device).view(-1, 1)
+            position_ids_r = torch.arange(seq_length, dtype=torch.long, device=hidden_states.device).view(1, -1)
+            distance = position_ids_l - position_ids_r
+            positional_embedding = self.distance_embedding(distance + self.max_position_embeddings - 1)
+            positional_embedding = positional_embedding.to(dtype=query_layer.dtype)  # fp16 compatibility
+
+            if self.position_embedding_type == "relative_key":
+                relative_position_scores = torch.einsum("bhld,lrd->bhlr", query_layer, positional_embedding)
+                attention_scores = attention_scores + relative_position_scores
+            elif self.position_embedding_type == "relative_key_query":
+                relative_position_scores_query = torch.einsum("bhld,lrd->bhlr", query_layer, positional_embedding)
+                relative_position_scores_key = torch.einsum("bhrd,lrd->bhlr", key_layer, positional_embedding)
+                attention_scores = attention_scores + relative_position_scores_query + relative_position_scores_key
+
+        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
+        attention_scores_dtype = attention_scores.dtype
+
+        if attention_mask is not None:
+            # Apply the attention mask is (precomputed for all layers in BertModel forward() function)
+            attention_scores = attention_scores + attention_mask
+
+        # Normalize the attention scores to probabilities.
+        attention_probs = nn.Softmax(dim=-1)(attention_scores).to(attention_scores_dtype)
+
+        if is_cross_attention and self.save_attention:
+            self.save_attention_map(attention_probs)
+            attention_probs.register_hook(self.save_attn_gradients)
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        attention_probs_dropped = self.dropout(attention_probs)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attention_probs_dropped = attention_probs_dropped * head_mask
+
+        context_layer = torch.matmul(attention_probs_dropped, value_layer)
+
+        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
+        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
+        context_layer = context_layer.view(*new_context_layer_shape)
+
+        outputs = (context_layer, attention_probs) if output_attentions else (context_layer,)
+
+        outputs = outputs + (past_key_value,)
+        return outputs
+
+
+# Copied from transformers.models.bert.modeling_bert.BertSelfOutput with Bert->InstructBlipVideoQFormer
+class InstructBlipVideoQFormerSelfOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+# Copied from transformers.models.blip_2.modeling_blip_2.Blip2QFormerAttention with Blip2->InstructBlipVideo
+class InstructBlipVideoQFormerAttention(nn.Module):
+    def __init__(self, config, is_cross_attention=False):
+        super().__init__()
+        self.attention = InstructBlipVideoQFormerMultiHeadAttention(config, is_cross_attention)
+        self.output = InstructBlipVideoQFormerSelfOutput(config)
+        self.pruned_heads = set()
+
+    def prune_heads(self, heads):
+        if len(heads) == 0:
+            return
+        heads, index = find_pruneable_heads_and_indices(
+            heads, self.attention.num_attention_heads, self.attention.attention_head_size, self.pruned_heads
+        )
+
+        # Prune linear layers
+        self.attention.query = prune_linear_layer(self.attention.query, index)
+        self.attention.key = prune_linear_layer(self.attention.key, index)
+        self.attention.value = prune_linear_layer(self.attention.value, index)
+        self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)
+
+        # Update hyper params and store pruned heads
+        self.attention.num_attention_heads = self.attention.num_attention_heads - len(heads)
+        self.attention.all_head_size = self.attention.attention_head_size * self.attention.num_attention_heads
+        self.pruned_heads = self.pruned_heads.union(heads)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.Tensor]:
+        self_outputs = self.attention(
+            hidden_states,
+            attention_mask,
+            head_mask,
+            encoder_hidden_states,
+            encoder_attention_mask,
+            past_key_value,
+            output_attentions,
+        )
+        attention_output = self.output(self_outputs[0], hidden_states)
+        outputs = (attention_output,) + self_outputs[1:]  # add attentions if we output them
+        return outputs
+
+
+# Copied from transformers.models.bert.modeling_bert.BertIntermediate with Bert->InstructBlipVideoQFormer
+class InstructBlipVideoQFormerIntermediate(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.intermediate_size)
+        if isinstance(config.hidden_act, str):
+            self.intermediate_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.intermediate_act_fn = config.hidden_act
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.intermediate_act_fn(hidden_states)
+        return hidden_states
+
+
+# Copied from transformers.models.bert.modeling_bert.BertOutput with Bert->InstructBlipVideoQFormer
+class InstructBlipVideoQFormerOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+class InstructBlipVideoQFormerLayer(nn.Module):
+    def __init__(self, config, layer_idx):
+        super().__init__()
+        self.chunk_size_feed_forward = config.chunk_size_feed_forward
+        self.seq_len_dim = 1
+        self.attention = InstructBlipVideoQFormerAttention(config)
+
+        self.layer_idx = layer_idx
+
+        if layer_idx % config.cross_attention_frequency == 0:
+            self.crossattention = InstructBlipVideoQFormerAttention(config, is_cross_attention=True)
+            self.has_cross_attention = True
+        else:
+            self.has_cross_attention = False
+
+        self.intermediate = InstructBlipVideoQFormerIntermediate(config)
+        self.output = InstructBlipVideoQFormerOutput(config)
+
+        self.intermediate_query = InstructBlipVideoQFormerIntermediate(config)
+        self.output_query = InstructBlipVideoQFormerOutput(config)
