Add files using upload-large-folder tool

lib/python3.12/site-packages/transformers/models/aya_vision/__init__.py

@@ -0,0 +1,28 @@
+# 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 _LazyModule
+from ...utils.import_utils import define_import_structure
+
+
+if TYPE_CHECKING:
+    from .configuration_aya_vision import *
+    from .modeling_aya_vision import *
+    from .processing_aya_vision import *
+else:
+    import sys
+
+    _file = globals()["__file__"]
+    sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)

lib/python3.12/site-packages/transformers/models/aya_vision/configuration_aya_vision.py

@@ -0,0 +1,109 @@
+# coding=utf-8
+# Copyright 2025 Cohere 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.
+"""AyaVision model configuration"""
+
+from ...configuration_utils import PretrainedConfig
+from ...utils import logging
+from ..auto import CONFIG_MAPPING, AutoConfig
+
+
+logger = logging.get_logger(__name__)
+
+
+class AyaVisionConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`AyaVisionForConditionalGeneration`]. It is used to instantiate an
+    AyaVision model according to the specified arguments, defining the model architecture. Instantiating a configuration
+    with the defaults will yield a similar configuration to that of AyaVision.
+    e.g. [CohereForAI/aya-vision-8b](https://huggingface.co/CohereForAI/aya-vision-8b)
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        vision_config (`Union[AutoConfig, dict]`,  *optional*, defaults to `CLIPVisionConfig`):
+            The config object or dictionary of the vision backbone.
+        text_config (`Union[AutoConfig, dict]`, *optional*, defaults to `LlamaConfig`):
+            The config object or dictionary of the text backbone.
+        vision_feature_select_strategy (`str`, *optional*, defaults to `"full"`):
+            The feature selection strategy used to select the vision feature from the vision backbone.
+            Can be one of `"default"` or `"full"`. If `"default"`, the CLS token is removed from the vision features.
+            If `"full"`, the full vision features are used.
+        vision_feature_layer (`int`, *optional*, defaults to -1):
+            The index of the layer to select the vision feature.
+        downsample_factor (`int`, *optional*, defaults to 2):
+            The downsample factor to apply to the vision features.
+        adapter_layer_norm_eps (`float`, *optional*, defaults to 1e-06):
+            The epsilon value used for layer normalization in the adapter.
+        image_token_index (`int`, *optional*, defaults to 255036):
+            The image token index to encode the image prompt.
+    """
+
+    model_type = "aya_vision"
+    sub_configs = {"text_config": AutoConfig, "vision_config": AutoConfig}
+
+    def __init__(
+        self,
+        vision_config=None,
+        text_config=None,
+        vision_feature_select_strategy="full",
+        vision_feature_layer=-1,
+        downsample_factor=2,
+        adapter_layer_norm_eps=1e-6,
+        image_token_index=255036,
+        **kwargs,
+    ):
+        self.image_token_index = image_token_index
+        self.downsample_factor = downsample_factor
+        self.adapter_layer_norm_eps = adapter_layer_norm_eps
+        if vision_feature_select_strategy not in ["default", "full"]:
+            raise ValueError(
+                "vision_feature_select_strategy should be one of 'default', 'full'."
+                f"Got: {vision_feature_select_strategy}"
+            )
+
+        self.vision_feature_select_strategy = vision_feature_select_strategy
+        self.vision_feature_layer = vision_feature_layer
+
+        if isinstance(vision_config, dict):
+            vision_config["model_type"] = (
+                vision_config["model_type"] if "model_type" in vision_config else "clip_vision_model"
+            )
+            vision_config = CONFIG_MAPPING[vision_config["model_type"]](**vision_config)
+        elif vision_config is None:
+            vision_config = CONFIG_MAPPING["siglip_vision_model"](
+                hidden_size=1152,
+                intermediate_size=4304,
+                patch_size=14,
+                image_size=384,
+                num_hidden_layers=26,
+                num_attention_heads=14,
+                vision_use_head=False,
+            )
+
+        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 "llama"
+            text_config = CONFIG_MAPPING[text_config["model_type"]](**text_config)
+        elif text_config is None:
+            text_config = CONFIG_MAPPING["cohere2"]()
+
+        self.text_config = text_config
+
+        super().__init__(**kwargs)
+
+
+__all__ = ["AyaVisionConfig"]

lib/python3.12/site-packages/transformers/models/aya_vision/modeling_aya_vision.py

@@ -0,0 +1,555 @@
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+#           This file was automatically generated from src/transformers/models/aya_vision/modular_aya_vision.py.
+#               Do NOT edit this file manually as any edits will be overwritten by the generation of
+#             the file from the modular. If any change should be done, please apply the change to the
+#                          modular_aya_vision.py file directly. One of our CI enforces this.
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+# coding=utf-8
+# Copyright 2025 the Cohere Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+from typing import List, Optional, Tuple, Union
+
+import torch
+from torch import nn
+
+from ...activations import ACT2FN
+from ...generation import GenerationMixin
+from ...modeling_outputs import ModelOutput
+from ...modeling_utils import PreTrainedModel
+from ...utils import (
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    is_torchdynamo_compiling,
+    replace_return_docstrings,
+)
+from ...utils.deprecation import deprecate_kwarg
+from ..auto import AutoModel, AutoModelForCausalLM
+from .configuration_aya_vision import AyaVisionConfig
+
+
+_CONFIG_FOR_DOC = "AyaVisionConfig"
+
+
+class AyaVisionMultiModalProjector(nn.Module):
+    def __init__(self, config: AyaVisionConfig):
+        super().__init__()
+        self.config = config
+        self.downsample_factor = config.downsample_factor
+        self.alignment_intermediate_size = getattr(
+            config, "alignment_intermediate_size", config.text_config.hidden_size
+        )
+        self.layernorm = nn.LayerNorm(
+            config.vision_config.hidden_size * (config.downsample_factor**2), eps=config.adapter_layer_norm_eps
+        )
+
+        self.linear_1 = nn.Linear(
+            config.vision_config.hidden_size * (config.downsample_factor**2),
+            self.alignment_intermediate_size,
+            bias=True,
+        )
+
+        self.act = ACT2FN["silu"]  # SwiGLU uses SiLU activation
+        # For SwiGLU, project down to half size since we split intermediate dim
+        self.linear_2 = nn.Linear(self.alignment_intermediate_size // 2, config.text_config.hidden_size, bias=True)
+
+    def forward(self, image_features):
+        image_features = self.pixel_shuffle(image_features)
+        image_features = self.layernorm(image_features)
+        hidden_states = self.linear_1(image_features)
+
+        # Split along last dimension and apply SwiGLU
+        x, gate = hidden_states.chunk(2, dim=-1)
+        hidden_states = self.act(gate) * x
+
+        hidden_states = self.linear_2(hidden_states)
+        return hidden_states
+
+    def pixel_shuffle(self, image_features):  # B, S, D
+        batch_size, seq_length, feature_dim = image_features.shape
+        height = width = int(seq_length**0.5)
+        image_features = image_features.reshape(image_features.shape[0], width, height, -1)
+        channels = image_features.shape[-1]
+        image_features = image_features.reshape(
+            batch_size, width, int(height / self.downsample_factor), int(channels * self.downsample_factor)
+        )
+        image_features = image_features.permute(0, 2, 1, 3)
+        image_features = image_features.reshape(
+            batch_size, int(height / self.downsample_factor), int(width / self.downsample_factor), -1
+        )
+        image_features = image_features.permute(0, 2, 1, 3)
+        return image_features
+
+
+AYA_VISION_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 ([`AyaVisionConfig`] or [`AyaVisionVisionConfig`]):
+            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 Aya Vision Model outputting raw hidden-states without any specific head on top.",
+    AYA_VISION_START_DOCSTRING,
+)
+class AyaVisionPreTrainedModel(PreTrainedModel):
+    config_class = AyaVisionConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["AyaVisionVisionAttention"]
+    _skip_keys_device_placement = "past_key_values"
+    _supports_cache_class = True
+    _supports_flash_attn_2 = True
+    _supports_sdpa = True
+    _supports_quantized_cache = False
+    _supports_static_cache = False
+
+    def _init_weights(self, module):
+        # important: this ported version of AyaVision 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/AyaVision/tree/main/aya_vision should serve for that purpose
+        std = (
+            self.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_()
+
+
+@dataclass
+class AyaVisionCausalLMOutputWithPast(ModelOutput):
+    """
+    Base class for AyaVision 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 (`torch.FloatTensor`, *optional*):
+            A `torch.FloatTensor` of size (batch_size, num_images, sequence_length, hidden_size)`.
+            image_hidden_states of the model produced by the vision encoder and after projecting the last hidden state.
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    logits: Optional[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[torch.FloatTensor] = None
+
+
+AYA_VISION_INPUTS_DOCSTRING = """
+    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)
+        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 [`GotOcr2ImageProcessor.__call__`] for details. [`AyaVisionProcessor`] uses
+            [`GotOcr2ImageProcessor`] for processing images.
+        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.
+        vision_feature_layer (`Union[int, List[int]], *optional*, defaults to -2`):
+            The index of the layer to select the vision feature. If multiple indices are provided,
+            the vision feature of the corresponding indices will be concatenated to form the
+            vision features.
+        vision_feature_select_strategy (`str`, *optional*, defaults to `"default"`):
+            The feature selection strategy used to select the vision feature from the vision backbone.
+            Can be one of `"default"` or `"full"`.
+        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.
+        cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
+            Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`,
+            this tensor is not affected by padding. It is used to update the cache in the correct position and to infer
+            the complete sequence length.
+"""
+
+
+@add_start_docstrings(
+    """The AyaVision model which consists of a vision backbone and a language model.""",
+    AYA_VISION_START_DOCSTRING,
+)
+class AyaVisionForConditionalGeneration(AyaVisionPreTrainedModel, GenerationMixin):
+    def __init__(self, config: AyaVisionConfig):
+        super().__init__(config)
+        self.vision_tower = AutoModel.from_config(config.vision_config)
+
+        self.multi_modal_projector = AyaVisionMultiModalProjector(config)
+        self.vocab_size = config.text_config.vocab_size
+        self.language_model = AutoModelForCausalLM.from_config(config.text_config)
+
+        if self.language_model._tied_weights_keys is not None:
+            self._tied_weights_keys = [f"language_model.{k}" for k in self.language_model._tied_weights_keys]
+
+        self.pad_token_id = self.config.pad_token_id if self.config.pad_token_id is not None else -1
+
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.language_model.get_input_embeddings()
+
+    def set_input_embeddings(self, value):
+        self.language_model.set_input_embeddings(value)
+
+    def get_output_embeddings(self):
+        return self.language_model.get_output_embeddings()
+
+    def set_output_embeddings(self, new_embeddings):
+        self.language_model.set_output_embeddings(new_embeddings)
+
+    def set_decoder(self, decoder):
+        self.language_model.set_decoder(decoder)
+
+    def get_decoder(self):
+        return self.language_model.get_decoder()
+
+    def get_image_features(
+        self,
+        pixel_values: torch.FloatTensor,
+        vision_feature_layer: Union[int, List[int]],
+        vision_feature_select_strategy: str,
+        **kwargs,
+    ):
+        """
+        Obtains image last hidden states from the vision tower and apply multimodal projection.
+
+        Args:
+            pixel_values (`torch.FloatTensor]` of shape `(batch_size, channels, height, width)`)
+               The tensors corresponding to the input images.
+            vision_feature_layer (`Union[int, List[int]]`):
+                The index of the layer to select the vision feature. If multiple indices are provided,
+                the vision feature of the corresponding indices will be concatenated to form the
+                vision features.
+            vision_feature_select_strategy (`str`):
+                The feature selection strategy used to select the vision feature from the vision backbone.
+                Can be one of `"default"` or `"full"`
+        Returns:
+            image_features (`torch.Tensor`): Image feature tensor of shape `(num_images, image_length, embed_dim)`).
+        """
+        if vision_feature_select_strategy not in ["default", "full"]:
+            raise ValueError(f"Unexpected select feature strategy: {self.config.vision_feature_select_strategy}")
+
+        kwargs = {k: v for k, v in kwargs.items() if v is not None}
+        # this is not memory efficient at all (output_hidden_states=True) will save all the hidden states.
+        image_outputs = self.vision_tower(pixel_values, output_hidden_states=True, **kwargs)
+
+        # If we have one vision feature layer, return the corresponding hidden states,
+        # otherwise, select the hidden states of each feature layer and concatenate them
+        if isinstance(vision_feature_layer, int):
+            selected_image_feature = image_outputs.hidden_states[vision_feature_layer]
+            if vision_feature_select_strategy == "default":
+                selected_image_feature = selected_image_feature[:, 1:]
+        else:
+            hs_pool = [image_outputs.hidden_states[layer_idx] for layer_idx in vision_feature_layer]
+            # For default; crop CLS from each hidden state in the hidden state pool
+            if vision_feature_select_strategy == "default":
+                hs_pool = [hs[:, 1:] for hs in hs_pool]
+            selected_image_feature = torch.cat(hs_pool, dim=-1)
+
+        image_features = self.multi_modal_projector(selected_image_feature)
+        return image_features
+
+    @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+    @add_start_docstrings_to_model_forward(AYA_VISION_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=AyaVisionCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        pixel_values: Optional[torch.FloatTensor] = 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,
+        vision_feature_layer: Optional[Union[int, List[int]]] = None,
+        vision_feature_select_strategy: Optional[str] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        logits_to_keep: Union[int, torch.Tensor] = 0,
+        image_sizes: Optional[torch.Tensor] = None,
+        **lm_kwargs,
+    ) -> Union[Tuple, AyaVisionCausalLMOutputWithPast]:
+        r"""
+            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+                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]`.
+
+            logits_to_keep (`int` or `torch.Tensor`, *optional*):
+                If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all
+                `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
+                token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
+                If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension.
+                This is useful when using packed tensor format (single dimension for batch and sequence length).
+
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from transformers import AutoProcessor, AyaVisionForConditionalGeneration
+        >>> import torch
+
+        >>> torch_device = "cuda:0"
+        >>> processor = AutoProcessor.from_pretrained("CohereForAI/aya-vision-8b", use_fast=True)
+        >>> model = AyaVisionForConditionalGeneration.from_pretrained("CohereForAI/aya-vision-8b", device_map=torch_device)
+
+        >>> messages = [
+        ...     {
+        ...         "role": "user",
+        ...         "content": [
+        ...             {
+        ...                 "type": "image",
+        ...                 "url": "https://pbs.twimg.com/media/Fx7YvfQWYAIp6rZ?format=jpg&name=medium",
+        ...             },
+        ...             {"type": "text", "text": "चित्र में लिखा पाठ क्या कहता है?"},
+        ...         ],
+        ...     }
+        ... ]
+
+        >>> inputs = processor.apply_chat_template(
+        ...     messages, padding=True, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt", device=torch_device
+        ... ).to(model.device)
+
+        >>> gen_tokens = model.generate(**inputs, max_new_tokens=300, do_sample=True, temperature=0.3)
+        >>> processor.tokenizer.decode(gen_tokens[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
+        ```"""
+
+        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
+        vision_feature_layer = (
+            vision_feature_layer if vision_feature_layer is not None else self.config.vision_feature_layer
+        )
+        vision_feature_select_strategy = (
+            vision_feature_select_strategy
+            if vision_feature_select_strategy is not None
+            else self.config.vision_feature_select_strategy
+        )
+
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+        if pixel_values is not None and inputs_embeds is not None:
+            raise ValueError(
+                "You cannot specify both pixel_values and inputs_embeds at the same time, and must specify either one"
+            )
+
+        if inputs_embeds is None:
+            inputs_embeds = self.get_input_embeddings()(input_ids)
+
+        if pixel_values is not None:
+            image_features = self.get_image_features(
+                pixel_values=pixel_values,
+                vision_feature_layer=vision_feature_layer,
+                vision_feature_select_strategy=vision_feature_select_strategy,
+                image_sizes=image_sizes,
+            )
+
+            special_image_mask = (input_ids == self.config.image_token_index).unsqueeze(-1)
+            special_image_mask = special_image_mask.expand_as(inputs_embeds).to(inputs_embeds.device)
+            if not is_torchdynamo_compiling() and inputs_embeds[special_image_mask].numel() != image_features.numel():
+                n_image_tokens = (input_ids == self.config.image_token_index).sum()
+                n_image_features = image_features.shape[0] * image_features.shape[1]
+                raise ValueError(
+                    f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {n_image_features}"
+                )
+            image_features = image_features.to(inputs_embeds.device, inputs_embeds.dtype)
+            inputs_embeds = inputs_embeds.masked_scatter(special_image_mask, image_features)
+
+        outputs = self.language_model(
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            cache_position=cache_position,
+            logits_to_keep=logits_to_keep,
+            **lm_kwargs,
+        )
+
+        logits = outputs[0]
+
+        loss = None
+        if labels is not None:
+            # Shift so that tokens < n predict n
+            if attention_mask is not None:
+                # we use the input attention mask to shift the logits and labels, because it is 2D.
+                # we also crop attn mask in case it is longer, which happens in PrefixTuning with peft
+                shift_attention_mask = attention_mask[:, -(logits.shape[1] - 1) :].to(logits.device)
+                shift_logits = logits[..., :-1, :][shift_attention_mask.to(logits.device) != 0].contiguous()
+                shift_labels = labels[..., 1:][shift_attention_mask.to(labels.device) != 0].contiguous()
+            else:
+                shift_logits = logits[..., :-1, :].contiguous()
+                shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = nn.CrossEntropyLoss()
+            loss = loss_fct(
+                shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1).to(shift_logits.device)
+            )
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return AyaVisionCausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            image_hidden_states=image_features if pixel_values is not None else None,
+        )
+
+    def prepare_inputs_for_generation(
+        self,
+        input_ids,
+        past_key_values=None,
+        inputs_embeds=None,
+        pixel_values=None,
+        attention_mask=None,
+        cache_position=None,
+        logits_to_keep=None,
+        **kwargs,
+    ):
+        # Overwritten -- in specific circumstances we don't want to forward image inputs to the model
+
+        model_inputs = self.language_model.prepare_inputs_for_generation(
+            input_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            attention_mask=attention_mask,
+            cache_position=cache_position,
+            logits_to_keep=logits_to_keep,
+            **kwargs,
+        )
+
+        if cache_position[0] == 0:
+            # If we're in cached decoding stage, pixel values should be None because input ids do not contain special image token anymore
+            # Otherwise we need pixel values to be passed to model
+            model_inputs["pixel_values"] = pixel_values
+
+        return model_inputs
+
+    def tie_weights(self):
+        return self.language_model.tie_weights()
+
+    def resize_token_embeddings(self, new_num_tokens: Optional[int] = None, pad_to_multiple_of=None) -> nn.Embedding:
+        model_embeds = self.language_model.resize_token_embeddings(new_num_tokens, pad_to_multiple_of)
+        # update vocab size
+        self.config.text_config.vocab_size = model_embeds.num_embeddings
+        self.vocab_size = model_embeds.num_embeddings
+        return model_embeds
+
+
+__all__ = ["AyaVisionForConditionalGeneration", "AyaVisionPreTrainedModel"]

lib/python3.12/site-packages/transformers/models/aya_vision/modular_aya_vision.py

@@ -0,0 +1,300 @@
+# coding=utf-8
+# Copyright 2025 the Cohere 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 AyaVision model."""
+
+from typing import List, Optional, Tuple, Union
+
+import torch
+from torch import nn
+
+from transformers.models.llava.modeling_llava import (
+    LlavaCausalLMOutputWithPast,
+    LlavaForConditionalGeneration,
+    LlavaPreTrainedModel,
+)
+
+from ...activations import ACT2FN
+from ...utils import (
+    add_start_docstrings,
+    logging,
+)
+from .configuration_aya_vision import AyaVisionConfig
+
+
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = "AyaVisionConfig"
+
+
+class AyaVisionMultiModalProjector(nn.Module):
+    def __init__(self, config: AyaVisionConfig):
+        super().__init__()
+        self.config = config
+        self.downsample_factor = config.downsample_factor
+        self.alignment_intermediate_size = getattr(
+            config, "alignment_intermediate_size", config.text_config.hidden_size
+        )
+        self.layernorm = nn.LayerNorm(
+            config.vision_config.hidden_size * (config.downsample_factor**2), eps=config.adapter_layer_norm_eps
+        )
+
+        self.linear_1 = nn.Linear(
+            config.vision_config.hidden_size * (config.downsample_factor**2),
+            self.alignment_intermediate_size,
+            bias=True,
+        )
+
+        self.act = ACT2FN["silu"]  # SwiGLU uses SiLU activation
+        # For SwiGLU, project down to half size since we split intermediate dim
+        self.linear_2 = nn.Linear(self.alignment_intermediate_size // 2, config.text_config.hidden_size, bias=True)
+
+    def forward(self, image_features):
+        image_features = self.pixel_shuffle(image_features)
+        image_features = self.layernorm(image_features)
+        hidden_states = self.linear_1(image_features)
+
+        # Split along last dimension and apply SwiGLU
+        x, gate = hidden_states.chunk(2, dim=-1)
+        hidden_states = self.act(gate) * x
+
+        hidden_states = self.linear_2(hidden_states)
+        return hidden_states
+
+    def pixel_shuffle(self, image_features):  # B, S, D
+        batch_size, seq_length, feature_dim = image_features.shape
+        height = width = int(seq_length**0.5)
+        image_features = image_features.reshape(image_features.shape[0], width, height, -1)
+        channels = image_features.shape[-1]
+        image_features = image_features.reshape(
+            batch_size, width, int(height / self.downsample_factor), int(channels * self.downsample_factor)
+        )
+        image_features = image_features.permute(0, 2, 1, 3)
+        image_features = image_features.reshape(
+            batch_size, int(height / self.downsample_factor), int(width / self.downsample_factor), -1
+        )
+        image_features = image_features.permute(0, 2, 1, 3)
+        return image_features
+
+
+AYA_VISION_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 ([`AyaVisionConfig`] or [`AyaVisionVisionConfig`]):
+            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 Aya Vision Model outputting raw hidden-states without any specific head on top.",
+    AYA_VISION_START_DOCSTRING,
+)
+class AyaVisionPreTrainedModel(LlavaPreTrainedModel):
+    _supports_quantized_cache = False
+    _supports_static_cache = False
+
+
+class AyaVisionCausalLMOutputWithPast(LlavaCausalLMOutputWithPast):
+    pass
+
+
+AYA_VISION_INPUTS_DOCSTRING = """
+    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)
+        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 [`GotOcr2ImageProcessor.__call__`] for details. [`AyaVisionProcessor`] uses
+            [`GotOcr2ImageProcessor`] for processing images.
+        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.
+        vision_feature_layer (`Union[int, List[int]], *optional*, defaults to -2`):
+            The index of the layer to select the vision feature. If multiple indices are provided,
+            the vision feature of the corresponding indices will be concatenated to form the
+            vision features.
+        vision_feature_select_strategy (`str`, *optional*, defaults to `"default"`):
+            The feature selection strategy used to select the vision feature from the vision backbone.
+            Can be one of `"default"` or `"full"`.
+        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.
+        cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
+            Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`,
+            this tensor is not affected by padding. It is used to update the cache in the correct position and to infer
+            the complete sequence length.
+"""
+
+
+@add_start_docstrings(
+    """The AyaVision model which consists of a vision backbone and a language model.""",
+    AYA_VISION_START_DOCSTRING,
+)
+class AyaVisionForConditionalGeneration(LlavaForConditionalGeneration):
+    def tie_weights(self):
+        return self.language_model.tie_weights()
+
+    def resize_token_embeddings(self, new_num_tokens: Optional[int] = None, pad_to_multiple_of=None) -> nn.Embedding:
+        model_embeds = self.language_model.resize_token_embeddings(new_num_tokens, pad_to_multiple_of)
+        # update vocab size
+        self.config.text_config.vocab_size = model_embeds.num_embeddings
+        self.vocab_size = model_embeds.num_embeddings
+        return model_embeds
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        pixel_values: Optional[torch.FloatTensor] = 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,
+        vision_feature_layer: Optional[Union[int, List[int]]] = None,
+        vision_feature_select_strategy: Optional[str] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        logits_to_keep: Union[int, torch.Tensor] = 0,
+        image_sizes: Optional[torch.Tensor] = None,
+        **lm_kwargs,
+    ) -> Union[Tuple, AyaVisionCausalLMOutputWithPast]:
+        r"""
+            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+                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]`.
+
+            logits_to_keep (`int` or `torch.Tensor`, *optional*):
+                If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all
+                `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
+                token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
+                If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension.
+                This is useful when using packed tensor format (single dimension for batch and sequence length).
+
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from transformers import AutoProcessor, AyaVisionForConditionalGeneration
+        >>> import torch
+
+        >>> torch_device = "cuda:0"
+        >>> processor = AutoProcessor.from_pretrained("CohereForAI/aya-vision-8b", use_fast=True)
+        >>> model = AyaVisionForConditionalGeneration.from_pretrained("CohereForAI/aya-vision-8b", device_map=torch_device)
+
+        >>> messages = [
+        ...     {
+        ...         "role": "user",
+        ...         "content": [
+        ...             {
+        ...                 "type": "image",
+        ...                 "url": "https://pbs.twimg.com/media/Fx7YvfQWYAIp6rZ?format=jpg&name=medium",
+        ...             },
+        ...             {"type": "text", "text": "चित्र में लिखा पाठ क्या कहता है?"},
+        ...         ],
+        ...     }
+        ... ]
+
+        >>> inputs = processor.apply_chat_template(
+        ...     messages, padding=True, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt", device=torch_device
+        ... ).to(model.device)
+
+        >>> gen_tokens = model.generate(**inputs, max_new_tokens=300, do_sample=True, temperature=0.3)
+        >>> processor.tokenizer.decode(gen_tokens[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
+        ```"""
+        super().forward(
+            input_ids=input_ids,
+            pixel_values=pixel_values,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            vision_feature_layer=vision_feature_layer,
+            vision_feature_select_strategy=vision_feature_select_strategy,
+            labels=labels,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            cache_position=cache_position,
+            logits_to_keep=logits_to_keep,
+            image_sizes=image_sizes,
+            **lm_kwargs,
+        )
+
+
+__all__ = ["AyaVisionForConditionalGeneration", "AyaVisionPreTrainedModel"]

lib/python3.12/site-packages/transformers/models/aya_vision/processing_aya_vision.py

