Add Idefics2 (#30253)

* Initial add model additions

* Test

* All weights loading

* Can perform full forward pass

* Local and remote the same

* Matching local and remote

* Fixup

* Idefics2Model importable; fixup docstrings

* Don't skip by default

* Remove deprecated use_resampler arg

* Remove self.config

* DecoupledLinear takes config

* Tidy up

* Enable eager attention and tidy up

* Most tests passing

* Update for batch of processed images

* Add image processor

* Update doc pages

* Update conversion script

* Remove erroneous breakpoint

* Remove accidendtal spelling change

* Update to reflect changes on hub - make generate work

* Fix up

* Image processor tests

* Update tests

* Add a processor

* Add a processor

* Update convert script

* Update modeling file - remove fixmes

* Bug fix

* Add processing test

* Use processor

* Fix up

* Update src/transformers/models/idefics2/modeling_idefics2.py
Co-authored-by: Victor SANH <victorsanh@gmail.com>

* Update src/transformers/models/idefics2/modeling_idefics2.py
Co-authored-by: Victor SANH <victorsanh@gmail.com>

* Fix test

* Update config - PR comments and defaults align with checkpoint

* Reviewer comments

* Add copied froms for flahs attention

* Update src/transformers/models/idefics2/modeling_idefics2.py
Co-authored-by: Victor SANH <victorsanh@gmail.com>

* Apply suggestions from code review
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* Remove qk_layer_norm and freeze_layers functionality

* Fix

* Remove freeze_layer options from config

* Sync with upstream main

* Fix attention shapes siglip

* Remove Llava-next refs - TO REBASE

* Use AutoModel for text model

* Add comment to explain vision embeddings

* Fix issue with tie_word_embeddings

* Address review comments

* Fix and fix up

* Chat templates for idefics

* Fix copies

* Fix

* Add layer norms to FA2

* Fix tests

* Apply suggestions from code review
Co-authored-by: Victor SANH <victorsanh@gmail.com>

* Fix

* Review comments

* Update src/transformers/models/idefics2/modeling_idefics2.py
Co-authored-by: Victor SANH <victorsanh@gmail.com>

* Update inputs merger

* Merge weights in correct order

* Update convert script

* Update src/transformers/models/idefics2/processing_idefics2.py
Co-authored-by: Victor SANH <victorsanh@gmail.com>

* Update template

* Model code examples (fix idefics too)

* More review comments

* Tidy up

* Update processing

* Fix attention mask preparation

* Update inputs_merger inputs

* Vectorize inputs_merger

* Update src/transformers/models/idefics2/__init__.py
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* Update src/transformers/models/idefics2/modeling_idefics2.py

* Review comments

* saying bye to the `qk_layer_norms`

* Simplify

* Update latents

* Remove erroneuous readme changes

* Return images when applying chat template

* Fix bug - prompt images are for a single sample

* Update src/transformers/models/idefics2/modeling_idefics2.py

* image splitting

* fix test

* some more comment

* some comment

* Apply suggestions from code review
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/idefics2/image_processing_idefics2.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update processor

* Update model tests

* Update src/transformers/models/idefics2/processing_idefics2.py
Co-authored-by: Victor SANH <victorsanh@gmail.com>

* Update src/transformers/models/idefics2/processing_idefics2.py
Co-authored-by: Victor SANH <victorsanh@gmail.com>

* Don't add BOS in template

* Update src/transformers/models/idefics2/processing_idefics2.py
Co-authored-by: Victor SANH <victorsanh@gmail.com>

* Remove index in examples

* Update tests to reflect #13

* Update src/transformers/models/idefics2/processing_idefics2.py
Co-authored-by: Victor SANH <victorsanh@gmail.com>

* PR comment - consistent typing

* Update readme and model doc

* Update docs

* Update checkpoint references

* Update examples

* Fix and update tests

* Small addition

* Update tests - remove copied from as no ignore placement copy could be found

* Update example

* small fixes

* Update docs/source/en/model_doc/idefics2.md
Co-authored-by: Victor SANH <victorsanh@gmail.com>

* Update docs/source/en/model_doc/idefics2.md
Co-authored-by: Victor SANH <victorsanh@gmail.com>

* Update README.md
Co-authored-by: Victor SANH <victorsanh@gmail.com>

* Connector model as bridge

* Fix up

* Fix up

* Don't pass model inputs for generation kwargs update

* IDEFICS-2 -> Idefics2

* Remove config archive name

* IDEFICS-2 -> Idefics2

* Add back llava-next

* Update readmes

* Add requirements for processor tester

* Use custom convert_to_rgb to avoid possible BC

* Fix doc example

* Fix doc example

* Skip model doc tests - as model to large

* More doc example - account for image splitting

* Update src/transformers/image_transforms.py

* Fix config doctest

---------
Co-authored-by: Pablo Montalvo <39954772+molbap@users.noreply.github.com>
Co-authored-by: ArthurZucker <arthur.zucker@gmail.com>
Co-authored-by: Victor SANH <victorsanh@gmail.com>
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

src/transformers/models/__init__.py

@@ -111,6 +111,7 @@ from . import (
     hubert,
     ibert,
     idefics,
+    idefics2,
     imagegpt,
     informer,
     instructblip,

src/transformers/models/auto/configuration_auto.py

@@ -125,6 +125,7 @@ CONFIG_MAPPING_NAMES = OrderedDict(
         ("hubert", "HubertConfig"),
         ("ibert", "IBertConfig"),
         ("idefics", "IdeficsConfig"),
+        ("idefics2", "Idefics2Config"),
         ("imagegpt", "ImageGPTConfig"),
         ("informer", "InformerConfig"),
         ("instructblip", "InstructBlipConfig"),
@@ -283,6 +284,7 @@ CONFIG_MAPPING_NAMES = OrderedDict(
     ]
 )
 
+
 MODEL_NAMES_MAPPING = OrderedDict(
     [
         # Add full (and cased) model names here
@@ -390,6 +392,7 @@ MODEL_NAMES_MAPPING = OrderedDict(
         ("hubert", "Hubert"),
         ("ibert", "I-BERT"),
         ("idefics", "IDEFICS"),
+        ("idefics2", "Idefics2"),
         ("imagegpt", "ImageGPT"),
         ("informer", "Informer"),
         ("instructblip", "InstructBLIP"),

src/transformers/models/auto/image_processing_auto.py

@@ -71,6 +71,7 @@ IMAGE_PROCESSOR_MAPPING_NAMES = OrderedDict(
         ("grounding-dino", "GroundingDinoImageProcessor"),
         ("groupvit", "CLIPImageProcessor"),
         ("idefics", "IdeficsImageProcessor"),
+        ("idefics2", "Idefics2ImageProcessor"),
         ("imagegpt", "ImageGPTImageProcessor"),
         ("instructblip", "BlipImageProcessor"),
         ("kosmos-2", "CLIPImageProcessor"),

src/transformers/models/auto/modeling_auto.py

@@ -120,6 +120,7 @@ MODEL_MAPPING_NAMES = OrderedDict(
         ("hubert", "HubertModel"),
         ("ibert", "IBertModel"),
         ("idefics", "IdeficsModel"),
+        ("idefics2", "Idefics2Model"),
         ("imagegpt", "ImageGPTModel"),
         ("informer", "InformerModel"),
         ("jukebox", "JukeboxModel"),
@@ -287,6 +288,7 @@ MODEL_FOR_PRETRAINING_MAPPING_NAMES = OrderedDict(
         ("gptsan-japanese", "GPTSanJapaneseForConditionalGeneration"),
         ("ibert", "IBertForMaskedLM"),
         ("idefics", "IdeficsForVisionText2Text"),
+        ("idefics2", "Idefics2ForConditionalGeneration"),
         ("layoutlm", "LayoutLMForMaskedLM"),
         ("llava", "LlavaForConditionalGeneration"),
         ("llava_next", "LlavaNextForConditionalGeneration"),
@@ -678,6 +680,7 @@ MODEL_FOR_VISION_2_SEQ_MAPPING_NAMES = OrderedDict(
         ("blip", "BlipForConditionalGeneration"),
         ("blip-2", "Blip2ForConditionalGeneration"),
         ("git", "GitForCausalLM"),
+        ("idefics2", "Idefics2ForConditionalGeneration"),
         ("instructblip", "InstructBlipForConditionalGeneration"),
         ("kosmos-2", "Kosmos2ForConditionalGeneration"),
         ("llava", "LlavaForConditionalGeneration"),

src/transformers/models/auto/processing_auto.py

@@ -61,6 +61,7 @@ PROCESSOR_MAPPING_NAMES = OrderedDict(
         ("groupvit", "CLIPProcessor"),
         ("hubert", "Wav2Vec2Processor"),
         ("idefics", "IdeficsProcessor"),
+        ("idefics2", "Idefics2Processor"),
         ("instructblip", "InstructBlipProcessor"),
         ("kosmos-2", "Kosmos2Processor"),
         ("layoutlmv2", "LayoutLMv2Processor"),

src/transformers/models/auto/tokenization_auto.py

@@ -201,6 +201,7 @@ else:
             ("hubert", ("Wav2Vec2CTCTokenizer", None)),
             ("ibert", ("RobertaTokenizer", "RobertaTokenizerFast" if is_tokenizers_available() else None)),
             ("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)),
             ("jukebox", ("JukeboxTokenizer", None)),
             (

src/transformers/models/idefics/modeling_idefics.py

@@ -1458,18 +1458,27 @@ class IdeficsForVisionText2Text(IdeficsPreTrainedModel):
         Example:
 
         ```python
-        >>> from transformers import AutoTokenizer, IdeficsForVisionText2Text
+        >>> from transformers import AutoProcessor, IdeficsForVisionText2Text
 
         >>> model = IdeficsForVisionText2Text.from_pretrained("HuggingFaceM4/idefics-9b")
-        >>> tokenizer = AutoTokenizer.from_pretrained("HuggingFaceM4/idefics-9b")
-
-        >>> prompt = "Hey, are you consciours? Can you talk to me?"
-        >>> inputs = tokenizer(prompt, return_tensors="pt")
-
-        >>> # Generate
-        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
-        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
-        "Hey, are you consciours? Can you talk to me?\nI'm not consciours, but I can talk to you."
+        >>> processor = AutoProcessor.from_pretrained("HuggingFaceM4/idefics-9b")
+
+        >>> dogs_image_url_1 = "https://huggingface.co/datasets/hf-internal-testing/fixtures_nlvr2/raw/main/image1.jpeg"
+        >>> dogs_image_url_2 = "https://huggingface.co/datasets/hf-internal-testing/fixtures_nlvr2/raw/main/image2.jpeg"
+
+        >>> prompts = [
+        ...     [
+        ...         "User:",
+        ...         dogs_image_url_1,
+        ...         "Describe this image.\nAssistant: An image of two dogs.\n",
+        ...         "User:",
+        ...         dogs_image_url_2,
+        ...         "Describe this image.\nAssistant:",
+        ...     ]
+        ... ]
+        >>> inputs = processor(prompts, return_tensors="pt")
+        >>> generate_ids = model.generate(**inputs, max_new_tokens=6)
+        >>> processor.batch_decode(generate_ids, skip_special_tokens=True)
         ```"""
 
         output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions

src/transformers/models/idefics2/__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_idefics2": ["Idefics2Config"]}
+
+
+try:
+    if not is_vision_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["image_processing_idefics2"] = ["Idefics2ImageProcessor"]
+
+
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_idefics2"] = [
+        "IDEFICS2_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "Idefics2ForConditionalGeneration",
+        "Idefics2PreTrainedModel",
+        "Idefics2Model",
+    ]
+    _import_structure["processing_idefics2"] = ["Idefics2Processor"]
+
+if TYPE_CHECKING:
+    from .configuration_idefics2 import Idefics2Config
+
+    try:
+        if not is_vision_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .image_processing_idefics2 import Idefics2ImageProcessor
+
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_idefics2 import (
+            IDEFICS2_PRETRAINED_MODEL_ARCHIVE_LIST,
+            Idefics2ForConditionalGeneration,
+            Idefics2Model,
+            Idefics2PreTrainedModel,
+        )
+        from .processing_idefics2 import Idefics2Processor
+
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure)

src/transformers/models/idefics2/configuration_idefics2.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.
+"""Idefics2 model configuration"""
+
+import os
+from typing import Union
+
+from ...configuration_utils import PretrainedConfig
+from ...utils import logging
+from ..auto import CONFIG_MAPPING
+
+
+logger = logging.get_logger(__name__)
+
+
+class Idefics2VisionConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`Idefics2VisionModel`]. It is used to instantiate a
+    Idefics2 vision encoder according to the specified arguments, defining the model architecture. Instantiating a
+    configuration with the defaults will yield a similar configuration to that of the SigLIP checkpoint
+    [google/siglip-base-patch16-224](https://huggingface.co/google/siglip-base-patch16-224) used in the Idefics2 model
+    [HuggingFaceM4/idefics2-8b](https://huggingface.co/HuggingFaceM4/idefics2-8b).
+
+    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 768):
+            Dimensionality of the encoder layers and the pooler layer.
+        intermediate_size (`int`, *optional*, defaults to 3072):
+            Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
+        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.
+        num_channels (`int`, *optional*, defaults to 3):
+            Number of channels in the input images.
+        image_size (`int`, *optional*, defaults to 224):
+            The size (resolution) of each image.
+        patch_size (`int`, *optional*, defaults to 32):
+            The size (resolution) of each patch.
+        hidden_act (`str` or `function`, *optional*, defaults to `"gelu_pytorch_tanh"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"selu"` and `"gelu_new"` ``"quick_gelu"` are supported.
+        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.
+        intializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation for initializing all weight matrices in the model.
+
+    Example:
+
+    ```python
+    >>> from transformers.models.idefics2.modeling_idefics2 import Idefics2VisionTransformer
+    >>> from transformers.models.idefics2.configuration_idefics2 import Idefics2VisionConfig
+
+    >>> # Initializing a Idefics2VisionConfig with google/siglip-base-patch16-224 style configuration
+    >>> configuration = Idefics2VisionConfig()
+
+    >>> # Initializing a Idefics2VisionTransformer (with random weights) from the google/siglip-base-patch16-224 style configuration
+    >>> model = Idefics2VisionTransformer(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "idefics2"
+
+    def __init__(
+        self,
+        hidden_size=768,
+        intermediate_size=3072,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        num_channels=3,
+        image_size=224,
+        patch_size=32,
+        hidden_act="gelu_pytorch_tanh",
+        layer_norm_eps=1e-6,
+        attention_dropout=0.0,
+        initializer_range=0.02,
+        **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.num_channels = num_channels
+        self.patch_size = patch_size
+        self.image_size = image_size
+        self.attention_dropout = attention_dropout
+        self.layer_norm_eps = layer_norm_eps
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+
+    @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 Idefics2Config
+        if config_dict.get("model_type") == "idefics2":
+            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 Idefics2PerceiverConfig(PretrainedConfig):
+    r"""
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
+            The non-linear activation function (function or string) in the perceiver block.
+        resampler_n_latents (`int`, *optional*, defaults to 64):
+            Number of latent embeddings to resample ("compress") the input sequence to (usually < 128).
+        resampler_depth (`int`, *optional*, defaults to 3):
+            Depth of the Perceiver Resampler (Transformer w/ cross attention). Should be shallow (<= 3).
+        resampler_n_heads (`int`, *optional*, defaults to 16):
+            Number of heads in each Transformer block (for multi-headed self-attention).
+        resampler_head_dim (`int`, *optional*, defaults to 96):
+            Dimensionality of each head projection in the Transformer block.
+        num_key_value_heads (`int`, *optional*, defaults to 4):
+            Number of key-value heads in the perceiver attention block.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+    """
+
+    model_type = "idefics2"
+
+    def __init__(
+        self,
+        hidden_act="silu",
+        resampler_n_latents=64,
+        resampler_depth=3,
+        resampler_n_heads=16,
+        resampler_head_dim=96,
+        num_key_value_heads=4,
+        attention_dropout=0.0,
+        **kwargs,
+    ):
+        self.hidden_act = hidden_act
+        self.resampler_n_latents = resampler_n_latents
+        self.resampler_depth = resampler_depth
+        self.resampler_n_heads = resampler_n_heads
+        self.num_key_value_heads = num_key_value_heads
+        self.resampler_head_dim = resampler_head_dim
+        self.attention_dropout = attention_dropout
+        if self.num_key_value_heads > self.resampler_n_heads:
+            raise ValueError(
+                f"num_key_value_heads={self.num_key_value_heads} must be less than or equal to"
+                f" resampler_n_heads={self.resampler_n_heads}"
+            )
+        super().__init__(**kwargs)
+
+
+class Idefics2Config(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`Idefics2Model`]. It is used to instantiate a
+    Idefics2 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 model of the Idefics2
+    [HuggingFaceM4/idefics2-8b](https://huggingface.co/HuggingFaceM4/idefics2-8b) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should cache the key/value pairs of the attention mechanism.
+        image_token_id (`int`, *optional*, defaults to 32001):
+            The id of the "image" token.
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether or not to tie the word embeddings with the token embeddings.
+        vision_config (`IdeficsVisionConfig` or `dict`, *optional*):
+            Custom vision config or dict
+        perceiver_config (`IdeficsPerceiverConfig` or `dict`, *optional*):
+            Custom perceiver config or dict
+        text_config (`MistralConfig` or `dict`, *optional*):
+            Custom text config or dict for the text model
+
+    Example:
+    ```python
+    >>> from transformers import Idefics2Model, Idefics2Config
+    >>> # Initializing configuration
+    >>> configuration = Idefics2Config()
+    >>> # Initializing a model from the configuration
+    >>> model = Idefics2Model(configuration)
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "idefics2"
+    is_composition = True
+
+    def __init__(
+        self,
+        use_cache=True,
+        image_token_id=32_001,
+        tie_word_embeddings=False,
+        vision_config=None,
+        perceiver_config=None,
+        text_config=None,
+        **kwargs,
+    ):
+        self.image_token_id = image_token_id
+        self.use_cache = use_cache
+        self.tie_word_embeddings = tie_word_embeddings
+
+        if perceiver_config is None:
+            self.perceiver_config = Idefics2PerceiverConfig()
+            logger.info("perciver_config is None, using default perceiver config")
+        elif isinstance(perceiver_config, dict):
+            self.perceiver_config = Idefics2PerceiverConfig(**perceiver_config)
+        elif isinstance(perceiver_config, Idefics2PerceiverConfig):
+            self.perceiver_config = perceiver_config
+
+        if vision_config is None:
+            self.vision_config = Idefics2VisionConfig()
+            logger.info("vision_config is None, using default vision config")
+        elif isinstance(vision_config, dict):
+            self.vision_config = Idefics2VisionConfig(**vision_config)
+        elif isinstance(vision_config, Idefics2VisionConfig):
+            self.vision_config = vision_config
+
+        if isinstance(text_config, dict):
+            text_config["model_type"] = text_config["model_type"] if "model_type" in text_config else "mistral"
+            text_config = CONFIG_MAPPING[text_config["model_type"]](**text_config)
+        elif text_config is None:
+            logger.info("text_config is None, using default text config")
+            text_config = CONFIG_MAPPING["mistral"](
+                max_position_embeddings=4096 * 8,
+                rms_norm_eps=1e-5,
+                # None in the original configuration_mistral, we set it to the unk_token_id
+                pad_token_id=0,
+                tie_word_embeddings=False,
+            )
+
+        self.text_config = text_config
+
+        super().__init__(**kwargs, tie_word_embeddings=tie_word_embeddings)