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_value=None,
+        output_attentions=False,
+        query_length=0,
+    ):
+        # decoder uni-directional self-attention cached key/values tuple is at positions 1,2
+        self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None
+        self_attention_outputs = self.attention(
+            hidden_states,
+            attention_mask,
+            head_mask,
+            output_attentions=output_attentions,
+            past_key_value=self_attn_past_key_value,
+        )
+        attention_output = self_attention_outputs[0]
+        outputs = self_attention_outputs[1:-1]
+
+        present_key_value = self_attention_outputs[-1]
+
+        if query_length > 0:
+            query_attention_output = attention_output[:, :query_length, :]
+
+            if self.has_cross_attention:
+                if encoder_hidden_states is None:
+                    raise ValueError("encoder_hidden_states must be given for cross-attention layers")
+                cross_attention_outputs = self.crossattention(
+                    query_attention_output,
+                    attention_mask,
+                    head_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    output_attentions=output_attentions,
+                )
+                query_attention_output = cross_attention_outputs[0]
+                # add cross attentions if we output attention weights
+                outputs = outputs + cross_attention_outputs[1:-1]
+
+            layer_output = apply_chunking_to_forward(
+                self.feed_forward_chunk_query,
+                self.chunk_size_feed_forward,
+                self.seq_len_dim,
+                query_attention_output,
+            )
+
+            if attention_output.shape[1] > query_length:
+                layer_output_text = apply_chunking_to_forward(
+                    self.feed_forward_chunk,
+                    self.chunk_size_feed_forward,
+                    self.seq_len_dim,
+                    attention_output[:, query_length:, :],
+                )
+                layer_output = torch.cat([layer_output, layer_output_text], dim=1)
+        else:
+            layer_output = apply_chunking_to_forward(
+                self.feed_forward_chunk,
+                self.chunk_size_feed_forward,
+                self.seq_len_dim,
+                attention_output,
+            )
+        outputs = (layer_output,) + outputs
+
+        outputs = outputs + (present_key_value,)
+
+        return outputs
+
+    def feed_forward_chunk(self, attention_output):
+        intermediate_output = self.intermediate(attention_output)
+        layer_output = self.output(intermediate_output, attention_output)
+        return layer_output
+
+    def feed_forward_chunk_query(self, attention_output):
+        intermediate_output = self.intermediate_query(attention_output)
+        layer_output = self.output_query(intermediate_output, attention_output)
+        return layer_output
+
+
+# Copied from transformers.models.blip_2.modeling_blip_2.Blip2QFormerEncoder with Blip2->InstructBlipVideo
+class InstructBlipVideoQFormerEncoder(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.layer = nn.ModuleList(
+            [InstructBlipVideoQFormerLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+        )
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_values=None,
+        use_cache=None,
+        output_attentions=False,
+        output_hidden_states=False,
+        return_dict=True,
+        query_length=0,
+    ):
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attentions = () if output_attentions else None
+        all_cross_attentions = () if output_attentions else None
+
+        next_decoder_cache = () if use_cache else None
+
+        for i in range(self.config.num_hidden_layers):
+            layer_module = self.layer[i]
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            layer_head_mask = head_mask[i] if head_mask is not None else None
+            past_key_value = past_key_values[i] if past_key_values is not None else None
+
+            if getattr(self.config, "gradient_checkpointing", False) and self.training:
+                if use_cache:
+                    logger.warning(
+                        "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                    )
+                    use_cache = False
+                layer_outputs = self._gradient_checkpointing_func(
+                    layer_module.__call__,
+                    hidden_states,
+                    attention_mask,
+                    layer_head_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                )
+            else:
+                layer_outputs = layer_module(
+                    hidden_states,
+                    attention_mask,
+                    layer_head_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    past_key_value,
+                    output_attentions,
+                    query_length,
+                )
+
+            hidden_states = layer_outputs[0]
+            if use_cache:
+                next_decoder_cache += (layer_outputs[-1],)
+            if output_attentions:
+                all_self_attentions = all_self_attentions + (layer_outputs[1],)
+                if layer_module.has_cross_attention:
+                    all_cross_attentions = all_cross_attentions + (layer_outputs[2],)
+
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(
+                v
+                for v in [
+                    hidden_states,
+                    next_decoder_cache,
+                    all_hidden_states,
+                    all_self_attentions,
+                    all_cross_attentions,
+                ]
+                if v is not None
+            )
+        return BaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            past_key_values=next_decoder_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attentions,
+            cross_attentions=all_cross_attentions,
+        )
+
+
+class InstructBlipVideoQFormerEmbeddings(nn.Module):
+    """Construct the embeddings from word and position embeddings."""