@@ -0,0 +1,252 @@
+# coding=utf-8
+# Copyright 2025 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 List, Optional, Union
+
+from transformers.processing_utils import (
+    ImagesKwargs,
+    ProcessingKwargs,
+    ProcessorMixin,
+    Unpack,
+)
+from transformers.tokenization_utils_base import PreTokenizedInput, TextInput
+
+from ...image_processing_utils import BatchFeature
+from ...image_utils import (
+    ImageInput,
+    make_flat_list_of_images,
+)
+
+
+class AyaVisionImagesKwargs(ImagesKwargs, total=False):
+    crop_to_patches: Optional[bool]
+    min_patches: Optional[int]
+    max_patches: Optional[int]
+
+
+class AyaVisionProcessorKwargs(ProcessingKwargs, total=False):
+    images_kwargs: AyaVisionImagesKwargs
+    _defaults = {
+        "text_kwargs": {
+            "padding_side": "left",
+            "padding": True,
+        },
+        "images_kwargs": {
+            "crop_to_patches": True,
+        },
+    }
+
+
+class AyaVisionProcessor(ProcessorMixin):
+    r"""
+    Constructs a AyaVision processor which wraps a [`AutoImageProcessor`] and
+    [`PretrainedTokenizerFast`] tokenizer into a single processor that inherits both the image processor and
+    tokenizer functionalities. See the [`~AyaVisionProcessor.__call__`] and [`~AyaVisionProcessor.decode`] for more information.
+    Args:
+        image_processor ([`AutoImageProcessor`], *optional*):
+            The image processor is a required input.
+        tokenizer ([`PreTrainedTokenizer`, `PreTrainedTokenizerFast`], *optional*):
+            The tokenizer is a required input.
+        patch_size (`int`, *optional*, defaults to 28):
+            The size of image patches for tokenization.
+        img_size (`int`, *optional*, defaults to 364):
+            The size of the image to be tokenized. This should correspond to the size given to the image processor.
+        image_token (`str`, *optional*, defaults to `"<image>"`):
+            The token to be used to represent an image in the text.
+        downsample_factor (`int`, *optional*, defaults to 1):
+            The factor by which to scale the patch size.
+        start_of_img_token (`str`, *optional*, defaults to `"<|START_OF_IMG|>"`):
+            The token to be used to represent the start of an image in the text.
+        end_of_img_token (`str`, *optional*, defaults to `"<|END_OF_IMG|>"`):
+            The token to be used to represent the end of an image in the text.
+        img_patch_token (`str`, *optional*, defaults to `"<|IMG_PATCH|>"`):
+            The token to be used to represent an image patch in the text.
+        img_line_break_token (`str`, *optional*, defaults to `"<|IMG_LINE_BREAK|>"`):
+            The token to be used to represent a line break in the text.
+        tile_token (`str`, *optional*, defaults to `"TILE"`):
+            The token to be used to represent an image patch in the text.
+        tile_global_token (`str`, *optional*, defaults to `"TILE_GLOBAL"`):
+            The token to be used to represent the cover image in the text.
+        chat_template (`str`, *optional*): A Jinja template which will be used to convert lists of messages
+            in a chat into a tokenizable string.
+    """
+
+    attributes = ["image_processor", "tokenizer"]
+    valid_kwargs = [
+        "chat_template",
+        "image_token",
+        "patch_size",
+        "img_size",
+        "downsample_factor",
+        "start_of_img_token",
+        "end_of_img_token",
+        "img_patch_token",
+        "img_line_break_token",
+        "tile_token",
+        "tile_global_token",
+    ]
+    image_processor_class = "AutoImageProcessor"
+    tokenizer_class = "AutoTokenizer"
+
+    def __init__(
+        self,
+        image_processor=None,
+        tokenizer=None,
+        patch_size: int = 28,
+        img_size: int = 364,
+        image_token="<image>",  # set the default and let users change if they have peculiar special tokens in rare cases
+        downsample_factor: int = 1,
+        start_of_img_token="<|START_OF_IMG|>",
+        end_of_img_token="<|END_OF_IMG|>",
+        img_patch_token="<|IMG_PATCH|>",
+        img_line_break_token="<|IMG_LINE_BREAK|>",
+        tile_token="TILE",
+        tile_global_token="TILE_GLOBAL",
+        chat_template=None,
+        **kwargs,
+    ):
+        super().__init__(image_processor, tokenizer, chat_template=chat_template)
+
+        self.image_token = image_token
+        self.patch_size = patch_size * downsample_factor
+        self.img_size = img_size
+
+        self.start_of_img_token = start_of_img_token
+        self.end_of_img_token = end_of_img_token
+        self.img_patch_token = img_patch_token
+        self.img_line_break_token = img_line_break_token
+        self.tile_token = tile_token
+        self.tile_global_token = tile_global_token
+
+    def _prompt_split_image(self, num_patches):
+        """
+        Create a structured string representation of image tokens
+
+        Args:
+           num_patches: Number of patches in the image
+
+        Returns:
+            String with appropriate image tokens
+        """
+
+        img_patches_per_tile = (self.img_size // self.patch_size) ** 2
+        img_string = f"{self.start_of_img_token}"
+        if num_patches > 1:
+            for idx in range(1, num_patches):
+                img_string += f"{self.tile_token}_{idx}" + f"{self.img_patch_token}" * img_patches_per_tile
+
+        img_string += f"{self.tile_global_token}" + f"{self.img_patch_token}" * img_patches_per_tile
+        img_string += f"{self.end_of_img_token}"
+        return img_string
+
+    def __call__(
+        self,
+        images: Optional[ImageInput] = None,
+        text: Optional[Union[TextInput, PreTokenizedInput, List[TextInput], List[PreTokenizedInput]]] = None,
+        audio=None,
+        videos=None,
+        **kwargs: Unpack[AyaVisionProcessorKwargs],
+    ) -> BatchFeature:
+        """
+        Main method to prepare for the model one or several sequences(s) and image(s). This method forwards the `text`
+        and `kwargs` arguments to PreTrainedTokenizerFast's [`~PreTrainedTokenizerFast.__call__`] to encode the text.
+        To prepare the vision inputs, this method forwards the `images` and `kwargs` arguments to
+        GotOcr2ImageProcessor's [`~GotOcr2ImageProcessor.__call__`] if `images` is not `None`.
+
+        Args:
+            images (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, `List[PIL.Image.Image]`, `List[np.ndarray]`, `List[torch.Tensor]`):
+                The image or batch of images to be prepared. Each image can be a PIL image, NumPy array or PyTorch
+                tensor. Both channels-first and channels-last formats are supported.
+            text (`str`, `List[str]`, `List[List[str]]`):
+                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).
+            return_tensors (`str` or [`~utils.TensorType`], *optional*):
+                If set, will return tensors of a particular framework. Acceptable values are:
+                - `'tf'`: Return TensorFlow `tf.constant` objects.
+                - `'pt'`: Return PyTorch `torch.Tensor` objects.
+                - `'np'`: Return NumPy `np.ndarray` objects.
+                - `'jax'`: Return JAX `jnp.ndarray` objects.
+
+        Returns:
+            [`BatchFeature`]: A [`BatchFeature`] with the following fields:
+
+            - **input_ids** -- List of token ids to be fed to a model. Returned when `text` is not `None`.
+            - **attention_mask** -- List of indices specifying which tokens should be attended to by the model (when
+              `return_attention_mask=True` or if *"attention_mask"* is in `self.model_input_names` and if `text` is not
+              `None`).
+            - **pixel_values** -- Pixel values to be fed to a model. Returned when `images` is not `None`.
+        """
+        if text is None:
+            raise ValueError("You have to specify text.")
+
+        output_kwargs = self._merge_kwargs(
+            AyaVisionProcessorKwargs,
+            tokenizer_init_kwargs=self.tokenizer.init_kwargs,
+            **kwargs,
+        )
+
+        if not isinstance(text, (list, tuple)):
+            text = [text]
+
+        # Process images
+        image_inputs = {}
+        if images is not None:
+            images = make_flat_list_of_images(images)
+            image_inputs = self.image_processor(images=images, **output_kwargs["images_kwargs"])
+            num_patches = image_inputs.pop("num_patches")
+            image_index = 0
+            processed_text = []
+            for prompt in text:
+                new_prompt = prompt
+                while "<image>" in new_prompt:
+                    # Replace the image placeholder with structured image tokens
+                    image_tokens = self._prompt_split_image(num_patches[image_index])
+                    new_prompt = new_prompt.replace("<image>", image_tokens, 1)
+                    image_index += 1
+                processed_text.append(new_prompt)
+
+            if image_index != len(images):
+                raise ValueError("Number of image placeholders in the prompt does not match the number of images.")
+
+            text = processed_text
+
+        text_inputs = self.tokenizer(text, **output_kwargs["text_kwargs"])
+
+        return BatchFeature(data={**text_inputs, **image_inputs})
+
+    def batch_decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to PreTrainedTokenizerFast'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 PreTrainedTokenizerFast'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(tokenizer_input_names) + list(image_processor_input_names)
+
+
+__all__ = ["AyaVisionProcessor"]

lib/python3.12/site-packages/transformers/models/bert_japanese/__init__.py

@@ -0,0 +1,26 @@
+# 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 _LazyModule
+from ...utils.import_utils import define_import_structure
+
+
+if TYPE_CHECKING:
+    from .tokenization_bert_japanese import *
+else:
+    import sys
+
+    _file = globals()["__file__"]
+    sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)

lib/python3.12/site-packages/transformers/models/bert_japanese/tokenization_bert_japanese.py