src/transformers/models/idefics2/convert_idefics2_weights_to_hf.py

+# 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.
+
+import argparse
+import copy
+
+import torch
+from accelerate import init_empty_weights
+
+from transformers import (
+    AutoConfig,
+    AutoModelForCausalLM,
+    AutoTokenizer,
+    Idefics2Config,
+    Idefics2ForConditionalGeneration,
+    Idefics2ImageProcessor,
+    Idefics2Processor,
+    MistralConfig,
+)
+
+
+EPILOG_TXT = """Example:
+    python transformers/src/transformers/models/idefics2/convert_idefics2_weights_to_hf.py --original_model_id HuggingFaceM4/idefics2-8b --output_hub_path org/idefics2
+"""
+
+
+KEYS_TO_MODIFY_MAPPING = {
+    "lm_head.weight": "lm_head.linear.weight",
+    "model.layers": "model.text_model.layers",
+    "model.norm": "model.text_model.norm",
+    "model.perceiver_resampler": "model.connector.perceiver_resampler",
+    "model.modality_projection": "model.connector.modality_projection",
+}
+
+
+WEIGHTS_TO_MERGE_MAPPING = (
+    # (weights to merge in merging order), (new weight name)
+    (
+        ("model.embed_tokens.weight", "model.embed_tokens.additional_embedding.weight"),
+        "model.text_model.embed_tokens.weight",
+    ),
+    (("lm_head.linear.weight", "additional_fc.weight"), "lm_head.weight"),
+)
+
+
+def convert_state_dict_to_hf(state_dict):
+    new_state_dict = {}
+    for key, value in state_dict.items():
+        if key.endswith(".inv_freq"):
+            continue
+        for key_to_modify, new_key in KEYS_TO_MODIFY_MAPPING.items():
+            if key_to_modify in key:
+                key = key.replace(key_to_modify, new_key)
+
+        new_state_dict[key] = value
+    return new_state_dict
+
+
+def merge_weights(state_dict):
+    new_state_dict = copy.deepcopy(state_dict)
+
+    # Merge the weights
+    for weights_to_merge, new_weight_name in WEIGHTS_TO_MERGE_MAPPING:
+        for weight in weights_to_merge:
+            assert weight in state_dict, f"Weight {weight} is missing in the state dict"
+            if new_weight_name not in new_state_dict:
+                new_state_dict[new_weight_name] = [state_dict[weight]]
+            else:
+                new_state_dict[new_weight_name].append(state_dict[weight])
+        new_state_dict[new_weight_name] = torch.cat(new_state_dict[new_weight_name], dim=0)
+
+    # Remove the weights that were merged
+    for weights_to_merge, new_weight_name in WEIGHTS_TO_MERGE_MAPPING:
+        for weight in weights_to_merge:
+            if weight in new_state_dict and weight != new_weight_name:
+                new_state_dict.pop(weight)
+
+    return new_state_dict
+
+
+def get_config(checkpoint):
+    if checkpoint == "HuggingFaceM4/idefics2":
+        # We load the config then recreate to use the text_config
+        config = AutoConfig.from_pretrained(checkpoint)
+        text_config = MistralConfig(
+            vocab_size=config.vocab_size + config.additional_vocab_size,
+            hidden_size=config.hidden_size,
+            intermediate_size=config.intermediate_size,
+            num_hidden_layers=config.num_hidden_layers,
+            num_attention_heads=config.num_attention_heads,
+            num_key_value_heads=config.num_key_value_heads,
+            hidden_act=config.hidden_act,
+            max_position_embeddings=config.max_position_embeddings,
+            initializer_range=config.initializer_range,
+            rms_norm_eps=config.rms_norm_eps,
+            tie_word_embeddings=config.tie_word_embeddings,
+            rope_theta=config.rope_theta,
+            sliding_window=config.sliding_window,
+            attention_dropout=config.attention_dropout,
+            pad_token_id=config.pad_token_id,
+            bos_token_id=config.bos_token_id,
+            eos_token_id=config.eos_token_id,
+        )
+        perceiver_config = config.perceiver_config.to_dict()
+        config = Idefics2Config(
+            text_config=text_config.to_dict(),
+            vision_config=config.vision_config,
+            perceiver_config=perceiver_config,
+            use_cache=config.use_cache,
+            image_token_id=config.image_token_id,
+            tie_word_embeddings=config.tie_word_embeddings,
+        )
+        return config
+
+    return AutoConfig.from_pretrained(checkpoint)
+
+
+def convert_idefics2_hub_to_hf(original_model_id, output_hub_path, push_to_hub):
+    # The original model maps to AutoModelForCausalLM, converted we map to Idefics2ForConditionalGeneration
+    original_model = AutoModelForCausalLM.from_pretrained(original_model_id, trust_remote_code=True)
+    # The original model doesn't use the idefics2 processing objects
+    image_seq_len = original_model.config.perceiver_config.resampler_n_latents
+    image_processor = Idefics2ImageProcessor()
+    tokenizer = AutoTokenizer.from_pretrained(original_model_id)
+    processor = Idefics2Processor(
+        image_processor=image_processor,
+        tokenizer=tokenizer,
+        image_seq_len=image_seq_len,
+    )
+    state_dict = original_model.state_dict()
+    state_dict = convert_state_dict_to_hf(state_dict)
+
+    # Merge weights
+    state_dict = merge_weights(state_dict)
+
+    config = get_config(original_model_id)
+
+    with init_empty_weights():
+        model = Idefics2ForConditionalGeneration(config)
+
+    model.load_state_dict(state_dict, strict=True, assign=True)
+
+    model.save_pretrained(output_hub_path)
+    processor.save_pretrained(output_hub_path)
+
+    if push_to_hub:
+        model.push_to_hub(output_hub_path, private=True)
+        processor.push_to_hub(output_hub_path, private=True)
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        epilog=EPILOG_TXT,
+        formatter_class=argparse.RawDescriptionHelpFormatter,
+    )
+    parser.add_argument(
+        "--original_model_id",
+        help="Hub location of the text model",
+    )
+    parser.add_argument(
+        "--output_hub_path",
+        help="Location on the hub of the converted model",
+    )
+    parser.add_argument(
+        "--push_to_hub",
+        action="store_true",
+        help="If set, the model will be pushed to the hub after conversion.",
+    )
+    args = parser.parse_args()
+    convert_idefics2_hub_to_hf(args.original_model_id, args.output_hub_path, args.push_to_hub)
+
+
+if __name__ == "__main__":
+    main()

src/transformers/models/idefics2/image_processing_idefics2.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.
+
+
+from typing import Any, Dict, Iterable, List, Optional, Tuple, Union
+
+import numpy as np
+
+from ...image_processing_utils import BaseImageProcessor, BatchFeature
+from ...image_transforms import PaddingMode, pad, resize, to_channel_dimension_format
+from ...image_utils import (
+    IMAGENET_STANDARD_MEAN,
+    IMAGENET_STANDARD_STD,
+    ChannelDimension,
+    ImageInput,
+    PILImageResampling,
+    get_image_size,
+    infer_channel_dimension_format,
+    is_scaled_image,
+    is_valid_image,
+    to_numpy_array,
+    valid_images,
+    validate_preprocess_arguments,
+)
+from ...utils import TensorType, is_vision_available, logging
+
+
+logger = logging.get_logger(__name__)
+
+
+if is_vision_available():
+    import PIL
+    from PIL import Image
+
+
+def get_resize_output_image_size(image, size, input_data_format) -> Tuple[int, int]:
+    """
+    Get the output size of the image after resizing given a dictionary specifying the max and min sizes.
+
+    Args:
+        image (`np.ndarray`):
+            Image to resize.
+        size (`Dict[str, int]`):
+            Size of the output image containing the keys "shortest_edge" and "longest_edge".
+        input_data_format (`ChannelDimension` or `str`):
+            The channel dimension format of the input image.
+
+    Returns:
+        The output size of the image after resizing.
+    """
+    height, width = get_image_size(image, channel_dim=input_data_format)
+
+    min_len = size["shortest_edge"]
+    max_len = size["longest_edge"]
+    aspect_ratio = width / height
+
+    if width >= height and width > max_len:
+        width = max_len
+        height = int(width / aspect_ratio)
+    elif height > width and height > max_len:
+        height = max_len
+        width = int(height * aspect_ratio)
+    height = max(height, min_len)
+    width = max(width, min_len)
+    return height, width
+
+
+def make_list_of_images(images: ImageInput) -> List[List[np.ndarray]]:
+    """
+    Convert a single image or a list of images to a list of numpy arrays.
+
+    Args:
+        images (`ImageInput`):
+            A single image or a list of images.
+
+    Returns:
+        A list of numpy arrays.
+    """
+    # If it's a single image, convert it to a list of lists
+    if is_valid_image(images):
+        images = [[images]]
+    # If it's a list of images, it's a single batch, so convert it to a list of lists
+    elif isinstance(images, (list, tuple)) and len(images) > 0 and is_valid_image(images[0]):
+        images = [images]
+    # If it's a list of batches, it's already in the right format
+    elif (
+        isinstance(images, (list, tuple))
+        and len(images) > 0
+        and isinstance(images[0], (list, tuple))
+        and is_valid_image(images[0][0])
+    ):
+        pass
+    else:
+        raise ValueError(
+            "Invalid input type. Must be a single image, a list of images, or a list of batches of images."
+        )
+    return images
+
+
+# Copied from transformers.models.detr.image_processing_detr.max_across_indices
+def max_across_indices(values: Iterable[Any]) -> List[Any]:
+    """
+    Return the maximum value across all indices of an iterable of values.
+    """
+    return [max(values_i) for values_i in zip(*values)]
+
+
+def get_max_height_width(
+    images_list: List[List[np.ndarray]], input_data_format: Optional[Union[str, ChannelDimension]] = None
+) -> List[int]:
+    """
+    Get the maximum height and width across all images in a batch.
+    """
+    if input_data_format is None:
+        input_data_format = infer_channel_dimension_format(images_list[0][0])
+
+    image_sizes = []
+    for images in images_list:
+        for image in images:
+            image_sizes.append(get_image_size(image, channel_dim=input_data_format))
+
+    max_height, max_width = max_across_indices(image_sizes)
+    return (max_height, max_width)
+
+
+# Copied from transformers.models.detr.image_processing_detr.make_pixel_mask
+def make_pixel_mask(
+    image: np.ndarray, output_size: Tuple[int, int], input_data_format: Optional[Union[str, ChannelDimension]] = None
+) -> np.ndarray:
+    """
+    Make a pixel mask for the image, where 1 indicates a valid pixel and 0 indicates padding.
+
+    Args:
+        image (`np.ndarray`):
+            Image to make the pixel mask for.
+        output_size (`Tuple[int, int]`):
+            Output size of the mask.
+    """
+    input_height, input_width = get_image_size(image, channel_dim=input_data_format)
+    mask = np.zeros(output_size, dtype=np.int64)
+    mask[:input_height, :input_width] = 1
+    return mask
+
+
+# FIXME Amy: merge this function with the one in image_transforms.py
+def convert_to_rgb(image: ImageInput) -> ImageInput:
+    """
+    Converts an image to RGB format. Only converts if the image is of type PIL.Image.Image, otherwise returns the image
+    as is.
+    Args:
+        image (Image):
+            The image to convert.
+    """
+    if not isinstance(image, PIL.Image.Image):
+        return image
+
+    # `image.convert("RGB")` would only work for .jpg images, as it creates a wrong background
+    # for transparent images. The call to `alpha_composite` handles this case
+    if image.mode == "RGB":
+        return image
+
+    image_rgba = image.convert("RGBA")
+    background = Image.new("RGBA", image_rgba.size, (255, 255, 255))
+    alpha_composite = Image.alpha_composite(background, image_rgba)
+    alpha_composite = alpha_composite.convert("RGB")
+    return alpha_composite
+
+
+class Idefics2ImageProcessor(BaseImageProcessor):
+    r"""
+    Constructs a Idefics image processor.
+
+    Args:
+        do_convert_rgb (`bool`, *optional*, defaults to `True`):
+            Whether to convert the image to RGB. This is useful if the input image is of a different format e.g. RGBA.
+            Only has an effect if the input image is in the PIL format.
+        do_resize (`bool`, *optional*, defaults to `True`):
+            Whether to resize the image. The longest edge of the image is resized to  be <= `size["longest_edge"]`, with the
+            shortest edge resized to keep the input aspect ratio, with a minimum size of `size["shortest_edge"]`.
+        size (`Dict`, *optional*):
+            Controls the size of the output image. This is a dictionary containing the keys "shortest_edge" and "longest_edge".
+        resample (`Resampling`, *optional*, defaults to `Resampling.BILINEAR`):
+            Resampling filter to use when resizing the image.
+        do_rescale (`bool`, *optional*, defaults to `True`):
+            Whether to rescale the image. If set to `True`, the image is rescaled to have pixel values between 0 and 1.
+        rescale_factor (`float`, *optional*, defaults to `1/255`):
+            Rescale factor to rescale the image by if `do_rescale` is set to `True`.
+        do_normalize (`bool`, *optional*, defaults to `True`):
+            Whether to normalize the image. If set to `True`, the image is normalized to have a mean of `image_mean` and
+            a standard deviation of `image_std`.
+        image_mean (`float` or `List[float]`, *optional*, defaults to `IDEFICS_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 `IDEFICS_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_pad (`bool`, *optional*, defaults to `True`):
+            Whether or not to pad the images to the largest height and width in the batch and number of images per
+            sample in the batch, such that the returned tensor is of shape (batch_size, max_num_images, num_channels, max_height, max_width).
+        do_image_splitting (`bool`, *optional*, defaults to `False`):
+            Whether to split the image into a sequence 4 equal sub-images concatenated with the original image. That
+            strategy was first introduced in https://arxiv.org/abs/2311.06607.
+    """
+
+    model_input_names = ["pixel_values"]
+
+    def __init__(
+        self,
+        do_convert_rgb: bool = True,
+        do_resize: bool = True,
+        size: Dict[str, int] = None,
+        resample: PILImageResampling = PILImageResampling.BILINEAR,
+        do_rescale: bool = True,
+        rescale_factor: float = 1 / 255,
+        do_normalize: bool = True,
+        image_mean: Optional[Union[float, List[float]]] = None,
+        image_std: Optional[Union[float, List[float]]] = None,
+        do_pad: bool = True,
+        do_image_splitting: bool = False,
+        **kwargs,
+    ) -> None:
+        super().__init__(**kwargs)
+        self.do_convert_rgb = do_convert_rgb
+        self.do_resize = do_resize
+        self.size = size if size is not None else {"shortest_edge": 378, "longest_edge": 980}
+        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 IMAGENET_STANDARD_MEAN
+        self.image_std = image_std if image_std is not None else IMAGENET_STANDARD_STD
+        self.do_pad = do_pad
+        self.do_image_splitting = do_image_splitting
+
+    def resize(
+        self,
+        image: np.ndarray,
+        size: Dict[str, int],
+        resample: PILImageResampling = PILImageResampling.BILINEAR,
+        data_format: Optional[Union[str, ChannelDimension]] = None,
+        input_data_format: Optional[Union[str, ChannelDimension]] = None,
+        **kwargs,
+    ) -> np.ndarray:
+        """
+        Resize an image. The shortest edge of the image is resized to size["shortest_edge"], with the longest edge
+        resized to keep the input aspect ratio.
+
+        Args:
+            image (`np.ndarray`):
+                Image to resize.
+            size (`Dict[str, int]`):
+                Size of the output image.
+            resample (`PILImageResampling`, *optional*, defaults to `PILImageResampling.BICUBIC`):
+                Resampling filter to use when resiizing the image.
+            data_format (`str` or `ChannelDimension`, *optional*):
+                The channel dimension format of the image. If not provided, it will be the same as the input image.
+            input_data_format (`ChannelDimension` or `str`, *optional*):
+                The channel dimension format of the input image. If not provided, it will be inferred.
+        """
+        if "shortest_edge" in size and "longest_edge" in size:
+            size = get_resize_output_image_size(image, size, input_data_format)
+        elif "height" in size and "width" in size:
+            size = (size["height"], size["width"])
+        else:
+            raise ValueError(
+                "size must be a dictionary with keys 'shortest_edge' and 'longest_edge' or 'height' and 'width'."
+            )
+        return resize(
+            image, size, resample=resample, data_format=data_format, input_data_format=input_data_format, **kwargs
+        )
+
+    # Copied from transformers.models.vilt.image_processing_vilt.ViltImageProcessor._pad_image
+    def _pad_image(
+        self,
+        image: np.ndarray,
+        output_size: Tuple[int, int],
+        constant_values: Union[float, Iterable[float]] = 0,
+        data_format: Optional[ChannelDimension] = None,
+        input_data_format: Optional[Union[str, ChannelDimension]] = None,
+    ) -> np.ndarray:
+        """
+        Pad an image with zeros to the given size.
+        """
+        input_height, input_width = get_image_size(image, channel_dim=input_data_format)
+        output_height, output_width = output_size
+
+        pad_bottom = output_height - input_height
+        pad_right = output_width - input_width
+        padding = ((0, pad_bottom), (0, pad_right))
+        padded_image = pad(
+            image,
+            padding,
+            mode=PaddingMode.CONSTANT,
+            constant_values=constant_values,
+            data_format=data_format,
+            input_data_format=input_data_format,
+        )
+        return padded_image
+
+    def pad(
+        self,
+        images: List[np.ndarray],
+        constant_values: Union[float, Iterable[float]] = 0,
+        return_pixel_mask: bool = True,
+        return_tensors: Optional[Union[str, TensorType]] = None,
+        data_format: Optional[ChannelDimension] = None,
+        input_data_format: Optional[Union[str, ChannelDimension]] = None,
+    ) -> BatchFeature:
+        """
+        For a list of images, for each images, pads a batch of images to the bottom and right of the image with zeros to the size of largest height and width.
+        For each sample in the batch, pads the sample with empty images to the max_number of images per sample in the batch. Optionally returns a pixel mask.
+
+        Args:
+            images (`np.ndarray`):
+                List of list of images to pad. Pads to the largest height and width in the batch.
+            constant_values (`float` or `Iterable[float]`, *optional*):
+                The value to use for the padding if `mode` is `"constant"`.
+            return_pixel_mask (`bool`, *optional*, defaults to `True`):
+                Whether to return a pixel mask.
+            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 (`str` or `ChannelDimension`, *optional*):
+                The channel dimension format of the image. If not provided, it will be the same as the input image.
+            input_data_format (`ChannelDimension` or `str`, *optional*):
+                The channel dimension format of the input image. If not provided, it will be inferred.
+        """
+        pad_size = get_max_height_width(images, input_data_format=input_data_format)
+
+        batch_size = len(images)
+        max_num_images = max(len(images_) for images_ in images)
+        input_data_format = (
+            infer_channel_dimension_format(images[0][0]) if input_data_format is None else input_data_format
+        )
+        data_format = input_data_format if data_format is None else data_format
+
+        def empty_image(size, input_data_format):
+            if input_data_format == ChannelDimension.FIRST:
+                return np.zeros((3, *size), dtype=np.uint8)
+            elif input_data_format == ChannelDimension.LAST:
+                return np.zeros((*size, 3), dtype=np.uint8)
+            raise ValueError("Invalid channel dimension format.")
+
+        padded_images_list = [
+            [empty_image(pad_size, data_format) for _ in range(max_num_images)] for _ in range(batch_size)
+        ]
+        padded_masks = [[np.zeros(pad_size) for _ in range(max_num_images)] for _ in range(batch_size)]
+
+        for batch_idx in range(batch_size):
+            for sample_idx, image in enumerate(images[batch_idx]):
+                padded_images_list[batch_idx][sample_idx] = self._pad_image(
+                    image,
+                    pad_size,
+                    constant_values=constant_values,
+                    data_format=data_format,
+                    input_data_format=input_data_format,
+                )
+                padded_masks[batch_idx][sample_idx] = make_pixel_mask(
+                    image, output_size=pad_size, input_data_format=input_data_format
+                )
+
+        padded_masks = padded_masks if return_pixel_mask else None
+        return padded_images_list, padded_masks
+
+    def _crop(
+        self,
+        im: np.ndarray,
+        w1: int,
+        h1: int,
+        w2: int,
+        h2: int,
+        input_data_format: Optional[Union[str, ChannelDimension]] = None,
+    ) -> np.ndarray:
+        if input_data_format == ChannelDimension.FIRST:
+            return im[:, h1:h2, w1:w2]
+        elif input_data_format == ChannelDimension.LAST:
+            return im[h1:h2, w1:w2, :]
+
+    def split_image(
+        self,
+        image: np.ndarray,
+        input_data_format: Optional[Union[str, ChannelDimension]] = None,
+    ):
+        """
+        Split an image into 4 equal sub-images, and the concatenate that sequence with the original image.
+        That means that a single image becomes a sequence of 5 images.
+        This is a "trick" to spend more compute on each image with no changes in the vision encoder.
+
+        Args:
+            image (`np.ndarray`):
+                Images to split.
+            input_data_format (`ChannelDimension` or `str`, *optional*):
+                The channel dimension format of the input image. If not provided, it will be inferred.
+        """
+        height, width = get_image_size(image, input_data_format)
+
+        mid_width = width // 2
+        mid_height = height // 2
+        return [
+            self._crop(image, 0, 0, mid_width, mid_height, input_data_format),
+            self._crop(image, mid_width, 0, width, mid_height, input_data_format),
+            self._crop(image, 0, mid_height, mid_width, height, input_data_format),
+            self._crop(image, mid_width, mid_height, width, height, input_data_format),
+            image,
+        ]
+
+    def preprocess(
+        self,
+        images: ImageInput,
+        do_convert_rgb: Optional[bool] = 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_pad: Optional[bool] = None,
+        do_image_splitting: Optional[bool] = None,
+        return_tensors: Optional[Union[str, TensorType]] = None,
+        input_data_format: Optional[ChannelDimension] = None,
+        data_format: Optional[ChannelDimension] = ChannelDimension.FIRST,
+    ):
+        """
+        Preprocess a batch of images.
+
+        Args:
+            images (`ImageInput`):
+                A list of images to preprocess.
+            do_convert_rgb (`bool`, *optional*, defaults to `self.do_convert_rgb`):
+                Whether to convert the image to RGB.
+            do_resize (`bool`, *optional*, defaults to `self.do_resize`):
+                Whether to resize the image.
+            size (`Dict[str, int]`, *optional*, defaults to `self.size`):
+                Size of the image after resizing. Shortest edge of the image is resized to size["shortest_edge"], with
+                the longest edge resized to keep the input aspect ratio.
+            resample (`int`, *optional*, defaults to `self.resample`):
+                Resampling filter to use if resizing the image. This can be one of the enum `PILImageResampling`. Only
+                has an effect if `do_resize` is set to `True`.
+            do_rescale (`bool`, *optional*, defaults to `self.do_rescale`):
+                Whether to rescale the image.
+            rescale_factor (`float`, *optional*, defaults to `self.rescale_factor`):
+                Rescale factor to rescale the image by if `do_rescale` is set to `True`.
+            do_normalize (`bool`, *optional*, defaults to `self.do_normalize`):
+                Whether to normalize the image.
+            image_mean (`float` or `List[float]`, *optional*, defaults to `self.image_mean`):
+                Image mean to use for normalization. Only has an effect if `do_normalize` is set to `True`.
+            image_std (`float` or `List[float]`, *optional*, defaults to `self.image_std`):
+                Image standard deviation to use for normalization. Only has an effect if `do_normalize` is set to
+                `True`.
+            do_pad (`bool`, *optional*, defaults to `self.do_pad`):
+                Whether or not to pad the images to the largest height and width in the batch.
+            do_image_splitting (`bool`, *optional*, defaults to `self.do_image_splitting`):
+                Whether to split the image into a sequence 4 equal sub-images concatenated with the original image. That
+                strategy was first introduced in https://arxiv.org/abs/2311.06607.
+            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
+        size = size if size is not None else self.size
+        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
+        do_pad = do_pad if do_pad is not None else self.do_pad
+        do_image_splitting = do_image_splitting if do_image_splitting is not None else self.do_image_splitting
+
+        images_list = make_list_of_images(images)
+
+        if not valid_images(images_list[0]):
+            raise ValueError(
+                "Invalid image type. Must be of type PIL.Image.Image, numpy.ndarray, "
+                "torch.Tensor, tf.Tensor or jax.ndarray."
+            )
+
+        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 do_convert_rgb:
+            images_list = [[convert_to_rgb(image) for image in images] for images in images_list]
+
+        # All transformations expect numpy arrays.
+        images_list = [[to_numpy_array(image) for image in images] for images in images_list]
+
+        if is_scaled_image(images_list[0][0]) and do_rescale:
+            logger.warning_once(
+                "It looks like you are trying to rescale already rescaled images. 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(images_list[0][0])
+
+        if do_image_splitting:
+            new_images_list = []
+            for images in images_list:
+                new_images = []
+                for image in images:
+                    new_images.extend(self.split_image(image, input_data_format))
+                new_images_list.append(new_images)
+            images_list = new_images_list
+
+        if do_resize:
+            images_list = [
+                [
+                    self.resize(image=image, size=size, resample=resample, input_data_format=input_data_format)
+                    for image in images
+                ]
+                for images in images_list
+            ]
+
+        if do_rescale:
+            images_list = [
+                [
+                    self.rescale(image=image, scale=rescale_factor, input_data_format=input_data_format)
+                    for image in images
+                ]
+                for images in images_list
+            ]
+
+        if do_normalize:
+            images_list = [
+                [
+                    self.normalize(image=image, mean=image_mean, std=image_std, input_data_format=input_data_format)
+                    for image in images
+                ]
+                for images in images_list
+            ]
+
+        pixel_attention_mask = None
+        if do_pad:
+            images_list, pixel_attention_mask = self.pad(
+                images_list, return_pixel_mask=True, return_tensors=return_tensors, input_data_format=input_data_format
+            )
+
+        if data_format is not None:
+            images_list = [
+                [
+                    to_channel_dimension_format(image, data_format, input_channel_dim=input_data_format)
+                    for image in images
+                ]
+                for images in images_list
+            ]
+
+        data = {"pixel_values": np.array(images_list) if do_pad else images_list}  # Faster tensor conversion
+        if pixel_attention_mask is not None:
+            data["pixel_attention_mask"] = np.array(pixel_attention_mask) if do_pad else pixel_attention_mask
+
+        return BatchFeature(data=data, tensor_type=return_tensors)