+
+    def __init__(self, config):
+        super().__init__()
+        self.word_embeddings = nn.Embedding(config.vocab_size, config.hidden_size, padding_idx=config.pad_token_id)
+        self.position_embeddings = nn.Embedding(config.max_position_embeddings, config.hidden_size)
+
+        self.layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+        # position_ids (1, len position emb) is contiguous in memory and exported when serialized
+        self.register_buffer(
+            "position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)), persistent=False
+        )
+        self.position_embedding_type = getattr(config, "position_embedding_type", "absolute")
+
+        self.config = config
+
+    def forward(
+        self,
+        input_ids=None,
+        position_ids=None,
+        query_embeds=None,
+        past_key_values_length=0,
+    ):
+        if input_ids is not None:
+            seq_length = input_ids.size()[1]
+        else:
+            seq_length = 0
+
+        if position_ids is None:
+            position_ids = self.position_ids[:, past_key_values_length : seq_length + past_key_values_length].clone()
+
+        if input_ids is not None:
+            embeddings = self.word_embeddings(input_ids)
+            if self.position_embedding_type == "absolute":
+                position_embeddings = self.position_embeddings(position_ids.to(embeddings.device))
+                embeddings = embeddings + position_embeddings
+
+            if query_embeds is not None:
+                embeddings = torch.cat((query_embeds, embeddings), dim=1)
+        else:
+            embeddings = query_embeds
+
+        embeddings = embeddings.to(self.layernorm.weight.dtype)
+        embeddings = self.layernorm(embeddings)
+        embeddings = self.dropout(embeddings)
+        return embeddings
+
+
+class InstructBlipVideoQFormerModel(InstructBlipVideoPreTrainedModel):
+    """
+    Querying Transformer (Q-Former), used in Instructblipvideo. Slightly modified from BLIP-2 as it also takes the
+    instruction as input.
+    """
+
+    def __init__(self, config: InstructBlipVideoQFormerConfig):
+        super().__init__(config)
+        self.config = config
+
+        self.embeddings = InstructBlipVideoQFormerEmbeddings(config)
+
+        self.encoder = InstructBlipVideoQFormerEncoder(config)
+
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embeddings.word_embeddings
+
+    def set_input_embeddings(self, value):
+        self.embeddings.word_embeddings = value
+
+    def _prune_heads(self, heads_to_prune):
+        """
+        Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
+        class PreTrainedModel
+        """
+        for layer, heads in heads_to_prune.items():
+            self.encoder.layer[layer].attention.prune_heads(heads)
+
+    def get_extended_attention_mask(
+        self,
+        attention_mask: torch.Tensor,
+        input_shape: Tuple[int],
+        device: torch.device,
+        has_query: bool = False,
+    ) -> torch.Tensor:
+        """
+        Makes broadcastable attention and causal masks so that future and masked tokens are ignored.
+
+        Arguments:
+            attention_mask (`torch.Tensor`):
+                Mask with ones indicating tokens to attend to, zeros for tokens to ignore.
+            input_shape (`Tuple[int]`):
+                The shape of the input to the model.
+            device: (`torch.device`):
+                The device of the input to the model.