@@ -0,0 +1,982 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and 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.
+"""Tokenization classes."""
+
+import collections
+import copy
+import os
+import unicodedata
+from typing import Any, Dict, List, Optional, Tuple
+
+from ...tokenization_utils import PreTrainedTokenizer, _is_control, _is_punctuation, _is_whitespace
+from ...utils import is_sentencepiece_available, is_sudachi_projection_available, logging
+
+
+if is_sentencepiece_available():
+    import sentencepiece as spm
+else:
+    spm = None
+
+logger = logging.get_logger(__name__)
+
+VOCAB_FILES_NAMES = {"vocab_file": "vocab.txt", "spm_file": "spiece.model"}
+
+SPIECE_UNDERLINE = "▁"
+
+
+# Copied from transformers.models.bert.tokenization_bert.load_vocab
+def load_vocab(vocab_file):
+    """Loads a vocabulary file into a dictionary."""
+    vocab = collections.OrderedDict()
+    with open(vocab_file, "r", encoding="utf-8") as reader:
+        tokens = reader.readlines()
+    for index, token in enumerate(tokens):
+        token = token.rstrip("\n")
+        vocab[token] = index
+    return vocab
+
+
+# Copied from transformers.models.bert.tokenization_bert.whitespace_tokenize
+def whitespace_tokenize(text):
+    """Runs basic whitespace cleaning and splitting on a piece of text."""
+    text = text.strip()
+    if not text:
+        return []
+    tokens = text.split()
+    return tokens
+
+
+class BertJapaneseTokenizer(PreTrainedTokenizer):
+    r"""
+    Construct a BERT tokenizer for Japanese text.
+
+    This tokenizer inherits from [`PreTrainedTokenizer`] which contains most of the main methods. Users should refer
+    to: this superclass for more information regarding those methods.
+
+    Args:
+        vocab_file (`str`):
+            Path to a one-wordpiece-per-line vocabulary file.
+        spm_file (`str`, *optional*):
+            Path to [SentencePiece](https://github.com/google/sentencepiece) file (generally has a .spm or .model
+            extension) that contains the vocabulary.
+        do_lower_case (`bool`, *optional*, defaults to `True`):
+            Whether to lower case the input. Only has an effect when do_basic_tokenize=True.
+        do_word_tokenize (`bool`, *optional*, defaults to `True`):
+            Whether to do word tokenization.
+        do_subword_tokenize (`bool`, *optional*, defaults to `True`):
+            Whether to do subword tokenization.
+        word_tokenizer_type (`str`, *optional*, defaults to `"basic"`):
+            Type of word tokenizer. Choose from ["basic", "mecab", "sudachi", "jumanpp"].
+        subword_tokenizer_type (`str`, *optional*, defaults to `"wordpiece"`):
+            Type of subword tokenizer. Choose from ["wordpiece", "character", "sentencepiece",].
+        mecab_kwargs (`dict`, *optional*):
+            Dictionary passed to the `MecabTokenizer` constructor.
+        sudachi_kwargs (`dict`, *optional*):
+            Dictionary passed to the `SudachiTokenizer` constructor.
+        jumanpp_kwargs (`dict`, *optional*):
+            Dictionary passed to the `JumanppTokenizer` constructor.
+    """
+
+    vocab_files_names = VOCAB_FILES_NAMES
+
+    def __init__(
+        self,
+        vocab_file,
+        spm_file=None,
+        do_lower_case=False,
+        do_word_tokenize=True,
+        do_subword_tokenize=True,
+        word_tokenizer_type="basic",
+        subword_tokenizer_type="wordpiece",
+        never_split=None,
+        unk_token="[UNK]",
+        sep_token="[SEP]",
+        pad_token="[PAD]",
+        cls_token="[CLS]",
+        mask_token="[MASK]",
+        mecab_kwargs=None,
+        sudachi_kwargs=None,
+        jumanpp_kwargs=None,
+        **kwargs,
+    ):
+        if subword_tokenizer_type == "sentencepiece":
+            if not os.path.isfile(spm_file):
+                raise ValueError(
+                    f"Can't find a vocabulary file at path '{spm_file}'. To load the vocabulary from a Google"
+                    " pretrained model use `tokenizer = AutoTokenizer.from_pretrained(PRETRAINED_MODEL_NAME)`"
+                )
+            self.spm_file = spm_file
+        else:
+            if not os.path.isfile(vocab_file):
+                raise ValueError(
+                    f"Can't find a vocabulary file at path '{vocab_file}'. To load the vocabulary from a Google"
+                    " pretrained model use `tokenizer = AutoTokenizer.from_pretrained(PRETRAINED_MODEL_NAME)`"
+                )
+            self.vocab = load_vocab(vocab_file)
+            self.ids_to_tokens = collections.OrderedDict([(ids, tok) for tok, ids in self.vocab.items()])
+
+        self.do_word_tokenize = do_word_tokenize
+        self.word_tokenizer_type = word_tokenizer_type
+        self.lower_case = do_lower_case
+        self.never_split = never_split
+        self.mecab_kwargs = copy.deepcopy(mecab_kwargs)
+        self.sudachi_kwargs = copy.deepcopy(sudachi_kwargs)
+        self.jumanpp_kwargs = copy.deepcopy(jumanpp_kwargs)
+        if do_word_tokenize:
+            if word_tokenizer_type == "basic":
+                self.word_tokenizer = BasicTokenizer(
+                    do_lower_case=do_lower_case, never_split=never_split, tokenize_chinese_chars=False
+                )
+            elif word_tokenizer_type == "mecab":
+                self.word_tokenizer = MecabTokenizer(
+                    do_lower_case=do_lower_case, never_split=never_split, **(mecab_kwargs or {})
+                )
+            elif word_tokenizer_type == "sudachi":
+                self.word_tokenizer = SudachiTokenizer(
+                    do_lower_case=do_lower_case, never_split=never_split, **(sudachi_kwargs or {})
+                )
+            elif word_tokenizer_type == "jumanpp":
+                self.word_tokenizer = JumanppTokenizer(
+                    do_lower_case=do_lower_case, never_split=never_split, **(jumanpp_kwargs or {})
+                )
+            else:
+                raise ValueError(f"Invalid word_tokenizer_type '{word_tokenizer_type}' is specified.")
+
+        self.do_subword_tokenize = do_subword_tokenize
+        self.subword_tokenizer_type = subword_tokenizer_type
+        if do_subword_tokenize:
+            if subword_tokenizer_type == "wordpiece":
+                self.subword_tokenizer = WordpieceTokenizer(vocab=self.vocab, unk_token=str(unk_token))
+            elif subword_tokenizer_type == "character":
+                self.subword_tokenizer = CharacterTokenizer(vocab=self.vocab, unk_token=str(unk_token))
+            elif subword_tokenizer_type == "sentencepiece":
+                self.subword_tokenizer = SentencepieceTokenizer(vocab=self.spm_file, unk_token=str(unk_token))
+            else:
+                raise ValueError(f"Invalid subword_tokenizer_type '{subword_tokenizer_type}' is specified.")
+        super().__init__(
+            spm_file=spm_file,
+            unk_token=unk_token,
+            sep_token=sep_token,
+            pad_token=pad_token,
+            cls_token=cls_token,
+            mask_token=mask_token,
+            do_lower_case=do_lower_case,
+            do_word_tokenize=do_word_tokenize,
+            do_subword_tokenize=do_subword_tokenize,
+            word_tokenizer_type=word_tokenizer_type,
+            subword_tokenizer_type=subword_tokenizer_type,
+            never_split=never_split,
+            mecab_kwargs=mecab_kwargs,
+            sudachi_kwargs=sudachi_kwargs,
+            jumanpp_kwargs=jumanpp_kwargs,
+            **kwargs,
+        )
+
+    @property
+    def do_lower_case(self):
+        return self.lower_case
+
+    def __getstate__(self):
+        state = dict(self.__dict__)
+        if self.word_tokenizer_type in ["mecab", "sudachi", "jumanpp"]:
+            del state["word_tokenizer"]
+        return state
+
+    def __setstate__(self, state):
+        self.__dict__ = state
+        if self.word_tokenizer_type == "mecab":
+            self.word_tokenizer = MecabTokenizer(
+                do_lower_case=self.do_lower_case, never_split=self.never_split, **(self.mecab_kwargs or {})
+            )
+        elif self.word_tokenizer_type == "sudachi":
+            self.word_tokenizer = SudachiTokenizer(
+                do_lower_case=self.do_lower_case, never_split=self.never_split, **(self.sudachi_kwargs or {})
+            )
+        elif self.word_tokenizer_type == "jumanpp":
+            self.word_tokenizer = JumanppTokenizer(
+                do_lower_case=self.do_lower_case, never_split=self.never_split, **(self.jumanpp_kwargs or {})
+            )
+
+    def _tokenize(self, text):
+        if self.do_word_tokenize:
+            tokens = self.word_tokenizer.tokenize(text, never_split=self.all_special_tokens)
+        else:
+            tokens = [text]
+
+        if self.do_subword_tokenize:
+            split_tokens = [sub_token for token in tokens for sub_token in self.subword_tokenizer.tokenize(token)]
+        else:
+            split_tokens = tokens
+
+        return split_tokens
+
+    @property
+    def vocab_size(self):
+        if self.subword_tokenizer_type == "sentencepiece":
+            return len(self.subword_tokenizer.sp_model)
+        return len(self.vocab)
+
+    def get_vocab(self):
+        if self.subword_tokenizer_type == "sentencepiece":
+            vocab = {self.convert_ids_to_tokens(i): i for i in range(self.vocab_size)}
+            vocab.update(self.added_tokens_encoder)
+            return vocab
+        return dict(self.vocab, **self.added_tokens_encoder)
+
+    def _convert_token_to_id(self, token):
+        """Converts a token (str) in an id using the vocab."""
+        if self.subword_tokenizer_type == "sentencepiece":
+            return self.subword_tokenizer.sp_model.PieceToId(token)
+        return self.vocab.get(token, self.vocab.get(self.unk_token))
+
+    def _convert_id_to_token(self, index):
+        """Converts an index (integer) in a token (str) using the vocab."""
+        if self.subword_tokenizer_type == "sentencepiece":
+            return self.subword_tokenizer.sp_model.IdToPiece(index)
+        return self.ids_to_tokens.get(index, self.unk_token)
+
+    def convert_tokens_to_string(self, tokens):
+        """Converts a sequence of tokens (string) in a single string."""
+        if self.subword_tokenizer_type == "sentencepiece":
+            return self.subword_tokenizer.sp_model.decode(tokens)
+        out_string = " ".join(tokens).replace(" ##", "").strip()
+        return out_string
+
+    # Copied from transformers.models.bert.tokenization_bert.BertTokenizer.build_inputs_with_special_tokens
+    def build_inputs_with_special_tokens(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        Build model inputs from a sequence or a pair of sequence for sequence classification tasks by concatenating and
+        adding special tokens. A BERT sequence has the following format:
+
+        - single sequence: `[CLS] X [SEP]`
+        - pair of sequences: `[CLS] A [SEP] B [SEP]`
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs to which the special tokens will be added.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            `List[int]`: List of [input IDs](../glossary#input-ids) with the appropriate special tokens.
+        """
+        if token_ids_1 is None:
+            return [self.cls_token_id] + token_ids_0 + [self.sep_token_id]
+        cls = [self.cls_token_id]
+        sep = [self.sep_token_id]
+        return cls + token_ids_0 + sep + token_ids_1 + sep
+
+    # Copied from transformers.models.bert.tokenization_bert.BertTokenizer.get_special_tokens_mask
+    def get_special_tokens_mask(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None, already_has_special_tokens: bool = False
+    ) -> List[int]:
+        """
+        Retrieve sequence ids from a token list that has no special tokens added. This method is called when adding
+        special tokens using the tokenizer `prepare_for_model` method.
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+            already_has_special_tokens (`bool`, *optional*, defaults to `False`):
+                Whether or not the token list is already formatted with special tokens for the model.
+
+        Returns:
+            `List[int]`: A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
+        """
+
+        if already_has_special_tokens:
+            return super().get_special_tokens_mask(
+                token_ids_0=token_ids_0, token_ids_1=token_ids_1, already_has_special_tokens=True
+            )
+
+        if token_ids_1 is not None:
+            return [1] + ([0] * len(token_ids_0)) + [1] + ([0] * len(token_ids_1)) + [1]
+        return [1] + ([0] * len(token_ids_0)) + [1]
+
+    # Copied from transformers.models.bert.tokenization_bert.BertTokenizer.create_token_type_ids_from_sequences
+    def create_token_type_ids_from_sequences(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        Create a mask from the two sequences passed to be used in a sequence-pair classification task. A BERT sequence
+        pair mask has the following format:
+
+        ```
+        0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1
+        | first sequence    | second sequence |
+        ```
+
+        If `token_ids_1` is `None`, this method only returns the first portion of the mask (0s).
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            `List[int]`: List of [token type IDs](../glossary#token-type-ids) according to the given sequence(s).
+        """
+        sep = [self.sep_token_id]
+        cls = [self.cls_token_id]
+        if token_ids_1 is None:
+            return len(cls + token_ids_0 + sep) * [0]
+        return len(cls + token_ids_0 + sep) * [0] + len(token_ids_1 + sep) * [1]
+
+    def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
+        if os.path.isdir(save_directory):
+            if self.subword_tokenizer_type == "sentencepiece":
+                vocab_file = os.path.join(
+                    save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["spm_file"]
+                )
+            else:
+                vocab_file = os.path.join(
+                    save_directory,
+                    (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"],
+                )
+        else:
+            vocab_file = (filename_prefix + "-" if filename_prefix else "") + save_directory
+
+        if self.subword_tokenizer_type == "sentencepiece":
+            with open(vocab_file, "wb") as writer:
+                content_spiece_model = self.subword_tokenizer.sp_model.serialized_model_proto()
+                writer.write(content_spiece_model)
+        else:
+            with open(vocab_file, "w", encoding="utf-8") as writer:
+                index = 0
+                for token, token_index in sorted(self.vocab.items(), key=lambda kv: kv[1]):
+                    if index != token_index:
+                        logger.warning(
+                            f"Saving vocabulary to {vocab_file}: vocabulary indices are not consecutive."
+                            " Please check that the vocabulary is not corrupted!"
+                        )
+                        index = token_index
+                    writer.write(token + "\n")
+                    index += 1
+        return (vocab_file,)
+
+
+class MecabTokenizer:
+    """Runs basic tokenization with MeCab morphological parser."""
+
+    def __init__(
+        self,
+        do_lower_case=False,
+        never_split=None,
+        normalize_text=True,
+        mecab_dic: Optional[str] = "unidic_lite",
+        mecab_option: Optional[str] = None,
+    ):
+        """
+        Constructs a MecabTokenizer.
+
+        Args:
+            **do_lower_case**: (*optional*) boolean (default True)
+                Whether to lowercase the input.
+            **never_split**: (*optional*) list of str
+                Kept for backward compatibility purposes. Now implemented directly at the base class level (see
+                [`PreTrainedTokenizer.tokenize`]) List of tokens not to split.
+            **normalize_text**: (*optional*) boolean (default True)
+                Whether to apply unicode normalization to text before tokenization.
+            **mecab_dic**: (*optional*) string (default "ipadic")
+                Name of dictionary to be used for MeCab initialization. If you are using a system-installed dictionary,
+                set this option to `None` and modify *mecab_option*.
+            **mecab_option**: (*optional*) string
+                String passed to MeCab constructor.
+        """
+        self.do_lower_case = do_lower_case
+        self.never_split = never_split if never_split is not None else []
+        self.normalize_text = normalize_text
+
+        try:
+            import fugashi
+        except ModuleNotFoundError as error:
+            raise error.__class__(
+                "You need to install fugashi to use MecabTokenizer. "
+                "See https://pypi.org/project/fugashi/ for installation."
+            )
+
+        mecab_option = mecab_option or ""
+
+        if mecab_dic is not None:
+            if mecab_dic == "ipadic":
+                try:
+                    import ipadic
+                except ModuleNotFoundError as error:
+                    raise error.__class__(
+                        "The ipadic dictionary is not installed. "
+                        "See https://github.com/polm/ipadic-py for installation."
+                    )
+
+                dic_dir = ipadic.DICDIR
+
+            elif mecab_dic == "unidic_lite":
+                try:
+                    import unidic_lite
+                except ModuleNotFoundError as error:
+                    raise error.__class__(
+                        "The unidic_lite dictionary is not installed. "
+                        "See https://github.com/polm/unidic-lite for installation."
+                    )
+
+                dic_dir = unidic_lite.DICDIR
+
+            elif mecab_dic == "unidic":
+                try:
+                    import unidic
+                except ModuleNotFoundError as error:
+                    raise error.__class__(
+                        "The unidic dictionary is not installed. "
+                        "See https://github.com/polm/unidic-py for installation."
+                    )
+
+                dic_dir = unidic.DICDIR
+                if not os.path.isdir(dic_dir):
+                    raise RuntimeError(
+                        "The unidic dictionary itself is not found. "
+                        "See https://github.com/polm/unidic-py for installation."
+                    )
+
+            else:
+                raise ValueError("Invalid mecab_dic is specified.")
+
+            mecabrc = os.path.join(dic_dir, "mecabrc")
+            mecab_option = f'-d "{dic_dir}" -r "{mecabrc}" ' + mecab_option
+
+        self.mecab = fugashi.GenericTagger(mecab_option)
+
+    def tokenize(self, text, never_split=None, **kwargs):
+        """Tokenizes a piece of text."""
+        if self.normalize_text:
+            text = unicodedata.normalize("NFKC", text)
+
+        never_split = self.never_split + (never_split if never_split is not None else [])
+        tokens = []
+
+        for word in self.mecab(text):
+            token = word.surface
+
+            if self.do_lower_case and token not in never_split:
+                token = token.lower()
+
+            tokens.append(token)
+
+        return tokens
+
+
+class SudachiTokenizer:
+    """Runs basic tokenization with Sudachi morphological parser."""
+
+    def __init__(
+        self,
+        do_lower_case=False,
+        never_split=None,
+        normalize_text=True,
+        trim_whitespace=False,
+        sudachi_split_mode="A",
+        sudachi_config_path=None,
+        sudachi_resource_dir=None,
+        sudachi_dict_type="core",
+        sudachi_projection=None,
+    ):
+        """
+        Constructs a SudachiTokenizer.
+
+        Args:
+            **do_lower_case**: (*optional*) boolean (default True)
+                Whether to lowercase the input.
+            **never_split**: (*optional*) list of str
+                Kept for backward compatibility purposes. Now implemented directly at the base class level (see
+                [`PreTrainedTokenizer.tokenize`]) List of tokens not to split.
+            **normalize_text**: (*optional*) boolean (default True)
+                Whether to apply unicode normalization to text before tokenization.
+            **trim_whitespace**: (*optional*) boolean (default False)
+                Whether to trim all whitespace, tab, newline from tokens.
+            **sudachi_split_mode**: (*optional*) string
+                Split mode of sudachi, choose from `["A", "B", "C"]`.
+            **sudachi_config_path**: (*optional*) string
+            **sudachi_resource_dir**: (*optional*) string
+            **sudachi_dict_type**: (*optional*) string
+                dict type of sudachi, choose from `["small", "core", "full"]`.
+            **sudachi_projection**: (*optional*) string
+                Word projection mode of sudachi, choose from `["surface", "normalized", "reading", "dictionary", "dictionary_and_surface", "normalized_and_surface", "normalized_nouns"]`.
+        """
+
+        self.do_lower_case = do_lower_case
+        self.never_split = never_split if never_split is not None else []
+        self.normalize_text = normalize_text
+        self.trim_whitespace = trim_whitespace
+
+        try:
+            from sudachipy import dictionary, tokenizer
+        except ImportError:
+            raise ImportError(
+                "You need to install sudachipy to use SudachiTokenizer. "
+                "See https://github.com/WorksApplications/SudachiPy for installation."
+            )
+
+        if sudachi_split_mode == "A":
+            self.split_mode = tokenizer.Tokenizer.SplitMode.A
+        elif sudachi_split_mode == "B":
+            self.split_mode = tokenizer.Tokenizer.SplitMode.B
+        elif sudachi_split_mode == "C":
+            self.split_mode = tokenizer.Tokenizer.SplitMode.C
+        else:
+            raise ValueError("Invalid sudachi_split_mode is specified.")
+
+        self.projection = sudachi_projection
+
+        sudachi_dictionary = dictionary.Dictionary(
+            config_path=sudachi_config_path, resource_dir=sudachi_resource_dir, dict=sudachi_dict_type
+        )
+        if is_sudachi_projection_available():
+            self.sudachi = sudachi_dictionary.create(self.split_mode, projection=self.projection)
+        elif self.projection is not None:
+            raise ImportError("You need to install sudachipy>=0.6.8 to specify `projection` field in sudachi_kwargs.")
+        else:
+            self.sudachi = sudachi_dictionary.create(self.split_mode)
+
+    def tokenize(self, text, never_split=None, **kwargs):
+        """Tokenizes a piece of text."""
+        if self.normalize_text:
+            text = unicodedata.normalize("NFKC", text)
+
+        never_split = self.never_split + (never_split if never_split is not None else [])
+        tokens = []
+
+        for word in self.sudachi.tokenize(text):
+            token = word.surface()
+
+            if self.do_lower_case and token not in never_split:
+                token = token.lower()
+
+            if self.trim_whitespace:
+                if token.strip() == "":
+                    continue
+                else:
+                    token = token.strip()
+
+            tokens.append(token)
+
+        return tokens
+
+
+class JumanppTokenizer:
+    """Runs basic tokenization with jumanpp morphological parser."""
+
+    def __init__(
+        self,
+        do_lower_case=False,
+        never_split=None,
+        normalize_text=True,
+        trim_whitespace=False,
+    ):
+        """
+        Constructs a JumanppTokenizer.
+
+        Args:
+            **do_lower_case**: (*optional*) boolean (default True)
+                Whether to lowercase the input.
+            **never_split**: (*optional*) list of str
+                Kept for backward compatibility purposes. Now implemented directly at the base class level (see
+                [`PreTrainedTokenizer.tokenize`]) List of tokens not to split.
+            **normalize_text**: (*optional*) boolean (default True)
+                Whether to apply unicode normalization to text before tokenization.
+            **trim_whitespace**: (*optional*) boolean (default False)
+                Whether to trim all whitespace, tab, newline from tokens.
+        """
+
+        self.do_lower_case = do_lower_case
+        self.never_split = never_split if never_split is not None else []
+        self.normalize_text = normalize_text
+        self.trim_whitespace = trim_whitespace
+
+        try:
+            import rhoknp
+        except ImportError:
+            raise ImportError(
+                "You need to install rhoknp to use JumanppTokenizer. "
+                "See https://github.com/ku-nlp/rhoknp for installation."
+            )
+
+        self.juman = rhoknp.Jumanpp()
+
+    def tokenize(self, text, never_split=None, **kwargs):
+        """Tokenizes a piece of text."""
+        if self.normalize_text:
+            text = unicodedata.normalize("NFKC", text)
+
+        text = text.strip()
+
+        never_split = self.never_split + (never_split if never_split is not None else [])
+        tokens = []
+
+        for mrph in self.juman.apply_to_sentence(text).morphemes:
+            token = mrph.text
+
+            if self.do_lower_case and token not in never_split:
+                token = token.lower()
+
+            if self.trim_whitespace:
+                if token.strip() == "":
+                    continue
+                else:
+                    token = token.strip()
+
+            tokens.append(token)
+
+        return tokens
+
+
+class CharacterTokenizer:
+    """Runs Character tokenization."""
+
+    def __init__(self, vocab, unk_token, normalize_text=True):
+        """
+        Constructs a CharacterTokenizer.
+
+        Args:
+            **vocab**:
+                Vocabulary object.
+            **unk_token**: str
+                A special symbol for out-of-vocabulary token.
+            **normalize_text**: (`optional`) boolean (default True)
+                Whether to apply unicode normalization to text before tokenization.
+        """
+        self.vocab = vocab
+        self.unk_token = unk_token
+        self.normalize_text = normalize_text
+
+    def tokenize(self, text):
+        """
+        Tokenizes a piece of text into characters.
+
+        For example, `input = "apple""` wil return as output `["a", "p", "p", "l", "e"]`.
+
+        Args:
+            text: A single token or whitespace separated tokens.
+                This should have already been passed through *BasicTokenizer*.
+
+        Returns:
+            A list of characters.
+        """
+        if self.normalize_text:
+            text = unicodedata.normalize("NFKC", text)
+
+        output_tokens = []
+        for char in text:
+            if char not in self.vocab:
+                output_tokens.append(self.unk_token)
+                continue
+
+            output_tokens.append(char)
+
+        return output_tokens
+
+
+# Copied from transformers.models.bert.tokenization_bert.BasicTokenizer
+class BasicTokenizer:
+    """
+    Constructs a BasicTokenizer that will run basic tokenization (punctuation splitting, lower casing, etc.).
+
+    Args:
+        do_lower_case (`bool`, *optional*, defaults to `True`):
+            Whether or not to lowercase the input when tokenizing.
+        never_split (`Iterable`, *optional*):
+            Collection of tokens which will never be split during tokenization. Only has an effect when
+            `do_basic_tokenize=True`
+        tokenize_chinese_chars (`bool`, *optional*, defaults to `True`):
+            Whether or not to tokenize Chinese characters.
+
+            This should likely be deactivated for Japanese (see this
+            [issue](https://github.com/huggingface/transformers/issues/328)).
+        strip_accents (`bool`, *optional*):
+            Whether or not to strip all accents. If this option is not specified, then it will be determined by the
+            value for `lowercase` (as in the original BERT).
+        do_split_on_punc (`bool`, *optional*, defaults to `True`):
+            In some instances we want to skip the basic punctuation splitting so that later tokenization can capture
+            the full context of the words, such as contractions.
+    """
+
+    def __init__(
+        self,
+        do_lower_case=True,
+        never_split=None,
+        tokenize_chinese_chars=True,
+        strip_accents=None,
+        do_split_on_punc=True,
+    ):
+        if never_split is None:
+            never_split = []
+        self.do_lower_case = do_lower_case
+        self.never_split = set(never_split)
+        self.tokenize_chinese_chars = tokenize_chinese_chars
+        self.strip_accents = strip_accents
+        self.do_split_on_punc = do_split_on_punc
+
+    def tokenize(self, text, never_split=None):
+        """
+        Basic Tokenization of a piece of text. For sub-word tokenization, see WordPieceTokenizer.
+
+        Args:
+            never_split (`List[str]`, *optional*)
+                Kept for backward compatibility purposes. Now implemented directly at the base class level (see
+                [`PreTrainedTokenizer.tokenize`]) List of token not to split.
+        """
+        # union() returns a new set by concatenating the two sets.
+        never_split = self.never_split.union(set(never_split)) if never_split else self.never_split
+        text = self._clean_text(text)
+
+        # This was added on November 1st, 2018 for the multilingual and Chinese
+        # models. This is also applied to the English models now, but it doesn't
+        # matter since the English models were not trained on any Chinese data
+        # and generally don't have any Chinese data in them (there are Chinese
+        # characters in the vocabulary because Wikipedia does have some Chinese
+        # words in the English Wikipedia.).
+        if self.tokenize_chinese_chars:
+            text = self._tokenize_chinese_chars(text)
+        # prevents treating the same character with different unicode codepoints as different characters
+        unicode_normalized_text = unicodedata.normalize("NFC", text)
+        orig_tokens = whitespace_tokenize(unicode_normalized_text)
+        split_tokens = []
+        for token in orig_tokens:
+            if token not in never_split:
+                if self.do_lower_case:
+                    token = token.lower()
+                    if self.strip_accents is not False:
+                        token = self._run_strip_accents(token)
+                elif self.strip_accents:
+                    token = self._run_strip_accents(token)
+            split_tokens.extend(self._run_split_on_punc(token, never_split))
+
+        output_tokens = whitespace_tokenize(" ".join(split_tokens))
+        return output_tokens
+
+    def _run_strip_accents(self, text):
+        """Strips accents from a piece of text."""
+        text = unicodedata.normalize("NFD", text)
+        output = []
+        for char in text:
+            cat = unicodedata.category(char)
+            if cat == "Mn":
+                continue
+            output.append(char)
+        return "".join(output)
+
+    def _run_split_on_punc(self, text, never_split=None):
+        """Splits punctuation on a piece of text."""
+        if not self.do_split_on_punc or (never_split is not None and text in never_split):
+            return [text]
+        chars = list(text)
+        i = 0
+        start_new_word = True
+        output = []
+        while i < len(chars):
+            char = chars[i]
+            if _is_punctuation(char):
+                output.append([char])
+                start_new_word = True
+            else:
+                if start_new_word:
+                    output.append([])
+                start_new_word = False
+                output[-1].append(char)
+            i += 1
+
+        return ["".join(x) for x in output]
+
+    def _tokenize_chinese_chars(self, text):
+        """Adds whitespace around any CJK character."""
+        output = []
+        for char in text:
+            cp = ord(char)
+            if self._is_chinese_char(cp):
+                output.append(" ")
+                output.append(char)
+                output.append(" ")
+            else:
+                output.append(char)
+        return "".join(output)
+
+    def _is_chinese_char(self, cp):
+        """Checks whether CP is the codepoint of a CJK character."""
+        # This defines a "chinese character" as anything in the CJK Unicode block:
+        #   https://en.wikipedia.org/wiki/CJK_Unified_Ideographs_(Unicode_block)
+        #
+        # Note that the CJK Unicode block is NOT all Japanese and Korean characters,
+        # despite its name. The modern Korean Hangul alphabet is a different block,
+        # as is Japanese Hiragana and Katakana. Those alphabets are used to write
+        # space-separated words, so they are not treated specially and handled
+        # like the all of the other languages.
+        if (
+            (cp >= 0x4E00 and cp <= 0x9FFF)
+            or (cp >= 0x3400 and cp <= 0x4DBF)  #
+            or (cp >= 0x20000 and cp <= 0x2A6DF)  #
+            or (cp >= 0x2A700 and cp <= 0x2B73F)  #
+            or (cp >= 0x2B740 and cp <= 0x2B81F)  #
+            or (cp >= 0x2B820 and cp <= 0x2CEAF)  #
+            or (cp >= 0xF900 and cp <= 0xFAFF)
+            or (cp >= 0x2F800 and cp <= 0x2FA1F)  #
+        ):  #
+            return True
+
+        return False
+
+    def _clean_text(self, text):
+        """Performs invalid character removal and whitespace cleanup on text."""
+        output = []
+        for char in text:
+            cp = ord(char)
+            if cp == 0 or cp == 0xFFFD or _is_control(char):
+                continue
+            if _is_whitespace(char):
+                output.append(" ")
+            else:
+                output.append(char)
+        return "".join(output)
+
+
+# Copied from transformers.models.bert.tokenization_bert.WordpieceTokenizer
+class WordpieceTokenizer:
+    """Runs WordPiece tokenization."""
+
+    def __init__(self, vocab, unk_token, max_input_chars_per_word=100):
+        self.vocab = vocab
+        self.unk_token = unk_token
+        self.max_input_chars_per_word = max_input_chars_per_word
+
+    def tokenize(self, text):
+        """
+        Tokenizes a piece of text into its word pieces. This uses a greedy longest-match-first algorithm to perform
+        tokenization using the given vocabulary.
+
+        For example, `input = "unaffable"` wil return as output `["un", "##aff", "##able"]`.
+
+        Args:
+            text: A single token or whitespace separated tokens. This should have
+                already been passed through *BasicTokenizer*.
+
+        Returns:
+            A list of wordpiece tokens.
+        """
+
+        output_tokens = []
+        for token in whitespace_tokenize(text):
+            chars = list(token)
+            if len(chars) > self.max_input_chars_per_word:
+                output_tokens.append(self.unk_token)
+                continue
+
+            is_bad = False
+            start = 0
+            sub_tokens = []
+            while start < len(chars):
+                end = len(chars)
+                cur_substr = None
+                while start < end:
+                    substr = "".join(chars[start:end])
+                    if start > 0:
+                        substr = "##" + substr
+                    if substr in self.vocab:
+                        cur_substr = substr
+                        break
+                    end -= 1
+                if cur_substr is None:
+                    is_bad = True
+                    break
+                sub_tokens.append(cur_substr)
+                start = end
+
+            if is_bad:
+                output_tokens.append(self.unk_token)
+            else:
+                output_tokens.extend(sub_tokens)
+        return output_tokens
+
+
+class SentencepieceTokenizer:
+    """
+    Runs sentencepiece tokenization. Based on transformers.models.albert.tokenization_albert.AlbertTokenizer.
+    """
+
+    def __init__(
+        self,
+        vocab,
+        unk_token,
+        do_lower_case=False,
+        remove_space=True,
+        keep_accents=True,
+        sp_model_kwargs: Optional[Dict[str, Any]] = None,
+    ):
+        self.vocab = vocab
+        self.unk_token = unk_token
+        self.do_lower_case = do_lower_case
+        self.remove_space = remove_space
+        self.keep_accents = keep_accents
+
+        self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs
+        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
+        self.sp_model.Load(self.vocab)
+
+    def preprocess_text(self, inputs):
+        if self.remove_space:
+            outputs = " ".join(inputs.strip().split())
+        else:
+            outputs = inputs
+        outputs = outputs.replace("``", '"').replace("''", '"')
+
+        if not self.keep_accents:
+            outputs = unicodedata.normalize("NFKD", outputs)
+            outputs = "".join([c for c in outputs if not unicodedata.combining(c)])
+        if self.do_lower_case:
+            outputs = outputs.lower()
+
+        return outputs
+
+    def tokenize(self, text):
+        """
+        Tokenizes text by sentencepiece. Based on [SentencePiece](https://github.com/google/sentencepiece).
+        Tokenization needs the given vocabulary.
+
+        Args:
+            text: A string needs to be tokenized.
+
+        Returns:
+            A list of sentencepiece tokens.
+        """
+        text = self.preprocess_text(text)
+        pieces = self.sp_model.encode(text, out_type=str)
+        new_pieces = []
+        for piece in pieces:
+            if len(piece) > 1 and piece[-1] == str(",") and piece[-2].isdigit():
+                cur_pieces = self.sp_model.EncodeAsPieces(piece[:-1].replace(SPIECE_UNDERLINE, ""))
+                if piece[0] != SPIECE_UNDERLINE and cur_pieces[0][0] == SPIECE_UNDERLINE:
+                    if len(cur_pieces[0]) == 1:
+                        cur_pieces = cur_pieces[1:]
+                    else:
+                        cur_pieces[0] = cur_pieces[0][1:]
+                cur_pieces.append(piece[-1])
+                new_pieces.extend(cur_pieces)
+            else:
+                new_pieces.append(piece)
+
+        return new_pieces
+
+
+__all__ = ["BertJapaneseTokenizer", "CharacterTokenizer", "MecabTokenizer"]

lib/python3.12/site-packages/transformers/models/bigbird_pegasus/__init__.py

@@ -0,0 +1,27 @@
+# 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 _LazyModule
+from ...utils.import_utils import define_import_structure
+
+
+if TYPE_CHECKING:
+    from .configuration_bigbird_pegasus import *
+    from .modeling_bigbird_pegasus import *
+else:
+    import sys
+
+    _file = globals()["__file__"]
+    sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)

lib/python3.12/site-packages/transformers/models/bigbird_pegasus/configuration_bigbird_pegasus.py