src/transformers/models/idefics2/modeling_idefics2.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.
+"""PyTorch Idefics2 model."""
+
+import inspect
+import math
+from dataclasses import dataclass
+from typing import Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.nn.functional as F
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import CrossEntropyLoss
+
+from ... import PreTrainedModel
+from ...activations import ACT2FN
+from ...cache_utils import Cache, DynamicCache
+from ...modeling_attn_mask_utils import _prepare_4d_attention_mask
+from ...modeling_outputs import BaseModelOutput, ModelOutput
+from ...utils import (
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    is_flash_attn_2_available,
+    is_flash_attn_greater_or_equal_2_10,
+    logging,
+    replace_return_docstrings,
+)
+from ..auto import AutoModel
+from .configuration_idefics2 import Idefics2Config, Idefics2VisionConfig
+
+
+if is_flash_attn_2_available():
+    from flash_attn import flash_attn_func, flash_attn_varlen_func
+    from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input  # noqa
+
+    _flash_supports_window_size = "window_size" in list(inspect.signature(flash_attn_func).parameters)
+
+
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = "Idefics2Config"
+
+IDEFICS2_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "HuggingFaceM4/idefics2-8b",
+    # See all IDEFICS2 models at https://huggingface.co/models?filter=idefics2
+]
+
+
+@dataclass
+class Idefics2BaseModelOutputWithPast(ModelOutput):
+    """
+    Base class for Idefics2 model's outputs that may also contain a past key/values (to speed up sequential decoding).
+    Args:
+        last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the model.
+            If `past_key_values` is used only the last hidden-state of the sequences of shape `(batch_size, 1,
+            hidden_size)` is output.
+        past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
+            Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
+            `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and optionally if
+            `config.is_encoder_decoder=True` 2 additional tensors of shape `(batch_size, num_heads,
+            encoder_sequence_length, embed_size_per_head)`.
+            Contains pre-computed hidden-states (key and values in the self-attention blocks and optionally if
+            `config.is_encoder_decoder=True` in the cross-attention blocks) that can be used (see `past_key_values`
+            input) to speed up sequential decoding.
+        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
+            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
+            Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
+        attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
+            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
+            sequence_length)`.
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+        image_hidden_states (`tuple(torch.FloatTensor)`, *optional*):
+            Tuple of `torch.FloatTensor` (one for the output of the image embeddings, `(batch_size, num_images,
+            sequence_length, hidden_size)`.
+            image_hidden_states of the model produced by the vision encoder, and optionally by the perceiver
+    """
+
+    last_hidden_state: torch.FloatTensor = None
+    past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+    image_hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+
+
+@dataclass
+# Copied from transformers.models.idefics.modeling_idefics.IdeficsCausalLMOutputWithPast with Idefics->Idefics2
+class Idefics2CausalLMOutputWithPast(ModelOutput):
+    """
+    Base class for Idefics2 causal language model (or autoregressive) outputs.
+    Args:
+        loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
+            Language modeling loss (for next-token prediction).
+        logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
+            Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
+            `(batch_size, num_heads, sequence_length, embed_size_per_head)`)
+            Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
+            `past_key_values` input) to speed up sequential decoding.
+        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
+            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
+            Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
+        attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
+            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
+            sequence_length)`.
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+        image_hidden_states (`tuple(torch.FloatTensor)`, *optional*):
+            Tuple of `torch.FloatTensor` (one for the output of the image embeddings, `(batch_size, num_images,
+            sequence_length, hidden_size)`.
+            image_hidden_states of the model produced by the vision encoder, and optionally by the perceiver
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    logits: torch.FloatTensor = None
+    past_key_values: Optional[List[torch.FloatTensor]] = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+    image_hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+
+
+class Idefics2VisionEmbeddings(nn.Module):
+    """
+    This is a modified version of `siglip.modelign_siglip.SiglipVisionEmbeddings` to enable images of variable
+    resolution.
+
+    The modifications are adapted from [Patch n' Pack: NaViT, a Vision Transformer for any Aspect Ratio and Resolution](https://arxiv.org/abs/2307.06304)
+    which allows treating images in their native aspect ratio and without the need to resize them to the same
+    fixed size. In particular, we start from the original pre-trained SigLIP model
+    (which uses images of fixed-size square images) and adapt it by training on images of variable resolutions.
+    """
+
+    def __init__(self, config: Idefics2VisionConfig):
+        super().__init__()
+        self.embed_dim = config.hidden_size
+        self.image_size = config.image_size
+        self.patch_size = config.patch_size
+
+        self.patch_embedding = nn.Conv2d(
+            in_channels=config.num_channels,
+            out_channels=self.embed_dim,
+            kernel_size=self.patch_size,
+            stride=self.patch_size,
+            padding="valid",
+        )
+
+        self.num_patches_per_side = self.image_size // self.patch_size
+        self.num_patches = self.num_patches_per_side**2
+        self.num_positions = self.num_patches
+        self.position_embedding = nn.Embedding(self.num_positions, self.embed_dim)
+
+    def forward(self, pixel_values: torch.FloatTensor, patch_attention_mask: torch.BoolTensor) -> torch.Tensor:
+        batch_size, _, max_im_h, max_im_w = pixel_values.shape
+
+        patch_embeds = self.patch_embedding(pixel_values)
+        embeddings = patch_embeds.flatten(2).transpose(1, 2)
+
+        max_nb_patches_h, max_nb_patches_w = max_im_h // self.patch_size, max_im_w // self.patch_size
+        boundaries = torch.arange(1 / self.num_patches_per_side, 1.0, 1 / self.num_patches_per_side)
+        position_ids = torch.full(size=(batch_size, max_nb_patches_h * max_nb_patches_w), fill_value=0)
+
+        for batch_idx, p_attn_mask in enumerate(patch_attention_mask):
+            nb_patches_h = p_attn_mask[:, 0].sum()
+            nb_patches_w = p_attn_mask[0].sum()
+
+            fractional_coords_h = torch.arange(0, 1 - 1e-6, 1 / nb_patches_h)
+            fractional_coords_w = torch.arange(0, 1 - 1e-6, 1 / nb_patches_w)
+
+            bucket_coords_h = torch.bucketize(fractional_coords_h, boundaries, right=True)
+            bucket_coords_w = torch.bucketize(fractional_coords_w, boundaries, right=True)
+
+            pos_ids = (bucket_coords_h[:, None] * self.num_patches_per_side + bucket_coords_w).flatten()
+            position_ids[batch_idx][p_attn_mask.view(-1).cpu()] = pos_ids
+
+        position_ids = position_ids.to(self.position_embedding.weight.device)
+        embeddings = embeddings + self.position_embedding(position_ids)
+        return embeddings
+
+
+# Copied from transformers.models.siglip.modeling_siglip.SiglipAttention with Siglip->Idefics2Vision
+class Idefics2VisionAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    # Copied from transformers.models.clip.modeling_clip.CLIPAttention.__init__
+    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 = config.attention_dropout
+
+        self.k_proj = nn.Linear(self.embed_dim, self.embed_dim)
+        self.v_proj = nn.Linear(self.embed_dim, self.embed_dim)
+        self.q_proj = nn.Linear(self.embed_dim, self.embed_dim)
+        self.out_proj = nn.Linear(self.embed_dim, self.embed_dim)
+
+        # Ignore copy
+        self.is_causal = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_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"""
+
+        batch_size, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        query_states = query_states.view(batch_size, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(batch_size, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(batch_size, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+
+        k_v_seq_len = key_states.shape[-2]
+        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) * self.scale
+
+        if attn_weights.size() != (batch_size, self.num_heads, q_len, k_v_seq_len):
+            raise ValueError(
+                f"Attention weights should be of size {(batch_size, self.num_heads, q_len, k_v_seq_len)}, but is"
+                f" {attn_weights.size()}"
+            )
+
+        if attention_mask is not None:
+            if attention_mask.size() != (batch_size, 1, q_len, k_v_seq_len):
+                raise ValueError(
+                    f"Attention mask should be of size {(batch_size, 1, q_len, k_v_seq_len)}, but is {attention_mask.size()}"
+                )
+            attn_weights = attn_weights + attention_mask
+
+        # upcast attention to fp32
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
+        attn_weights = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)
+        attn_output = torch.matmul(attn_weights, value_states)
+
+        if attn_output.size() != (batch_size, self.num_heads, q_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output` should be of size {(batch_size, self.num_heads, q_len, self.head_dim)}, but is"
+                f" {attn_output.size()}"
+            )
+
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.reshape(batch_size, q_len, self.embed_dim)
+
+        attn_output = self.out_proj(attn_output)
+
+        return attn_output, attn_weights
+
+
+class Idefics2VisionFlashAttention2(Idefics2VisionAttention):
+    """
+    Idefics2Vision flash attention module. This module inherits from `Idefics2VisionAttention` as the weights of the module stays
+    untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
+    flash attention and deal with padding tokens in case the input contains any of them.
+    """
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2.__init__
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
+        # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
+        # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
+        self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        **kwargs,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        output_attentions = False
+
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        # Flash attention requires the input to have the shape
+        # batch_size x seq_length x head_dim x hidden_dim
+        # therefore we just need to keep the original shape
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+
+        kv_seq_len = key_states.shape[-2]
+        if past_key_value is not None:
+            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
+
+        # TODO: These transpose are quite inefficient but Flash Attention requires the layout [batch_size, sequence_length, num_heads, head_dim]. We would need to refactor the KV cache
+        # to be able to avoid many of these transpose/reshape/view.
+        query_states = query_states.transpose(1, 2)
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+
+        dropout_rate = self.dropout if self.training else 0.0
+
+        # In PEFT, usually we cast the layer norms in float32 for training stability reasons
+        # therefore the input hidden states gets silently casted in float32. Hence, we need
+        # cast them back in the correct dtype just to be sure everything works as expected.
+        # This might slowdown training & inference so it is recommended to not cast the LayerNorms
+        # in fp32. (Idefics2VisionRMSNorm handles it correctly)
+
+        input_dtype = query_states.dtype
+        if input_dtype == torch.float32:
+            if torch.is_autocast_enabled():
+                target_dtype = torch.get_autocast_gpu_dtype()
+            # Handle the case where the model is quantized
+            elif hasattr(self.config, "_pre_quantization_dtype"):
+                target_dtype = self.config._pre_quantization_dtype
+            else:
+                target_dtype = self.q_proj.weight.dtype
+
+            logger.warning_once(
+                f"The input hidden states seems to be silently casted in float32, this might be related to"
+                f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
+                f" {target_dtype}."
+            )
+
+            query_states = query_states.to(target_dtype)
+            key_states = key_states.to(target_dtype)
+            value_states = value_states.to(target_dtype)
+
+        attn_output = self._flash_attention_forward(
+            query_states, key_states, value_states, attention_mask, q_len, dropout=dropout_rate
+        )
+
+        attn_output = attn_output.reshape(bsz, q_len, self.embed_dim).contiguous()
+        attn_output = self.out_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2._flash_attention_forward
+    def _flash_attention_forward(
+        self, query_states, key_states, value_states, attention_mask, query_length, dropout=0.0, softmax_scale=None
+    ):
+        """
+        Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token
+        first unpad the input, then computes the attention scores and pad the final attention scores.
+
+        Args:
+            query_states (`torch.Tensor`):
+                Input query states to be passed to Flash Attention API
+            key_states (`torch.Tensor`):
+                Input key states to be passed to Flash Attention API
+            value_states (`torch.Tensor`):
+                Input value states to be passed to Flash Attention API
+            attention_mask (`torch.Tensor`):
+                The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
+                position of padding tokens and 1 for the position of non-padding tokens.
+            dropout (`float`):
+                Attention dropout
+            softmax_scale (`float`, *optional*):
+                The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
+        """
+        if not self._flash_attn_uses_top_left_mask:
+            causal = self.is_causal
+        else:
+            # TODO: Remove the `query_length != 1` check once Flash Attention for RoCm is bumped to 2.1. For details, please see the comment in LlamaFlashAttention2 __init__.
+            causal = self.is_causal and query_length != 1
+
+        # Contains at least one padding token in the sequence
+        if attention_mask is not None:
+            batch_size = query_states.shape[0]
+            query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input(
+                query_states, key_states, value_states, attention_mask, query_length
+            )
+
+            cu_seqlens_q, cu_seqlens_k = cu_seq_lens
+            max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens
+
+            attn_output_unpad = flash_attn_varlen_func(
+                query_states,
+                key_states,
+                value_states,
+                cu_seqlens_q=cu_seqlens_q,
+                cu_seqlens_k=cu_seqlens_k,
+                max_seqlen_q=max_seqlen_in_batch_q,
+                max_seqlen_k=max_seqlen_in_batch_k,
+                dropout_p=dropout,
+                softmax_scale=softmax_scale,
+                causal=causal,
+            )
+
+            attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
+        else:
+            attn_output = flash_attn_func(
+                query_states, key_states, value_states, dropout, softmax_scale=softmax_scale, causal=causal
+            )
+
+        return attn_output
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2._upad_input
+    def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length):
+        indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)
+        batch_size, kv_seq_len, num_key_value_heads, head_dim = key_layer.shape
+
+        key_layer = index_first_axis(
+            key_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
+        )
+        value_layer = index_first_axis(
+            value_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
+        )
+        if query_length == kv_seq_len:
+            query_layer = index_first_axis(
+                query_layer.reshape(batch_size * kv_seq_len, self.num_heads, head_dim), indices_k
+            )
+            cu_seqlens_q = cu_seqlens_k
+            max_seqlen_in_batch_q = max_seqlen_in_batch_k
+            indices_q = indices_k
+        elif query_length == 1:
+            max_seqlen_in_batch_q = 1
+            cu_seqlens_q = torch.arange(
+                batch_size + 1, dtype=torch.int32, device=query_layer.device
+            )  # There is a memcpy here, that is very bad.
+            indices_q = cu_seqlens_q[:-1]
+            query_layer = query_layer.squeeze(1)
+        else:
+            # The -q_len: slice assumes left padding.
+            attention_mask = attention_mask[:, -query_length:]
+            query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask)
+
+        return (
+            query_layer,
+            key_layer,
+            value_layer,
+            indices_q,
+            (cu_seqlens_q, cu_seqlens_k),
+            (max_seqlen_in_batch_q, max_seqlen_in_batch_k),
+        )
+
+
+IDEFICS_VISION_ATTENTION_CLASSES = {
+    "eager": Idefics2VisionAttention,
+    "flash_attention_2": Idefics2VisionFlashAttention2,
+}
+
+
+# Copied from transformers.models.siglip.modeling_siglip.SiglipMLP with Siglip->Idefics2Vision
+class Idefics2VisionMLP(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
+
+
+class Idefics2MLP(nn.Module):
+    def __init__(
+        self,
+        hidden_size: int,
+        intermediate_size: int,
+        output_size: int,
+        hidden_act: str,
+    ):
+        super().__init__()
+        self.gate_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
+        self.up_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
+        self.down_proj = nn.Linear(intermediate_size, output_size, bias=False)
+        self.act_fn = ACT2FN[hidden_act]
+
+    def forward(self, x):
+        return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+
+
+# Copied from transformers.models.siglip.modeling_siglip.SiglipMultiheadAttentionPoolingHead with Siglip->Idefics2
+class Idefics2MultiheadAttentionPoolingHead(nn.Module):
+    """Multihead Attention Pooling."""
+
+    def __init__(self, config: Idefics2VisionConfig):
+        super().__init__()
+
+        self.probe = nn.Parameter(torch.randn(1, 1, config.hidden_size))
+        self.attention = torch.nn.MultiheadAttention(config.hidden_size, config.num_attention_heads, batch_first=True)
+        self.layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        # Ignore copy
+        self.mlp = Idefics2MLP(
+            hidden_size=config.hidden_size,
+            intermediate_size=config.intermediate_size,
+            hidden_act=config.hidden_act,
+            output_size=config.hidden_size,
+        )
+
+    def forward(self, hidden_state):
+        batch_size = hidden_state.shape[0]
+        probe = self.probe.repeat(batch_size, 1, 1)
+
+        hidden_state = self.attention(probe, hidden_state, hidden_state)[0]
+
+        residual = hidden_state
+        hidden_state = self.layernorm(hidden_state)
+        hidden_state = residual + self.mlp(hidden_state)
+
+        return hidden_state[:, 0]
+
+
+class Idefics2EncoderLayer(nn.Module):
+    def __init__(self, config: Idefics2Config):
+        super().__init__()
+        self.embed_dim = config.hidden_size
+        self.self_attn = IDEFICS_VISION_ATTENTION_CLASSES[config._attn_implementation](config)
+        self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
+        self.mlp = Idefics2VisionMLP(config)
+        self.layer_norm2 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
+
+    # Copied from transformers.models.siglip.modeling_siglip.SiglipEncoderLayer.forward
+    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 shape `(batch, 1, q_len, k_v_seq_len)` where padding elements are indicated by very large negative values.
+            output_attentions (`bool`, *optional*, defaults to `False`):
+                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,
+            attention_mask=attention_mask,
+            output_attentions=output_attentions,
+        )
+        hidden_states = residual + hidden_states
+
+        residual = hidden_states
+        hidden_states = self.layer_norm2(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (attn_weights,)
+
+        return outputs
+
+
+# Copied from transformers.models.siglip.modeling_siglip.SiglipEncoder with Siglip->Idefics2
+class Idefics2Encoder(nn.Module):
+    """
+    Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a
+    [`Idefics2EncoderLayer`].
+
+    Args:
+        config: Idefics2Config
+    """
+
+    def __init__(self, config: Idefics2Config):
+        super().__init__()
+        self.config = config
+        self.layers = nn.ModuleList([Idefics2EncoderLayer(config) for _ in range(config.num_hidden_layers)])
+        self.gradient_checkpointing = False
+
+    # Ignore copy
+    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)`):
+                Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
+                This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+                than the model's internal embedding lookup matrix.
+            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 encoder_layer in 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
+        )