+
+        Returns:
+            `torch.Tensor` The extended attention mask, with a the same dtype as `attention_mask.dtype`.
+        """
+        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+        # ourselves in which case we just need to make it broadcastable to all heads.
+        if attention_mask.dim() == 3:
+            extended_attention_mask = attention_mask[:, None, :, :]
+        elif attention_mask.dim() == 2:
+            # Provided a padding mask of dimensions [batch_size, seq_length]
+            # - the model is an encoder, so make the mask broadcastable to [batch_size, num_heads, seq_length, seq_length]
+            extended_attention_mask = attention_mask[:, None, None, :]
+        else:
+            raise ValueError(
+                f"Wrong shape for input_ids (shape {input_shape}) or attention_mask (shape {attention_mask.shape})",
+            )
+
+        # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
+        # masked positions, this operation will create a tensor which is 0.0 for
+        # positions we want to attend and -10000.0 for masked positions.
+        # Since we are adding it to the raw scores before the softmax, this is
+        # effectively the same as removing these entirely.
+        extended_attention_mask = extended_attention_mask.to(dtype=self.dtype)  # fp16 compatibility
+        extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0
+        return extended_attention_mask
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        query_embeds: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple[torch.FloatTensor], BaseModelOutputWithPoolingAndCrossAttentions]:
+        r"""
+        encoder_hidden_states  (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if
+            the model is configured as a decoder.
+        encoder_attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in
+            the cross-attention if the model is configured as a decoder. Mask values selected in `[0, 1]`:
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+        past_key_values (`tuple(tuple(torch.FloatTensor))` of length `config.n_layers` with each tuple having 4 tensors of:
+            shape `(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`): Contains precomputed key and
+            value hidden states of the attention blocks. Can be used to speed up decoding. If `past_key_values` are
+            used, the user can optionally input only the last `decoder_input_ids` (those that don't have their past key
+            value states given to this model) of shape `(batch_size, 1)` instead of all `decoder_input_ids` of shape
+            `(batch_size, sequence_length)`.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+            `past_key_values`).
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            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
+
+        if input_ids is None and query_embeds is None:
+            raise ValueError("You have to specify query_embeds when input_ids is None")
+
+        # past_key_values_length
+        past_key_values_length = (
+            past_key_values[0][0].shape[2] - self.config.query_length if past_key_values is not None else 0
+        )
+
+        query_length = query_embeds.shape[1] if query_embeds is not None else 0
+
+        embedding_output = self.embeddings(
+            input_ids=input_ids,
+            position_ids=position_ids,
+            query_embeds=query_embeds,
+            past_key_values_length=past_key_values_length,
+        )
+
+        input_shape = embedding_output.size()[:-1]
+        batch_size, seq_length = input_shape
+        device = embedding_output.device
+
+        if attention_mask is None:
+            attention_mask = torch.ones(((batch_size, seq_length + past_key_values_length)), device=device)
+
+        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+        # ourselves in which case we just need to make it broadcastable to all heads.
+        extended_attention_mask = self.get_extended_attention_mask(attention_mask, input_shape, device)
+
+        # If a 2D or 3D attention mask is provided for the cross-attention
+        # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
+        if encoder_hidden_states is not None:
+            if isinstance(encoder_hidden_states, list):
+                encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states[0].size()
+            else:
+                encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size()
+            encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
+
+            if isinstance(encoder_attention_mask, list):
+                encoder_extended_attention_mask = [self.invert_attention_mask(mask) for mask in encoder_attention_mask]
+            elif encoder_attention_mask is None:
+                encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)
+                encoder_extended_attention_mask = self.invert_attention_mask(encoder_attention_mask)
+            else:
+                encoder_extended_attention_mask = self.invert_attention_mask(encoder_attention_mask)
+        else:
+            encoder_extended_attention_mask = None
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
+
+        encoder_outputs = self.encoder(
+            embedding_output,
+            attention_mask=extended_attention_mask,
+            head_mask=head_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_extended_attention_mask,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            query_length=query_length,
+        )
+        sequence_output = encoder_outputs[0]
+        pooled_output = sequence_output[:, 0, :]
+
+        if not return_dict:
+            return (sequence_output, pooled_output) + encoder_outputs[1:]
+
+        return BaseModelOutputWithPoolingAndCrossAttentions(
+            last_hidden_state=sequence_output,
+            pooler_output=pooled_output,
+            past_key_values=encoder_outputs.past_key_values,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+            cross_attentions=encoder_outputs.cross_attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    Instructblipvideo Model for generating text given an image and an optional text prompt. The model consists of a vision
+    encoder, Querying Transformer (Q-Former) and a language model.