@@ -0,0 +1,412 @@
+# coding=utf-8
+# Copyright Google Research and 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.
+"""BigBirdPegasus model configuration"""
+
+from collections import OrderedDict
+from typing import Any, Mapping, Optional
+
+from ... import PreTrainedTokenizer
+from ...configuration_utils import PretrainedConfig
+from ...onnx import OnnxConfig, OnnxConfigWithPast, OnnxSeq2SeqConfigWithPast
+from ...onnx.utils import compute_effective_axis_dimension
+from ...utils import TensorType, is_torch_available, logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class BigBirdPegasusConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`BigBirdPegasusModel`]. It is used to instantiate
+    an BigBirdPegasus 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 BigBirdPegasus
+    [google/bigbird-pegasus-large-arxiv](https://huggingface.co/google/bigbird-pegasus-large-arxiv) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 96103):
+            Vocabulary size of the BigBirdPegasus model. Defines the number of different tokens that can be represented
+            by the `inputs_ids` passed when calling [`BigBirdPegasusModel`].
+        d_model (`int`, *optional*, defaults to 1024):
+            Dimension of the layers and the pooler layer.
+        encoder_layers (`int`, *optional*, defaults to 16):
+            Number of encoder layers.
+        decoder_layers (`int`, *optional*, defaults to 16):
+            Number of decoder layers.
+        encoder_attention_heads (`int`, *optional*, defaults to 16):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        decoder_attention_heads (`int`, *optional*, defaults to 16):
+            Number of attention heads for each attention layer in the Transformer decoder.
+        decoder_ffn_dim (`int`, *optional*, defaults to 4096):
+            Dimension of the "intermediate" (often named feed-forward) layer in decoder.
+        encoder_ffn_dim (`int`, *optional*, defaults to 4096):
+            Dimension of the "intermediate" (often named feed-forward) layer in decoder.
+        activation_function (`str` or `function`, *optional*, defaults to `"gelu_new"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"silu"` and `"gelu_new"` are supported.
+        dropout (`float`, *optional*, defaults to 0.1):
+            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        activation_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for activations inside the fully connected layer.
+        classifier_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for classifier.
+        max_position_embeddings (`int`, *optional*, defaults to 4096):
+            The maximum sequence length that this model might ever be used with. Typically set this to something large
+            just in case (e.g., 1024 or 2048 or 4096).
+        init_std (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        encoder_layerdrop (`float`, *optional*, defaults to 0.0):
+            The LayerDrop probability for the encoder. See the [LayerDrop paper](see https://arxiv.org/abs/1909.11556)
+            for more details.
+        decoder_layerdrop (`float`, *optional*, defaults to 0.0):
+            The LayerDrop probability for the decoder. See the [LayerDrop paper](see https://arxiv.org/abs/1909.11556)
+            for more details.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models).
+        attention_type (`str`, *optional*, defaults to `"block_sparse"`)
+            Whether to use block sparse attention (with n complexity) as introduced in paper or original attention
+            layer (with n^2 complexity) in encoder. Possible values are `"original_full"` and `"block_sparse"`.
+        use_bias (`bool`, *optional*, defaults to `False`)
+            Whether to use bias in query, key, value.
+        block_size (`int`, *optional*, defaults to 64)
+            Size of each block. Useful only when `attention_type == "block_sparse"`.
+        num_random_blocks (`int`, *optional*, defaults to 3)
+            Each query is going to attend these many number of random blocks. Useful only when `attention_type ==
+            "block_sparse"`.
+        scale_embeddings (`bool`, *optional*, defaults to `True`)
+            Whether to rescale embeddings with (hidden_size ** 0.5).
+
+    Example:
+
+    ```python
+    >>> from transformers import BigBirdPegasusConfig, BigBirdPegasusModel
+
+    >>> # Initializing a BigBirdPegasus bigbird-pegasus-base style configuration
+    >>> configuration = BigBirdPegasusConfig()
+
+    >>> # Initializing a model (with random weights) from the bigbird-pegasus-base style configuration
+    >>> model = BigBirdPegasusModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "bigbird_pegasus"
+    keys_to_ignore_at_inference = ["past_key_values"]
+    attribute_map = {
+        "num_attention_heads": "encoder_attention_heads",
+        "hidden_size": "d_model",
+        "attention_probs_dropout_prob": "attention_dropout",
+    }
+
+    def __init__(
+        self,
+        vocab_size=96103,
+        max_position_embeddings=4096,
+        encoder_layers=16,
+        encoder_ffn_dim=4096,
+        encoder_attention_heads=16,
+        decoder_layers=16,
+        decoder_ffn_dim=4096,
+        decoder_attention_heads=16,
+        encoder_layerdrop=0.0,
+        decoder_layerdrop=0.0,
+        use_cache=True,
+        is_encoder_decoder=True,
+        activation_function="gelu_new",
+        d_model=1024,
+        dropout=0.1,
+        attention_dropout=0.0,
+        activation_dropout=0.0,
+        init_std=0.02,
+        decoder_start_token_id=2,
+        classifier_dropout=0.0,
+        scale_embedding=True,
+        pad_token_id=0,
+        bos_token_id=2,
+        eos_token_id=1,
+        attention_type="block_sparse",  # only for encoder
+        block_size=64,
+        num_random_blocks=3,
+        use_bias=False,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.d_model = d_model
+        self.encoder_ffn_dim = encoder_ffn_dim
+        self.encoder_layers = encoder_layers
+        self.encoder_attention_heads = encoder_attention_heads
+        self.decoder_ffn_dim = decoder_ffn_dim
+        self.decoder_layers = decoder_layers
+        self.decoder_attention_heads = decoder_attention_heads
+        self.dropout = dropout
+        self.attention_dropout = attention_dropout
+        self.activation_dropout = activation_dropout
+        self.activation_function = activation_function
+        self.init_std = init_std
+        self.encoder_layerdrop = encoder_layerdrop
+        self.decoder_layerdrop = decoder_layerdrop
+        self.classifier_dropout = classifier_dropout
+        self.use_cache = use_cache
+        self.num_hidden_layers = encoder_layers
+        self.scale_embedding = scale_embedding  # scale factor will be sqrt(d_model) if True
+
+        # extra config
+        self.attention_type = attention_type
+        self.block_size = block_size
+        self.num_random_blocks = num_random_blocks
+        self.use_bias = use_bias
+
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            is_encoder_decoder=is_encoder_decoder,
+            decoder_start_token_id=decoder_start_token_id,
+            **kwargs,
+        )
+
+
+# Copied from transformers.models.bart.configuration_bart.BartOnnxConfig
+class BigBirdPegasusOnnxConfig(OnnxSeq2SeqConfigWithPast):
+    @property
+    def inputs(self) -> Mapping[str, Mapping[int, str]]:
+        if self.task in ["default", "seq2seq-lm"]:
+            common_inputs = OrderedDict(
+                [
+                    ("input_ids", {0: "batch", 1: "encoder_sequence"}),
+                    ("attention_mask", {0: "batch", 1: "encoder_sequence"}),
+                ]
+            )
+
+            if self.use_past:
+                common_inputs["decoder_input_ids"] = {0: "batch"}
+                common_inputs["decoder_attention_mask"] = {0: "batch", 1: "past_decoder_sequence + sequence"}
+            else:
+                common_inputs["decoder_input_ids"] = {0: "batch", 1: "decoder_sequence"}
+                common_inputs["decoder_attention_mask"] = {0: "batch", 1: "decoder_sequence"}
+
+            if self.use_past:
+                self.fill_with_past_key_values_(common_inputs, direction="inputs")
+        elif self.task == "causal-lm":
+            # TODO: figure this case out.
+            common_inputs = OrderedDict(
+                [
+                    ("input_ids", {0: "batch", 1: "encoder_sequence"}),
+                    ("attention_mask", {0: "batch", 1: "encoder_sequence"}),
+                ]
+            )
+            if self.use_past:
+                num_encoder_layers, _ = self.num_layers
+                for i in range(num_encoder_layers):
+                    common_inputs[f"past_key_values.{i}.key"] = {0: "batch", 2: "past_sequence + sequence"}
+                    common_inputs[f"past_key_values.{i}.value"] = {0: "batch", 2: "past_sequence + sequence"}
+        else:
+            common_inputs = OrderedDict(
+                [
+                    ("input_ids", {0: "batch", 1: "encoder_sequence"}),
+                    ("attention_mask", {0: "batch", 1: "encoder_sequence"}),
+                    ("decoder_input_ids", {0: "batch", 1: "decoder_sequence"}),
+                    ("decoder_attention_mask", {0: "batch", 1: "decoder_sequence"}),
+                ]
+            )
+
+        return common_inputs
+
+    @property
+    def outputs(self) -> Mapping[str, Mapping[int, str]]:
+        if self.task in ["default", "seq2seq-lm"]:
+            common_outputs = super().outputs
+        else:
+            common_outputs = super(OnnxConfigWithPast, self).outputs
+            if self.use_past:
+                num_encoder_layers, _ = self.num_layers
+                for i in range(num_encoder_layers):
+                    common_outputs[f"present.{i}.key"] = {0: "batch", 2: "past_sequence + sequence"}
+                    common_outputs[f"present.{i}.value"] = {0: "batch", 2: "past_sequence + sequence"}
+        return common_outputs
+
+    def _generate_dummy_inputs_for_default_and_seq2seq_lm(
+        self,
+        tokenizer: PreTrainedTokenizer,
+        batch_size: int = -1,
+        seq_length: int = -1,
+        is_pair: bool = False,
+        framework: Optional[TensorType] = None,
+    ) -> Mapping[str, Any]:
+        encoder_inputs = self._generate_dummy_inputs_for_sequence_classification_and_question_answering(
+            tokenizer, batch_size, seq_length, is_pair, framework
+        )
+
+        # Generate decoder inputs
+        decoder_seq_length = seq_length if not self.use_past else 1
+        decoder_inputs = self._generate_dummy_inputs_for_sequence_classification_and_question_answering(
+            tokenizer, batch_size, decoder_seq_length, is_pair, framework
+        )
+        decoder_inputs = {f"decoder_{name}": tensor for name, tensor in decoder_inputs.items()}
+        common_inputs = dict(**encoder_inputs, **decoder_inputs)
+
+        if self.use_past:
+            if not is_torch_available():
+                raise ValueError("Cannot generate dummy past_keys inputs without PyTorch installed.")
+            else:
+                import torch
+            batch, encoder_seq_length = common_inputs["input_ids"].shape
+            decoder_seq_length = common_inputs["decoder_input_ids"].shape[1]
+            num_encoder_attention_heads, num_decoder_attention_heads = self.num_attention_heads
+            encoder_shape = (
+                batch,
+                num_encoder_attention_heads,
+                encoder_seq_length,
+                self._config.hidden_size // num_encoder_attention_heads,
+            )
+            decoder_past_length = decoder_seq_length + 3
+            decoder_shape = (
+                batch,
+                num_decoder_attention_heads,
+                decoder_past_length,
+                self._config.hidden_size // num_decoder_attention_heads,
+            )
+
+            common_inputs["decoder_attention_mask"] = torch.cat(
+                [common_inputs["decoder_attention_mask"], torch.ones(batch, decoder_past_length)], dim=1
+            )
+
+            common_inputs["past_key_values"] = []
+            # If the number of encoder and decoder layers are present in the model configuration, both are considered
+            num_encoder_layers, num_decoder_layers = self.num_layers
+            min_num_layers = min(num_encoder_layers, num_decoder_layers)
+            max_num_layers = max(num_encoder_layers, num_decoder_layers) - min_num_layers
+            remaining_side_name = "encoder" if num_encoder_layers > num_decoder_layers else "decoder"
+
+            for _ in range(min_num_layers):
+                common_inputs["past_key_values"].append(
+                    (
+                        torch.zeros(decoder_shape),
+                        torch.zeros(decoder_shape),
+                        torch.zeros(encoder_shape),
+                        torch.zeros(encoder_shape),
+                    )
+                )
+            # TODO: test this.
+            shape = encoder_shape if remaining_side_name == "encoder" else decoder_shape
+            for _ in range(min_num_layers, max_num_layers):
+                common_inputs["past_key_values"].append((torch.zeros(shape), torch.zeros(shape)))
+        return common_inputs
+
+    def _generate_dummy_inputs_for_causal_lm(
+        self,
+        tokenizer: PreTrainedTokenizer,
+        batch_size: int = -1,
+        seq_length: int = -1,
+        is_pair: bool = False,
+        framework: Optional[TensorType] = None,
+    ) -> Mapping[str, Any]:
+        common_inputs = self._generate_dummy_inputs_for_sequence_classification_and_question_answering(
+            tokenizer, batch_size, seq_length, is_pair, framework
+        )
+
+        if self.use_past:
+            if not is_torch_available():
+                raise ValueError("Cannot generate dummy past_keys inputs without PyTorch installed.")
+            else:
+                import torch
+            batch, seqlen = common_inputs["input_ids"].shape
+            # Not using the same length for past_key_values
+            past_key_values_length = seqlen + 2
+            num_encoder_layers, _ = self.num_layers
+            num_encoder_attention_heads, _ = self.num_attention_heads
+            past_shape = (
+                batch,
+                num_encoder_attention_heads,
+                past_key_values_length,
+                self._config.hidden_size // num_encoder_attention_heads,
+            )
+
+            mask_dtype = common_inputs["attention_mask"].dtype
+            common_inputs["attention_mask"] = torch.cat(
+                [common_inputs["attention_mask"], torch.ones(batch, past_key_values_length, dtype=mask_dtype)], dim=1
+            )
+            common_inputs["past_key_values"] = [
+                (torch.zeros(past_shape), torch.zeros(past_shape)) for _ in range(num_encoder_layers)
+            ]
+        return common_inputs
+
+    def _generate_dummy_inputs_for_sequence_classification_and_question_answering(
+        self,
+        tokenizer: PreTrainedTokenizer,
+        batch_size: int = -1,
+        seq_length: int = -1,
+        is_pair: bool = False,
+        framework: Optional[TensorType] = None,
+    ) -> Mapping[str, Any]:
+        # Copied from OnnxConfig.generate_dummy_inputs
+        # Did not use super(OnnxConfigWithPast, self).generate_dummy_inputs for code clarity.
+        # If dynamic axis (-1) we forward with a fixed dimension of 2 samples to avoid optimizations made by ONNX
+        batch_size = compute_effective_axis_dimension(
+            batch_size, fixed_dimension=OnnxConfig.default_fixed_batch, num_token_to_add=0
+        )
+
+        # If dynamic axis (-1) we forward with a fixed dimension of 8 tokens to avoid optimizations made by ONNX
+        token_to_add = tokenizer.num_special_tokens_to_add(is_pair)
+        seq_length = compute_effective_axis_dimension(
+            seq_length, fixed_dimension=OnnxConfig.default_fixed_sequence, num_token_to_add=token_to_add
+        )
+
+        # Generate dummy inputs according to compute batch and sequence
+        dummy_input = [" ".join([tokenizer.unk_token]) * seq_length] * batch_size
+        common_inputs = dict(tokenizer(dummy_input, return_tensors=framework))
+        return common_inputs
+
+    def generate_dummy_inputs(
+        self,
+        tokenizer: PreTrainedTokenizer,
+        batch_size: int = -1,
+        seq_length: int = -1,
+        is_pair: bool = False,
+        framework: Optional[TensorType] = None,
+    ) -> Mapping[str, Any]:
+        if self.task in ["default", "seq2seq-lm"]:
+            common_inputs = self._generate_dummy_inputs_for_default_and_seq2seq_lm(
+                tokenizer, batch_size=batch_size, seq_length=seq_length, is_pair=is_pair, framework=framework
+            )
+
+        elif self.task == "causal-lm":
+            common_inputs = self._generate_dummy_inputs_for_causal_lm(
+                tokenizer, batch_size=batch_size, seq_length=seq_length, is_pair=is_pair, framework=framework
+            )
+        else:
+            common_inputs = self._generate_dummy_inputs_for_sequence_classification_and_question_answering(
+                tokenizer, batch_size=batch_size, seq_length=seq_length, is_pair=is_pair, framework=framework
+            )
+
+        return common_inputs
+
+    def _flatten_past_key_values_(self, flattened_output, name, idx, t):
+        if self.task in ["default", "seq2seq-lm"]:
+            flattened_output = super()._flatten_past_key_values_(flattened_output, name, idx, t)
+        else:
+            flattened_output = super(OnnxSeq2SeqConfigWithPast, self)._flatten_past_key_values_(
+                flattened_output, name, idx, t
+            )
+
+
+__all__ = ["BigBirdPegasusConfig", "BigBirdPegasusOnnxConfig"]

lib/python3.12/site-packages/transformers/models/bros/__init__.py

@@ -0,0 +1,28 @@
+# 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 _LazyModule
+from ...utils.import_utils import define_import_structure
+
+
+if TYPE_CHECKING:
+    from .configuration_bros import *
+    from .modeling_bros import *
+    from .processing_bros import *
+else:
+    import sys
+
+    _file = globals()["__file__"]
+    sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)

lib/python3.12/site-packages/transformers/models/bros/configuration_bros.py

@@ -0,0 +1,138 @@
+# coding=utf-8
+# Copyright 2023-present NAVER Corp, The Microsoft Research Asia LayoutLM Team Authors and 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.
+"""Bros model configuration"""
+
+from ...configuration_utils import PretrainedConfig
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class BrosConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`BrosModel`] or a [`TFBrosModel`]. It is used to
+    instantiate a Bros 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 Bros
+    [jinho8345/bros-base-uncased](https://huggingface.co/jinho8345/bros-base-uncased) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 30522):
+            Vocabulary size of the Bros model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`BrosModel`] or [`TFBrosModel`].
+        hidden_size (`int`, *optional*, defaults to 768):
+            Dimensionality of the encoder layers and the pooler layer.
+        num_hidden_layers (`int`, *optional*, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 12):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        intermediate_size (`int`, *optional*, defaults to 3072):
+            Dimensionality of the "intermediate" (often named feed-forward) layer in the Transformer encoder.
+        hidden_act (`str` or `Callable`, *optional*, defaults to `"gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"silu"` and `"gelu_new"` are supported.
+        hidden_dropout_prob (`float`, *optional*, defaults to 0.1):
+            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_probs_dropout_prob (`float`, *optional*, defaults to 0.1):
+            The dropout ratio for the attention probabilities.
+        max_position_embeddings (`int`, *optional*, defaults to 512):
+            The maximum sequence length that this model might ever be used with. Typically set this to something large
+            just in case (e.g., 512 or 1024 or 2048).
+        type_vocab_size (`int`, *optional*, defaults to 2):
+            The vocabulary size of the `token_type_ids` passed when calling [`BrosModel`] or [`TFBrosModel`].
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        layer_norm_eps (`float`, *optional*, defaults to 1e-12):
+            The epsilon used by the layer normalization layers.
+        pad_token_id (`int`, *optional*, defaults to 0):
+            The index of the padding token in the token vocabulary.
+        dim_bbox (`int`, *optional*, defaults to 8):
+            The dimension of the bounding box coordinates. (x0, y1, x1, y0, x1, y1, x0, y1)
+        bbox_scale (`float`, *optional*, defaults to 100.0):
+            The scale factor of the bounding box coordinates.
+        n_relations (`int`, *optional*, defaults to 1):
+            The number of relations for SpadeEE(entity extraction), SpadeEL(entity linking) head.
+        classifier_dropout_prob (`float`, *optional*, defaults to 0.1):
+            The dropout ratio for the classifier head.
+
+
+    Examples:
+
+    ```python
+    >>> from transformers import BrosConfig, BrosModel
+
+    >>> # Initializing a BROS jinho8345/bros-base-uncased style configuration
+    >>> configuration = BrosConfig()
+
+    >>> # Initializing a model from the jinho8345/bros-base-uncased style configuration
+    >>> model = BrosModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "bros"
+
+    def __init__(
+        self,
+        vocab_size=30522,
+        hidden_size=768,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        intermediate_size=3072,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        max_position_embeddings=512,
+        type_vocab_size=2,
+        initializer_range=0.02,
+        layer_norm_eps=1e-12,
+        pad_token_id=0,
+        dim_bbox=8,
+        bbox_scale=100.0,
+        n_relations=1,
+        classifier_dropout_prob=0.1,
+        **kwargs,
+    ):
+        super().__init__(
+            vocab_size=vocab_size,
+            hidden_size=hidden_size,
+            num_hidden_layers=num_hidden_layers,
+            num_attention_heads=num_attention_heads,
+            intermediate_size=intermediate_size,
+            hidden_act=hidden_act,
+            hidden_dropout_prob=hidden_dropout_prob,
+            attention_probs_dropout_prob=attention_probs_dropout_prob,
+            max_position_embeddings=max_position_embeddings,
+            type_vocab_size=type_vocab_size,
+            initializer_range=initializer_range,
+            layer_norm_eps=layer_norm_eps,
+            pad_token_id=pad_token_id,
+            **kwargs,
+        )
+
+        self.dim_bbox = dim_bbox
+        self.bbox_scale = bbox_scale
+        self.n_relations = n_relations
+        self.dim_bbox_sinusoid_emb_2d = self.hidden_size // 4
+        self.dim_bbox_sinusoid_emb_1d = self.dim_bbox_sinusoid_emb_2d // self.dim_bbox
+        self.dim_bbox_projection = self.hidden_size // self.num_attention_heads
+        self.classifier_dropout_prob = classifier_dropout_prob
+
+
+__all__ = ["BrosConfig"]