+
+
+class Idefics2VisionTransformer(nn.Module):
+    def __init__(self, config: Idefics2VisionConfig):
+        super().__init__()
+        embed_dim = config.hidden_size
+
+        self.config = config
+        self.embeddings = Idefics2VisionEmbeddings(config)
+        self.encoder = Idefics2Encoder(config)
+        self.post_layernorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
+        self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"
+
+    def get_input_embeddings(self):
+        return self.embeddings
+
+    def set_input_embeddings(self, value):
+        self.embeddings = value
+
+    def forward(
+        self,
+        pixel_values,
+        patch_attention_mask: Optional[torch.BoolTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutput]:
+        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
+
+        batch_size = pixel_values.size(0)
+        if patch_attention_mask is None:
+            patch_size = self.config.patch_size
+            patch_attention_mask = torch.ones(
+                (
+                    batch_size,
+                    pixel_values.size(2) // patch_size,
+                    pixel_values.size(3) // patch_size,
+                )
+            )
+            patch_attention_mask = patch_attention_mask.to(dtype=torch.bool, device=pixel_values.device)
+
+        hidden_states = self.embeddings(pixel_values=pixel_values, patch_attention_mask=patch_attention_mask)
+
+        patch_attention_mask = patch_attention_mask.view(batch_size, -1)
+        # The call to `_upad_input` in `_flash_attention_forward` is expensive
+        # So when the `patch_attention_mask` is full of 1s (i.e. attending to the whole sequence),
+        # avoiding passing the attention_mask, which is equivalent to attending to the full sequence
+        if not torch.any(~patch_attention_mask):
+            patch_attention_mask = None
+        elif not self._use_flash_attention_2:
+            patch_attention_mask = _prepare_4d_attention_mask(patch_attention_mask, hidden_states.dtype)
+
+        encoder_outputs = self.encoder(
+            inputs_embeds=hidden_states,
+            attention_mask=patch_attention_mask,
+            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)
+
+        if not return_dict:
+            return (last_hidden_state,) + encoder_outputs[1:]
+
+        return BaseModelOutput(
+            last_hidden_state=last_hidden_state,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )
+
+
+# Copied from transformers.models.llama.modeling_llama.repeat_kv
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """
+    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
+    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
+    """
+    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
+    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+
+# Copied from transformers.models.llama.modeling_llama._get_unpad_data
+def _get_unpad_data(attention_mask):
+    seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
+    indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
+    max_seqlen_in_batch = seqlens_in_batch.max().item()
+    cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.int32), (1, 0))
+    return (
+        indices,
+        cu_seqlens,
+        max_seqlen_in_batch,
+    )
+
+
+# Copied from transformers.models.llama.modeling_llama.LlamaRMSNorm with Llama->Idefics2
+class Idefics2RMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        Idefics2RMSNorm is equivalent to T5LayerNorm
+        """
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+        return self.weight * hidden_states.to(input_dtype)
+
+
+class Idefics2PerceiverAttention(nn.Module):
+    def __init__(self, config, layer_idx: Optional[int] = None) -> None:
+        """Perceiver Cross-Attention Module --> let long-form inputs be `context`, resampled embeddings be `latents`"""
+        super().__init__()
+
+        self.layer_idx = None
+        self.hidden_size = config.text_config.hidden_size
+        self.num_heads = config.perceiver_config.resampler_n_heads
+        self.head_dim = config.perceiver_config.resampler_head_dim
+        self.num_key_value_heads = config.perceiver_config.num_key_value_heads
+        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
+        self.attention_dropout = config.perceiver_config.attention_dropout
+
+        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
+        self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
+        self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
+        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
+
+        self.is_causal = False
+
+    def forward(
+        self,
+        latents: torch.Tensor,
+        context: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        """
+        Runs Perceiver Self-Attention, with special (context, latents) appended along the `seq` dimension!
+
+        Args:
+            latents (`torch.Tensor`): Tensor of shape [bsz, n_latents, embed_dim] representing fixed length latents to compress to.
+            context (`torch.Tensor`): Tensor of shape [bsz, seq, embed_dim] representing long-form context to resample.
+            attention_mask (`torch.Tensor`, *optional*): Tensor of shape [bsz, 1, seq, n_latents] representing attention mask.
+            position_ids (`torch.LongTensor`, *optional*): Tensor of shape [bsz, seq] representing position indices of each input token.
+            past_key_value (`Tuple[torch.Tensor]`, *optional*): Tuple of tensors containing cached key and value states.
+            output_attentions (`bool`, *optional*, defaults to `False`): Whether to return attention weights.
+            use_cache (`bool`, *optional*, defaults to `False`): Whether to use past_key_value for caching.
+        """
+        bsz, q_len, _ = latents.size()
+        kv_seq_len = q_len + context.size()[1]
+
+        hidden_states = torch.concat([context, latents], dim=-2)
+
+        query_states = self.q_proj(latents)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, kv_seq_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, kv_seq_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+        past_key_value = getattr(self, "past_key_value", past_key_value)
+
+        if past_key_value is not None:
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx)
+
+        # repeat k/v heads if n_kv_heads < n_heads
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
+
+        if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
+            raise ValueError(
+                f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
+                f" {attn_weights.size()}"
+            )
+
+        if attention_mask is not None:
+            if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
+                raise ValueError(
+                    f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
+                )
+
+            attn_weights = attn_weights + attention_mask
+
+        # upcast attention to fp32
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
+        attn_output = torch.matmul(attn_weights, value_states)
+
+        if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
+                f" {attn_output.size()}"
+            )
+
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.reshape(bsz, q_len, self.num_heads * self.head_dim)
+
+        attn_output = self.o_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+
+# Copied from transformers.models.mistral.modeling_mistral.MistralFlashAttention2 with MistralAttention->Idefics2PerceiverAttention,MistralFlashAttention->Idefics2PerceiverFlashAttention,Mistral->Idefics2
+class Idefics2PerceiverFlashAttention2(Idefics2PerceiverAttention):
+    """
+    Idefics2 flash attention module. This module inherits from `Idefics2PerceiverAttention` as the weights of the module stays
+    untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
+    flash attention and deal with padding tokens in case the input contains any of them.
+    """
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2.__init__
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
+        # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
+        # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
+        self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
+
+    # Ignore copy
+    def forward(
+        self,
+        latents: torch.Tensor,
+        context: torch.Tensor,
+        attention_mask: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        **kwargs,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        bsz, q_len, _ = latents.size()
+        kv_seq_len = q_len + context.size()[1]
+
+        # Query, Key, Value Projections --> Note that in Flamingo, latents are *concatenated* with context prior to attn!
+        #   Note: This results in queries w/ `seq = n_latents`, and keys, values with `seq = len(context) + n_latents`
+        query_states = self.q_proj(latents)
+        key_states = self.k_proj(torch.cat([context, latents], dim=-2))
+        value_states = self.v_proj(torch.cat([context, latents], dim=-2))
+
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, kv_seq_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, kv_seq_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+        kv_seq_len = key_states.shape[-2]
+        if past_key_value is not None:
+            kv_seq_len += past_key_value[0].shape[-2]
+
+        if past_key_value is not None:
+            # Activate slicing cache only if the config has a value `sliding_windows` attribute
+            if hasattr(self.config, "sliding_window") and kv_seq_len > self.config.sliding_window:
+                slicing_tokens = kv_seq_len - self.config.sliding_window
+
+                past_key = past_key_value[0]
+                past_value = past_key_value[1]
+
+                past_key = past_key[:, :, slicing_tokens:, :].contiguous()
+                past_value = past_value[:, :, slicing_tokens:, :].contiguous()
+
+                if past_key.shape[-2] != self.config.sliding_window - 1:
+                    raise ValueError(
+                        "past key must have a shape of (`batch_size, num_heads, self.config.sliding_window-1,"
+                        f" head_dim`), got {past_key.shape}"
+                    )
+
+                past_key_value = (past_key, past_value)
+
+                if attention_mask is not None:
+                    attention_mask = attention_mask[:, slicing_tokens:]
+                    attention_mask = torch.cat([attention_mask, torch.ones_like(attention_mask[:, -1:])], dim=-1)
+
+            key_states = torch.cat([past_key_value[0], key_states], dim=2)
+            value_states = torch.cat([past_key_value[1], value_states], dim=2)
+
+        past_key_value = (key_states, value_states) if use_cache else None
+
+        # repeat k/v heads if n_kv_heads < n_heads
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+        dropout_rate = 0.0 if not self.training else self.attention_dropout
+
+        # In PEFT, usually we cast the layer norms in float32 for training stability reasons
+        # therefore the input hidden states gets silently casted in float32. Hence, we need
+        # cast them back in float16 just to be sure everything works as expected.
+        input_dtype = query_states.dtype
+        if input_dtype == torch.float32:
+            if torch.is_autocast_enabled():
+                target_dtype = torch.get_autocast_gpu_dtype()
+            # Handle the case where the model is quantized
+            elif hasattr(self.config, "_pre_quantization_dtype"):
+                target_dtype = self.config._pre_quantization_dtype
+            else:
+                target_dtype = self.q_proj.weight.dtype
+
+            logger.warning_once(
+                f"The input hidden states seems to be silently casted in float32, this might be related to"
+                f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
+                f" {target_dtype}."
+            )
+
+            query_states = query_states.to(target_dtype)
+            key_states = key_states.to(target_dtype)
+            value_states = value_states.to(target_dtype)
+
+        # Reashape to the expected shape for Flash Attention
+        query_states = query_states.transpose(1, 2)
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+
+        attn_output = self._flash_attention_forward(
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            q_len,
+            dropout=dropout_rate,
+            use_sliding_windows=False,
+        )
+
+        attn_output = attn_output.reshape(bsz, q_len, self.num_heads * self.head_dim).contiguous()
+        attn_output = self.o_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+    def _flash_attention_forward(
+        self,
+        query_states,
+        key_states,
+        value_states,
+        attention_mask,
+        query_length,
+        dropout=0.0,
+        softmax_scale=None,
+        use_sliding_windows=False,
+    ):
+        """
+        Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token
+        first unpad the input, then computes the attention scores and pad the final attention scores.
+
+        Args:
+            query_states (`torch.Tensor`):
+                Input query states to be passed to Flash Attention API
+            key_states (`torch.Tensor`):
+                Input key states to be passed to Flash Attention API
+            value_states (`torch.Tensor`):
+                Input value states to be passed to Flash Attention API
+            attention_mask (`torch.Tensor`):
+                The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
+                position of padding tokens and 1 for the position of non-padding tokens.
+            dropout (`float`):
+                Attention dropout
+            softmax_scale (`float`, *optional*):
+                The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
+            use_sliding_windows (`bool`, *optional*):
+                Whether to activate sliding window attention.
+        """
+        if not self._flash_attn_uses_top_left_mask:
+            causal = self.is_causal
+        else:
+            # TODO: Remove the `query_length != 1` check once Flash Attention for RoCm is bumped to 2.1. For details, please see the comment in LlamaFlashAttention2 __init__.
+            causal = self.is_causal and query_length != 1
+
+        # Contains at least one padding token in the sequence
+        if attention_mask is not None:
+            batch_size = query_states.shape[0]
+            query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input(
+                query_states, key_states, value_states, attention_mask, query_length
+            )
+
+            cu_seqlens_q, cu_seqlens_k = cu_seq_lens
+            max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens
+
+            if not use_sliding_windows:
+                attn_output_unpad = flash_attn_varlen_func(
+                    query_states,
+                    key_states,
+                    value_states,
+                    cu_seqlens_q=cu_seqlens_q,
+                    cu_seqlens_k=cu_seqlens_k,
+                    max_seqlen_q=max_seqlen_in_batch_q,
+                    max_seqlen_k=max_seqlen_in_batch_k,
+                    dropout_p=dropout,
+                    softmax_scale=softmax_scale,
+                    causal=causal,
+                )
+            else:
+                attn_output_unpad = flash_attn_varlen_func(
+                    query_states,
+                    key_states,
+                    value_states,
+                    cu_seqlens_q=cu_seqlens_q,
+                    cu_seqlens_k=cu_seqlens_k,
+                    max_seqlen_q=max_seqlen_in_batch_q,
+                    max_seqlen_k=max_seqlen_in_batch_k,
+                    dropout_p=dropout,
+                    softmax_scale=softmax_scale,
+                    causal=causal,
+                    window_size=(self.config.sliding_window, self.config.sliding_window),
+                )
+
+            attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
+        else:
+            if not use_sliding_windows:
+                attn_output = flash_attn_func(
+                    query_states,
+                    key_states,
+                    value_states,
+                    dropout,
+                    softmax_scale=softmax_scale,
+                    causal=causal,
+                )
+            else:
+                attn_output = flash_attn_func(
+                    query_states,
+                    key_states,
+                    value_states,
+                    dropout,
+                    softmax_scale=softmax_scale,
+                    causal=causal,
+                    window_size=(self.config.sliding_window, self.config.sliding_window),
+                )
+
+        return attn_output
+
+    def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length):
+        batch_size, kv_seq_len, num_heads, head_dim = key_layer.shape
+
+        # On the first iteration we need to properly re-create the padding mask
+        # by slicing it on the proper place
+        if kv_seq_len != attention_mask.shape[-1]:
+            attention_mask_num_tokens = attention_mask.shape[-1]
+            attention_mask = attention_mask[:, attention_mask_num_tokens - kv_seq_len :]
+
+        indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)
+
+        key_layer = index_first_axis(key_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k)
+        value_layer = index_first_axis(value_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k)
+
+        if query_length == kv_seq_len:
+            query_layer = index_first_axis(
+                query_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k
+            )
+            cu_seqlens_q = cu_seqlens_k
+            max_seqlen_in_batch_q = max_seqlen_in_batch_k
+            indices_q = indices_k
+        elif query_length == 1:
+            max_seqlen_in_batch_q = 1
+            cu_seqlens_q = torch.arange(
+                batch_size + 1, dtype=torch.int32, device=query_layer.device
+            )  # There is a memcpy here, that is very bad.
+            indices_q = cu_seqlens_q[:-1]
+            query_layer = query_layer.squeeze(1)
+        else:
+            # The -q_len: slice assumes left padding.
+            attention_mask = attention_mask[:, -query_length:]
+            query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask)
+
+        return (
+            query_layer,
+            key_layer,
+            value_layer,
+            indices_q,
+            (cu_seqlens_q, cu_seqlens_k),
+            (max_seqlen_in_batch_q, max_seqlen_in_batch_k),
+        )
+
+
+IDEFICS2_PERCEIVER_ATTENTION_CLASSES = {
+    "eager": Idefics2PerceiverAttention,
+    "flash_attention_2": Idefics2PerceiverFlashAttention2,
+}
+
+
+class Idefics2PerceiverLayer(nn.Module):
+    def __init__(self, config, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.text_config.hidden_size
+        self.n_latents = config.perceiver_config.resampler_n_latents
+        self.depth = config.perceiver_config.resampler_depth
+        self.rms_norm_eps = config.text_config.rms_norm_eps
+
+        self.input_latents_norm = Idefics2RMSNorm(self.hidden_size, eps=self.rms_norm_eps)
+        self.input_context_norm = Idefics2RMSNorm(self.hidden_size, eps=self.rms_norm_eps)
+        self.self_attn = IDEFICS2_PERCEIVER_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx=layer_idx)
+        self.post_attention_layernorm = Idefics2RMSNorm(self.hidden_size, eps=self.rms_norm_eps)
+        self.mlp = Idefics2MLP(
+            hidden_size=config.text_config.hidden_size,
+            intermediate_size=config.text_config.hidden_size * 4,
+            output_size=config.text_config.hidden_size,
+            hidden_act=config.perceiver_config.hidden_act,
+        )
+
+    def forward(
+        self,
+        latents: torch.Tensor,
+        context: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+        **kwargs,
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+        """
+        Args:
+            latents (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+            context (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+            attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
+                `(batch, sequence_length)` where padding elements are indicated by 0.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            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`).
+            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
+        """
+        residual = latents
+
+        latents = self.input_latents_norm(latents)
+        context = self.input_context_norm(context)
+
+        latents, self_attn_weights, present_key_value = self.self_attn(
+            latents=latents,
+            context=context,
+            attention_mask=attention_mask,
+        )
+        latents = residual + latents
+        residual = latents
+
+        latents = self.post_attention_layernorm(latents)
+        latents = self.mlp(latents)
+        latents = residual + latents
+
+        outputs = (latents,)
+
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        if use_cache:
+            outputs += (present_key_value,)
+
+        return outputs
+
+
+class Idefics2PerceiverResampler(nn.Module):
+    def __init__(self, config) -> None:
+        """
+        Instantiates a Perceiver Resampler that operates over a sequence of embeddings (say from a ResNet or ViT or
+        MAE) of a given dimension, performs `depth` blocks of cross-attention with a fixed `n_latents` inputs, then
+        returns a Tensor of shape [bsz, n_latents, embed_dim]. The Resampler acts as a form of learned pooling and
+        is derived from [Perceiver: General Perception with Iterative Attention](https://arxiv.org/abs/2103.03206).
+        """
+        super().__init__()
+        self.hidden_size = config.text_config.hidden_size
+        self.hidden_act = config.perceiver_config.hidden_act
+        self.n_latents = config.perceiver_config.resampler_n_latents
+        self.depth = config.perceiver_config.resampler_depth
+        self.rms_norm_eps = config.text_config.rms_norm_eps
+
+        # Create Latents for Perceiver
+        self.latents = nn.Parameter(torch.ones(self.n_latents, self.hidden_size))
+
+        # Create Transformer Blocks
+        self.layers = nn.ModuleList([Idefics2PerceiverLayer(config, idx) for idx in range(self.depth)])
+        self.norm = Idefics2RMSNorm(self.hidden_size, eps=self.rms_norm_eps)