+
+    One can optionally pass `input_ids` to the model, which serve as a text prompt, to make the language model continue
+    the prompt. Otherwise, the language model starts generating text from the [BOS] (beginning-of-sequence) token.
+    """,
+    INSTRUCTBLIPVIDEO_START_DOCSTRING,
+)
+class InstructBlipVideoForConditionalGeneration(InstructBlipVideoPreTrainedModel):
+    config_class = InstructBlipVideoConfig
+    main_input_name = "pixel_values"
+
+    def __init__(self, config: InstructBlipVideoConfig):
+        super().__init__(config)
+
+        self.vision_model = InstructBlipVideoVisionModel(config.vision_config)
+
+        self.query_tokens = nn.Parameter(torch.zeros(1, config.num_query_tokens, config.qformer_config.hidden_size))
+        self.qformer = InstructBlipVideoQFormerModel(config.qformer_config)
+
+        self.language_projection = nn.Linear(config.qformer_config.hidden_size, config.text_config.hidden_size)
+
+        if config.use_decoder_only_language_model:
+            language_model = AutoModelForCausalLM.from_config(
+                config.text_config, attn_implementation=config._attn_implementation
+            )
+        else:
+            language_model = AutoModelForSeq2SeqLM.from_config(
+                config.text_config, attn_implementation=config._attn_implementation
+            )
+
+        if language_model._no_split_modules is not None:
+            self._no_split_modules.extend(language_model._no_split_modules)
+
+        if language_model._keep_in_fp32_modules is not None:
+            self._keep_in_fp32_modules.extend(language_model._keep_in_fp32_modules)
+
+        self.language_model = language_model
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.language_model.get_input_embeddings()
+
+    def set_input_embeddings(self, value):
+        self.language_model.set_input_embeddings(value)
+
+    def set_output_embeddings(self, new_embeddings):
+        self.language_model.set_output_embeddings(new_embeddings)
+
+    def get_output_embeddings(self) -> nn.Module:
+        return self.language_model.get_output_embeddings()
+
+    def get_encoder(self):
+        return self.language_model.get_encoder()
+
+    def get_decoder(self):
+        return self.language_model.get_decoder()
+
+    def _tie_weights(self):
+        if not self.config.use_decoder_only_language_model:
+            self.language_model.encoder.embed_tokens = self.language_model.shared
+            self.language_model.decoder.embed_tokens = self.language_model.shared
+
+    def _preprocess_accelerate(self):
+        r"""
+        Some pre-processing hacks to make the model `accelerate` compatible. Check
+        https://github.com/huggingface/transformers/pull/21707 for more details.
+        """
+        hf_device_map = self.hf_device_map
+
+        if len(hf_device_map) > 1 and "language_model" not in hf_device_map and torch.cuda.device_count() > 1:
+            # warn users about unexpected behavior when using multi-GPU + Instructblipvideo + `accelerate`.
+            logger.warning(
+                "The `language_model` is not in the `hf_device_map` dictionary and you are running your script"
+                " in a multi-GPU environment. this may lead to unexpected behavior when using `accelerate`."
+                " Please pass a `device_map` that contains `language_model` to remove this warning."
+                " Please refer to https://github.com/huggingface/blog/blob/main/accelerate-large-models.md for"
+                " more details on creating a `device_map` for large models.",
+            )
+
+        if hasattr(self.language_model, "_hf_hook"):
+            self.language_model._hf_hook.io_same_device = True  # For `generate` compatibility
+
+    @add_start_docstrings_to_model_forward(INSTRUCTBLIPVIDEO_INPUTS_DOCSTRING)
+    @replace_return_docstrings(
+        output_type=InstructBlipVideoForConditionalGenerationModelOutput, config_class=InstructBlipVideoVisionConfig
+    )
+    def forward(
+        self,
+        pixel_values: torch.FloatTensor,
+        qformer_input_ids: torch.FloatTensor,
+        qformer_attention_mask: Optional[torch.LongTensor] = None,
+        input_ids: Optional[torch.FloatTensor] = None,
+        attention_mask: Optional[torch.LongTensor] = None,
+        decoder_input_ids: Optional[torch.LongTensor] = None,
+        decoder_attention_mask: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        labels: Optional[torch.LongTensor] = None,
+        return_dict: Optional[bool] = None,
+        interpolate_pos_encoding: bool = False,
+    ) -> Union[Tuple, InstructBlipVideoForConditionalGenerationModelOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the language modeling loss. Indices should be in `[-100, 0, ..., config.vocab_size -
+            1]`. All labels set to `-100` are ignored (masked), the loss is only computed for labels in `[0, ...,
+            config.vocab_size]`
+
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from transformers import InstructBlipVideoProcessor, InstructBlipVideoForConditionalGeneration
+        >>> import torch
+        >>> from huggingface_hub import hf_hub_download
+        >>> from av
+
+        >>> def read_video_pyav(container, indices):
+        ...     '''
+        ...     Decode the video with PyAV decoder.