lib/python3.12/site-packages/transformers/models/bros/modeling_bros.py

@@ -0,0 +1,1323 @@
+# coding=utf-8
+# Copyright 2023-present NAVER Corp, The Microsoft Research Asia LayoutLM Team Authors and 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.
+"""PyTorch Bros model."""
+
+import math
+from dataclasses import dataclass
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import CrossEntropyLoss
+
+from ...activations import ACT2FN
+from ...modeling_outputs import (
+    BaseModelOutputWithPastAndCrossAttentions,
+    BaseModelOutputWithPoolingAndCrossAttentions,
+    TokenClassifierOutput,
+)
+from ...modeling_utils import PreTrainedModel
+from ...pytorch_utils import apply_chunking_to_forward, find_pruneable_heads_and_indices, prune_linear_layer
+from ...utils import (
+    ModelOutput,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    logging,
+    replace_return_docstrings,
+)
+from .configuration_bros import BrosConfig
+
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "jinho8345/bros-base-uncased"
+_CONFIG_FOR_DOC = "BrosConfig"
+
+
+BROS_START_DOCSTRING = r"""
+    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 ([`BrosConfig`]): 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.
+"""
+
+BROS_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `({0})`):
+            Indices of input sequence tokens in the vocabulary.
+
+            Indices can be obtained using [`BrosProcessor`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+
+        bbox ('torch.FloatTensor' of shape '(batch_size, num_boxes, 4)'):
+            Bounding box coordinates for each token in the input sequence. Each bounding box is a list of four values
+            (x1, y1, x2, y2), where (x1, y1) is the top left corner, and (x2, y2) is the bottom right corner of the
+            bounding box.
+
+        attention_mask (`torch.FloatTensor` of shape `({0})`, *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)
+
+        bbox_first_token_mask (`torch.FloatTensor` of shape `({0})`, *optional*):
+            Mask to indicate the first token of each bounding box. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+        token_type_ids (`torch.LongTensor` of shape `({0})`, *optional*):
+            Segment token indices to indicate first and second portions of the inputs. Indices are selected in `[0,
+            1]`:
+
+            - 0 corresponds to a *sentence A* token,
+            - 1 corresponds to a *sentence B* token.
+
+            [What are token type IDs?](../glossary#token-type-ids)
+
+        position_ids (`torch.LongTensor` of shape `({0})`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.max_position_embeddings - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+
+        head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
+            Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        inputs_embeds (`torch.FloatTensor` of shape `({0}, 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.
+
+        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 [`~file_utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+@dataclass
+class BrosSpadeOutput(ModelOutput):
+    """
+    Base class for outputs of token classification models.
+
+    Args:
+        loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided) :
+            Classification loss.
+        initial_token_logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.num_labels)`):
+            Classification scores for entity initial tokens (before SoftMax).
+        subsequent_token_logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, sequence_length+1)`):
+            Classification scores for entity sequence tokens (before SoftMax).
+        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.
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    initial_token_logits: Optional[torch.FloatTensor] = None
+    subsequent_token_logits: Optional[torch.FloatTensor] = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+class BrosPositionalEmbedding1D(nn.Module):
+    # Reference: https://github.com/kimiyoung/transformer-xl/blob/master/pytorch/mem_transformer.py#L15
+
+    def __init__(self, config):
+        super(BrosPositionalEmbedding1D, self).__init__()
+
+        self.dim_bbox_sinusoid_emb_1d = config.dim_bbox_sinusoid_emb_1d
+
+        inv_freq = 1 / (
+            10000 ** (torch.arange(0.0, self.dim_bbox_sinusoid_emb_1d, 2.0) / self.dim_bbox_sinusoid_emb_1d)
+        )
+        self.register_buffer("inv_freq", inv_freq)
+
+    def forward(self, pos_seq: torch.Tensor) -> torch.Tensor:
+        seq_size = pos_seq.size()
+        b1, b2, b3 = seq_size
+        sinusoid_inp = pos_seq.view(b1, b2, b3, 1) * self.inv_freq.view(1, 1, 1, self.dim_bbox_sinusoid_emb_1d // 2)
+        pos_emb = torch.cat([sinusoid_inp.sin(), sinusoid_inp.cos()], dim=-1)
+        return pos_emb
+
+
+class BrosPositionalEmbedding2D(nn.Module):
+    def __init__(self, config):
+        super(BrosPositionalEmbedding2D, self).__init__()
+
+        self.dim_bbox = config.dim_bbox
+        self.x_pos_emb = BrosPositionalEmbedding1D(config)
+        self.y_pos_emb = BrosPositionalEmbedding1D(config)
+
+    def forward(self, bbox: torch.Tensor) -> torch.Tensor:
+        stack = []
+        for i in range(self.dim_bbox):
+            if i % 2 == 0:
+                stack.append(self.x_pos_emb(bbox[..., i]))
+            else:
+                stack.append(self.y_pos_emb(bbox[..., i]))
+        bbox_pos_emb = torch.cat(stack, dim=-1)
+        return bbox_pos_emb
+
+
+class BrosBboxEmbeddings(nn.Module):
+    def __init__(self, config):
+        super(BrosBboxEmbeddings, self).__init__()
+        self.bbox_sinusoid_emb = BrosPositionalEmbedding2D(config)
+        self.bbox_projection = nn.Linear(config.dim_bbox_sinusoid_emb_2d, config.dim_bbox_projection, bias=False)
+
+    def forward(self, bbox: torch.Tensor):
+        bbox_t = bbox.transpose(0, 1)
+        bbox_pos = bbox_t[None, :, :, :] - bbox_t[:, None, :, :]
+        bbox_pos_emb = self.bbox_sinusoid_emb(bbox_pos)
+        bbox_pos_emb = self.bbox_projection(bbox_pos_emb)
+
+        return bbox_pos_emb
+
+
+class BrosTextEmbeddings(nn.Module):
+    """Construct the embeddings from word, position and token_type embeddings."""
+
+    def __init__(self, config):
+        super().__init__()
+
+        self.word_embeddings = nn.Embedding(config.vocab_size, config.hidden_size, padding_idx=config.pad_token_id)
+        self.position_embeddings = nn.Embedding(config.max_position_embeddings, config.hidden_size)
+        self.token_type_embeddings = nn.Embedding(config.type_vocab_size, config.hidden_size)
+
+        # self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load
+        # any TensorFlow checkpoint file
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        # position_ids (1, len position emb) is contiguous in memory and exported when serialized
+        self.position_embedding_type = getattr(config, "position_embedding_type", "absolute")
+        self.register_buffer("position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)))
+        self.register_buffer(
+            "token_type_ids",
+            torch.zeros(
+                self.position_ids.size(),
+                dtype=torch.long,
+                device=self.position_ids.device,
+            ),
+            persistent=False,
+        )
+
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        past_key_values_length: int = 0,
+    ) -> torch.Tensor:
+        if input_ids is not None:
+            input_shape = input_ids.size()
+        else:
+            input_shape = inputs_embeds.size()[:-1]
+
+        seq_length = input_shape[1]
+
+        if position_ids is None:
+            position_ids = self.position_ids[:, past_key_values_length : seq_length + past_key_values_length]
+
+        if token_type_ids is None:
+            if hasattr(self, "token_type_ids"):
+                buffered_token_type_ids = self.token_type_ids[:, :seq_length]
+                buffered_token_type_ids_expanded = buffered_token_type_ids.expand(input_shape[0], seq_length)
+                token_type_ids = buffered_token_type_ids_expanded
+            else:
+                token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=self.position_ids.device)
+
+        if inputs_embeds is None:
+            inputs_embeds = self.word_embeddings(input_ids)
+        token_type_embeddings = self.token_type_embeddings(token_type_ids)
+
+        embeddings = inputs_embeds + token_type_embeddings
+        if self.position_embedding_type == "absolute":
+            position_embeddings = self.position_embeddings(position_ids)
+            embeddings += position_embeddings
+        embeddings = self.LayerNorm(embeddings)
+        embeddings = self.dropout(embeddings)
+        return embeddings
+
+
+class BrosSelfAttention(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        if config.hidden_size % config.num_attention_heads != 0 and not hasattr(config, "embedding_size"):
+            raise ValueError(
+                f"The hidden size ({config.hidden_size}) is not a multiple of the number of attention "
+                f"heads ({config.num_attention_heads})"
+            )
+
+        self.num_attention_heads = config.num_attention_heads
+        self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
+        self.all_head_size = self.num_attention_heads * self.attention_head_size
+
+        self.query = nn.Linear(config.hidden_size, self.all_head_size)
+        self.key = nn.Linear(config.hidden_size, self.all_head_size)
+        self.value = nn.Linear(config.hidden_size, self.all_head_size)
+
+        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
+        self.position_embedding_type = getattr(config, "position_embedding_type", "absolute")
+        if self.position_embedding_type == "relative_key" or self.position_embedding_type == "relative_key_query":
+            self.max_position_embeddings = config.max_position_embeddings
+            self.distance_embedding = nn.Embedding(2 * config.max_position_embeddings - 1, self.attention_head_size)
+
+        self.is_decoder = config.is_decoder
+
+    def transpose_for_scores(self, x: torch.Tensor):
+        new_x_shape = x.size()[:-1] + (
+            self.num_attention_heads,
+            self.attention_head_size,
+        )
+        x = x.view(*new_x_shape)
+        return x.permute(0, 2, 1, 3)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        bbox_pos_emb: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        output_attentions: Optional[torch.Tensor] = False,
+    ) -> Tuple[torch.Tensor]:
+        mixed_query_layer = self.query(hidden_states)
+
+        # If this is instantiated as a cross-attention module, the keys
+        # and values come from an encoder; the attention mask needs to be
+        # such that the encoder's padding tokens are not attended to.
+        is_cross_attention = encoder_hidden_states is not None
+
+        if is_cross_attention and past_key_value is not None:
+            # reuse k,v, cross_attentions
+            key_layer = past_key_value[0]
+            value_layer = past_key_value[1]
+            attention_mask = encoder_attention_mask
+        elif is_cross_attention:
+            key_layer = self.transpose_for_scores(self.key(encoder_hidden_states))
+            value_layer = self.transpose_for_scores(self.value(encoder_hidden_states))
+            attention_mask = encoder_attention_mask
+        elif past_key_value is not None:
+            key_layer = self.transpose_for_scores(self.key(hidden_states))
+            value_layer = self.transpose_for_scores(self.value(hidden_states))
+            key_layer = torch.cat([past_key_value[0], key_layer], dim=2)
+            value_layer = torch.cat([past_key_value[1], value_layer], dim=2)
+        else:
+            key_layer = self.transpose_for_scores(self.key(hidden_states))
+            value_layer = self.transpose_for_scores(self.value(hidden_states))
+
+        query_layer = self.transpose_for_scores(mixed_query_layer)
+
+        if self.is_decoder:
+            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
+            # Further calls to cross_attention layer can then reuse all cross-attention
+            # key/value_states (first "if" case)
+            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
+            # all previous decoder key/value_states. Further calls to uni-directional self-attention
+            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
+            # if encoder bi-directional self-attention `past_key_value` is always `None`
+            past_key_value = (key_layer, value_layer)
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
+
+        if self.position_embedding_type == "relative_key" or self.position_embedding_type == "relative_key_query":
+            seq_length = hidden_states.size()[1]
+            position_ids_l = torch.arange(seq_length, dtype=torch.long, device=hidden_states.device).view(-1, 1)
+            position_ids_r = torch.arange(seq_length, dtype=torch.long, device=hidden_states.device).view(1, -1)
+            distance = position_ids_l - position_ids_r
+            positional_embedding = self.distance_embedding(distance + self.max_position_embeddings - 1)
+            positional_embedding = positional_embedding.to(dtype=query_layer.dtype)  # fp16 compatibility
+
+            if self.position_embedding_type == "relative_key":
+                relative_position_scores = torch.einsum("bhld,lrd->bhlr", query_layer, positional_embedding)
+                attention_scores = attention_scores + relative_position_scores
+            elif self.position_embedding_type == "relative_key_query":
+                relative_position_scores_query = torch.einsum("bhld,lrd->bhlr", query_layer, positional_embedding)
+                relative_position_scores_key = torch.einsum("bhrd,lrd->bhlr", key_layer, positional_embedding)
+
+                attention_scores = attention_scores + relative_position_scores_query + relative_position_scores_key
+
+        # bbox positional encoding
+        batch_size, n_head, seq_length, d_head = query_layer.shape
+        bbox_pos_emb = bbox_pos_emb.view(seq_length, seq_length, batch_size, d_head)
+        bbox_pos_emb = bbox_pos_emb.permute([2, 0, 1, 3])
+        bbox_pos_scores = torch.einsum("bnid,bijd->bnij", (query_layer, bbox_pos_emb))
+
+        attention_scores = attention_scores + bbox_pos_scores
+
+        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
+        if attention_mask is not None:
+            # Apply the attention mask is (precomputed for all layers in BrosModel forward() function)
+            attention_scores = attention_scores + attention_mask
+
+        # Normalize the attention scores to probabilities.
+        attention_probs = nn.Softmax(dim=-1)(attention_scores)
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        attention_probs = self.dropout(attention_probs)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attention_probs = attention_probs * head_mask
+
+        context_layer = torch.matmul(attention_probs, value_layer)
+
+        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
+        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
+        context_layer = context_layer.view(*new_context_layer_shape)
+
+        outputs = (context_layer, attention_probs) if output_attentions else (context_layer,)
+
+        if self.is_decoder:
+            outputs = outputs + (past_key_value,)
+        return outputs
+
+
+# Copied from transformers.models.bert.modeling_bert.BertSelfOutput with Bert->Bros
+class BrosSelfOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+class BrosAttention(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.self = BrosSelfAttention(config)
+        self.output = BrosSelfOutput(config)
+        self.pruned_heads = set()
+
+    def prune_heads(self, heads):
+        if len(heads) == 0:
+            return
+        heads, index = find_pruneable_heads_and_indices(
+            heads,
+            self.self.num_attention_heads,
+            self.self.attention_head_size,
+            self.pruned_heads,
+        )
+
+        # Prune linear layers
+        self.self.query = prune_linear_layer(self.self.query, index)
+        self.self.key = prune_linear_layer(self.self.key, index)
+        self.self.value = prune_linear_layer(self.self.value, index)
+        self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)
+
+        # Update hyper params and store pruned heads
+        self.self.num_attention_heads = self.self.num_attention_heads - len(heads)
+        self.self.all_head_size = self.self.attention_head_size * self.self.num_attention_heads
+        self.pruned_heads = self.pruned_heads.union(heads)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        bbox_pos_emb: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.Tensor]:
+        self_outputs = self.self(
+            hidden_states=hidden_states,
+            bbox_pos_emb=bbox_pos_emb,
+            attention_mask=attention_mask,
+            head_mask=head_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            past_key_value=past_key_value,
+            output_attentions=output_attentions,
+        )
+        attention_output = self.output(self_outputs[0], hidden_states)
+        outputs = (attention_output,) + self_outputs[1:]  # add attentions if we output them
+        return outputs
+
+
+# Copied from transformers.models.bert.modeling_bert.BertIntermediate with Bert->Bros
+class BrosIntermediate(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.intermediate_size)
+        if isinstance(config.hidden_act, str):
+            self.intermediate_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.intermediate_act_fn = config.hidden_act
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.intermediate_act_fn(hidden_states)
+        return hidden_states
+
+
+class BrosOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+class BrosLayer(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.chunk_size_feed_forward = config.chunk_size_feed_forward
+        self.seq_len_dim = 1
+        self.attention = BrosAttention(config)
+        self.is_decoder = config.is_decoder
+        self.add_cross_attention = config.add_cross_attention
+        if self.add_cross_attention:
+            if not self.is_decoder:
+                raise Exception(f"{self} should be used as a decoder model if cross attention is added")
+            self.crossattention = BrosAttention(config)
+        self.intermediate = BrosIntermediate(config)
+        self.output = BrosOutput(config)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        bbox_pos_emb: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.Tensor]:
+        # decoder uni-directional self-attention cached key/values tuple is at positions 1,2
+        self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None
+        self_attention_outputs = self.attention(
+            hidden_states,
+            bbox_pos_emb=bbox_pos_emb,
+            attention_mask=attention_mask,
+            head_mask=head_mask,
+            output_attentions=output_attentions,
+            past_key_value=self_attn_past_key_value,
+        )
+        attention_output = self_attention_outputs[0]
+
+        # if decoder, the last output is tuple of self-attn cache
+        if self.is_decoder:
+            outputs = self_attention_outputs[1:-1]
+            present_key_value = self_attention_outputs[-1]
+        else:
+            outputs = self_attention_outputs[1:]  # add self attentions if we output attention weights
+
+        cross_attn_present_key_value = None
+        if self.is_decoder and encoder_hidden_states is not None:
+            if hasattr(self, "crossattention"):
+                raise Exception(
+                    f"If `encoder_hidden_states` are passed, {self} has to be instantiated with cross-attention layers by setting `config.add_cross_attention=True`"
+                )
+
+            # cross_attn cached key/values tuple is at positions 3,4 of past_key_value tuple
+            cross_attn_past_key_value = past_key_value[-2:] if past_key_value is not None else None
+            cross_attention_outputs = self.crossattention(
+                attention_output,
+                attention_mask,
+                head_mask,
+                encoder_hidden_states,
+                encoder_attention_mask,
+                cross_attn_past_key_value,
+                output_attentions,
+            )
+            attention_output = cross_attention_outputs[0]
+            outputs = outputs + cross_attention_outputs[1:-1]  # add cross attentions if we output attention weights
+
+            # add cross-attn cache to positions 3,4 of present_key_value tuple
+            cross_attn_present_key_value = cross_attention_outputs[-1]
+            present_key_value = present_key_value + cross_attn_present_key_value
+
+        layer_output = apply_chunking_to_forward(
+            self.feed_forward_chunk,
+            self.chunk_size_feed_forward,
+            self.seq_len_dim,
+            attention_output,
+        )
+        outputs = (layer_output,) + outputs
+
+        # if decoder, return the attn key/values as the last output
+        if self.is_decoder:
+            outputs = outputs + (present_key_value,)
+
+        return outputs
+
+    def feed_forward_chunk(self, attention_output):
+        intermediate_output = self.intermediate(attention_output)
+        layer_output = self.output(intermediate_output, attention_output)
+        return layer_output
+
+
+class BrosEncoder(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.layer = nn.ModuleList([BrosLayer(config) for _ in range(config.num_hidden_layers)])
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        bbox_pos_emb: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = False,
+        output_hidden_states: Optional[bool] = False,
+        return_dict: Optional[bool] = True,
+    ) -> Union[Tuple[torch.Tensor], BaseModelOutputWithPastAndCrossAttentions]:
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attentions = () if output_attentions else None
+        all_cross_attentions = () if output_attentions and self.config.add_cross_attention else None
+
+        next_decoder_cache = () if use_cache else None
+        for i, layer_module in enumerate(self.layer):
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            layer_head_mask = head_mask[i] if head_mask is not None else None
+            past_key_value = past_key_values[i] if past_key_values is not None else None
+
+            if getattr(self.config, "gradient_checkpointing", False) and self.training:
+                if use_cache:
+                    logger.warning(
+                        "`use_cache=True` is incompatible with `config.gradient_checkpointing=True`. Setting "
+                        "`use_cache=False`..."
+                    )
+                    use_cache = False
+                layer_outputs = self._gradient_checkpointing_func(
+                    layer_module.__call__,
+                    hidden_states,
+                    bbox_pos_emb,
+                    attention_mask,
+                    layer_head_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    output_attentions,
+                )
+            else:
+                layer_outputs = layer_module(
+                    hidden_states=hidden_states,
+                    bbox_pos_emb=bbox_pos_emb,
+                    attention_mask=attention_mask,
+                    head_mask=layer_head_mask,
+                    encoder_hidden_states=encoder_hidden_states,
+                    encoder_attention_mask=encoder_attention_mask,
+                    past_key_value=past_key_value,
+                    output_attentions=output_attentions,
+                )
+
+            hidden_states = layer_outputs[0]
+            if use_cache:
+                next_decoder_cache += (layer_outputs[-1],)
+            if output_attentions:
+                all_self_attentions = all_self_attentions + (layer_outputs[1],)
+                if self.config.add_cross_attention:
+                    all_cross_attentions = all_cross_attentions + (layer_outputs[2],)
+
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(
+                v
+                for v in [
+                    hidden_states,
+                    next_decoder_cache,
+                    all_hidden_states,
+                    all_self_attentions,
+                    all_cross_attentions,
+                ]
+                if v is not None
+            )
+        return BaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            past_key_values=next_decoder_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attentions,
+            cross_attentions=all_cross_attentions,
+        )
+
+
+# Copied from transformers.models.bert.modeling_bert.BertPooler with Bert->Bros
+class BrosPooler(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.activation = nn.Tanh()
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        # We "pool" the model by simply taking the hidden state corresponding
+        # to the first token.
+        first_token_tensor = hidden_states[:, 0]
+        pooled_output = self.dense(first_token_tensor)
+        pooled_output = self.activation(pooled_output)
+        return pooled_output
+
+
+class BrosRelationExtractor(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.n_relations = config.n_relations
+        self.backbone_hidden_size = config.hidden_size
+        self.head_hidden_size = config.hidden_size
+        self.classifier_dropout_prob = config.classifier_dropout_prob
+
+        self.drop = nn.Dropout(self.classifier_dropout_prob)
+        self.query = nn.Linear(self.backbone_hidden_size, self.n_relations * self.head_hidden_size)
+
+        self.key = nn.Linear(self.backbone_hidden_size, self.n_relations * self.head_hidden_size)
+
+        self.dummy_node = nn.Parameter(torch.zeros(1, self.backbone_hidden_size))
+
+    def forward(self, query_layer: torch.Tensor, key_layer: torch.Tensor):
+        query_layer = self.query(self.drop(query_layer))
+
+        dummy_vec = self.dummy_node.unsqueeze(0).repeat(1, key_layer.size(1), 1)
+        key_layer = torch.cat([key_layer, dummy_vec], axis=0)
+        key_layer = self.key(self.drop(key_layer))
+
+        query_layer = query_layer.view(
+            query_layer.size(0), query_layer.size(1), self.n_relations, self.head_hidden_size
+        )
+        key_layer = key_layer.view(key_layer.size(0), key_layer.size(1), self.n_relations, self.head_hidden_size)
+
+        relation_score = torch.matmul(
+            query_layer.permute(2, 1, 0, 3), key_layer.permute(2, 1, 3, 0)
+        )  # equivalent to torch.einsum("ibnd,jbnd->nbij", (query_layer, key_layer))
+
+        return relation_score
+
+
+class BrosPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = BrosConfig
+    base_model_prefix = "bros"
+
+    def _init_weights(self, module):
+        """Initialize the weights"""
+        if isinstance(module, nn.Linear):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+
+
+@add_start_docstrings(
+    "The bare Bros Model transformer outputting raw hidden-states without any specific head on top.",
+    BROS_START_DOCSTRING,
+)
+class BrosModel(BrosPreTrainedModel):
+    def __init__(self, config, add_pooling_layer=True):
+        super().__init__(config)
+        self.config = config
+
+        self.embeddings = BrosTextEmbeddings(config)
+        self.bbox_embeddings = BrosBboxEmbeddings(config)
+        self.encoder = BrosEncoder(config)
+
+        self.pooler = BrosPooler(config) if add_pooling_layer else None
+
+        self.init_weights()
+
+    def get_input_embeddings(self):
+        return self.embeddings.word_embeddings
+
+    def set_input_embeddings(self, value):
+        self.embeddings.word_embeddings = value
+
+    def _prune_heads(self, heads_to_prune):
+        """
+        Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
+        class PreTrainedModel
+        """
+        for layer, heads in heads_to_prune.items():
+            self.encoder.layer[layer].attention.prune_heads(heads)
+
+    @add_start_docstrings_to_model_forward(BROS_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @replace_return_docstrings(output_type=BaseModelOutputWithPoolingAndCrossAttentions, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        bbox: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple[torch.Tensor], BaseModelOutputWithPoolingAndCrossAttentions]:
+        r"""
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> import torch
+        >>> from transformers import BrosProcessor, BrosModel
+
+        >>> processor = BrosProcessor.from_pretrained("jinho8345/bros-base-uncased")
+
+        >>> model = BrosModel.from_pretrained("jinho8345/bros-base-uncased")
+
+        >>> encoding = processor("Hello, my dog is cute", add_special_tokens=False, return_tensors="pt")
+        >>> bbox = torch.tensor([[[0, 0, 1, 1]]]).repeat(1, encoding["input_ids"].shape[-1], 1)
+        >>> encoding["bbox"] = bbox
+
+        >>> outputs = model(**encoding)
+        >>> last_hidden_states = outputs.last_hidden_state
+        ```"""
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if self.config.is_decoder:
+            use_cache = use_cache if use_cache is not None else self.config.use_cache
+        else:
+            use_cache = False
+
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        if bbox is None:
+            raise ValueError("You have to specify bbox")
+
+        batch_size, seq_length = input_shape
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+
+        # past_key_values_length
+        past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0
+
+        if attention_mask is None:
+            attention_mask = torch.ones(input_shape, device=device)
+
+        if token_type_ids is None:
+            if hasattr(self.embeddings, "token_type_ids"):
+                buffered_token_type_ids = self.embeddings.token_type_ids[:, :seq_length]
+                buffered_token_type_ids_expanded = buffered_token_type_ids.expand(batch_size, seq_length)
+                token_type_ids = buffered_token_type_ids_expanded
+            else:
+                token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
+
+        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+        # ourselves in which case we just need to make it broadcastable to all heads.
+        extended_attention_mask: torch.Tensor = self.get_extended_attention_mask(attention_mask, input_shape, device)
+
+        # If a 2D or 3D attention mask is provided for the cross-attention
+        # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
+        if self.config.is_decoder and encoder_hidden_states is not None:
+            encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size()
+            encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
+            if encoder_attention_mask is None:
+                encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)
+            encoder_extended_attention_mask = self.invert_attention_mask(encoder_attention_mask)
+        else:
+            encoder_extended_attention_mask = None
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
+
+        embedding_output = self.embeddings(
+            input_ids=input_ids,
+            position_ids=position_ids,
+            token_type_ids=token_type_ids,
+            inputs_embeds=inputs_embeds,
+            past_key_values_length=past_key_values_length,
+        )
+
+        # if bbox has 2 points (4 float tensors) per token, convert it to 4 points (8 float tensors) per token
+        if bbox.shape[-1] == 4:
+            bbox = bbox[:, :, [0, 1, 2, 1, 2, 3, 0, 3]]
+        scaled_bbox = bbox * self.config.bbox_scale
+        bbox_position_embeddings = self.bbox_embeddings(scaled_bbox)
+
+        encoder_outputs = self.encoder(
+            embedding_output,
+            bbox_pos_emb=bbox_position_embeddings,
+            attention_mask=extended_attention_mask,
+            head_mask=head_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_extended_attention_mask,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        sequence_output = encoder_outputs[0]
+        pooled_output = self.pooler(sequence_output) if self.pooler is not None else None
+
+        if not return_dict:
+            return (sequence_output, pooled_output) + encoder_outputs[1:]
+
+        return BaseModelOutputWithPoolingAndCrossAttentions(
+            last_hidden_state=sequence_output,
+            pooler_output=pooled_output,
+            past_key_values=encoder_outputs.past_key_values,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+            cross_attentions=encoder_outputs.cross_attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    Bros Model with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for
+    Named-Entity-Recognition (NER) tasks.
+    """,
+    BROS_START_DOCSTRING,
+)
+class BrosForTokenClassification(BrosPreTrainedModel):
+    _keys_to_ignore_on_load_unexpected = [r"pooler"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.bros = BrosModel(config)
+        classifier_dropout = (
+            config.classifier_dropout if hasattr(config, "classifier_dropout") else config.hidden_dropout_prob
+        )
+        self.dropout = nn.Dropout(classifier_dropout)
+        self.classifier = nn.Linear(config.hidden_size, config.num_labels)
+
+        self.init_weights()
+
+    @add_start_docstrings_to_model_forward(BROS_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @replace_return_docstrings(output_type=TokenClassifierOutput, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        bbox: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        bbox_first_token_mask: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        labels: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple[torch.Tensor], TokenClassifierOutput]:
+        r"""
+
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> import torch
+        >>> from transformers import BrosProcessor, BrosForTokenClassification
+
+        >>> processor = BrosProcessor.from_pretrained("jinho8345/bros-base-uncased")
+
+        >>> model = BrosForTokenClassification.from_pretrained("jinho8345/bros-base-uncased")
+
+        >>> encoding = processor("Hello, my dog is cute", add_special_tokens=False, return_tensors="pt")
+        >>> bbox = torch.tensor([[[0, 0, 1, 1]]]).repeat(1, encoding["input_ids"].shape[-1], 1)
+        >>> encoding["bbox"] = bbox
+
+        >>> outputs = model(**encoding)
+        ```"""
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.bros(
+            input_ids,
+            bbox=bbox,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        sequence_output = outputs[0]
+
+        sequence_output = self.dropout(sequence_output)
+        logits = self.classifier(sequence_output)
+
+        loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            if bbox_first_token_mask is not None:
+                bbox_first_token_mask = bbox_first_token_mask.view(-1)
+                loss = loss_fct(
+                    logits.view(-1, self.num_labels)[bbox_first_token_mask], labels.view(-1)[bbox_first_token_mask]
+                )
+            else:
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+
+        if not return_dict:
+            output = (logits,) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TokenClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    Bros Model with a token classification head on top (initial_token_layers and subsequent_token_layer on top of the
+    hidden-states output) e.g. for Named-Entity-Recognition (NER) tasks. The initial_token_classifier is used to
+    predict the first token of each entity, and the subsequent_token_classifier is used to predict the subsequent
+    tokens within an entity. Compared to BrosForTokenClassification, this model is more robust to serialization errors
+    since it predicts next token from one token.
+    """,
+    BROS_START_DOCSTRING,
+)
+class BrosSpadeEEForTokenClassification(BrosPreTrainedModel):
+    _keys_to_ignore_on_load_unexpected = [r"pooler"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.config = config
+        self.num_labels = config.num_labels
+        self.n_relations = config.n_relations
+        self.backbone_hidden_size = config.hidden_size
+
+        self.bros = BrosModel(config)
+        classifier_dropout = (
+            config.classifier_dropout if hasattr(config, "classifier_dropout") else config.hidden_dropout_prob
+        )
+
+        # Initial token classification for Entity Extraction (NER)
+        self.initial_token_classifier = nn.Sequential(
+            nn.Dropout(classifier_dropout),
+            nn.Linear(config.hidden_size, config.hidden_size),
+            nn.Dropout(classifier_dropout),
+            nn.Linear(config.hidden_size, config.num_labels),
+        )
+
+        # Subsequent token classification for Entity Extraction (NER)
+        self.subsequent_token_classifier = BrosRelationExtractor(config)
+
+        self.init_weights()
+
+    @add_start_docstrings_to_model_forward(BROS_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @replace_return_docstrings(output_type=BrosSpadeOutput, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        bbox: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        bbox_first_token_mask: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        initial_token_labels: Optional[torch.Tensor] = None,
+        subsequent_token_labels: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple[torch.Tensor], BrosSpadeOutput]:
+        r"""
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> import torch
+        >>> from transformers import BrosProcessor, BrosSpadeEEForTokenClassification
+
+        >>> processor = BrosProcessor.from_pretrained("jinho8345/bros-base-uncased")
+
+        >>> model = BrosSpadeEEForTokenClassification.from_pretrained("jinho8345/bros-base-uncased")
+
+        >>> encoding = processor("Hello, my dog is cute", add_special_tokens=False, return_tensors="pt")
+        >>> bbox = torch.tensor([[[0, 0, 1, 1]]]).repeat(1, encoding["input_ids"].shape[-1], 1)
+        >>> encoding["bbox"] = bbox
+
+        >>> outputs = model(**encoding)
+        ```"""
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.bros(
+            input_ids=input_ids,
+            bbox=bbox,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        last_hidden_states = outputs[0]
+        last_hidden_states = last_hidden_states.transpose(0, 1).contiguous()
+        initial_token_logits = self.initial_token_classifier(last_hidden_states).transpose(0, 1).contiguous()
+        subsequent_token_logits = self.subsequent_token_classifier(last_hidden_states, last_hidden_states).squeeze(0)
+
+        # make subsequent token (sequence token classification) mask
+        inv_attention_mask = 1 - attention_mask
+        batch_size, max_seq_length = inv_attention_mask.shape
+        device = inv_attention_mask.device
+        invalid_token_mask = torch.cat([inv_attention_mask, torch.zeros([batch_size, 1]).to(device)], axis=1).bool()
+        subsequent_token_logits = subsequent_token_logits.masked_fill(
+            invalid_token_mask[:, None, :], torch.finfo(subsequent_token_logits.dtype).min
+        )
+        self_token_mask = torch.eye(max_seq_length, max_seq_length + 1).to(device=device, dtype=torch.bool)
+        subsequent_token_logits = subsequent_token_logits.masked_fill(
+            self_token_mask[None, :, :], torch.finfo(subsequent_token_logits.dtype).min
+        )
+        subsequent_token_mask = attention_mask.view(-1).bool()
+
+        loss = None
+        if initial_token_labels is not None and subsequent_token_labels is not None:
+            loss_fct = CrossEntropyLoss()
+
+            # get initial token loss
+            initial_token_labels = initial_token_labels.view(-1)
+            if bbox_first_token_mask is not None:
+                bbox_first_token_mask = bbox_first_token_mask.view(-1)
+                initial_token_loss = loss_fct(
+                    initial_token_logits.view(-1, self.num_labels)[bbox_first_token_mask],
+                    initial_token_labels[bbox_first_token_mask],
+                )
+            else:
+                initial_token_loss = loss_fct(initial_token_logits.view(-1, self.num_labels), initial_token_labels)
+
+            subsequent_token_labels = subsequent_token_labels.view(-1)
+            subsequent_token_loss = loss_fct(
+                subsequent_token_logits.view(-1, max_seq_length + 1)[subsequent_token_mask],
+                subsequent_token_labels[subsequent_token_mask],
+            )
+
+            loss = initial_token_loss + subsequent_token_loss
+
+        if not return_dict:
+            output = (initial_token_logits, subsequent_token_logits) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return BrosSpadeOutput(
+            loss=loss,
+            initial_token_logits=initial_token_logits,
+            subsequent_token_logits=subsequent_token_logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    Bros Model with a token classification head on top (a entity_linker layer on top of the hidden-states output) e.g.
+    for Entity-Linking. The entity_linker is used to predict intra-entity links (one entity to another entity).
+    """,
+    BROS_START_DOCSTRING,
+)
+class BrosSpadeELForTokenClassification(BrosPreTrainedModel):
+    _keys_to_ignore_on_load_unexpected = [r"pooler"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.config = config
+        self.num_labels = config.num_labels
+        self.n_relations = config.n_relations
+        self.backbone_hidden_size = config.hidden_size
+
+        self.bros = BrosModel(config)
+        (config.classifier_dropout if hasattr(config, "classifier_dropout") else config.hidden_dropout_prob)
+
+        self.entity_linker = BrosRelationExtractor(config)
+
+        self.init_weights()
+
+    @add_start_docstrings_to_model_forward(BROS_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @replace_return_docstrings(output_type=TokenClassifierOutput, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        bbox: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        bbox_first_token_mask: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        labels: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple[torch.Tensor], TokenClassifierOutput]:
+        r"""
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> import torch
+        >>> from transformers import BrosProcessor, BrosSpadeELForTokenClassification
+
+        >>> processor = BrosProcessor.from_pretrained("jinho8345/bros-base-uncased")
+
+        >>> model = BrosSpadeELForTokenClassification.from_pretrained("jinho8345/bros-base-uncased")
+
+        >>> encoding = processor("Hello, my dog is cute", add_special_tokens=False, return_tensors="pt")
+        >>> bbox = torch.tensor([[[0, 0, 1, 1]]]).repeat(1, encoding["input_ids"].shape[-1], 1)
+        >>> encoding["bbox"] = bbox
+
+        >>> outputs = model(**encoding)
+        ```"""
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.bros(
+            input_ids=input_ids,
+            bbox=bbox,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        last_hidden_states = outputs[0]
+        last_hidden_states = last_hidden_states.transpose(0, 1).contiguous()
+
+        logits = self.entity_linker(last_hidden_states, last_hidden_states).squeeze(0)
+
+        loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+
+            batch_size, max_seq_length = attention_mask.shape
+            device = attention_mask.device
+
+            self_token_mask = torch.eye(max_seq_length, max_seq_length + 1).to(device=device, dtype=torch.bool)
+
+            mask = bbox_first_token_mask.view(-1)
+            bbox_first_token_mask = torch.cat(
+                [
+                    ~bbox_first_token_mask,
+                    torch.zeros([batch_size, 1], dtype=torch.bool, device=device),
+                ],
+                axis=1,
+            )
+            logits = logits.masked_fill(bbox_first_token_mask[:, None, :], torch.finfo(logits.dtype).min)
+            logits = logits.masked_fill(self_token_mask[None, :, :], torch.finfo(logits.dtype).min)
+
+            loss = loss_fct(logits.view(-1, max_seq_length + 1)[mask], labels.view(-1)[mask])
+
+        if not return_dict:
+            output = (logits,) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TokenClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+__all__ = [
+    "BrosPreTrainedModel",
+    "BrosModel",
+    "BrosForTokenClassification",
+    "BrosSpadeEEForTokenClassification",
+    "BrosSpadeELForTokenClassification",
+]