+
+        self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"
+
+    def forward(
+        self,
+        context: torch.Tensor,
+        attention_mask,
+    ) -> torch.Tensor:
+        # seq embed -> bsz seq embed
+        latents = self.latents.unsqueeze(0).expand((context.shape[0], *self.latents.size()))
+
+        latent_attention_mask = torch.ones(
+            (attention_mask.size(0), latents.size(1)), dtype=attention_mask.dtype, device=attention_mask.device
+        )
+        attention_mask = torch.cat([attention_mask, latent_attention_mask], dim=-1)
+        attention_mask = (
+            _prepare_4d_attention_mask(attention_mask, latents.dtype, tgt_len=self.n_latents)
+            if not self._use_flash_attention_2
+            else attention_mask
+        )
+
+        compressed_context = latents
+        for perceiver_layer in self.layers:
+            layer_outputs = perceiver_layer(
+                compressed_context,
+                context,
+                attention_mask=attention_mask,
+                position_ids=None,
+                past_key_value=None,
+                output_attentions=False,
+                use_cache=False,
+            )
+
+            compressed_context = layer_outputs[0]
+
+        compressed_context = self.norm(compressed_context)
+
+        return compressed_context
+
+
+class Idefics2Connector(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.modality_projection = Idefics2MLP(
+            hidden_size=config.vision_config.hidden_size,
+            intermediate_size=config.text_config.intermediate_size,
+            output_size=config.text_config.hidden_size,
+            hidden_act=config.text_config.hidden_act,
+        )
+        self.perceiver_resampler = Idefics2PerceiverResampler(config)
+
+    def forward(self, image_hidden_states, attention_mask):
+        image_hidden_states = self.modality_projection(image_hidden_states)
+        image_hidden_states = self.perceiver_resampler(context=image_hidden_states, attention_mask=attention_mask)
+        return image_hidden_states
+
+
+IDEFICS2_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 ([`Idefics2Config`] or [`Idefics2VisionConfig`]):
+            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.
+"""
+
+
+@add_start_docstrings(
+    "The bare Idefics2 Model outputting raw hidden-states without any specific head on top.",
+    IDEFICS2_START_DOCSTRING,
+)
+class Idefics2PreTrainedModel(PreTrainedModel):
+    config_class = Idefics2Config
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["Idefics2VisionAttention", "Idefics2MLP", "Idefics2PerceiverLayer", "Idefics2DecoderLayer"]
+    _skip_keys_device_placement = "past_key_values"
+    _supports_flash_attn_2 = True
+
+    def _init_weights(self, module):
+        # important: this ported version of Idefics2 isn't meant for training from scratch - only
+        # inference and fine-tuning - so the proper init weights code has been removed - the original codebase
+        # https://github.com/haotian-liu/LLaVA/tree/main/idefics2 should serve for that purpose
+        std = (
+            self.config.text_config.initializer_range
+            if hasattr(self.config, "initializer_range")
+            else self.config.text_config.initializer_range
+        )
+
+        if hasattr(module, "class_embedding"):
+            module.class_embedding.data.normal_(mean=0.0, std=std)
+
+        if isinstance(module, (nn.Linear, nn.Conv2d)):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+    @classmethod
+    def _autoset_attn_implementation(
+        cls,
+        config,
+        use_flash_attention_2: bool = False,
+        torch_dtype: Optional[torch.dtype] = None,
+        device_map: Optional[Union[str, Dict[str, int]]] = None,
+        check_device_map: bool = True,
+        **kwargs,
+    ):
+        """
+        Overrides the method in `PreTrainedModel` to update the vision config with the correct attention implementation
+        """
+        config = super()._autoset_attn_implementation(
+            config=config,
+            use_flash_attention_2=use_flash_attention_2,
+            torch_dtype=torch_dtype,
+            device_map=device_map,
+            check_device_map=check_device_map,
+            **kwargs,
+        )
+        config.vision_config._attn_implementation = config._attn_implementation
+        return config
+
+
+IDEFICS2_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__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)
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
+            `past_key_values`).
+
+            If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
+            and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
+            information on the default strategy.
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.n_positions - 1]`. [What are position IDs?](../glossary#position-ids)
+        past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
+            Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
+            `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
+            `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
+
+            Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
+            blocks) that can be used (see `past_key_values` input) to speed up sequential 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)`.
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
+            model's internal embedding lookup matrix.
+        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)):
+            The tensors corresponding to the input images. Pixel values can be obtained using
+            [`AutoImageProcessor`]. See [`CLIPImageProcessor.__call__`] for details ([]`LlavaProcessor`] uses
+            [`CLIPImageProcessor`] for processing images).
+        pixel_attention_mask (`torch.Tensor` of shape `(batch_size, image_size, image_size)`, *optional*):
+            Mask to avoid performing attention on padding pixel indices.
+        image_hidden_states (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)`):
+            The hidden states of the image encoder after modality projection and perceiver resampling.
+        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 (`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.
+"""
+
+
+@add_start_docstrings(
+    """Idefics2 model consisting of a SIGLIP vision encoder and Mistral language decoder""",
+    IDEFICS2_START_DOCSTRING,
+)
+class Idefics2Model(Idefics2PreTrainedModel):
+    def __init__(self, config: Idefics2Config):
+        super().__init__(config)
+        self.padding_idx = self.config.text_config.pad_token_id
+        self.vocab_size = self.config.text_config.vocab_size
+
+        self.vision_model = Idefics2VisionTransformer(config.vision_config)
+        self.connector = Idefics2Connector(config)
+        self.text_model = AutoModel.from_config(config.text_config)
+
+        self.image_seq_len = config.perceiver_config.resampler_n_latents
+        self.image_token_id = self.config.image_token_id
+
+        self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"
+
+        self.post_init()
+
+    def enable_input_require_grads(self):
+        """
+        Enables the gradients for the input embeddings.
+
+        This is useful for lora when using gradient checkpointing.
+        c.f. https://github.com/huggingface/peft/issues/1402#issuecomment-1913675032
+
+        Override to set output.requires_grad = True for both the decoder's and vision model's embeddings.
+        """
+
+        def get_lowest_module(module):
+            if len(list(module.children())) == 0:
+                # If the module has no children, it is a leaf module (e.g., Linear, Conv2d, etc.)
+                return module
+            else:
+                # Recursively call the function on each child module
+                return get_lowest_module(list(module.children())[0])
+
+        def make_inputs_require_grads(module, input, output):
+            output.requires_grad_(True)
+
+        self._text_require_grads_hook = self.get_input_embeddings().register_forward_hook(make_inputs_require_grads)
+        self._vision_require_grads_hook = get_lowest_module(self.vision_model).register_forward_hook(
+            make_inputs_require_grads
+        )
+
+    def get_input_embeddings(self):
+        return self.text_model.get_input_embeddings()
+
+    def set_input_embeddings(self, value):
+        self.text_model.set_input_embeddings(value)
+
+    def resize_token_embeddings(self, new_num_tokens: Optional[int] = None, pad_to_multiple_of=None) -> nn.Embedding:
+        model_embeds = self.text_model.resize_token_embeddings(
+            new_num_tokens=new_num_tokens, pad_to_multiple_of=pad_to_multiple_of
+        )
+        self.config.text_config.vocab_size = model_embeds.num_embeddings
+        return model_embeds
+
+    def inputs_merger(
+        self,
+        input_ids: torch.LongTensor,
+        inputs_embeds: Optional[torch.Tensor],
+        image_hidden_states: Optional[torch.Tensor],
+    ):
+        """
+        This method aims at merging the token embeddings with the image hidden states into one single sequence of vectors that are fed to the transformer LM.
+        The merging happens as follows:
+        - The text token sequence is: `tok_1 tok_2 tok_3 <fake_token_around_image> <image> <image> ... <image> <fake_token_around_image> tok_4`.
+        - We get the image hidden states for the image through the vision encoder (and potentially the perceiver), and that hidden state is then projected into the text embedding space.
+        We thus have a sequence of image hidden states of size (1, image_seq_len, hidden_dim), where 1 is for batch_size of 1 image and hidden_dim is the hidden_dim of the LM transformer.
+        - The merging happens so that we obtain the following sequence: `vector_tok_1 vector_tok_2 vector_tok_3 vector_fake_tok_around_image {sequence of image_seq_len image hidden states} vector_fake_toke_around_image vector_tok_4`. That sequence is fed to the LM.
+        - To fit the format of that sequence, `input_ids`, `input_embeds`, `attention_mask` are all 3 adapted to insert the image hidden states.
+        """
+        num_images, _, vision_hidden_size = image_hidden_states.shape
+        special_image_token_mask = input_ids == self.image_token_id
+        new_inputs_embeds = inputs_embeds.clone()
+        reshaped_image_hidden_states = image_hidden_states.view(-1, vision_hidden_size)
+        new_inputs_embeds[special_image_token_mask] = reshaped_image_hidden_states
+        return new_inputs_embeds
+
+    @add_start_docstrings_to_model_forward(
+        """
+        Inputs fed to the model can have an arbitrary number of images. To account for this, pixel_values fed to
+        the model have image padding -> (batch_size, max_num_images, 3, max_heights, max_widths) where
+        max_num_images is the maximum number of images among the batch_size samples in the batch.
+
+        Padding images are not needed beyond padding the pixel_values at the entrance of the model.
+        For efficiency, we only pass through the vision_model's forward the real images by
+        discarding the padding images i.e. pixel_values of size (image_batch_size, 3, height, width) where
+        image_batch_size would be 7 when num_images_per_sample=[1, 3, 1, 2] and max_num_images would be 3.
+        """,
+        IDEFICS2_INPUTS_DOCSTRING,
+    )
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        pixel_attention_mask: Optional[torch.BoolTensor] = None,
+        image_hidden_states: Optional[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, Idefics2BaseModelOutputWithPast]:
+        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
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # retrieve input_ids and inputs_embeds
+        if input_ids is not None:
+            batch_size, seq_length = input_ids.shape
+        elif inputs_embeds is not None:
+            batch_size, seq_length, _ = inputs_embeds.shape
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        past_seen_tokens = 0
+        if use_cache:
+            if not isinstance(past_key_values, Cache):
+                past_key_values = DynamicCache.from_legacy_cache(past_key_values)
+            past_seen_tokens = past_key_values.get_usable_length(seq_length)
+
+        if inputs_embeds is not None and input_ids is None and past_seen_tokens == 0:
+            raise ValueError("When first calling the model, if input_embeds are passed, input_ids should not be None.")
+
+        if inputs_embeds is None:
+            inputs_embeds = self.text_model.get_input_embeddings()(input_ids)
+
+        # START VISUAL INPUTS INTEGRATION
+        if pixel_values is not None and image_hidden_states is not None:
+            raise ValueError("You cannot specify both pixel_values and image_hidden_states at the same time")
+        elif pixel_values is not None:
+            batch_size, num_images, num_channels, height, width = pixel_values.shape
+            pixel_values = pixel_values.to(dtype=self.dtype)  # fp16 compatibility
+            pixel_values = pixel_values.view(batch_size * num_images, *pixel_values.shape[2:])
+
+            # Remove padding images - padding images are full 0.
+            nb_values_per_image = pixel_values.shape[1:].numel()
+            real_images_inds = (pixel_values == 0.0).sum(dim=(-1, -2, -3)) != nb_values_per_image
+            pixel_values = pixel_values[real_images_inds].contiguous()
+
+            # Handle the vision attention mask
+            if pixel_attention_mask is None:
+                pixel_attention_mask = torch.ones(
+                    size=(pixel_values.size(0), pixel_values.size(2), pixel_values.size(3)),
+                    dtype=torch.bool,
+                    device=pixel_values.device,
+                )
+            else:
+                # Remove padding images from the mask/pP p
+                pixel_attention_mask = pixel_attention_mask.view(
+                    batch_size * num_images, *pixel_attention_mask.shape[2:]
+                )
+                pixel_attention_mask = pixel_attention_mask[real_images_inds].contiguous()
+
+            patch_size = self.config.vision_config.patch_size
+            patches_subgrid = pixel_attention_mask.unfold(dimension=1, size=patch_size, step=patch_size)
+            patches_subgrid = patches_subgrid.unfold(dimension=2, size=patch_size, step=patch_size)
+            patch_attention_mask = (patches_subgrid.sum(dim=(-1, -2)) > 0).bool()
+
+            # Get sequence from the vision encoder
+            image_hidden_states = self.vision_model(
+                pixel_values=pixel_values,
+                patch_attention_mask=patch_attention_mask,
+            ).last_hidden_state
+
+            # Modality projection & resampling
+            image_hidden_states = self.connector(
+                image_hidden_states, attention_mask=patch_attention_mask.view(pixel_values.size(0), -1)
+            )
+
+        elif image_hidden_states is not None:
+            image_hidden_states = image_hidden_states.to(dtype=self.dtype, device=input_ids.device)
+
+        if past_seen_tokens == 0 and inputs_embeds is not None and image_hidden_states is not None:
+            # When we generate, we don't want to replace the potential image_token_id that we generated by images
+            # that simply don't exist
+            inputs_embeds = self.inputs_merger(
+                input_ids=input_ids,
+                inputs_embeds=inputs_embeds,
+                image_hidden_states=image_hidden_states,
+            )
+
+        outputs = self.text_model(
+            inputs_embeds=inputs_embeds,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        if not return_dict:
+            return tuple(v for v in [*outputs, image_hidden_states] if v is not None)
+
+        return Idefics2BaseModelOutputWithPast(
+            last_hidden_state=outputs.last_hidden_state,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            image_hidden_states=image_hidden_states,
+        )
+
+
+@add_start_docstrings(
+    """The Idefics2 Model with a language modeling head. It is made up a SigLIP vision encoder, with a language modeling head on top. """,
+    IDEFICS2_START_DOCSTRING,
+)
+class Idefics2ForConditionalGeneration(Idefics2PreTrainedModel):
+    _tied_weights_keys = ["lm_head.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = Idefics2Model(config)
+        self.image_token_id = self.config.image_token_id
+
+        self.lm_head = nn.Linear(config.text_config.hidden_size, config.text_config.vocab_size, bias=False)
+        self.vocab_size = config.text_config.vocab_size
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def enable_input_require_grads(self):
+        """
+        Enables the gradients for the input embeddings. This is useful for fine-tuning adapter weights while keeping
+        the model weights fixed.
+        """
+
+        def make_inputs_require_grads(module, input, output):
+            output.requires_grad_(True)
+
+        self._text_require_grads_hook = self.get_input_embeddings().register_forward_hook(make_inputs_require_grads)
+        self._vision_require_grads_hook = self.model.vision_model.get_input_embeddings().register_forward_hook(
+            make_inputs_require_grads
+        )
+
+    def get_input_embeddings(self):
+        return self.model.text_model.get_input_embeddings()
+
+    def set_input_embeddings(self, value):
+        self.model.text_model.set_input_embeddings(value)
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def resize_token_embeddings(self, new_num_tokens: Optional[int] = None, pad_to_multiple_of=None) -> nn.Embedding:
+        # model_embeds = self.model.resize_token_embeddings(new_num_tokens=new_num_tokens, pad_to_multiple_of=pad_to_multiple_of)
+        model_embeds = self._resize_token_embeddings(new_num_tokens, pad_to_multiple_of)
+        if new_num_tokens is None and pad_to_multiple_of is None:
+            return model_embeds
+
+        # Update base model and current model config
+        # Ignore copy
+        self.config.text_config.vocab_size = model_embeds.weight.shape[0]
+        self.vocab_size = self.config.text_config.vocab_size
+
+        # Tie weights again if needed
+        self.tie_weights()
+
+        return model_embeds
+
+    def tie_weights(self):
+        """
+        Overwrite `transformers.modeling_utils.PreTrainedModel.tie_weights` to handle the case of DecoupledLinear and DecoupledEmbedding.
+        """
+        output_embeddings = self.get_output_embeddings()
+        input_embeddings = self.get_input_embeddings()
+
+        if getattr(self.config, "tie_word_embeddings", True):
+            output_embeddings.weight = input_embeddings.weight
+
+    @add_start_docstrings_to_model_forward(IDEFICS2_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=Idefics2CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        pixel_attention_mask: Optional[torch.BoolTensor] = None,
+        image_hidden_states: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, Idefics2CausalLMOutputWithPast]:
+        r"""
+        Args:
+            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+                config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+                (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> import requests
+        >>> import torch
+        >>> from PIL import Image
+        >>> from io import BytesIO
+
+        >>> from transformers import AutoProcessor, AutoModelForVision2Seq
+        >>> from transformers.image_utils import load_image
+
+        >>> DEVICE = "cuda:0"
+
+        >>> # Note that passing the image urls (instead of the actual pil images) to the processor is also possible
+        >>> image1 = load_image("https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg")
+        >>> image2 = load_image("https://cdn.britannica.com/59/94459-050-DBA42467/Skyline-Chicago.jpg")
+        >>> image3 = load_image("https://cdn.britannica.com/68/170868-050-8DDE8263/Golden-Gate-Bridge-San-Francisco.jpg")
+
+        >>> processor = AutoProcessor.from_pretrained("HuggingFaceM4/idefics2-8b-base")
+        >>> model = AutoModelForVision2Seq.from_pretrained(
+        ...     "HuggingFaceM4/idefics2-8b-base",
+        >>> ).to(DEVICE)
+
+        >>> BAD_WORDS_IDS = processor.tokenizer(["<image>", "<fake_token_around_image>"], add_special_tokens=False).input_ids
+        >>> EOS_WORDS_IDS = [processor.tokenizer.eos_token_id]
+
+        >>> # Create inputs
+        >>> prompts = [
+        ...   "<image>In this image, we can see the city of New York, and more specifically the Statue of Liberty.<image>In this image,",
+        ...   "In which city is that bridge located?<image>",
+        >>> ]
+        >>> images = [[image1, image2], [image3]]
+        >>> inputs = processor(text=prompts, padding=True, return_tensors="pt").to(DEVICE)
+
+        >>> # Generate
+        >>> generated_ids = model.generate(**inputs, bad_words_ids=BAD_WORDS_IDS, max_new_tokens=500)
+        >>> generated_texts = processor.batch_decode(generated_ids, skip_special_tokens=True)
+
+        >>> print(generated_texts)
+        ```"""
+
+        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
+
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            pixel_values=pixel_values,
+            pixel_attention_mask=pixel_attention_mask,
+            image_hidden_states=image_hidden_states,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = outputs[0]
+        logits = self.lm_head(hidden_states)
+        logits = logits.float()
+
+        loss = None
+        if labels is not None:
+            labels = labels.to(logits.device)
+            # Shift so that tokens < n predict n
+            if attention_mask is not None:
+                shift_attention_mask = attention_mask[..., 1:].to(logits.device)
+                shift_logits = logits[..., :-1, :][shift_attention_mask != 0].contiguous()
+                shift_labels = labels[..., 1:][shift_attention_mask != 0].contiguous()
+            else:
+                shift_logits = logits[..., :-1, :].contiguous()
+                shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss(ignore_index=self.image_token_id)
+            loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return Idefics2CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            image_hidden_states=outputs.image_hidden_states,
+        )
+
+    def prepare_inputs_for_generation(
+        self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
+    ):