+        ...     Args:
+        ...         container (`av.container.input.InputContainer`): PyAV container.
+        ...         indices (`List[int]`): List of frame indices to decode.
+        ...     Returns:
+        ...         result (np.ndarray): np array of decoded frames of shape (num_frames, height, width, 3).
+        ...     '''
+        ...     frames = []
+        ...     container.seek(0)
+        ...     start_index = indices[0]
+        ...     end_index = indices[-1]
+        ...     for i, frame in enumerate(container.decode(video=0)):
+        ...         if i > end_index:
+        ...             break
+        ...         if i >= start_index and i in indices:
+        ...             frames.append(frame)
+        ...     return np.stack([x.to_ndarray(format="rgb24") for x in frames])
+
+        >>> model = InstructBlipVideoProcessor.from_pretrained("Salesforce/instructblip-vicuna-7b", device_map="auto")
+        >>> processor = InstructBlipVideoForConditionalGeneration.from_pretrained("Salesforce/instructblip-vicuna-7b")
+
+        >>> file_path = hf_hub_download(
+                repo_id="nielsr/video-demo", filename="eating_spaghetti.mp4", repo_type="dataset"
+            )
+        >>> container = av.open(video_path)
+        >>> # sample uniformly 4 frames from the videWhy is this video funny?o
+        >>> total_frames = container.streams.video[0].frames
+        >>> indices = np.arange(0, total_frames, total_frames / 4).astype(int)
+        >>> clip = read_video_pyav(container, indices)
+
+        >>> prompt = "What is happening in the video?"
+        >>> inputs = processor(videos=clip, text=prompt, return_tensors="pt").to(device)
+
+        >>> outputs = model.generate(
+        ...     **inputs,
+        ...     do_sample=False,
+        ...     num_beams=5,
+        ...     max_length=256,
+        ...     repetition_penalty=1.5,
+        ...     length_penalty=1.0,
+        ... )
+        >>> generated_text = processor.batch_decode(outputs, skip_special_tokens=True)[0].strip()
+        >>> print(generated_text)
+        "A person is eating a bowl of pasta, and they are using a fork to eat it. The person is sitting at a table, and the plate of pasta is on the table in front"
+        ```"""
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # step 1: forward the images through the vision encoder,
+        # we process in a batched way, later unbatch it back (video has frames=4 always)
+        batch_size, frames, channel, height, width = pixel_values.shape
+        pixel_values = pixel_values.reshape(batch_size * frames, channel, height, width)
+
+        vision_outputs = self.vision_model(
+            pixel_values=pixel_values,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            interpolate_pos_encoding=interpolate_pos_encoding,
+        )
+        image_embeds = vision_outputs[0]
+
+        # step 2: forward the query tokens through the QFormer, using the image embeddings for cross-attention
+        image_attention_mask = torch.ones(image_embeds.size()[:-1], dtype=torch.long, device=image_embeds.device)
+
+        # difference with BLIP-2 here: we also feed the instruction prompt to the Q-Former
+        query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)
+        query_attention_mask = torch.ones(query_tokens.size()[:-1], dtype=torch.long, device=image_embeds.device)
+
+        if qformer_attention_mask is None:
+            qformer_attention_mask = torch.ones_like(qformer_input_ids)
+
+        qformer_input_ids = qformer_input_ids.repeat_interleave(frames, dim=0)
+        qformer_attention_mask = qformer_attention_mask.repeat_interleave(frames, dim=0)
+        qformer_attention_mask = torch.cat([query_attention_mask, qformer_attention_mask], dim=1)
+        query_outputs = self.qformer(
+            input_ids=qformer_input_ids,
+            attention_mask=qformer_attention_mask,
+            query_embeds=query_tokens,
+            encoder_hidden_states=image_embeds,
+            encoder_attention_mask=image_attention_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        query_output = query_outputs[0][:, : query_tokens.size(1), :]
+
+        # step 3: use the language model, conditioned on the query outputs and the prompt
+        language_model_inputs = self.language_projection(query_output)
+