lib/python3.12/site-packages/transformers/models/bros/processing_bros.py

@@ -0,0 +1,112 @@
+# coding=utf-8
+# Copyright 2023 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 Bros.
+"""
+
+from typing import List, Optional, Union
+
+from ...processing_utils import ProcessorMixin
+from ...tokenization_utils_base import BatchEncoding, PaddingStrategy, PreTokenizedInput, TextInput, TruncationStrategy
+from ...utils import TensorType
+
+
+class BrosProcessor(ProcessorMixin):
+    r"""
+    Constructs a Bros processor which wraps a BERT tokenizer.
+
+    [`BrosProcessor`] offers all the functionalities of [`BertTokenizerFast`]. See the docstring of
+    [`~BrosProcessor.__call__`] and [`~BrosProcessor.decode`] for more information.
+
+    Args:
+        tokenizer (`BertTokenizerFast`, *optional*):
+            An instance of ['BertTokenizerFast`]. The tokenizer is a required input.
+    """
+
+    attributes = ["tokenizer"]
+    tokenizer_class = ("BertTokenizer", "BertTokenizerFast")
+
+    def __init__(self, tokenizer=None, **kwargs):
+        if tokenizer is None:
+            raise ValueError("You need to specify a `tokenizer`.")
+
+        super().__init__(tokenizer)
+
+    def __call__(
+        self,
+        text: Union[TextInput, PreTokenizedInput, List[TextInput], List[PreTokenizedInput]] = None,
+        add_special_tokens: bool = True,
+        padding: Union[bool, str, PaddingStrategy] = False,
+        truncation: Union[bool, str, TruncationStrategy] = None,
+        max_length: Optional[int] = None,
+        stride: int = 0,
+        pad_to_multiple_of: Optional[int] = None,
+        return_token_type_ids: Optional[bool] = None,
+        return_attention_mask: Optional[bool] = None,
+        return_overflowing_tokens: bool = False,
+        return_special_tokens_mask: bool = False,
+        return_offsets_mapping: bool = False,
+        return_length: bool = False,
+        verbose: bool = True,
+        return_tensors: Optional[Union[str, TensorType]] = None,
+        **kwargs,
+    ) -> BatchEncoding:
+        """
+        This method uses [`BertTokenizerFast.__call__`] to prepare text for the model.
+
+        Please refer to the docstring of the above two methods for more information.
+        """
+        encoding = self.tokenizer(
+            text=text,
+            add_special_tokens=add_special_tokens,
+            padding=padding,
+            truncation=truncation,
+            max_length=max_length,
+            stride=stride,
+            pad_to_multiple_of=pad_to_multiple_of,
+            return_token_type_ids=return_token_type_ids,
+            return_attention_mask=return_attention_mask,
+            return_overflowing_tokens=return_overflowing_tokens,
+            return_special_tokens_mask=return_special_tokens_mask,
+            return_offsets_mapping=return_offsets_mapping,
+            return_length=return_length,
+            verbose=verbose,
+            return_tensors=return_tensors,
+            **kwargs,
+        )
+
+        return encoding
+
+    def batch_decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to BertTokenizerFast'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 BertTokenizerFast'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
+        return list(dict.fromkeys(tokenizer_input_names))
+
+
+__all__ = ["BrosProcessor"]

lib/python3.12/site-packages/transformers/models/clip/__init__.py

@@ -0,0 +1,35 @@
+# 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 _LazyModule
+from ...utils.import_utils import define_import_structure
+
+
+if TYPE_CHECKING:
+    from .configuration_clip import *
+    from .feature_extraction_clip import *
+    from .image_processing_clip import *
+    from .image_processing_clip_fast import *
+    from .modeling_clip import *
+    from .modeling_flax_clip import *
+    from .modeling_tf_clip import *
+    from .processing_clip import *
+    from .tokenization_clip import *
+    from .tokenization_clip_fast import *
+else:
+    import sys
+
+    _file = globals()["__file__"]
+    sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)

lib/python3.12/site-packages/transformers/models/clip/configuration_clip.py

@@ -0,0 +1,422 @@
+# coding=utf-8
+# Copyright 2021 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.
+"""CLIP model configuration"""
+
+from collections import OrderedDict
+from typing import TYPE_CHECKING, Any, Mapping, Optional
+
+
+if TYPE_CHECKING:
+    from ...processing_utils import ProcessorMixin
+    from ...utils import TensorType
+
+from ...configuration_utils import PretrainedConfig
+from ...onnx import OnnxConfig
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class CLIPTextConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`CLIPTextModel`]. It is used to instantiate a CLIP
+    text 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 text encoder of the CLIP
+    [openai/clip-vit-base-patch32](https://huggingface.co/openai/clip-vit-base-patch32) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 49408):
+            Vocabulary size of the CLIP text model. Defines the number of different tokens that can be represented by
+            the `inputs_ids` passed when calling [`CLIPModel`].
+        hidden_size (`int`, *optional*, defaults to 512):
+            Dimensionality of the encoder layers and the pooler layer.
+        intermediate_size (`int`, *optional*, defaults to 2048):
+            Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
+        projection_dim (`int`, *optional*, defaults to 512):
+            Dimensionality of text and vision projection layers.
+        num_hidden_layers (`int`, *optional*, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 8):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        max_position_embeddings (`int`, *optional*, defaults to 77):
+            The maximum sequence length that this model might ever be used with. Typically set this to something large
+            just in case (e.g., 512 or 1024 or 2048).
+        hidden_act (`str` or `function`, *optional*, defaults to `"quick_gelu"`):
+            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-05):
+            The epsilon used by the layer normalization layers.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        initializer_factor (`float`, *optional*, defaults to 1.0):
+            A factor for initializing all weight matrices (should be kept to 1, used internally for initialization
+            testing).
+        pad_token_id (`int`, *optional*, defaults to 1):
+            Padding token id.
+        bos_token_id (`int`, *optional*, defaults to 49406):
+            Beginning of stream token id.
+        eos_token_id (`int`, *optional*, defaults to 49407):
+            End of stream token id.
+
+    Example:
+
+    ```python
+    >>> from transformers import CLIPTextConfig, CLIPTextModel
+
+    >>> # Initializing a CLIPTextConfig with openai/clip-vit-base-patch32 style configuration
+    >>> configuration = CLIPTextConfig()
+
+    >>> # Initializing a CLIPTextModel (with random weights) from the openai/clip-vit-base-patch32 style configuration
+    >>> model = CLIPTextModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "clip_text_model"
+    base_config_key = "text_config"
+
+    def __init__(
+        self,
+        vocab_size=49408,
+        hidden_size=512,
+        intermediate_size=2048,
+        projection_dim=512,
+        num_hidden_layers=12,
+        num_attention_heads=8,
+        max_position_embeddings=77,
+        hidden_act="quick_gelu",
+        layer_norm_eps=1e-5,
+        attention_dropout=0.0,
+        initializer_range=0.02,
+        initializer_factor=1.0,
+        # This differs from `CLIPTokenizer`'s default and from openai/clip
+        # See https://github.com/huggingface/transformers/pull/24773#issuecomment-1632287538
+        pad_token_id=1,
+        bos_token_id=49406,
+        eos_token_id=49407,
+        **kwargs,
+    ):
+        super().__init__(pad_token_id=pad_token_id, bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)
+
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.projection_dim = projection_dim
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.max_position_embeddings = max_position_embeddings
+        self.layer_norm_eps = layer_norm_eps
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.initializer_factor = initializer_factor
+        self.attention_dropout = attention_dropout
+
+
+class CLIPVisionConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`CLIPVisionModel`]. It is used to instantiate a
+    CLIP 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 vision encoder of the CLIP
+    [openai/clip-vit-base-patch32](https://huggingface.co/openai/clip-vit-base-patch32) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        hidden_size (`int`, *optional*, defaults to 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.
+        projection_dim (`int`, *optional*, defaults to 512):
+            Dimensionality of text and vision projection layers.
+        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):
+            The number of input channels.
+        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 `"quick_gelu"`):
+            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-05):
+            The epsilon used by the layer normalization layers.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        initializer_factor (`float`, *optional*, defaults to 1.0):
+            A factor for initializing all weight matrices (should be kept to 1, used internally for initialization
+            testing).
+
+    Example:
+
+    ```python
+    >>> from transformers import CLIPVisionConfig, CLIPVisionModel
+
+    >>> # Initializing a CLIPVisionConfig with openai/clip-vit-base-patch32 style configuration
+    >>> configuration = CLIPVisionConfig()
+
+    >>> # Initializing a CLIPVisionModel (with random weights) from the openai/clip-vit-base-patch32 style configuration
+    >>> model = CLIPVisionModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "clip_vision_model"
+    base_config_key = "vision_config"
+
+    def __init__(
+        self,
+        hidden_size=768,
+        intermediate_size=3072,
+        projection_dim=512,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        num_channels=3,
+        image_size=224,
+        patch_size=32,
+        hidden_act="quick_gelu",
+        layer_norm_eps=1e-5,
+        attention_dropout=0.0,
+        initializer_range=0.02,
+        initializer_factor=1.0,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.projection_dim = projection_dim
+        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.initializer_range = initializer_range
+        self.initializer_factor = initializer_factor
+        self.attention_dropout = attention_dropout
+        self.layer_norm_eps = layer_norm_eps
+        self.hidden_act = hidden_act
+
+
+class CLIPConfig(PretrainedConfig):
+    r"""
+    [`CLIPConfig`] is the configuration class to store the configuration of a [`CLIPModel`]. It is used to instantiate
+    a CLIP model according to the specified arguments, defining the text model and vision model configs. Instantiating
+    a configuration with the defaults will yield a similar configuration to that of the CLIP
+    [openai/clip-vit-base-patch32](https://huggingface.co/openai/clip-vit-base-patch32) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        text_config (`dict`, *optional*):
+            Dictionary of configuration options used to initialize [`CLIPTextConfig`].
+        vision_config (`dict`, *optional*):
+            Dictionary of configuration options used to initialize [`CLIPVisionConfig`].
+        projection_dim (`int`, *optional*, defaults to 512):
+            Dimensionality of text and vision projection layers.
+        logit_scale_init_value (`float`, *optional*, defaults to 2.6592):
+            The initial value of the *logit_scale* parameter. Default is used as per the original CLIP implementation.
+        kwargs (*optional*):
+            Dictionary of keyword arguments.
+
+    Example:
+
+    ```python
+    >>> from transformers import CLIPConfig, CLIPModel
+
+    >>> # Initializing a CLIPConfig with openai/clip-vit-base-patch32 style configuration
+    >>> configuration = CLIPConfig()
+
+    >>> # Initializing a CLIPModel (with random weights) from the openai/clip-vit-base-patch32 style configuration
+    >>> model = CLIPModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+
+    >>> # We can also initialize a CLIPConfig from a CLIPTextConfig and a CLIPVisionConfig
+    >>> from transformers import CLIPTextConfig, CLIPVisionConfig
+
+    >>> # Initializing a CLIPText and CLIPVision configuration
+    >>> config_text = CLIPTextConfig()
+    >>> config_vision = CLIPVisionConfig()
+
+    >>> config = CLIPConfig.from_text_vision_configs(config_text, config_vision)
+    ```"""
+
+    model_type = "clip"
+    sub_configs = {"text_config": CLIPTextConfig, "vision_config": CLIPVisionConfig}
+
+    def __init__(
+        self, text_config=None, vision_config=None, projection_dim=512, logit_scale_init_value=2.6592, **kwargs
+    ):
+        # If `_config_dict` exist, we use them for the backward compatibility.
+        # We pop out these 2 attributes before calling `super().__init__` to avoid them being saved (which causes a lot
+        # of confusion!).
+        text_config_dict = kwargs.pop("text_config_dict", None)
+        vision_config_dict = kwargs.pop("vision_config_dict", None)
+
+        super().__init__(**kwargs)
+
+        # Instead of simply assigning `[text|vision]_config_dict` to `[text|vision]_config`, we use the values in
+        # `[text|vision]_config_dict` to update the values in `[text|vision]_config`. The values should be same in most
+        # cases, but we don't want to break anything regarding `_config_dict` that existed before commit `8827e1b2`.
+        if text_config_dict is not None:
+            if text_config is None:
+                text_config = {}
+
+            # This is the complete result when using `text_config_dict`.
+            _text_config_dict = CLIPTextConfig(**text_config_dict).to_dict()
+
+            # Give a warning if the values exist in both `_text_config_dict` and `text_config` but being different.
+            for key, value in _text_config_dict.items():
+                if key in text_config and value != text_config[key] and key not in ["transformers_version"]:
+                    # If specified in `text_config_dict`
+                    if key in text_config_dict:
+                        message = (
+                            f"`{key}` is found in both `text_config_dict` and `text_config` but with different values. "
+                            f'The value `text_config_dict["{key}"]` will be used instead.'
+                        )
+                    # If inferred from default argument values (just to be super careful)
+                    else:
+                        message = (
+                            f"`text_config_dict` is provided which will be used to initialize `CLIPTextConfig`. The "
+                            f'value `text_config["{key}"]` will be overridden.'
+                        )
+                    logger.info(message)
+
+            # Update all values in `text_config` with the ones in `_text_config_dict`.
+            text_config.update(_text_config_dict)
+
+        if vision_config_dict is not None:
+            if vision_config is None:
+                vision_config = {}
+
+            # This is the complete result when using `vision_config_dict`.
+            _vision_config_dict = CLIPVisionConfig(**vision_config_dict).to_dict()
+            # convert keys to string instead of integer
+            if "id2label" in _vision_config_dict:
+                _vision_config_dict["id2label"] = {
+                    str(key): value for key, value in _vision_config_dict["id2label"].items()
+                }
+
+            # Give a warning if the values exist in both `_vision_config_dict` and `vision_config` but being different.
+            for key, value in _vision_config_dict.items():
+                if key in vision_config and value != vision_config[key] and key not in ["transformers_version"]:
+                    # If specified in `vision_config_dict`
+                    if key in vision_config_dict:
+                        message = (
+                            f"`{key}` is found in both `vision_config_dict` and `vision_config` but with different "
+                            f'values. The value `vision_config_dict["{key}"]` will be used instead.'
+                        )
+                    # If inferred from default argument values (just to be super careful)
+                    else:
+                        message = (
+                            f"`vision_config_dict` is provided which will be used to initialize `CLIPVisionConfig`. "
+                            f'The value `vision_config["{key}"]` will be overridden.'
+                        )
+                    logger.info(message)
+
+            # Update all values in `vision_config` with the ones in `_vision_config_dict`.
+            vision_config.update(_vision_config_dict)
+
+        if text_config is None:
+            text_config = {}
+            logger.info("`text_config` is `None`. Initializing the `CLIPTextConfig` with default values.")
+
+        if vision_config is None:
+            vision_config = {}
+            logger.info("`vision_config` is `None`. initializing the `CLIPVisionConfig` with default values.")
+
+        self.text_config = CLIPTextConfig(**text_config)
+        self.vision_config = CLIPVisionConfig(**vision_config)
+
+        self.projection_dim = projection_dim
+        self.logit_scale_init_value = logit_scale_init_value
+        self.initializer_factor = 1.0
+
+    @classmethod
+    def from_text_vision_configs(cls, text_config: CLIPTextConfig, vision_config: CLIPVisionConfig, **kwargs):
+        r"""
+        Instantiate a [`CLIPConfig`] (or a derived class) from clip text model configuration and clip vision model
+        configuration.
+
+        Returns:
+            [`CLIPConfig`]: An instance of a configuration object
+        """
+
+        return cls(text_config=text_config.to_dict(), vision_config=vision_config.to_dict(), **kwargs)
+
+
+class CLIPOnnxConfig(OnnxConfig):
+    @property
+    def inputs(self) -> Mapping[str, Mapping[int, str]]:
+        return OrderedDict(
+            [
+                ("input_ids", {0: "batch", 1: "sequence"}),
+                ("pixel_values", {0: "batch", 1: "num_channels", 2: "height", 3: "width"}),
+                ("attention_mask", {0: "batch", 1: "sequence"}),
+            ]
+        )
+
+    @property
+    def outputs(self) -> Mapping[str, Mapping[int, str]]:
+        return OrderedDict(
+            [
+                ("logits_per_image", {0: "batch"}),
+                ("logits_per_text", {0: "batch"}),
+                ("text_embeds", {0: "batch"}),
+                ("image_embeds", {0: "batch"}),
+            ]
+        )
+
+    @property
+    def atol_for_validation(self) -> float:
+        return 1e-4
+
+    def generate_dummy_inputs(
+        self,
+        processor: "ProcessorMixin",
+        batch_size: int = -1,
+        seq_length: int = -1,
+        framework: Optional["TensorType"] = None,
+    ) -> Mapping[str, Any]:
+        text_input_dict = super().generate_dummy_inputs(
+            processor.tokenizer, batch_size=batch_size, seq_length=seq_length, framework=framework
+        )
+        image_input_dict = super().generate_dummy_inputs(
+            processor.image_processor, batch_size=batch_size, framework=framework
+        )
+        return {**text_input_dict, **image_input_dict}
+
+    @property
+    def default_onnx_opset(self) -> int:
+        return 14
+
+
+__all__ = ["CLIPConfig", "CLIPOnnxConfig", "CLIPTextConfig", "CLIPVisionConfig"]