+        # Omit tokens covered by past_key_values
+        if past_key_values is not None:
+            if isinstance(past_key_values, Cache):
+                cache_length = past_key_values.get_seq_length()
+                past_length = past_key_values.seen_tokens
+                max_cache_length = past_key_values.get_max_length()
+            else:
+                cache_length = past_length = past_key_values[0][0].shape[2]
+                max_cache_length = None
+
+            # Keep only the unprocessed tokens:
+            # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where
+            # some of the inputs are exclusively passed as part of the cache (e.g. when passing input_embeds as
+            # input)
+            if attention_mask is not None and attention_mask.shape[1] > input_ids.shape[1]:
+                input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :]
+            # 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard
+            # input_ids based on the past_length.
+            elif past_length < input_ids.shape[1]:
+                input_ids = input_ids[:, past_length:]
+            # 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens.
+
+            # If we are about to go beyond the maximum cache length, we need to crop the input attention mask.
+            if (
+                max_cache_length is not None
+                and attention_mask is not None
+                and cache_length + input_ids.shape[1] > max_cache_length
+            ):
+                attention_mask = attention_mask[:, -max_cache_length:]
+
+        position_ids = kwargs.get("position_ids", None)
+        if attention_mask is not None and position_ids is None:
+            # create position_ids on the fly for batch generation
+            position_ids = attention_mask.long().cumsum(-1) - 1
+            position_ids.masked_fill_(attention_mask == 0, 1)
+            if past_key_values:
+                position_ids = position_ids[:, -input_ids.shape[1] :]
+
+        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
+        if inputs_embeds is not None and past_key_values is None:
+            model_inputs = {"inputs_embeds": inputs_embeds}
+        else:
+            model_inputs = {"input_ids": input_ids}
+
+        image_hidden_states = kwargs.get("image_hidden_states", None)
+        if image_hidden_states is not None:
+            pixel_values = None
+            pixel_attention_mask = None
+        else:
+            pixel_values = kwargs.get("pixel_values", None)
+            pixel_attention_mask = kwargs.get("pixel_attention_mask", None)
+        model_inputs.update(
+            {
+                "position_ids": position_ids,
+                "past_key_values": past_key_values,
+                "use_cache": kwargs.get("use_cache"),
+                "attention_mask": attention_mask,
+                "pixel_values": pixel_values,
+                "pixel_attention_mask": pixel_attention_mask,
+                "image_hidden_states": image_hidden_states,
+            }
+        )
+        return model_inputs
+
+    def _update_model_kwargs_for_generation(self, outputs, model_kwargs, is_encoder_decoder, **kwargs):
+        model_kwargs = super()._update_model_kwargs_for_generation(
+            outputs=outputs,
+            model_kwargs=model_kwargs,
+            is_encoder_decoder=is_encoder_decoder,
+            **kwargs,
+        )
+        # Get the precomputed image_hidden_states
+        model_kwargs["image_hidden_states"] = outputs.image_hidden_states
+        return model_kwargs
+
+    @staticmethod
+    # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM._reorder_cache
+    def _reorder_cache(past_key_values, beam_idx):
+        reordered_past = ()
+        for layer_past in past_key_values:
+            reordered_past += (
+                tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past),
+            )
+        return reordered_past

src/transformers/models/idefics2/processing_idefics2.py

+# coding=utf-8
+# Copyright 2024 The HuggingFace Inc. team.
+#
+# 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.
+"""
+Processor class for IDEFICS2.
+"""
+
+from typing import TYPE_CHECKING, Dict, List, Optional, Union
+
+from ...feature_extraction_utils import BatchFeature
+from ...image_utils import ImageInput, is_valid_image, load_image
+from ...processing_utils import ProcessorMixin
+from ...tokenization_utils_base import AddedToken, BatchEncoding, PaddingStrategy, TextInput, TruncationStrategy
+from ...utils import TensorType, logging
+
+
+if TYPE_CHECKING:
+    from ...pipelines.conversational import Conversation
+    from ...tokenization_utils_base import PreTokenizedInput
+
+
+logger = logging.get_logger(__name__)
+
+
+def is_url(val) -> bool:
+    return isinstance(val, str) and val.startswith("http")
+
+
+def is_image_or_image_url(elem):
+    return is_url(elem) or is_valid_image(elem)
+
+
+class Idefics2Processor(ProcessorMixin):
+    r"""
+    Constructs a IDEFICS2 processor which wraps a LLama tokenizer and IDEFICS2 image processor into a single processor.
+
+    [`IdeficsProcessor`] offers all the functionalities of [`Idefics2ImageProcessor`] and [`LlamaTokenizerFast`]. See
+    the docstring of [`~IdeficsProcessor.__call__`] and [`~IdeficsProcessor.decode`] for more information.
+
+    Args:
+        image_processor (`Idefics2ImageProcessor`):
+            An instance of [`Idefics2ImageProcessor`]. The image processor is a required input.
+        tokenizer (`PreTrainedTokenizerBase`, *optional*):
+            An instance of [`PreTrainedTokenizerBase`]. This should correspond with the model's text model. The tokenizer is a required input.
+        image_seq_len (`int`, *optional*, defaults to 64):
+            The length of the image sequence i.e. the number of <image> tokens per image in the input.
+            This parameter is used to build the string from the input prompt and image tokens and should match the
+            config.perceiver_config.resampler_n_latents value for the model used.
+    """
+
+    attributes = ["image_processor", "tokenizer"]
+    image_processor_class = "Idefics2ImageProcessor"
+    tokenizer_class = "AutoTokenizer"
+
+    def __init__(self, image_processor, tokenizer=None, image_seq_len: int = 64, **kwargs):
+        if image_processor is None:
+            raise ValueError("You need to specify an `image_processor`.")
+        if tokenizer is None:
+            raise ValueError("You need to specify a `tokenizer`.")
+
+        self.fake_image_token = AddedToken("<fake_token_around_image>", normalized=False, special=True)
+        self.image_token = AddedToken("<image>", normalized=False, special=True)
+        self.end_of_utterance_token = AddedToken("<end_of_utterance>", normalized=False, special=True)
+        self.image_seq_len = image_seq_len
+
+        tokens_to_add = {
+            "additional_special_tokens": [self.fake_image_token, self.image_token, self.end_of_utterance_token]
+        }
+        tokenizer.add_special_tokens(tokens_to_add)
+
+        # Stores a Jinja template that formats chat histories into tokenizable strings
+        self.chat_template = kwargs.pop("chat_template", None)
+
+        super().__init__(image_processor, tokenizer)
+
+    def _extract_images_from_prompts(self, prompts):
+        prompt_images = []
+        for prompt in prompts:
+            images = []
+            for elem in prompt:
+                if is_valid_image(elem):
+                    images.append(elem)
+                elif is_url(elem):
+                    images.append(load_image(elem))
+            prompt_images.append(images)
+        return prompt_images
+
+    def __call__(
+        self,
+        text: Union[TextInput, "PreTokenizedInput", List[TextInput], List["PreTokenizedInput"]] = None,
+        images: Union[ImageInput, List[ImageInput], List[List[ImageInput]]] = None,
+        image_seq_len: Optional[int] = None,
+        padding: Union[bool, str, PaddingStrategy] = False,
+        truncation: Union[bool, str, TruncationStrategy] = None,
+        max_length: Optional[int] = None,
+        is_split_into_words: bool = False,
+        add_special_tokens: bool = True,
+        return_tensors: Optional[Union[str, TensorType]] = None,
+    ) -> BatchEncoding:
+        """
+        Processes the input prompts and returns a BatchEncoding.
+
+        Example:
+
+        ```python
+        >>> import requests
+        >>> from transformers import Idefics2Processor
+        >>> from transformers.image_utils import load_image
+
+        >>> processor = Idefics2Processor.from_pretrained("HuggingFaceM4/idefics2-8b", image_seq_len=2)
+        >>> processor.image_processor.do_image_splitting = False  # Force as False to simplify the example
+
+        >>> url1 = "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
+        >>> url2 = "https://cdn.britannica.com/59/94459-050-DBA42467/Skyline-Chicago.jpg"
+
+        >>> image1, image2 = load_image(url1), load_image(url2)
+        >>> images = [[image1], [image2]]
+
+        >>> text = [
+        ...     "<image>In this image, we see",
+        ...     "bla bla bla<image>",
+        ... ]
+        >>> outputs = processor(text=text, images=images, return_tensors="pt", padding=True)
+        >>> input_ids = outputs.input_ids
+        >>> input_tokens = processor.tokenizer.batch_decode(input_ids)
+        >>> print(input_tokens)
+        ['<s><fake_token_around_image><image><image><fake_token_around_image> In this image, we see', '<s> bla bla bla<fake_token_around_image><image><image><fake_token_around_image>']
+        ```
+
+        Args:
+            text (`Union[TextInput, PreTokenizedInput, List[TextInput], List[PreTokenizedInput]]`, *optional*):
+                The sequence or batch of sequences to be encoded. Each sequence can be a string or a list of strings
+                (pretokenized string). If the sequences are provided as list of strings (pretokenized), you must set
+                `is_split_into_words=True` (to lift the ambiguity with a batch of sequences).
+
+                Wherever an image token, `<image>` is encountered it is expanded to
+                `<fake_token_around_image>` + `<image>` * `image_seq_len` * <fake_token_around_image>`.
+            images (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, `List[PIL.Image.Image]`, `List[np.ndarray]`, `List[torch.Tensor]`, *optional*):
+                The image or batch of images to be prepared. Each image can be a PIL image, NumPy array or PyTorch
+                tensor. If is of type `List[ImageInput]`, it's assumed that this is for a single prompt i.e. of batch size 1.
+            image_seq_len (`int`, *optional*):
+                The length of the image sequence. If not provided, the default value is used.
+            padding (`Union[bool, str, PaddingStrategy]`, *optional*, defaults to `False`):
+                Padding strategy applied to the input ids. See [`PreTrainedTokenizerFast.pad`] for more information.
+            truncation (`Union[bool, str, TruncationStrategy]`, *optional*):
+                Truncation strategy applied to the input ids. See [`PreTrainedTokenizerFast.truncate`] for more information.
+            max_length (`int`, *optional*):
+                Maximum length of the returned list and optionally padding/truncation length. See
+                [`PreTrainedTokenizerFast.__call__`] for more information.
+            is_split_into_words (`bool`, *optional*, defaults to `False`):
+                Whether the input text is split into words or not. If set to `True`, the tokenizer will skip the
+                tokenization process and assume the input is already tokenized.
+            add_special_tokens (`bool`, *optional*, defaults to `True`):
+                Whether to add special tokens or not. See [`PreTrainedTokenizerFast.__call__`] for more information.
+            return_tensors (`Union[str, TensorType]`, *optional*):
+                If set, will return tensors of a particular framework. See [`PreTrainedTokenizerFast.__call__`] for more
+                information.
+        """
+        image_seq_len = image_seq_len if image_seq_len is not None else self.image_seq_len
+
+        n_images_in_text = []
+        inputs = BatchFeature()
+
+        if text is not None:
+            if isinstance(text, str):
+                text = [text]
+            elif not isinstance(text, list) and not isinstance(text[0], str):
+                raise ValueError("Invalid input text. Please provide a string, or a list of strings")
+
+            # Replace the image token with fake tokens around the expanded image token sequence of length `image_seq_len`
+            fake_image_token = self.fake_image_token.content
+            image_token = self.image_token.content
+            image_str = f"{fake_image_token}{image_token * image_seq_len}{fake_image_token}"
+
+            if self.image_processor.do_image_splitting:
+                # A single image token is split into 4 patches + 1 original image
+                image_str = image_str * 5
+
+            prompt_strings = []
+            for sample in text:
+                n_images_in_text.append(sample.count(image_token))
+                sample = sample.replace(image_token, image_str)
+                # Remove any double fake tokens if images are adjacent
+                sample = sample.replace(f"{fake_image_token}{fake_image_token}", f"{fake_image_token}")
+                prompt_strings.append(sample)
+
+            text_inputs = self.tokenizer(
+                text=prompt_strings,
+                add_special_tokens=add_special_tokens,
+                padding=padding,
+                truncation=truncation,
+                max_length=max_length,
+                is_split_into_words=is_split_into_words,
+                return_tensors=return_tensors,
+            )
+            inputs.update(text_inputs)
+
+        if images is not None:
+            if is_image_or_image_url(images):
+                images = [[images]]
+            elif isinstance(images, list) and is_image_or_image_url(images[0]):
+                images = [images]
+            elif (
+                not isinstance(images, list)
+                and not isinstance(images[0], list)
+                and not is_image_or_image_url(images[0][0])
+            ):
+                raise ValueError(
+                    "Invalid input images. Please provide a single image or a list of images or a list of list of images."
+                )
+
+            n_images_in_images = [len(sample) for sample in images]
+            if text is not None and not n_images_in_images == n_images_in_text:
+                raise ValueError(
+                    f"The number of images in the text {n_images_in_text} and images  {n_images_in_images} should be the same."
+                )
+
+            # Load images if they are URLs
+            images = [[load_image(im) for im in sample] for sample in images]
+            image_inputs = self.image_processor(images, return_tensors=return_tensors)
+            inputs.update(image_inputs)
+
+        return inputs
+
+    def batch_decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to LlamaTokenizerFast's [`~PreTrainedTokenizer.batch_decode`]. Please
+        refer to the docstring of this method for more information.
+        """
+        return self.tokenizer.batch_decode(*args, **kwargs)
+
+    def decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to LlamaTokenizerFast's [`~PreTrainedTokenizer.decode`]. Please refer to
+        the docstring of this method for more information.
+        """
+        return self.tokenizer.decode(*args, **kwargs)
+
+    @property
+    def model_input_names(self):
+        tokenizer_input_names = self.tokenizer.model_input_names
+        image_processor_input_names = self.image_processor.model_input_names
+        return list(dict.fromkeys(tokenizer_input_names + image_processor_input_names))
+
+    def apply_chat_template(
+        self,
+        conversation: Union[List[Dict[str, str]], "Conversation"],
+        chat_template: Optional[str] = None,
+        tokenize: bool = False,
+        **kwargs,
+    ) -> str:
+        """
+        Overrides the tokenizer's `apply_chat_template` method to apply the IDEFICS2 chat template by default
+        if no chat template is provided.
+
+        By default, the output isn't tokenized. This is because the IDEFICS2 chat template is designed to insert
+        the image token <image> into the sequence according to the message, but does not handle expanding the image
+        tokens to the sequence length or adding the surrounding tokens e.g. <fake_image_token>.
+
+        Args:
+            conversation (`Union[List[Dict, str, str], "Conversation"]`):
+                The conversation to format.
+            chat_template (`Optional[str]`, *optional*):
+                The Jinja template to use for formatting the conversation. If not provided, the default chat template
+                is used.
+            tokenize (`bool`, *optional*, defaults to `False`):
+                Whether to tokenize the output or not.
+            **kwargs:
+                Additional keyword arguments for the tokenizer's `apply_chat_template` method.
+        """
+
+        if chat_template is None:
+            if self.chat_template is not None:
+                chat_template = self.chat_template
+            else:
+                chat_template = self.default_chat_template
+
+        return self.tokenizer.apply_chat_template(
+            conversation, chat_template=chat_template, tokenize=tokenize, **kwargs
+        )
+
+    @property
+    def default_chat_template(self):
+        """
+        This template formats inputs in the form of a chat history. For each message in the chat history:
+        * the template will output the role of the speaker followed by the content of the message.
+        * content can be a single string or a list of strings and images.
+        * If the content element is an image, the template will output a sequence of <image> tokens and <fake_token_around_image> token before and after each image
+        * The template will output an <end_of_utterance> token at the end of each message.
+
+        Example:
+
+        ```python
+        messages = [{
+            "role": "user",
+            "content": [
+                {"type": "text", "text": "What’s in this image?"},
+                {"type": "image"},
+                {"type": "image"},
+                ],
+        },
+        {
+            "role": "assistant",
+            "content": [{"type": "text", "text": "This picture depicts Idefix, the dog of Obelix in Asterix and Obelix. Idefix is running on the ground."},]
+        }]
+        ```
+
+        Will create outputs like:
+        ```
+        User: What is in this Image?<image><image><end_of_utterance>
+        Assistant: This picture depicts Idefix, the dog of Obelix in Asterix and Obelix. Idefix is running on the ground.<end_of_utterance>
+        ```
+        """
+        # fmt: off
+        return (
+            "{% for message in messages %}"
+                "{{message['role'].capitalize()}}"
+                "{% if message['content'][0]['type'] == 'image' %}"
+                    "{{':'}}"
+                "{% else %}"
+                    "{{': '}}"
+                "{% endif %}"
+                "{% for line in message['content'] %}"
+                    "{% if line['type'] == 'text' %}"
+                        "{{line['text']}}"
+                    "{% elif line['type'] == 'image' %}"
+                        "{{ '<image>' }}"
+                    "{% endif %}"
+                "{% endfor %}"
+                "<end_of_utterance>\n"
+            "{% endfor %}"
+
+            "{% if add_generation_prompt %}"
+                "{{ 'Assistant:' }}"
+            "{% endif %}"
+        )
+        # fmt: on

src/transformers/utils/dummy_pt_objects.py

@@ -4405,6 +4405,37 @@ class IdeficsProcessor(metaclass=DummyObject):
         requires_backends(self, ["torch"])
 
 
+IDEFICS2_PRETRAINED_MODEL_ARCHIVE_LIST = None
+
+
+class Idefics2ForConditionalGeneration(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class Idefics2Model(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class Idefics2PreTrainedModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class Idefics2Processor(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
 IMAGEGPT_PRETRAINED_MODEL_ARCHIVE_LIST = None
 
 

src/transformers/utils/dummy_vision_objects.py

@@ -261,6 +261,13 @@ class IdeficsImageProcessor(metaclass=DummyObject):
         requires_backends(self, ["vision"])
 
 
+class Idefics2ImageProcessor(metaclass=DummyObject):
+    _backends = ["vision"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["vision"])
+
+
 class ImageGPTFeatureExtractor(metaclass=DummyObject):
     _backends = ["vision"]
 

tests/models/idefics2/test_image_processing_idefics2.py

+# coding=utf-8
+# Copyright 2024 HuggingFace Inc.
+#
+# 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 unittest
+
+import numpy as np
+
+from transformers.testing_utils import require_torch, require_vision
+from transformers.utils import is_torch_available, is_vision_available
+
+from ...test_image_processing_common import ImageProcessingTestMixin
+
+
+if is_vision_available():
+    from PIL import Image
+
+    from transformers import Idefics2ImageProcessor
+
+
+if is_torch_available():
+    import torch
+
+
+class Idefics2ImageProcessingTester(unittest.TestCase):
+    def __init__(
+        self,
+        parent,
+        batch_size=7,
+        num_channels=3,
+        num_images=1,
+        image_size=18,
+        min_resolution=30,
+        max_resolution=400,
+        do_resize=True,
+        size=None,
+        do_rescale=True,
+        rescale_factor=1 / 255,
+        do_normalize=True,
+        image_mean=[0.5, 0.5, 0.5],
+        image_std=[0.5, 0.5, 0.5],
+        do_convert_rgb=True,
+        do_pad=True,
+        do_image_splitting=True,
+    ):
+        size = size if size is not None else {"shortest_edge": 378, "longest_edge": 980}
+        self.parent = parent
+        self.batch_size = batch_size
+        self.num_channels = num_channels
+        self.num_images = num_images
+        self.image_size = image_size
+        self.min_resolution = min_resolution
+        self.max_resolution = max_resolution
+        self.do_resize = do_resize
+        self.size = size
+        self.do_normalize = do_normalize
+        self.image_mean = image_mean
+        self.image_std = image_std
+        self.do_rescale = do_rescale