+        # unbatch inputs back, each video-frame gets `num_query_tokens` seq length
+        language_model_inputs = language_model_inputs.reshape(batch_size, self.config.num_query_tokens * frames, -1)
+        language_model_attention_mask = torch.ones(
+            language_model_inputs.size()[:-1], dtype=torch.long, device=language_model_inputs.device
+        )
+
+        inputs_embeds = self.language_model.get_input_embeddings()(input_ids)
+        inputs_embeds = torch.cat([language_model_inputs, inputs_embeds.to(language_model_inputs.device)], dim=1)
+
+        if attention_mask is None:
+            attention_mask = torch.ones_like(input_ids)
+        attention_mask = torch.cat([language_model_attention_mask.to(attention_mask.device), attention_mask], dim=1)
+
+        if self.config.use_decoder_only_language_model:
+            outputs = self.language_model(
+                inputs_embeds=inputs_embeds,
+                attention_mask=attention_mask,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+            )
+            logits = outputs.logits if return_dict else outputs[0]
+            loss = None
+            # we compute the loss here since we need to take into account the sequence length of the query embeds
+            if labels is not None:
+                labels = labels.to(logits.device)
+                logits = logits[:, -labels.size(1) :, :]
+                # Shift so that tokens < n predict n
+                shift_logits = logits[..., :-1, :].contiguous()
+                shift_labels = labels[..., 1:].contiguous().to(logits.device)
+
+                # Flatten the tokens
+                loss_fct = CrossEntropyLoss(reduction="mean")
+
+                loss = loss_fct(shift_logits.view(-1, self.config.text_config.vocab_size), shift_labels.view(-1))
+        else:
+            outputs = self.language_model(
+                inputs_embeds=inputs_embeds,
+                attention_mask=attention_mask,
+                decoder_input_ids=decoder_input_ids,
+                decoder_attention_mask=decoder_attention_mask,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+                labels=labels,
+            )
+            loss = outputs.loss if return_dict else outputs[0]
+            logits = outputs.logits if return_dict else outputs[1]
+
+        if not return_dict:
+            output = (logits, vision_outputs, query_outputs, outputs)
+            return ((loss,) + output) if loss is not None else output
+
+        return InstructBlipVideoForConditionalGenerationModelOutput(
+            loss=loss,
+            logits=logits,
+            vision_outputs=vision_outputs,
+            qformer_outputs=query_outputs,
+            language_model_outputs=outputs,
+        )
+
+    @torch.no_grad()
+    def generate(
+        self,
+        pixel_values: torch.FloatTensor,
+        qformer_input_ids: Optional[torch.LongTensor] = None,
+        qformer_attention_mask: Optional[torch.LongTensor] = None,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.LongTensor] = None,
+        interpolate_pos_encoding: bool = False,
+        **generate_kwargs,
+    ) -> torch.LongTensor:
+        """
+        Overrides `generate` function to be able to use the model as a conditional generator.
+
+        Args:
+           pixel_values (`torch.FloatTensor` of shape (batch_size, num_channels, height, width) or
+                (batch_size, num_frames, num_channels, height, width)): Input images or videos to be processed.
+            qformer_input_ids (`torch.LongTensor` of shape (batch_size, sequence_length), *optional*):
+                The sequence used as a prompt to be fed to the Q-Former module.
+            qformer_attention_mask (`torch.LongTensor` of shape (batch_size, sequence_length), *optional*):
+                Mask to avoid performing attention on padding token indices.
+            input_ids (`torch.LongTensor` of shape (batch_size, sequence_length), *optional*):
+                The sequence used as a prompt for the generation.
+            attention_mask (`torch.LongTensor` of shape (batch_size, sequence_length), *optional*):
+                Mask to avoid performing attention on padding token indices.
+            interpolate_pos_encoding (`bool`, *optional*, defaults to `False`):
+                Whether to interpolate the positional encoding of the image embeddings.