lib/python3.12/site-packages/transformers/models/clip/feature_extraction_clip.py

@@ -0,0 +1,36 @@
+# coding=utf-8
+# Copyright 2021 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.
+"""Feature extractor class for CLIP."""
+
+import warnings
+
+from ...utils import logging
+from .image_processing_clip import CLIPImageProcessor
+
+
+logger = logging.get_logger(__name__)
+
+
+class CLIPFeatureExtractor(CLIPImageProcessor):
+    def __init__(self, *args, **kwargs) -> None:
+        warnings.warn(
+            "The class CLIPFeatureExtractor is deprecated and will be removed in version 5 of Transformers. Please"
+            " use CLIPImageProcessor instead.",
+            FutureWarning,
+        )
+        super().__init__(*args, **kwargs)
+
+
+__all__ = ["CLIPFeatureExtractor"]

lib/python3.12/site-packages/transformers/models/clip/image_processing_clip.py

@@ -0,0 +1,348 @@
+# coding=utf-8
+# Copyright 2022 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Image processor class for CLIP."""
+
+from typing import Dict, List, Optional, Union
+
+import numpy as np
+
+from ...image_processing_utils import BaseImageProcessor, BatchFeature, get_size_dict
+from ...image_transforms import (
+    convert_to_rgb,
+    get_resize_output_image_size,
+    resize,
+    to_channel_dimension_format,
+)
+from ...image_utils import (
+    OPENAI_CLIP_MEAN,
+    OPENAI_CLIP_STD,
+    ChannelDimension,
+    ImageInput,
+    PILImageResampling,
+    infer_channel_dimension_format,
+    is_scaled_image,
+    make_flat_list_of_images,
+    to_numpy_array,
+    valid_images,
+    validate_kwargs,
+    validate_preprocess_arguments,
+)
+from ...utils import TensorType, is_vision_available, logging
+
+
+logger = logging.get_logger(__name__)
+
+
+if is_vision_available():
+    import PIL
+
+
+class CLIPImageProcessor(BaseImageProcessor):
+    r"""
+    Constructs a CLIP image processor.
+
+    Args:
+        do_resize (`bool`, *optional*, defaults to `True`):
+            Whether to resize the image's (height, width) dimensions to the specified `size`. Can be overridden by
+            `do_resize` in the `preprocess` method.
+        size (`Dict[str, int]` *optional*, defaults to `{"shortest_edge": 224}`):
+            Size of the image after resizing. The shortest edge of the image is resized to size["shortest_edge"], with
+            the longest edge resized to keep the input aspect ratio. Can be overridden by `size` in the `preprocess`
+            method.
+        resample (`PILImageResampling`, *optional*, defaults to `Resampling.BICUBIC`):
+            Resampling filter to use if resizing the image. Can be overridden by `resample` in the `preprocess` method.
+        do_center_crop (`bool`, *optional*, defaults to `True`):
+            Whether to center crop the image to the specified `crop_size`. Can be overridden by `do_center_crop` in the
+            `preprocess` method.
+        crop_size (`Dict[str, int]` *optional*, defaults to 224):
+            Size of the output image after applying `center_crop`. Can be overridden by `crop_size` in the `preprocess`
+            method.
+        do_rescale (`bool`, *optional*, defaults to `True`):
+            Whether to rescale the image by the specified scale `rescale_factor`. Can be overridden by `do_rescale` in
+            the `preprocess` method.
+        rescale_factor (`int` or `float`, *optional*, defaults to `1/255`):
+            Scale factor to use if rescaling the image. Can be overridden by `rescale_factor` in the `preprocess`
+            method.
+        do_normalize (`bool`, *optional*, defaults to `True`):
+            Whether to normalize the image. Can be overridden by `do_normalize` in the `preprocess` method.
+        image_mean (`float` or `List[float]`, *optional*, defaults to `[0.48145466, 0.4578275, 0.40821073]`):
+            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.
+        image_std (`float` or `List[float]`, *optional*, defaults to `[0.26862954, 0.26130258, 0.27577711]`):
+            Standard deviation to use if normalizing the image. This is a float or list of floats the length of the
+            number of channels in the image. Can be overridden by the `image_std` parameter in the `preprocess` method.
+            Can be overridden by the `image_std` parameter in the `preprocess` method.
+        do_convert_rgb (`bool`, *optional*, defaults to `True`):
+            Whether to convert the image to RGB.
+    """
+
+    model_input_names = ["pixel_values"]
+
+    def __init__(
+        self,
+        do_resize: bool = True,
+        size: Dict[str, int] = None,
+        resample: PILImageResampling = PILImageResampling.BICUBIC,
+        do_center_crop: bool = True,
+        crop_size: Dict[str, int] = None,
+        do_rescale: bool = True,
+        rescale_factor: Union[int, float] = 1 / 255,
+        do_normalize: bool = True,
+        image_mean: Optional[Union[float, List[float]]] = None,
+        image_std: Optional[Union[float, List[float]]] = None,
+        do_convert_rgb: bool = True,
+        **kwargs,
+    ) -> None:
+        super().__init__(**kwargs)
+        size = size if size is not None else {"shortest_edge": 224}
+        size = get_size_dict(size, default_to_square=False)
+        crop_size = crop_size if crop_size is not None else {"height": 224, "width": 224}
+        crop_size = get_size_dict(crop_size, default_to_square=True, param_name="crop_size")
+
+        self.do_resize = do_resize
+        self.size = size
+        self.resample = resample
+        self.do_center_crop = do_center_crop
+        self.crop_size = crop_size
+        self.do_rescale = do_rescale
+        self.rescale_factor = rescale_factor
+        self.do_normalize = do_normalize
+        self.image_mean = image_mean if image_mean is not None else OPENAI_CLIP_MEAN
+        self.image_std = image_std if image_std is not None else OPENAI_CLIP_STD
+        self.do_convert_rgb = do_convert_rgb
+        self._valid_processor_keys = [
+            "images",
+            "do_resize",
+            "size",
+            "resample",
+            "do_center_crop",
+            "crop_size",
+            "do_rescale",
+            "rescale_factor",
+            "do_normalize",
+            "image_mean",
+            "image_std",
+            "do_convert_rgb",
+            "return_tensors",
+            "data_format",
+            "input_data_format",
+        ]
+
+        # for backwards compatibility of KOSMOS-2
+        if "use_square_size" in kwargs and kwargs["use_square_size"]:
+            self.size = {"height": size["shortest_edge"], "width": size["shortest_edge"]}
+            # Let's remove `use_square_size` (as it is removed from #27690), so the future Kosmos-2 image processors
+            # won't have this attr. being saved. (otherwise, it will enter this if branch while there is no more
+            # `shortest_edge` key.
+            delattr(self, "use_square_size")
+
+    def resize(
+        self,
+        image: np.ndarray,
+        size: Dict[str, int],
+        resample: PILImageResampling = PILImageResampling.BICUBIC,
+        data_format: Optional[Union[str, ChannelDimension]] = None,
+        input_data_format: Optional[Union[str, ChannelDimension]] = None,
+        **kwargs,
+    ) -> np.ndarray:
+        """
+        Resize an image. 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.
+        """
+        default_to_square = True
+        if "shortest_edge" in size:
+            size = size["shortest_edge"]
+            default_to_square = False
+        elif "height" in size and "width" in size:
+            size = (size["height"], size["width"])
+        else:
+            raise ValueError("Size must contain either 'shortest_edge' or 'height' and 'width'.")
+
+        output_size = get_resize_output_image_size(
+            image,
+            size=size,
+            default_to_square=default_to_square,
+            input_data_format=input_data_format,
+        )
+        return resize(
+            image,
+            size=output_size,
+            resample=resample,
+            data_format=data_format,
+            input_data_format=input_data_format,
+            **kwargs,
+        )
+
+    def preprocess(
+        self,
+        images: ImageInput,
+        do_resize: Optional[bool] = None,
+        size: Dict[str, int] = None,
+        resample: PILImageResampling = None,
+        do_center_crop: Optional[bool] = None,
+        crop_size: Optional[int] = None,
+        do_rescale: Optional[bool] = None,
+        rescale_factor: Optional[float] = None,
+        do_normalize: Optional[bool] = None,
+        image_mean: Optional[Union[float, List[float]]] = None,
+        image_std: Optional[Union[float, List[float]]] = None,
+        do_convert_rgb: Optional[bool] = None,
+        return_tensors: Optional[Union[str, TensorType]] = None,
+        data_format: Optional[ChannelDimension] = ChannelDimension.FIRST,
+        input_data_format: Optional[Union[str, ChannelDimension]] = None,
+        **kwargs,
+    ) -> PIL.Image.Image:
+        """
+        Preprocess an image or batch of images.
+
+        Args:
+            images (`ImageInput`):
+                Image to preprocess. Expects a single or batch of images with pixel values ranging from 0 to 255. If
+                passing in images with pixel values between 0 and 1, set `do_rescale=False`.
+            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_center_crop (`bool`, *optional*, defaults to `self.do_center_crop`):
+                Whether to center crop the image.
+            crop_size (`Dict[str, int]`, *optional*, defaults to `self.crop_size`):
+                Size of the center crop. Only has an effect if `do_center_crop` 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_convert_rgb (`bool`, *optional*, defaults to `self.do_convert_rgb`):
+                Whether to convert the image to RGB.
+            return_tensors (`str` or `TensorType`, *optional*):
+                The type of tensors to return. Can be one of:
+                - Unset: Return a list of `np.ndarray`.
+                - `TensorType.TENSORFLOW` or `'tf'`: Return a batch of type `tf.Tensor`.
+                - `TensorType.PYTORCH` or `'pt'`: Return a batch of type `torch.Tensor`.
+                - `TensorType.NUMPY` or `'np'`: Return a batch of type `np.ndarray`.
+                - `TensorType.JAX` or `'jax'`: Return a batch of type `jax.numpy.ndarray`.
+            data_format (`ChannelDimension` or `str`, *optional*, defaults to `ChannelDimension.FIRST`):
+                The channel dimension format for the output image. Can be one of:
+                - `"channels_first"` or `ChannelDimension.FIRST`: image in (num_channels, height, width) format.
+                - `"channels_last"` or `ChannelDimension.LAST`: image in (height, width, num_channels) format.
+                - Unset: Use the channel dimension format of the input image.
+            input_data_format (`ChannelDimension` or `str`, *optional*):
+                The channel dimension format for the input image. If unset, the channel dimension format is inferred
+                from the input image. Can be one of:
+                - `"channels_first"` or `ChannelDimension.FIRST`: image in (num_channels, height, width) format.
+                - `"channels_last"` or `ChannelDimension.LAST`: image in (height, width, num_channels) format.
+                - `"none"` or `ChannelDimension.NONE`: image in (height, width) format.
+        """
+        do_resize = do_resize if do_resize is not None else self.do_resize
+        size = size if size is not None else self.size
+        size = get_size_dict(size, param_name="size", default_to_square=False)
+        resample = resample if resample is not None else self.resample
+        do_center_crop = do_center_crop if do_center_crop is not None else self.do_center_crop
+        crop_size = crop_size if crop_size is not None else self.crop_size
+        crop_size = get_size_dict(crop_size, param_name="crop_size", default_to_square=True)
+        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
+
+        validate_kwargs(captured_kwargs=kwargs.keys(), valid_processor_keys=self._valid_processor_keys)
+
+        images = make_flat_list_of_images(images)
+
+        if not valid_images(images):
+            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_center_crop=do_center_crop,
+            crop_size=crop_size,
+            do_resize=do_resize,
+            size=size,
+            resample=resample,
+        )
+
+        if do_convert_rgb:
+            images = [convert_to_rgb(image) for image in images]
+
+        # All transformations expect numpy arrays.
+        images = [to_numpy_array(image) for image in images]
+
+        if do_rescale and is_scaled_image(images[0]):
+            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[0])
+
+        all_images = []
+        for image in images:
+            if do_resize:
+                image = self.resize(image=image, size=size, resample=resample, input_data_format=input_data_format)
+
+            if do_center_crop:
+                image = self.center_crop(image=image, size=crop_size, input_data_format=input_data_format)
+
+            if do_rescale:
+                image = self.rescale(image=image, scale=rescale_factor, input_data_format=input_data_format)
+
+            if do_normalize:
+                image = self.normalize(
+                    image=image, mean=image_mean, std=image_std, input_data_format=input_data_format
+                )
+
+            all_images.append(image)
+        images = [
+            to_channel_dimension_format(image, data_format, input_channel_dim=input_data_format)
+            for image in all_images
+        ]
+
+        data = {"pixel_values": images}
+        return BatchFeature(data=data, tensor_type=return_tensors)
+
+
+__all__ = ["CLIPImageProcessor"]

lib/python3.12/site-packages/transformers/models/clip/image_processing_clip_fast.py

@@ -0,0 +1,42 @@
+# 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.
+"""Fast Image processor class for CLIP."""
+
+from ...image_processing_utils_fast import BASE_IMAGE_PROCESSOR_FAST_DOCSTRING, BaseImageProcessorFast
+from ...image_utils import OPENAI_CLIP_MEAN, OPENAI_CLIP_STD, PILImageResampling
+from ...utils import add_start_docstrings
+
+
+@add_start_docstrings(
+    "Constructs a fast CLIP image processor.",
+    BASE_IMAGE_PROCESSOR_FAST_DOCSTRING,
+)
+class CLIPImageProcessorFast(BaseImageProcessorFast):
+    # To be checked against the slow image processor
+    # None values left after checking can be removed
+    resample = PILImageResampling.BICUBIC
+    image_mean = OPENAI_CLIP_MEAN
+    image_std = OPENAI_CLIP_STD
+    size = {"shortest_edge": 224}
+    default_to_square = False
+    crop_size = {"height": 224, "width": 224}
+    do_resize = True
+    do_center_crop = True
+    do_rescale = True
+    do_normalize = True
+    do_convert_rgb = True
+
+
+__all__ = ["CLIPImageProcessorFast"]