+        self.rescale_factor = rescale_factor
+        self.do_convert_rgb = do_convert_rgb
+        self.do_pad = do_pad
+        self.do_image_splitting = do_image_splitting
+
+    def prepare_image_processor_dict(self):
+        return {
+            "do_convert_rgb": self.do_convert_rgb,
+            "do_resize": self.do_resize,
+            "size": self.size,
+            "do_rescale": self.do_rescale,
+            "rescale_factor": self.rescale_factor,
+            "do_normalize": self.do_normalize,
+            "image_mean": self.image_mean,
+            "image_std": self.image_std,
+            "do_pad": self.do_pad,
+            "do_image_splitting": self.do_image_splitting,
+        }
+
+    def get_expected_values(self, image_inputs, batched=False):
+        """
+        This function computes the expected height and width when providing images to BridgeTowerImageProcessor,
+        assuming do_resize is set to True with a scalar size and size_divisor.
+        """
+        if not batched:
+            shortest_edge = self.size["shortest_edge"]
+            longest_edge = self.size["longest_edge"]
+            image = image_inputs[0]
+            if isinstance(image, Image.Image):
+                w, h = image.size
+            else:
+                h, w = image.shape[1], image.shape[2]
+
+            aspect_ratio = w / h
+            if w > h and w >= longest_edge:
+                w = longest_edge
+                h = int(w / aspect_ratio)
+            elif h > w and h >= longest_edge:
+                h = longest_edge
+                w = int(h * aspect_ratio)
+            w = max(w, shortest_edge)
+            h = max(h, shortest_edge)
+            expected_height = h
+            expected_width = w
+        else:
+            expected_values = []
+            for images in image_inputs:
+                for image in images:
+                    expected_height, expected_width = self.get_expected_values([image])
+                    expected_values.append((expected_height, expected_width))
+            expected_height = max(expected_values, key=lambda item: item[0])[0]
+            expected_width = max(expected_values, key=lambda item: item[1])[1]
+
+        return expected_height, expected_width
+
+    def expected_output_image_shape(self, images):
+        height, width = self.get_expected_values(images, batched=True)
+        effective_nb_images = self.num_images * 5 if self.do_image_splitting else 1
+        return effective_nb_images, self.num_channels, height, width
+
+    def prepare_image_inputs(
+        self,
+        batch_size=None,
+        min_resolution=None,
+        max_resolution=None,
+        num_channels=None,
+        num_images=None,
+        size_divisor=None,
+        equal_resolution=False,
+        numpify=False,
+        torchify=False,
+    ):
+        """This function prepares a list of PIL images, or a list of numpy arrays if one specifies numpify=True,
+        or a list of PyTorch tensors if one specifies torchify=True.
+
+        One can specify whether the images are of the same resolution or not.
+        """
+        assert not (numpify and torchify), "You cannot specify both numpy and PyTorch tensors at the same time"
+
+        batch_size = batch_size if batch_size is not None else self.batch_size
+        min_resolution = min_resolution if min_resolution is not None else self.min_resolution
+        max_resolution = max_resolution if max_resolution is not None else self.max_resolution
+        num_channels = num_channels if num_channels is not None else self.num_channels
+        num_images = num_images if num_images is not None else self.num_images
+
+        images_list = []
+        for i in range(batch_size):
+            images = []
+            for j in range(num_images):
+                if equal_resolution:
+                    width = height = max_resolution
+                else:
+                    # To avoid getting image width/height 0
+                    if size_divisor is not None:
+                        # If `size_divisor` is defined, the image needs to have width/size >= `size_divisor`
+                        min_resolution = max(size_divisor, min_resolution)
+                    width, height = np.random.choice(np.arange(min_resolution, max_resolution), 2)
+                images.append(np.random.randint(255, size=(num_channels, width, height), dtype=np.uint8))
+            images_list.append(images)
+
+        if not numpify and not torchify:
+            # PIL expects the channel dimension as last dimension
+            images_list = [[Image.fromarray(np.moveaxis(image, 0, -1)) for image in images] for images in images_list]
+
+        if torchify:
+            images_list = [[torch.from_numpy(image) for image in images] for images in images_list]
+
+        return images_list
+
+
+@require_torch
+@require_vision
+class Idefics2ImageProcessingTest(ImageProcessingTestMixin, unittest.TestCase):
+    image_processing_class = Idefics2ImageProcessor if is_vision_available() else None
+
+    def setUp(self):
+        self.image_processor_tester = Idefics2ImageProcessingTester(self)
+
+    @property
+    def image_processor_dict(self):
+        return self.image_processor_tester.prepare_image_processor_dict()
+
+    def test_image_processor_properties(self):
+        image_processing = self.image_processing_class(**self.image_processor_dict)
+        self.assertTrue(hasattr(image_processing, "do_convert_rgb"))
+        self.assertTrue(hasattr(image_processing, "do_resize"))
+        self.assertTrue(hasattr(image_processing, "size"))
+        self.assertTrue(hasattr(image_processing, "do_rescale"))
+        self.assertTrue(hasattr(image_processing, "rescale_factor"))
+        self.assertTrue(hasattr(image_processing, "do_normalize"))
+        self.assertTrue(hasattr(image_processing, "image_mean"))
+        self.assertTrue(hasattr(image_processing, "image_std"))
+        self.assertTrue(hasattr(image_processing, "do_pad"))
+        self.assertTrue(hasattr(image_processing, "do_image_splitting"))
+
+    def test_call_numpy(self):
+        # Initialize image_processing
+        image_processing = self.image_processing_class(**self.image_processor_dict)
+        # create random numpy tensors
+        image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=False, numpify=True)
+        for sample_images in image_inputs:
+            for image in sample_images:
+                self.assertIsInstance(image, np.ndarray)
+
+        # Test not batched input
+        encoded_images = image_processing(image_inputs[0], return_tensors="pt").pixel_values
+        expected_output_image_shape = self.image_processor_tester.expected_output_image_shape([image_inputs[0]])
+        self.assertEqual(tuple(encoded_images.shape), (1, *expected_output_image_shape))
+
+        # Test batched
+        encoded_images = image_processing(image_inputs, return_tensors="pt").pixel_values
+        expected_output_image_shape = self.image_processor_tester.expected_output_image_shape(image_inputs)
+        self.assertEqual(
+            tuple(encoded_images.shape), (self.image_processor_tester.batch_size, *expected_output_image_shape)
+        )
+
+    def test_call_pil(self):
+        # Initialize image_processing
+        image_processing = self.image_processing_class(**self.image_processor_dict)
+        # create random PIL images
+        image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=False)
+        for images in image_inputs:
+            for image in images:
+                self.assertIsInstance(image, Image.Image)
+
+        # Test not batched input
+        encoded_images = image_processing(image_inputs[0], return_tensors="pt").pixel_values
+        expected_output_image_shape = self.image_processor_tester.expected_output_image_shape([image_inputs[0]])
+        self.assertEqual(tuple(encoded_images.shape), (1, *expected_output_image_shape))
+
+        # Test batched
+        encoded_images = image_processing(image_inputs, return_tensors="pt").pixel_values
+        expected_output_image_shape = self.image_processor_tester.expected_output_image_shape(image_inputs)
+        self.assertEqual(
+            tuple(encoded_images.shape), (self.image_processor_tester.batch_size, *expected_output_image_shape)
+        )
+
+    def test_call_pytorch(self):
+        # Initialize image_processing
+        image_processing = self.image_processing_class(**self.image_processor_dict)
+        # create random PyTorch tensors
+        image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=False, torchify=True)
+
+        for images in image_inputs:
+            for image in images:
+                self.assertIsInstance(image, torch.Tensor)
+
+        # Test not batched input
+        encoded_images = image_processing(image_inputs[0], return_tensors="pt").pixel_values
+        expected_output_image_shape = self.image_processor_tester.expected_output_image_shape([image_inputs[0]])
+        self.assertEqual(tuple(encoded_images.shape), (1, *expected_output_image_shape))
+
+        # Test batched
+        expected_output_image_shape = self.image_processor_tester.expected_output_image_shape(image_inputs)
+        encoded_images = image_processing(image_inputs, return_tensors="pt").pixel_values
+        self.assertEqual(
+            tuple(encoded_images.shape),
+            (self.image_processor_tester.batch_size, *expected_output_image_shape),
+        )

tests/models/idefics2/test_modeling_idefics2.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.
+"""Testing suite for the PyTorch Idefics2 model."""
+
+import copy
+import gc
+import unittest
+from io import BytesIO
+
+import requests
+
+from transformers import (
+    AutoProcessor,
+    Idefics2Config,
+    Idefics2ForConditionalGeneration,
+    Idefics2Model,
+    is_torch_available,
+    is_vision_available,
+)
+from transformers.testing_utils import require_bitsandbytes, require_torch, slow, torch_device
+
+from ...generation.test_utils import GenerationTesterMixin
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_common import ModelTesterMixin, floats_tensor, ids_tensor
+
+
+if is_torch_available():
+    import torch
+else:
+    is_torch_greater_or_equal_than_2_0 = False
+
+if is_vision_available():
+    from PIL import Image
+
+
+class Idefics2VisionText2TextModelTester:
+    def __init__(
+        self,
+        parent,
+        is_training=True,
+        batch_size=2,
+        num_images=2,
+        seq_length=10,
+        vision_config={
+            "image_size": 12,
+            "patch_size": 12,
+            "num_channels": 3,
+            "hidden_size": 32,
+            "num_hidden_layers": 2,
+            "num_attention_heads": 4,
+            "intermediate_size": 32,
+            "dropout": 0.1,
+            "attention_dropout": 0.1,
+            "initializer_range": 0.02,
+        },
+        perceiver_config={
+            "hidden_act": "silu",
+            "resampler_n_latents": 2,
+            "resampler_depth": 2,
+            "resampler_n_heads": 2,
+            "num_key_value_heads": 1,
+            "resampler_head_dim": 12,
+            "attention_dropout": 0.0,
+        },
+        text_config={
+            "vocab_size": 100,
+            "hidden_size": 64,
+            "intermediate_size": 56,
+            "num_hidden_layers": 3,
+            "num_attention_heads": 2,
+            "num_key_value_heads": 2,
+            "hidden_act": "silu",
+            "max_position_embeddings": 256,
+            "initializer_range": 0.02,
+            "rms_norm_eps": 1e-6,
+            "pad_token_id": 0,  # None in the original configuration_mistral, we set it to the unk_token_id
+            "bos_token_id": 1,
+            "eos_token_id": 2,
+            "image_token_id": 32_001,
+            "tie_word_embeddings": False,
+            "rope_theta": 10000.0,
+            "sliding_window": 32,
+            "attention_dropout": 0.0,
+        },
+        use_cache=False,
+        tie_word_embeddings=False,
+        image_token_id=99,
+    ):
+        self.parent = parent
+        self.is_training = is_training
+        self.batch_size = batch_size
+        self.num_images = num_images
+        self.num_channels = 3
+        self.seq_length = seq_length
+        self.use_cache = use_cache
+        self.image_token_id = image_token_id
+        self.tie_word_embeddings = tie_word_embeddings
+        # Hack - add properties here so use common tests
+        self.vocab_size = text_config["vocab_size"]
+        self.num_hidden_layers = text_config["num_hidden_layers"]
+        self.num_attention_heads = text_config["num_attention_heads"]
+        self.hidden_size = text_config["hidden_size"]
+
+        self.vision_config = vision_config
+        self.perceiver_config = perceiver_config
+        self.text_config = text_config
+
+    def get_config(self):
+        return Idefics2Config(
+            use_cache=self.use_cache,
+            image_token_id=self.image_token_id,
+            tie_word_embeddings=self.tie_word_embeddings,
+            vision_config=self.vision_config,
+            perceiver_config=self.perceiver_config,
+            text_config=self.text_config,
+            vocab_size=self.vocab_size,
+        )
+
+    def prepare_config_and_inputs(self):
+        pixel_values = floats_tensor(
+            [
+                self.batch_size,
+                self.num_images,
+                self.vision_config["num_channels"],
+                self.vision_config["image_size"],
+                self.vision_config["image_size"],
+            ]
+        )
+        config = self.get_config()
+
+        return config, pixel_values
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        config, pixel_values = config_and_inputs
+        input_ids = ids_tensor([self.batch_size, self.seq_length], config.text_config.vocab_size - 2) + 1
+
+        # For simplicity just set the last n tokens to the image token
+        n_image_tokens_per_batch = self.num_images * self.perceiver_config["resampler_n_latents"]
+        input_ids[:, -n_image_tokens_per_batch:] = self.image_token_id
+        attention_mask = input_ids.ne(1).to(torch_device)
+        inputs_dict = {
+            "pixel_values": pixel_values,
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+        }
+        return config, inputs_dict
+
+
+@require_torch
+class Idefics2ModelTest(ModelTesterMixin, unittest.TestCase):
+    """
+    Model tester for `Idefics2`.
+    """
+
+    all_model_classes = (Idefics2Model,) if is_torch_available() else ()
+    fx_compatible = False
+    test_torchscript = False
+    test_pruning = False
+    test_resize_embeddings = True
+    test_head_masking = False
+
+    def setUp(self):
+        self.model_tester = Idefics2VisionText2TextModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=Idefics2Config, has_text_modality=False)
+
+    @unittest.skip("input_embeds cannot be passed in without input_ids")
+    def test_inputs_embeds():
+        pass
+
+    @unittest.skip("Model does not support padding right")
+    def test_flash_attn_2_generate_padding_right(self):
+        pass
+
+    @unittest.skip("Model does not support padding right")
+    def test_flash_attn_2_inference_padding_right(self):
+        pass
+
+    # We need to override as we need to prepare such that the image token is the last token
+    def test_resize_tokens_embeddings(self):
+        (original_config, inputs_dict) = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            config = copy.deepcopy(original_config)
+            model = model_class(config)
+            model.to(torch_device)
+
+            if self.model_tester.is_training is False:
+                model.eval()
+
+            model_vocab_size = config.text_config.vocab_size
+            # Retrieve the embeddings and clone theme
+            model_embed = model.resize_token_embeddings(model_vocab_size)
+            cloned_embeddings = model_embed.weight.clone()
+
+            # Check that resizing the token embeddings with a larger vocab size increases the model's vocab size
+            model_embed = model.resize_token_embeddings(model_vocab_size + 10)
+            self.assertEqual(model.config.text_config.vocab_size, model_vocab_size + 10)
+            # Check that it actually resizes the embeddings matrix
+            self.assertEqual(model_embed.weight.shape[0], cloned_embeddings.shape[0] + 10)
+            # Check that the model can still do a forward pass successfully (every parameter should be resized)
+            model(**self._prepare_for_class(inputs_dict, model_class))
+
+            # Check that resizing the token embeddings with a smaller vocab size decreases the model's vocab size
+            model_embed = model.resize_token_embeddings(model_vocab_size - 15)
+            self.assertEqual(model.config.text_config.vocab_size, model_vocab_size - 15)
+            # Check that it actually resizes the embeddings matrix
+            self.assertEqual(model_embed.weight.shape[0], cloned_embeddings.shape[0] - 15)
+
+            # Ignore copy
+            # Check that the model can still do a forward pass successfully (every parameter should be resized)
+            # Input ids should be clamped to the maximum size of the vocabulary - 1 and the image token should be the last token
+            inputs_dict["input_ids"].clamp_(max=model_vocab_size - 15 - 2)
+            n_images = self.model_tester.num_images * self.model_tester.perceiver_config["resampler_n_latents"]
+            model.image_token_id = model_vocab_size - 15 - 1
+            inputs_dict["input_ids"][:, -n_images:] = model.image_token_id
+
+            # make sure that decoder_input_ids are resized as well
+            if "decoder_input_ids" in inputs_dict:
+                inputs_dict["decoder_input_ids"].clamp_(max=model_vocab_size - 15 - 1)
+            model(**self._prepare_for_class(inputs_dict, model_class))
+
+            # Check that adding and removing tokens has not modified the first part of the embedding matrix.
+            models_equal = True
+            for p1, p2 in zip(cloned_embeddings, model_embed.weight):
+                if p1.data.ne(p2.data).sum() > 0:
+                    models_equal = False
+
+            self.assertTrue(models_equal)
+
+            config = copy.deepcopy(original_config)
+            model = model_class(config)
+            model.to(torch_device)
+
+            model_vocab_size = config.text_config.vocab_size
+            model.resize_token_embeddings(model_vocab_size + 10, pad_to_multiple_of=1)
+            self.assertTrue(model.config.text_config.vocab_size + 10, model_vocab_size)
+
+            model_embed = model.resize_token_embeddings(model_vocab_size, pad_to_multiple_of=64)
+            self.assertTrue(model_embed.weight.shape[0] // 64, 0)
+
+            self.assertTrue(model_embed.weight.shape[0], model.config.text_config.vocab_size)
+            self.assertTrue(model.config.text_config.vocab_size, model.vocab_size)
+
+            model_embed = model.resize_token_embeddings(model_vocab_size + 13, pad_to_multiple_of=64)
+            self.assertTrue(model_embed.weight.shape[0] // 64, 0)
+
+            # Check that resizing a model to a multiple of pad_to_multiple leads to a model of exactly that size
+            target_dimension = 128
+            model_embed = model.resize_token_embeddings(target_dimension, pad_to_multiple_of=64)
+            self.assertTrue(model_embed.weight.shape[0], target_dimension)
+
+            with self.assertRaisesRegex(
+                ValueError,
+                "Asking to pad the embedding matrix to a multiple of `1.3`, which is not and integer. Please make sure to pass an integer",
+            ):
+                model.resize_token_embeddings(model_vocab_size, pad_to_multiple_of=1.3)
+
+    # We need to override as we need to prepare such that the image token is the last token
+    def test_resize_embeddings_untied(self):
+        (original_config, inputs_dict) = self.model_tester.prepare_config_and_inputs_for_common()
+
+        original_config.tie_word_embeddings = False
+
+        for model_class in self.all_model_classes:
+            config = copy.deepcopy(original_config)
+            model = model_class(config).to(torch_device)
+
+            # if no output embeddings -> leave test
+            if model.get_output_embeddings() is None:
+                continue
+
+            # Check that resizing the token embeddings with a larger vocab size increases the model's vocab size
+            model_vocab_size = config.text_config.vocab_size
+            model.resize_token_embeddings(model_vocab_size + 10)
+            self.assertEqual(model.config.text_config.vocab_size, model_vocab_size + 10)
+            output_embeds = model.get_output_embeddings()
+            self.assertEqual(output_embeds.weight.shape[0], model_vocab_size + 10)
+            # Check bias if present
+            if output_embeds.bias is not None:
+                self.assertEqual(output_embeds.bias.shape[0], model_vocab_size + 10)
+            # Check that the model can still do a forward pass successfully (every parameter should be resized)
+            model(**self._prepare_for_class(inputs_dict, model_class))
+
+            # Check that resizing the token embeddings with a smaller vocab size decreases the model's vocab size
+            model.resize_token_embeddings(model_vocab_size - 15)
+            self.assertEqual(model.config.text_config.vocab_size, model_vocab_size - 15)
+            # Check that it actually resizes the embeddings matrix
+            output_embeds = model.get_output_embeddings()
+            self.assertEqual(output_embeds.weight.shape[0], model_vocab_size - 15)
+            # Check bias if present
+            if output_embeds.bias is not None:
+                self.assertEqual(output_embeds.bias.shape[0], model_vocab_size - 15)
+
+            # Check that the model can still do a forward pass successfully (every parameter should be resized)
+            # Input ids should be clamped to the maximum size of the vocabulary - 1 and the image token should be the last token
+            inputs_dict["input_ids"].clamp_(max=model_vocab_size - 15 - 2)
+            n_images = self.model_tester.num_images * self.model_tester.perceiver_config["resampler_n_latents"]
+            model.image_token_id = model_vocab_size - 15 - 1
+            inputs_dict["input_ids"][:, -n_images:] = model.image_token_id
+
+            # Check that the model can still do a forward pass successfully (every parameter should be resized)
+            model(**self._prepare_for_class(inputs_dict, model_class))
+
+
+@require_torch