+
+        Returns:
+            captions (list): A list of strings of length batch_size * num_captions.
+        """
+        if hasattr(self, "hf_device_map"):
+            # preprocess for `accelerate`
+            self._preprocess_accelerate()
+
+        # we process in a batched way, later unbatch it back (video has frames=4)
+        batch_size, frames, channel, height, width = pixel_values.shape
+        pixel_values = pixel_values.reshape(batch_size * frames, channel, height, width)
+
+        image_embeds = self.vision_model(
+            pixel_values,
+            return_dict=True,
+            interpolate_pos_encoding=interpolate_pos_encoding,
+        ).last_hidden_state
+        image_attention_mask = torch.ones(image_embeds.size()[:-1], dtype=torch.long, device=image_embeds.device)
+
+        query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)
+        query_attention_mask = torch.ones(query_tokens.size()[:-1], dtype=torch.long, device=image_embeds.device)
+        if qformer_attention_mask is None:
+            qformer_attention_mask = torch.ones_like(qformer_input_ids)
+
+        qformer_input_ids = qformer_input_ids.repeat_interleave(frames, dim=0)
+        qformer_attention_mask = qformer_attention_mask.repeat_interleave(frames, dim=0)
+        qformer_attention_mask = torch.cat([query_attention_mask, qformer_attention_mask], dim=1)
+        query_outputs = self.qformer(
+            input_ids=qformer_input_ids,
+            attention_mask=qformer_attention_mask,
+            query_embeds=query_tokens,
+            encoder_hidden_states=image_embeds,
+            encoder_attention_mask=image_attention_mask,
+            return_dict=True,
+        )
+        query_output = query_outputs.last_hidden_state[:, : query_tokens.size(1), :]
+
+        language_model_inputs = self.language_projection(query_output)
+
+        # unbatch the embeddings back by moving frames to seq-len
+        language_model_inputs = language_model_inputs.reshape(batch_size, self.config.num_query_tokens * frames, -1)
+        language_attention_mask = torch.ones(
+            language_model_inputs.size()[:-1], dtype=torch.long, device=language_model_inputs.device
+        )
+
+        if input_ids is None:
+            input_ids = (
+                torch.LongTensor([[self.config.text_config.bos_token_id]])
+                .repeat(batch_size, 1)
+                .to(image_embeds.device)
+            )
+        if attention_mask is None:
+            attention_mask = torch.ones_like(input_ids)
+        attention_mask = torch.cat([language_attention_mask, attention_mask.to(language_attention_mask.device)], dim=1)
+
+        # concatenate query embeddings with prompt embeddings
+        inputs_embeds = self.get_input_embeddings()(input_ids)
+        inputs_embeds = torch.cat([language_model_inputs, inputs_embeds.to(language_model_inputs.device)], dim=1)
+
+        # add image_embeds length to max_length, so that the final max_length in counted only on token embeds
+        # -1 is to account for the prepended BOS after `generate.`
+        if not self.language_model.config.is_encoder_decoder:
+            generate_kwargs["max_length"] = generate_kwargs.get("max_length", 20) + language_model_inputs.shape[1] - 1
+            generate_kwargs["min_length"] = generate_kwargs.get("min_length", 0) + language_model_inputs.shape[1]
+
+        outputs = self.language_model.generate(
+            inputs_embeds=inputs_embeds,
+            attention_mask=attention_mask,
+            **generate_kwargs,
+        )
+
+        # this is a temporary workaround to be consistent with other generation models and
+        # have BOS as the first token, even though under the hood we are calling LM with embeds
+        if not self.language_model.config.is_encoder_decoder:
+            # the InstructBLIP authors used inconsistent tokenizer/model files during training,
+            # with the tokenizer's bos token being set to </s> which has ID=2,
+            # whereas the model's text config has bos token id = 0
+            bos_token_id = (
+                2
+                if self.config.text_config.architectures[0] == "LLaMAForCausalLM"
+                else self.config.text_config.bos_token_id
+            )
+            bos_tokens = torch.LongTensor([[bos_token_id]]).repeat(batch_size, 1).to(image_embeds.device)
+            if not isinstance(outputs, torch.Tensor):
+                outputs.sequences = torch.cat([bos_tokens, outputs.sequences], dim=-1)
+            else:
+                outputs = torch.cat([bos_tokens, outputs], dim=-1)
+
+        return outputs