+class Idefics2ForConditionalGenerationModelTest(GenerationTesterMixin, ModelTesterMixin, unittest.TestCase):
+    """
+    Model tester for `Idefics2ForConditionalGeneration`.
+    """
+
+    all_model_classes = (Idefics2ForConditionalGeneration,) if is_torch_available() else ()
+    fx_compatible = False
+    test_pruning = False
+    test_resize_embeddings = True
+    test_head_masking = False
+    test_torchscript = False
+
+    def setUp(self):
+        self.model_tester = Idefics2VisionText2TextModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=Idefics2Config, has_text_modality=False)
+
+    @unittest.skip("input_embeds cannot be passed in without input_ids")
+    def test_inputs_embeds():
+        pass
+
+    @unittest.skip("Model does not support padding right")
+    def test_flash_attn_2_generate_padding_right(self):
+        pass
+
+    @unittest.skip("Model does not support padding right")
+    def test_flash_attn_2_inference_padding_right(self):
+        pass
+
+    # We need to override as we need to prepare such that the image token is the last token
+    def test_resize_tokens_embeddings(self):
+        (original_config, inputs_dict) = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            config = copy.deepcopy(original_config)
+            model = model_class(config)
+            model.to(torch_device)
+
+            model_vocab_size = config.text_config.vocab_size
+            # Retrieve the embeddings and clone theme
+            model_embed = model.resize_token_embeddings(model_vocab_size)
+            cloned_embeddings = model_embed.weight.clone()
+
+            # Check that resizing the token embeddings with a larger vocab size increases the model's vocab size
+            model_embed = model.resize_token_embeddings(model_vocab_size + 10)
+            self.assertEqual(model.config.text_config.vocab_size, model_vocab_size + 10)
+            # Check that it actually resizes the embeddings matrix
+            self.assertEqual(model_embed.weight.shape[0], cloned_embeddings.shape[0] + 10)
+            # Check that the model can still do a forward pass successfully (every parameter should be resized)
+            model(**self._prepare_for_class(inputs_dict, model_class))
+
+            # Check that resizing the token embeddings with a smaller vocab size decreases the model's vocab size
+            model_embed = model.resize_token_embeddings(model_vocab_size - 15)
+            self.assertEqual(model.config.text_config.vocab_size, model_vocab_size - 15)
+            # Check that it actually resizes the embeddings matrix
+            self.assertEqual(model_embed.weight.shape[0], cloned_embeddings.shape[0] - 15)
+
+            # Check that the model can still do a forward pass successfully (every parameter should be resized)
+            # Input ids should be clamped to the maximum size of the vocabulary - 1 and the image token should be the last token
+            inputs_dict["input_ids"].clamp_(max=model_vocab_size - 15 - 2)
+            n_images = self.model_tester.num_images * self.model_tester.perceiver_config["resampler_n_latents"]
+            model.model.image_token_id = model_vocab_size - 15 - 1
+            inputs_dict["input_ids"][:, -n_images:] = model.model.image_token_id
+
+            model(**self._prepare_for_class(inputs_dict, model_class))
+
+            # Check that adding and removing tokens has not modified the first part of the embedding matrix.
+            models_equal = True
+            for p1, p2 in zip(cloned_embeddings, model_embed.weight):
+                if p1.data.ne(p2.data).sum() > 0:
+                    models_equal = False
+
+            self.assertTrue(models_equal)
+
+            config = copy.deepcopy(original_config)
+            model = model_class(config)
+            model.to(torch_device)
+
+            model_vocab_size = config.text_config.vocab_size
+            model.resize_token_embeddings(model_vocab_size + 10, pad_to_multiple_of=1)
+            self.assertTrue(model.config.text_config.vocab_size + 10, model_vocab_size)
+
+            model_embed = model.resize_token_embeddings(model_vocab_size, pad_to_multiple_of=64)
+            self.assertTrue(model_embed.weight.shape[0] // 64, 0)
+
+            self.assertTrue(model_embed.weight.shape[0], model.config.text_config.vocab_size)
+            self.assertTrue(model.config.text_config.vocab_size, model.vocab_size)
+
+            model_embed = model.resize_token_embeddings(model_vocab_size + 13, pad_to_multiple_of=64)
+            self.assertTrue(model_embed.weight.shape[0] // 64, 0)
+
+            # Check that resizing a model to a multiple of pad_to_multiple leads to a model of exactly that size
+            target_dimension = 128
+            model_embed = model.resize_token_embeddings(target_dimension, pad_to_multiple_of=64)
+            self.assertTrue(model_embed.weight.shape[0], target_dimension)
+
+            with self.assertRaisesRegex(
+                ValueError,
+                "Asking to pad the embedding matrix to a multiple of `1.3`, which is not and integer. Please make sure to pass an integer",
+            ):
+                model.resize_token_embeddings(model_vocab_size, pad_to_multiple_of=1.3)
+
+    # We need to override as we need to prepare such that the image token is the last token
+    def test_resize_embeddings_untied(self):
+        (original_config, inputs_dict) = self.model_tester.prepare_config_and_inputs_for_common()
+
+        original_config.tie_word_embeddings = False
+
+        for model_class in self.all_model_classes:
+            config = copy.deepcopy(original_config)
+            model = model_class(config).to(torch_device)
+
+            # Check that resizing the token embeddings with a larger vocab size increases the model's vocab size
+            model_vocab_size = config.text_config.vocab_size
+            model.resize_token_embeddings(model_vocab_size + 10)
+            self.assertEqual(model.config.text_config.vocab_size, model_vocab_size + 10)
+            output_embeds = model.get_output_embeddings()
+            self.assertEqual(output_embeds.weight.shape[0], model_vocab_size + 10)
+            # Check bias if present
+            if output_embeds.bias is not None:
+                self.assertEqual(output_embeds.bias.shape[0], model_vocab_size + 10)
+            # Check that the model can still do a forward pass successfully (every parameter should be resized)
+            model(**self._prepare_for_class(inputs_dict, model_class))
+
+            # Check that resizing the token embeddings with a smaller vocab size decreases the model's vocab size
+            model.resize_token_embeddings(model_vocab_size - 15)
+            self.assertEqual(model.config.text_config.vocab_size, model_vocab_size - 15)
+            # Check that it actually resizes the embeddings matrix
+            output_embeds = model.get_output_embeddings()
+            self.assertEqual(output_embeds.weight.shape[0], model_vocab_size - 15)
+            # Check bias if present
+            if output_embeds.bias is not None:
+                self.assertEqual(output_embeds.bias.shape[0], model_vocab_size - 15)
+
+            # Check that the model can still do a forward pass successfully (every parameter should be resized)
+            # Input ids should be clamped to the maximum size of the vocabulary - 1 and the image token should be the last token
+            inputs_dict["input_ids"].clamp_(max=model_vocab_size - 15 - 2)
+            n_images = self.model_tester.num_images * self.model_tester.perceiver_config["resampler_n_latents"]
+            model.model.image_token_id = model_vocab_size - 15 - 1
+            inputs_dict["input_ids"][:, -n_images:] = model.model.image_token_id
+
+            # Check that the model can still do a forward pass successfully (every parameter should be resized)
+            model(**self._prepare_for_class(inputs_dict, model_class))
+
+
+@require_torch
+class Idefics2ForConditionalGenerationIntegrationTest(unittest.TestCase):
+    def setUp(self):
+        self.processor = AutoProcessor.from_pretrained("HuggingFaceM4/idefics2-8b-base")
+        self.image1 = Image.open(
+            BytesIO(
+                requests.get(
+                    "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
+                ).content
+            )
+        )
+        self.image2 = Image.open(
+            BytesIO(requests.get("https://cdn.britannica.com/59/94459-050-DBA42467/Skyline-Chicago.jpg").content)
+        )
+        self.image3 = Image.open(
+            BytesIO(
+                requests.get(
+                    "https://thumbs.dreamstime.com/b/golden-gate-bridge-san-francisco-purple-flowers-california-echium-candicans-36805947.jpg"
+                ).content
+            )
+        )
+
+    def tearDown(self):
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    @slow
+    def test_integration_test(self):
+        model = Idefics2ForConditionalGeneration.from_pretrained(
+            "HuggingFaceM4/idefics2-8b-base",
+            torch_dtype=torch.bfloat16,
+            device_map="auto",
+        )
+        model.to(torch_device)
+
+        # Create inputs
+        text = "<image>In this image, we see"
+        images = self.image1
+        inputs = self.processor(text=text, images=images, return_tensors="pt", padding=True)
+        inputs.to(torch_device)
+
+        generated_ids = model.generate(**inputs, max_new_tokens=10)
+        generated_texts = self.processor.batch_decode(generated_ids, skip_special_tokens=True)
+
+        # Batch affects generated text. Single batch output: ['In this image, we see the Statue of Liberty in the foreground and']
+        expected_generated_text = "In this image, we see the Statue of Liberty, the New York City"
+        self.assertEqual(generated_texts[0], expected_generated_text)
+
+    @slow
+    @require_bitsandbytes
+    def test_integration_test_4bit(self):
+        # Let' s make sure we test the preprocessing to replace what is used
+        model = Idefics2ForConditionalGeneration.from_pretrained(
+            "HuggingFaceM4/idefics2-8b-base", load_in_4bit=True, device_map="auto"
+        )
+
+        # Create pixel inputs
+        text = ["<image>In this image, we see", "bla, bla <image><image>"]
+        images = [[self.image1], [self.image2, self.image3]]
+        inputs = self.processor(text=text, images=images, padding=True, return_tensors="pt")
+
+        generated_ids = model.generate(**inputs, max_new_tokens=10)
+        generated_texts = self.processor.batch_decode(generated_ids, skip_special_tokens=True)
+
+        expected_generated_text = "In this image, we see the Statue of Liberty, the Hudson River,"
+        self.assertEqual(generated_texts[0], expected_generated_text)

tests/models/idefics2/test_processing_idefics2.py

+# coding=utf-8
+# Copyright 2024 HuggingFace Inc.
+#
+# 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 unittest
+from io import BytesIO
+
+import requests
+
+from transformers import Idefics2Processor
+from transformers.testing_utils import require_torch, require_vision
+from transformers.utils import is_vision_available
+
+
+if is_vision_available():
+    from PIL import Image
+
+
+@require_torch
+@require_vision
+class Idefics2ProcessorTest(unittest.TestCase):
+    def setUp(self):
+        self.processor = Idefics2Processor.from_pretrained("HuggingFaceM4/idefics2-8b", image_seq_len=2)
+        self.image1 = Image.open(
+            BytesIO(
+                requests.get(
+                    "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
+                ).content
+            )
+        )
+        self.image2 = Image.open(
+            BytesIO(requests.get("https://cdn.britannica.com/59/94459-050-DBA42467/Skyline-Chicago.jpg").content)
+        )
+        self.image3 = Image.open(
+            BytesIO(
+                requests.get(
+                    "https://thumbs.dreamstime.com/b/golden-gate-bridge-san-francisco-purple-flowers-california-echium-candicans-36805947.jpg"
+                ).content
+            )
+        )
+        self.bos_token = self.processor.tokenizer.bos_token
+        self.image_token = self.processor.image_token.content
+        self.fake_image_token = self.processor.fake_image_token.content
+
+        self.bos_token_id = self.processor.tokenizer.convert_tokens_to_ids(self.bos_token)
+        self.image_token_id = self.processor.tokenizer.convert_tokens_to_ids(self.image_token)
+        self.fake_image_token_id = self.processor.tokenizer.convert_tokens_to_ids(self.fake_image_token)
+        self.image_seq_len = self.processor.image_seq_len
+
+    def test_process_interleaved_images_prompts_no_image_splitting(self):
+        old_image_splitting = self.processor.image_processor.do_image_splitting
+
+        self.processor.image_processor.do_image_splitting = False
+
+        # Test that a single image is processed correctly
+        inputs = self.processor(images=self.image1)
+        self.assertEqual(inputs["pixel_values"].shape, (1, 1, 3, 653, 980))
+        self.assertEqual(inputs["pixel_attention_mask"].shape, (1, 1, 653, 980))
+        # fmt: on
+
+        # Test a single sample with image and text
+        image_str = "<image>"
+        text_str = "In this image, we see"
+        text = image_str + text_str
+        inputs = self.processor(text=text, images=self.image1)
+
+        # fmt: off
+        tokenized_sentence = self.processor.tokenizer(text_str, add_special_tokens=False)
+        expected_input_ids = [[self.bos_token_id] + [self.fake_image_token_id] + [self.image_token_id] * self.image_seq_len + [self.fake_image_token_id] + tokenized_sentence["input_ids"]]
+        self.assertEqual(inputs["input_ids"], expected_input_ids)
+        self.assertEqual(inputs["attention_mask"], [[1] * len(expected_input_ids[0])])
+        self.assertEqual(inputs["pixel_values"].shape, (1, 1, 3, 653, 980))
+        self.assertEqual(inputs["pixel_attention_mask"].shape, (1, 1, 653, 980))
+        # fmt: on
+
+        # Test that batch is correctly processed
+        image_str = "<image>"
+        text_str_1 = "In this image, we see"
+        text_str_2 = "bla, bla"
+
+        text = [
+            image_str + text_str_1,
+            text_str_2 + image_str + image_str,
+        ]
+        images = [[self.image1], [self.image2, self.image3]]
+
+        inputs = self.processor(text=text, images=images, padding=True)
+
+        # fmt: off
+        tokenized_sentence_1 = self.processor.tokenizer(text_str_1, add_special_tokens=False)
+        tokenized_sentence_2 = self.processor.tokenizer(text_str_2, add_special_tokens=False)
+        expected_input_ids_1 = [self.bos_token_id] + [self.fake_image_token_id] + [self.image_token_id] * self.image_seq_len + [self.fake_image_token_id] + tokenized_sentence_1["input_ids"]
+        expected_input_ids_2 = [self.bos_token_id] + tokenized_sentence_2["input_ids"] + [self.fake_image_token_id] + [self.image_token_id] * self.image_seq_len + [self.fake_image_token_id] + [self.image_token_id] * self.image_seq_len + [self.fake_image_token_id]
+        # Pad the first input to match the second input
+        pad_len = len(expected_input_ids_2) - len(expected_input_ids_1)
+        padded_expected_input_ids_1 = [0] * pad_len + expected_input_ids_1
+
+        self.assertEqual(
+            inputs["input_ids"], [padded_expected_input_ids_1, expected_input_ids_2]
+        )
+        self.assertEqual(
+            inputs["attention_mask"],
+            [[0] * pad_len + [1] * len(expected_input_ids_1), [1] * len(expected_input_ids_2)]
+        )
+        self.assertEqual(inputs['pixel_values'].shape, (2, 2, 3, 767, 980))
+        self.assertEqual(inputs['pixel_attention_mask'].shape, (2, 2, 767, 980))
+        # fmt: on
+
+        self.processor.image_processor.do_image_splitting = old_image_splitting
+
+    def test_process_interleaved_images_prompts_image_splitting(self):
+        old_image_splitting = self.processor.image_processor.do_image_splitting
+
+        self.processor.image_processor.do_image_splitting = True
+
+        # Test that a single image is processed correctly
+        inputs = self.processor(images=self.image1)
+        self.assertEqual(inputs["pixel_values"].shape, (1, 5, 3, 653, 980))
+        self.assertEqual(inputs["pixel_attention_mask"].shape, (1, 5, 653, 980))
+        # fmt: on
+
+        # Test a single sample with image and text
+        image_str = "<image>"
+        text_str = "In this image, we see"
+        text = image_str + text_str
+        inputs = self.processor(text=text, images=self.image1)
+
+        # fmt: off
+        tokenized_sentence = self.processor.tokenizer(text_str, add_special_tokens=False)
+        expected_input_ids = [[self.bos_token_id] + ([self.fake_image_token_id] + [self.image_token_id] * self.image_seq_len) * 5 + [self.fake_image_token_id] + tokenized_sentence["input_ids"]]
+        self.assertEqual(inputs["input_ids"], expected_input_ids)
+        self.assertEqual(inputs["attention_mask"], [[1] * len(expected_input_ids[0])])
+        self.assertEqual(inputs["pixel_values"].shape, (1, 5, 3, 653, 980))
+        self.assertEqual(inputs["pixel_attention_mask"].shape, (1, 5, 653, 980))
+        # fmt: on
+
+        # Test that batch is correctly processed
+        image_str = "<image>"
+        text_str_1 = "In this image, we see"
+        text_str_2 = "bla, bla"
+
+        text = [
+            image_str + text_str_1,
+            text_str_2 + image_str + image_str,
+        ]
+        images = [[self.image1], [self.image2, self.image3]]
+
+        inputs = self.processor(text=text, images=images, padding=True)
+
+        # fmt: off
+        tokenized_sentence_1 = self.processor.tokenizer(text_str_1, add_special_tokens=False)
+        tokenized_sentence_2 = self.processor.tokenizer(text_str_2, add_special_tokens=False)
+        expected_input_ids_1 = [self.bos_token_id] + ([self.fake_image_token_id] + [self.image_token_id] * self.image_seq_len) * 5 + [self.fake_image_token_id] + tokenized_sentence_1["input_ids"]
+        expected_input_ids_2 = [self.bos_token_id] + tokenized_sentence_2["input_ids"] + ([self.fake_image_token_id] + [self.image_token_id] * self.image_seq_len) * 5 + ([self.fake_image_token_id] + [self.image_token_id] * self.image_seq_len) * 5 + [self.fake_image_token_id]
+        # Pad the first input to match the second input
+        pad_len = len(expected_input_ids_2) - len(expected_input_ids_1)
+        padded_expected_input_ids_1 = [0] * pad_len + expected_input_ids_1
+
+        self.assertEqual(
+            inputs["input_ids"], [padded_expected_input_ids_1, expected_input_ids_2]
+        )
+        self.assertEqual(
+            inputs["attention_mask"],
+            [[0] * pad_len + [1] * len(expected_input_ids_1), [1] * len(expected_input_ids_2)]
+        )
+        self.assertEqual(inputs['pixel_values'].shape, (2, 10, 3, 767, 980))
+        self.assertEqual(inputs['pixel_attention_mask'].shape, (2, 10, 767, 980))
+        # fmt: on
+
+        self.processor.image_processor.do_image_splitting = old_image_splitting
+
+    def test_add_special_tokens_processor(self):
+        image_str = "<image>"
+        text_str = "In this image, we see"
+        text = text_str + image_str
+
+        n_image_repeat = 5 if self.processor.image_processor.do_image_splitting else 1
+
+        # fmt: off
+        inputs = self.processor(text=text, images=self.image1, add_special_tokens=False)
+        tokenized_sentence = self.processor.tokenizer(text_str, add_special_tokens=False)
+        expected_input_ids = [tokenized_sentence["input_ids"] + ([self.fake_image_token_id] + [self.image_token_id] * self.image_seq_len) * n_image_repeat + [self.fake_image_token_id]]
+        self.assertEqual(inputs["input_ids"], expected_input_ids)
+
+        inputs = self.processor(text=text, images=self.image1)
+        expected_input_ids = [[self.bos_token_id] + tokenized_sentence["input_ids"] + ([self.fake_image_token_id] + [self.image_token_id] * self.image_seq_len) * n_image_repeat + [self.fake_image_token_id]]
+        self.assertEqual(inputs["input_ids"], expected_input_ids)
+        # fmt: on
+
+    def test_apply_chat_template(self):
+        # Message contains content which a mix of lists with images and image urls and string
+        messages = [
+            {
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": "What do these images show?"},
+                    {"type": "image"},
+                    {"type": "image"},
+                    "What do these images show?",
+                ],
+            },
+            {
+                "role": "assistant",
+                "content": [
+                    {
+                        "type": "text",
+                        "text": "The first image shows the statue of Liberty in New York. The second image picture depicts Idefix, the dog of Obelix in Asterix and Obelix.",
+                    }
+                ],
+            },
+            {"role": "user", "content": [{"type": "text", "text": "And who is that?"}]},
+        ]
+
+        processor = self.processor
+        # Make short sequence length to test that the fake tokens are added correctly
+        rendered = processor.apply_chat_template(messages, add_generation_prompt=True)
+
+        expected_rendered = (
+            "User: What do these images show?<image><image><end_of_utterance>\n"
+            "Assistant: The first image shows the statue of Liberty in New York. The second image picture depicts Idefix, the dog of Obelix in Asterix and Obelix.<end_of_utterance>\n"
+            "User: And who is that?<end_of_utterance>\n"
+            "Assistant:"
+        )
+        self.assertEqual(rendered, expected_rendered)

tests/test_modeling_common.py

@@ -60,6 +60,7 @@ from transformers.models.auto.modeling_auto import (
     MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES,
     MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING_NAMES,
     MODEL_FOR_VIDEO_CLASSIFICATION_MAPPING_NAMES,
+    MODEL_FOR_VISION_2_SEQ_MAPPING_NAMES,
     MODEL_MAPPING_NAMES,
 )
 from transformers.testing_utils import (
@@ -220,6 +221,7 @@ class ModelTesterMixin:
                 *get_values(MODEL_FOR_CAUSAL_IMAGE_MODELING_MAPPING_NAMES),
                 *get_values(MODEL_FOR_MASKED_LM_MAPPING_NAMES),
                 *get_values(MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING_NAMES),
+                *get_values(MODEL_FOR_VISION_2_SEQ_MAPPING_NAMES),
             ]:
                 inputs_dict["labels"] = torch.zeros(
                     (self.model_tester.batch_size, self.model_tester.seq_length), dtype=torch.long, device=torch_device