lib/python3.12/site-packages/transformers/models/clip/modeling_clip.py

@@ -0,0 +1,1645 @@
+# coding=utf-8
+# Copyright 2021 The OpenAI Team Authors and 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.
+"""PyTorch CLIP model."""
+
+from dataclasses import dataclass
+from typing import Any, Optional, Tuple
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+
+from ...activations import ACT2FN
+from ...modeling_attn_mask_utils import _create_4d_causal_attention_mask, _prepare_4d_attention_mask
+from ...modeling_flash_attention_utils import flash_attn_supports_top_left_mask, is_flash_attn_available
+from ...modeling_outputs import BaseModelOutput, BaseModelOutputWithPooling, ImageClassifierOutput
+from ...modeling_utils import PreTrainedModel
+from ...pytorch_utils import is_torch_greater_or_equal_than_2_2
+from ...utils import (
+    ModelOutput,
+    add_code_sample_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    can_return_tuple,
+    logging,
+    replace_return_docstrings,
+    torch_int,
+)
+from .configuration_clip import CLIPConfig, CLIPTextConfig, CLIPVisionConfig
+
+
+if is_flash_attn_available():
+    from ...modeling_flash_attention_utils import _flash_attention_forward
+
+
+logger = logging.get_logger(__name__)
+
+# General docstring
+_CONFIG_FOR_DOC = "CLIPConfig"
+_CHECKPOINT_FOR_DOC = "openai/clip-vit-base-patch32"
+
+# Image classification docstring
+_IMAGE_CLASS_CHECKPOINT = "openai/clip-vit-base-patch32"
+_IMAGE_CLASS_EXPECTED_OUTPUT = "LABEL_0"
+
+
+# contrastive loss function, adapted from
+# https://sachinruk.github.io/blog/2021-03-07-clip.html
+def contrastive_loss(logits: torch.Tensor) -> torch.Tensor:
+    return nn.functional.cross_entropy(logits, torch.arange(len(logits), device=logits.device))
+
+
+def clip_loss(similarity: torch.Tensor) -> torch.Tensor:
+    caption_loss = contrastive_loss(similarity)
+    image_loss = contrastive_loss(similarity.t())
+    return (caption_loss + image_loss) / 2.0
+
+
+def _get_vector_norm(tensor: torch.Tensor) -> torch.Tensor:
+    """
+    This method is equivalent to tensor.norm(p=2, dim=-1, keepdim=True) and used to make
+    model `executorch` exportable. See issue https://github.com/pytorch/executorch/issues/3566
+    """
+    square_tensor = torch.pow(tensor, 2)
+    sum_tensor = torch.sum(square_tensor, dim=-1, keepdim=True)
+    normed_tensor = torch.pow(sum_tensor, 0.5)
+    return normed_tensor
+
+
+@dataclass
+class CLIPVisionModelOutput(ModelOutput):
+    """
+    Base class for vision model's outputs that also contains image embeddings of the pooling of the last hidden states.
+
+    Args:
+        image_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim)` *optional* returned when model is initialized with `with_projection=True`):
+            The image embeddings obtained by applying the projection layer to the pooler_output.
+        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.
+        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_embeds: Optional[torch.FloatTensor] = None
+    last_hidden_state: Optional[torch.FloatTensor] = None
+    hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
+    attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
+
+
+@dataclass
+class CLIPTextModelOutput(ModelOutput):
+    """
+    Base class for text model's outputs that also contains a pooling of the last hidden states.
+
+    Args:
+        text_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim)` *optional* returned when model is initialized with `with_projection=True`):
+            The text embeddings obtained by applying the projection layer to the pooler_output.
+        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.
+        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.
+    """
+
+    text_embeds: Optional[torch.FloatTensor] = None
+    last_hidden_state: Optional[torch.FloatTensor] = None
+    hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
+    attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
+
+
+@dataclass
+class CLIPOutput(ModelOutput):
+    """
+    Args:
+        loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `return_loss` is `True`):
+            Contrastive loss for image-text similarity.
+        logits_per_image (`torch.FloatTensor` of shape `(image_batch_size, text_batch_size)`):
+            The scaled dot product scores between `image_embeds` and `text_embeds`. This represents the image-text
+            similarity scores.
+        logits_per_text (`torch.FloatTensor` of shape `(text_batch_size, image_batch_size)`):
+            The scaled dot product scores between `text_embeds` and `image_embeds`. This represents the text-image
+            similarity scores.
+        text_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim`):
+            The text embeddings obtained by applying the projection layer to the pooled output of [`CLIPTextModel`].
+        image_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim`):
+            The image embeddings obtained by applying the projection layer to the pooled output of [`CLIPVisionModel`].
+        text_model_output (`BaseModelOutputWithPooling`):
+            The output of the [`CLIPTextModel`].
+        vision_model_output (`BaseModelOutputWithPooling`):
+            The output of the [`CLIPVisionModel`].
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    logits_per_image: Optional[torch.FloatTensor] = None
+    logits_per_text: Optional[torch.FloatTensor] = None
+    text_embeds: Optional[torch.FloatTensor] = None
+    image_embeds: Optional[torch.FloatTensor] = None
+    text_model_output: BaseModelOutputWithPooling = None
+    vision_model_output: BaseModelOutputWithPooling = None
+
+    def to_tuple(self) -> Tuple[Any]:
+        return tuple(
+            self[k] if k not in ["text_model_output", "vision_model_output"] else getattr(self, k).to_tuple()
+            for k in self.keys()
+        )
+
+
+class CLIPVisionEmbeddings(nn.Module):
+    def __init__(self, config: CLIPVisionConfig):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.image_size = config.image_size
+        self.patch_size = config.patch_size
+
+        self.class_embedding = nn.Parameter(torch.randn(self.embed_dim))
+
+        self.patch_embedding = nn.Conv2d(
+            in_channels=config.num_channels,
+            out_channels=self.embed_dim,
+            kernel_size=self.patch_size,
+            stride=self.patch_size,
+            bias=False,
+        )
+
+        self.num_patches = (self.image_size // self.patch_size) ** 2
+        self.num_positions = self.num_patches + 1
+        self.position_embedding = nn.Embedding(self.num_positions, self.embed_dim)
+        self.register_buffer("position_ids", torch.arange(self.num_positions).expand((1, -1)), persistent=False)
+
+    def interpolate_pos_encoding(self, embeddings: torch.Tensor, height: int, width: int) -> torch.Tensor:
+        """
+        This method allows to interpolate the pre-trained position encodings, to be able to use the model on higher resolution
+        images. This method is also adapted to support torch.jit tracing.
+
+        Adapted from:
+        - https://github.com/facebookresearch/dino/blob/de9ee3df6cf39fac952ab558447af1fa1365362a/vision_transformer.py#L174-L194, and
+        - https://github.com/facebookresearch/dinov2/blob/e1277af2ba9496fbadf7aec6eba56e8d882d1e35/dinov2/models/vision_transformer.py#L179-L211
+        """
+
+        num_patches = embeddings.shape[1] - 1
+        position_embedding = self.position_embedding.weight.unsqueeze(0)
+        num_positions = position_embedding.shape[1] - 1
+
+        # always interpolate when tracing to ensure the exported model works for dynamic input shapes
+        if not torch.jit.is_tracing() and num_patches == num_positions and height == width:
+            return self.position_embedding(self.position_ids)
+
+        class_pos_embed = position_embedding[:, :1]
+        patch_pos_embed = position_embedding[:, 1:]
+
+        dim = embeddings.shape[-1]
+
+        new_height = height // self.patch_size
+        new_width = width // self.patch_size
+
+        sqrt_num_positions = torch_int(num_positions**0.5)
+        patch_pos_embed = patch_pos_embed.reshape(1, sqrt_num_positions, sqrt_num_positions, dim)
+        patch_pos_embed = patch_pos_embed.permute(0, 3, 1, 2)
+
+        patch_pos_embed = nn.functional.interpolate(
+            patch_pos_embed,
+            size=(new_height, new_width),
+            mode="bicubic",
+            align_corners=False,
+        )
+
+        patch_pos_embed = patch_pos_embed.permute(0, 2, 3, 1).view(1, -1, dim)
+
+        return torch.cat((class_pos_embed, patch_pos_embed), dim=1)
+
+    def forward(self, pixel_values: torch.FloatTensor, interpolate_pos_encoding=False) -> torch.Tensor:
+        batch_size, _, height, width = pixel_values.shape
+        if not interpolate_pos_encoding and (height != self.image_size or width != self.image_size):
+            raise ValueError(
+                f"Input image size ({height}*{width}) doesn't match model ({self.image_size}*{self.image_size})."
+            )
+        target_dtype = self.patch_embedding.weight.dtype
+        patch_embeds = self.patch_embedding(pixel_values.to(dtype=target_dtype))  # shape = [*, width, grid, grid]
+        patch_embeds = patch_embeds.flatten(2).transpose(1, 2)
+
+        class_embeds = self.class_embedding.expand(batch_size, 1, -1)
+        embeddings = torch.cat([class_embeds, patch_embeds], dim=1)
+        if interpolate_pos_encoding:
+            embeddings = embeddings + self.interpolate_pos_encoding(embeddings, height, width)
+        else:
+            embeddings = embeddings + self.position_embedding(self.position_ids)
+        return embeddings
+
+
+class CLIPTextEmbeddings(nn.Module):
+    def __init__(self, config: CLIPTextConfig):
+        super().__init__()
+        embed_dim = config.hidden_size
+
+        self.token_embedding = nn.Embedding(config.vocab_size, embed_dim)
+        self.position_embedding = nn.Embedding(config.max_position_embeddings, embed_dim)
+
+        # position_ids (1, len position emb) is contiguous in memory and exported when serialized
+        self.register_buffer(
+            "position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)), persistent=False
+        )
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+    ) -> torch.Tensor:
+        seq_length = input_ids.shape[-1] if input_ids is not None else inputs_embeds.shape[-2]
+        max_position_embedding = self.position_embedding.weight.shape[0]
+
+        if seq_length > max_position_embedding:
+            raise ValueError(
+                f"Sequence length must be less than max_position_embeddings (got `sequence length`: "
+                f"{seq_length} and max_position_embeddings: {max_position_embedding}"
+            )
+
+        if position_ids is None:
+            position_ids = self.position_ids[:, :seq_length]
+
+        if inputs_embeds is None:
+            inputs_embeds = self.token_embedding(input_ids)
+
+        position_embeddings = self.position_embedding(position_ids)
+        embeddings = inputs_embeds + position_embeddings
+
+        return embeddings
+
+
+class CLIPAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.embed_dim // self.num_heads
+        if self.head_dim * self.num_heads != self.embed_dim:
+            raise ValueError(
+                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:"
+                f" {self.num_heads})."
+            )
+        self.scale = self.head_dim**-0.5
+        self.dropout = 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)
+
+    def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
+        return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        causal_attention_mask: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
+        """Input shape: Batch x Time x Channel"""
+
+        bsz, tgt_len, embed_dim = hidden_states.size()
+
+        # get query proj
+        query_states = self.q_proj(hidden_states) * self.scale
+        key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
+        value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
+
+        proj_shape = (bsz * self.num_heads, -1, self.head_dim)
+        query_states = self._shape(query_states, tgt_len, bsz).view(*proj_shape)
+        key_states = key_states.view(*proj_shape)
+        value_states = value_states.view(*proj_shape)
+
+        src_len = key_states.size(1)
+        attn_weights = torch.bmm(query_states, key_states.transpose(1, 2))
+
+        if attn_weights.size() != (bsz * self.num_heads, tgt_len, src_len):
+            raise ValueError(
+                f"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is"
+                f" {attn_weights.size()}"
+            )
+
+        # apply the causal_attention_mask first
+        if causal_attention_mask is not None:
+            if causal_attention_mask.size() != (bsz, 1, tgt_len, src_len):
+                raise ValueError(
+                    f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is"
+                    f" {causal_attention_mask.size()}"
+                )
+            attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + causal_attention_mask
+            attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
+
+        if attention_mask is not None:
+            if attention_mask.size() != (bsz, 1, tgt_len, src_len):
+                raise ValueError(
+                    f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.size()}"
+                )
+            attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + attention_mask
+            attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
+
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1)
+
+        if output_attentions:
+            # this operation is a bit awkward, but it's required to
+            # make sure that attn_weights keeps its gradient.
+            # In order to do so, attn_weights have to reshaped
+            # twice and have to be reused in the following
+            attn_weights_reshaped = attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
+            attn_weights = attn_weights_reshaped.view(bsz * self.num_heads, tgt_len, src_len)
+        else:
+            attn_weights_reshaped = None
+
+        attn_probs = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)
+
+        attn_output = torch.bmm(attn_probs, value_states)
+
+        if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
+                f" {attn_output.size()}"
+            )
+
+        attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
+        attn_output = attn_output.transpose(1, 2)
+        attn_output = attn_output.reshape(bsz, tgt_len, embed_dim)
+
+        attn_output = self.out_proj(attn_output)
+
+        return attn_output, attn_weights_reshaped
+
+
+class CLIPFlashAttention2(CLIPAttention):
+    """
+    CLIPAttention flash attention module. This module inherits from `CLIPAttention` 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.
+    """
+
+    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 alignment, 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 = flash_attn_supports_top_left_mask()
+
+    # Adapted from transformers.models.llama.modeling_llama.LlamaFlashAttention2.forward
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        causal_attention_mask: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
+        output_attentions = False
+
+        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)
+
+        # 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(batch_size, q_len, self.num_heads, self.head_dim)
+        key_states = key_states.view(batch_size, q_len, self.num_heads, self.head_dim)
+        value_states = value_states.view(batch_size, q_len, self.num_heads, self.head_dim)
+
+        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.
+
+        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 = _flash_attention_forward(
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            q_len,
+            dropout=dropout_rate,
+            is_causal=causal_attention_mask is not None,
+            use_top_left_mask=self._flash_attn_uses_top_left_mask,
+        )
+
+        attn_output = attn_output.reshape(batch_size, 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
+
+
+class CLIPSdpaAttention(CLIPAttention):
+    """
+    SDPA attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
+    `CLIPAttention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
+    SDPA API.
+    """
+
+    # Adapted from CLIPAttention.forward
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        causal_attention_mask: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
+        if output_attentions:
+            # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
+            logger.warning_once(
+                "CLIPModel is using CLIPSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not "
+                "support `output_attentions=True`. Falling back to the manual attention implementation, but specifying "
+                "the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can "
+                'be removed using the argument `attn_implementation="eager"` when loading the model.'
+            )
+            return super().forward(
+                hidden_states=hidden_states,
+                attention_mask=attention_mask,
+                causal_attention_mask=causal_attention_mask,
+                output_attentions=output_attentions,
+            )
+
+        # CLIP text model uses both `causal_attention_mask` and `attention_mask`
+        if attention_mask is not None and causal_attention_mask is not None:
+            attn_mask = attention_mask + causal_attention_mask
+        elif causal_attention_mask is not None:
+            attn_mask = causal_attention_mask
+        else:
+            attn_mask = attention_mask
+
+        bsz, tgt_len, embed_dim = 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(bsz, -1, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, -1, self.num_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, -1, self.num_heads, self.head_dim).transpose(1, 2)
+
+        # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
+        # Reference: https://github.com/pytorch/pytorch/issues/112577.
+        if not is_torch_greater_or_equal_than_2_2 and query_states.device.type == "cuda" and attn_mask is not None:
+            query_states = query_states.contiguous()
+            key_states = key_states.contiguous()
+            value_states = value_states.contiguous()
+
+        # CLIP text model uses both `causal_attention_mask` and `attention_mask` sequentially.
+        attn_output = torch.nn.functional.scaled_dot_product_attention(
+            query_states,
+            key_states,
+            value_states,
+            attn_mask=attn_mask,
+            dropout_p=self.dropout if self.training else 0.0,
+            scale=self.scale,
+        )
+
+        attn_output = attn_output.transpose(1, 2)
+        attn_output = attn_output.reshape(bsz, tgt_len, embed_dim)
+
+        attn_output = self.out_proj(attn_output)
+
+        return attn_output, None
+
+
+CLIP_ATTENTION_CLASSES = {
+    "eager": CLIPAttention,
+    "sdpa": CLIPSdpaAttention,
+    "flash_attention_2": CLIPFlashAttention2,
+}
+
+
+class CLIPMLP(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 CLIPEncoderLayer(nn.Module):
+    def __init__(self, config: CLIPConfig):
+        super().__init__()
+        self.embed_dim = config.hidden_size
+        self.self_attn = CLIP_ATTENTION_CLASSES[config._attn_implementation](config)
+        self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
+        self.mlp = CLIPMLP(config)
+        self.layer_norm2 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor,
+        causal_attention_mask: torch.Tensor,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.FloatTensor]:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+            attention_mask (`torch.FloatTensor`): attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+                `(config.encoder_attention_heads,)`.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+        """
+        residual = hidden_states
+
+        hidden_states = self.layer_norm1(hidden_states)
+        hidden_states, attn_weights = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            causal_attention_mask=causal_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
+
+
+class CLIPPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = CLIPConfig
+    base_model_prefix = "clip"
+    supports_gradient_checkpointing = True
+    _supports_sdpa = True
+    _supports_flash_attn_2 = True
+
+    def _init_weights(self, module):
+        """Initialize the weights"""
+        factor = self.config.initializer_factor
+        if isinstance(module, CLIPTextEmbeddings):
+            module.token_embedding.weight.data.normal_(mean=0.0, std=factor * 0.02)
+            module.position_embedding.weight.data.normal_(mean=0.0, std=factor * 0.02)
+        elif isinstance(module, CLIPVisionEmbeddings):
+            factor = self.config.initializer_factor
+            nn.init.normal_(module.class_embedding, mean=0.0, std=module.embed_dim**-0.5 * factor)
+            nn.init.normal_(module.patch_embedding.weight, std=module.config.initializer_range * factor)
+            nn.init.normal_(module.position_embedding.weight, std=module.config.initializer_range * factor)
+        elif isinstance(module, CLIPAttention):
+            factor = self.config.initializer_factor
+            in_proj_std = (module.embed_dim**-0.5) * ((2 * module.config.num_hidden_layers) ** -0.5) * factor
+            out_proj_std = (module.embed_dim**-0.5) * factor
+            nn.init.normal_(module.q_proj.weight, std=in_proj_std)
+            nn.init.normal_(module.k_proj.weight, std=in_proj_std)
+            nn.init.normal_(module.v_proj.weight, std=in_proj_std)
+            nn.init.normal_(module.out_proj.weight, std=out_proj_std)
+        elif isinstance(module, CLIPMLP):
+            factor = self.config.initializer_factor
+            in_proj_std = (module.config.hidden_size**-0.5) * ((2 * module.config.num_hidden_layers) ** -0.5) * factor
+            fc_std = (2 * module.config.hidden_size) ** -0.5 * factor
+            nn.init.normal_(module.fc1.weight, std=fc_std)
+            nn.init.normal_(module.fc2.weight, std=in_proj_std)
+        elif isinstance(module, CLIPModel):
+            nn.init.normal_(
+                module.text_projection.weight,
+                std=module.text_embed_dim**-0.5 * self.config.initializer_factor,
+            )
+            nn.init.normal_(
+                module.visual_projection.weight,
+                std=module.vision_embed_dim**-0.5 * self.config.initializer_factor,
+            )
+        elif isinstance(module, CLIPVisionModelWithProjection):
+            nn.init.normal_(
+                module.visual_projection.weight,
+                std=self.config.hidden_size**-0.5 * self.config.initializer_factor,
+            )
+        elif isinstance(module, CLIPTextModelWithProjection):
+            nn.init.normal_(
+                module.text_projection.weight,
+                std=self.config.hidden_size**-0.5 * self.config.initializer_factor,
+            )
+        elif isinstance(module, CLIPForImageClassification):
+            nn.init.normal_(
+                module.classifier.weight,
+                std=self.config.vision_config.hidden_size**-0.5 * self.config.initializer_factor,
+            )
+
+        if isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+        if isinstance(module, nn.Linear) and module.bias is not None:
+            module.bias.data.zero_()
+
+
+CLIP_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 ([`CLIPConfig`]): 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.
+"""
+
+CLIP_TEXT_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)
+        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.max_position_embeddings - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        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.
+"""
+
+CLIP_VISION_INPUTS_DOCSTRING = r"""
+    Args:
+        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            Pixel values. Padding will be ignored by default should you provide it. Pixel values can be obtained using
+            [`AutoImageProcessor`]. See [`CLIPImageProcessor.__call__`] for details.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        interpolate_pos_encoding (`bool`, *optional*, defaults `False`):
+            Whether to interpolate the pre-trained position encodings.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+CLIP_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)
+        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.max_position_embeddings - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            Pixel values. Padding will be ignored by default should you provide it. Pixel values can be obtained using
+            [`AutoImageProcessor`]. See [`CLIPImageProcessor.__call__`] for details.
+        return_loss (`bool`, *optional*):
+            Whether or not to return the contrastive loss.
+        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.
+        interpolate_pos_encoding (`bool`, *optional*, defaults `False`):
+            Whether to interpolate the pre-trained position encodings.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+class CLIPEncoder(nn.Module):
+    """
+    Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a
+    [`CLIPEncoderLayer`].
+
+    Args:
+        config: CLIPConfig
+    """
+
+    def __init__(self, config: CLIPConfig):
+        super().__init__()
+        self.config = config
+        self.layers = nn.ModuleList([CLIPEncoderLayer(config) for _ in range(config.num_hidden_layers)])
+        self.gradient_checkpointing = False
+
+    @can_return_tuple
+    def forward(
+        self,
+        inputs_embeds,
+        attention_mask: Optional[torch.Tensor] = None,
+        causal_attention_mask: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+    ) -> 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)
+            causal_attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Causal mask for the text model. 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
+        )
+
+        encoder_states = () if output_hidden_states else None
+        all_attentions = () if output_attentions else None
+
+        hidden_states = inputs_embeds
+        for idx, encoder_layer in enumerate(self.layers):
+            if output_hidden_states:
+                encoder_states = encoder_states + (hidden_states,)
+            if self.gradient_checkpointing and self.training:
+                layer_outputs = self._gradient_checkpointing_func(
+                    encoder_layer.__call__,
+                    hidden_states,
+                    attention_mask,
+                    causal_attention_mask,
+                    output_attentions,
+                )
+            else:
+                layer_outputs = encoder_layer(
+                    hidden_states,
+                    attention_mask,
+                    causal_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,)
+
+        return BaseModelOutput(
+            last_hidden_state=hidden_states,
+            hidden_states=encoder_states,
+            attentions=all_attentions,
+        )
+
+
+class CLIPTextTransformer(nn.Module):
+    def __init__(self, config: CLIPTextConfig):
+        super().__init__()
+        self.config = config
+        embed_dim = config.hidden_size
+        self.embeddings = CLIPTextEmbeddings(config)
+        self.encoder = CLIPEncoder(config)
+        self.final_layer_norm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
+
+        # For `pooled_output` computation
+        self.eos_token_id = config.eos_token_id
+
+        # For attention mask, it differs between `flash_attention_2` and other attention implementations
+        self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"
+
+    @can_return_tuple
+    @add_start_docstrings_to_model_forward(CLIP_TEXT_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=CLIPTextConfig)
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+    ) -> BaseModelOutputWithPooling:
+        r"""
+        Returns:
+
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+
+        if input_ids is None:
+            raise ValueError("You have to specify input_ids")
+
+        input_shape = input_ids.size()
+        input_ids = input_ids.view(-1, input_shape[-1])
+
+        hidden_states = self.embeddings(input_ids=input_ids, position_ids=position_ids)
+
+        # CLIP's text model uses causal mask, prepare it here.
+        # https://github.com/openai/CLIP/blob/cfcffb90e69f37bf2ff1e988237a0fbe41f33c04/clip/model.py#L324
+        causal_attention_mask = _create_4d_causal_attention_mask(
+            input_shape, hidden_states.dtype, device=hidden_states.device
+        )
+
+        # expand attention_mask
+        if attention_mask is not None and not self._use_flash_attention_2:
+            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+            attention_mask = _prepare_4d_attention_mask(attention_mask, hidden_states.dtype)
+
+        encoder_outputs: BaseModelOutput = self.encoder(
+            inputs_embeds=hidden_states,
+            attention_mask=attention_mask,
+            causal_attention_mask=causal_attention_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+        )
+
+        last_hidden_state = encoder_outputs.last_hidden_state
+        last_hidden_state = self.final_layer_norm(last_hidden_state)
+
+        if self.eos_token_id == 2:
+            # The `eos_token_id` was incorrect before PR #24773: Let's keep what have been done here.
+            # A CLIP model with such `eos_token_id` in the config can't work correctly with extra new tokens added
+            # ------------------------------------------------------------
+            # text_embeds.shape = [batch_size, sequence_length, transformer.width]
+            # take features from the eot embedding (eot_token is the highest number in each sequence)
+            # casting to torch.int for onnx compatibility: argmax doesn't support int64 inputs with opset 14
+            pooled_output = last_hidden_state[
+                torch.arange(last_hidden_state.shape[0], device=last_hidden_state.device),
+                input_ids.to(dtype=torch.int, device=last_hidden_state.device).argmax(dim=-1),
+            ]
+        else:
+            # The config gets updated `eos_token_id` from PR #24773 (so the use of exta new tokens is possible)
+            pooled_output = last_hidden_state[
+                torch.arange(last_hidden_state.shape[0], device=last_hidden_state.device),
+                # We need to get the first position of `eos_token_id` value (`pad_token_ids` might equal to `eos_token_id`)
+                # Note: we assume each sequence (along batch dim.) contains an  `eos_token_id` (e.g. prepared by the tokenizer)
+                (input_ids.to(dtype=torch.int, device=last_hidden_state.device) == self.eos_token_id)
+                .int()
+                .argmax(dim=-1),
+            ]
+
+        return BaseModelOutputWithPooling(
+            last_hidden_state=last_hidden_state,
+            pooler_output=pooled_output,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """The text model from CLIP without any head or projection on top.""",
+    CLIP_START_DOCSTRING,
+)
+class CLIPTextModel(CLIPPreTrainedModel):
+    config_class = CLIPTextConfig
+
+    _no_split_modules = ["CLIPTextEmbeddings", "CLIPEncoderLayer"]
+
+    def __init__(self, config: CLIPTextConfig):
+        super().__init__(config)
+        self.text_model = CLIPTextTransformer(config)
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self) -> nn.Module:
+        return self.text_model.embeddings.token_embedding
+
+    def set_input_embeddings(self, value):
+        self.text_model.embeddings.token_embedding = value
+
+    @can_return_tuple
+    @add_start_docstrings_to_model_forward(CLIP_TEXT_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=CLIPTextConfig)
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+    ) -> BaseModelOutputWithPooling:
+        r"""
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from transformers import AutoTokenizer, CLIPTextModel
+
+        >>> model = CLIPTextModel.from_pretrained("openai/clip-vit-base-patch32")
+        >>> tokenizer = AutoTokenizer.from_pretrained("openai/clip-vit-base-patch32")
+
+        >>> inputs = tokenizer(["a photo of a cat", "a photo of a dog"], padding=True, return_tensors="pt")
+
+        >>> outputs = model(**inputs)
+        >>> last_hidden_state = outputs.last_hidden_state
+        >>> pooled_output = outputs.pooler_output  # pooled (EOS token) states
+        ```"""
+
+        return self.text_model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+        )
+
+
+class CLIPVisionTransformer(nn.Module):
+    def __init__(self, config: CLIPVisionConfig):
+        super().__init__()
+        self.config = config
+        embed_dim = config.hidden_size
+
+        self.embeddings = CLIPVisionEmbeddings(config)
+        self.pre_layrnorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
+        self.encoder = CLIPEncoder(config)
+        self.post_layernorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
+
+    @can_return_tuple
+    @add_start_docstrings_to_model_forward(CLIP_VISION_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=CLIPVisionConfig)
+    def forward(
+        self,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        interpolate_pos_encoding: Optional[bool] = False,
+    ) -> BaseModelOutputWithPooling:
+        r"""
+        Returns:
+
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+
+        if pixel_values is None:
+            raise ValueError("You have to specify pixel_values")
+
+        hidden_states = self.embeddings(pixel_values, interpolate_pos_encoding=interpolate_pos_encoding)
+        hidden_states = self.pre_layrnorm(hidden_states)
+
+        encoder_outputs: BaseModelOutput = self.encoder(
+            inputs_embeds=hidden_states,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+        )
+
+        last_hidden_state = encoder_outputs.last_hidden_state
+        pooled_output = last_hidden_state[:, 0, :]
+        pooled_output = self.post_layernorm(pooled_output)
+
+        return BaseModelOutputWithPooling(
+            last_hidden_state=last_hidden_state,
+            pooler_output=pooled_output,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """The vision model from CLIP without any head or projection on top.""",
+    CLIP_START_DOCSTRING,
+)
+class CLIPVisionModel(CLIPPreTrainedModel):
+    config_class = CLIPVisionConfig
+    main_input_name = "pixel_values"
+    _no_split_modules = ["CLIPEncoderLayer"]
+
+    def __init__(self, config: CLIPVisionConfig):
+        super().__init__(config)
+        self.vision_model = CLIPVisionTransformer(config)
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self) -> nn.Module:
+        return self.vision_model.embeddings.patch_embedding
+
+    @can_return_tuple
+    @add_start_docstrings_to_model_forward(CLIP_VISION_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=CLIPVisionConfig)
+    def forward(
+        self,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        interpolate_pos_encoding: bool = False,
+    ) -> BaseModelOutputWithPooling:
+        r"""
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from PIL import Image
+        >>> import requests
+        >>> from transformers import AutoProcessor, CLIPVisionModel
+
+        >>> model = CLIPVisionModel.from_pretrained("openai/clip-vit-base-patch32")
+        >>> processor = AutoProcessor.from_pretrained("openai/clip-vit-base-patch32")
+
+        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> image = Image.open(requests.get(url, stream=True).raw)
+
+        >>> inputs = processor(images=image, return_tensors="pt")
+
+        >>> outputs = model(**inputs)
+        >>> last_hidden_state = outputs.last_hidden_state
+        >>> pooled_output = outputs.pooler_output  # pooled CLS states
+        ```"""
+
+        return self.vision_model(
+            pixel_values=pixel_values,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            interpolate_pos_encoding=interpolate_pos_encoding,
+        )
+
+
+@add_start_docstrings(CLIP_START_DOCSTRING)
+class CLIPModel(CLIPPreTrainedModel):
+    config_class = CLIPConfig
+    _no_split_modules = ["CLIPTextEmbeddings", "CLIPEncoderLayer", "CLIPVisionEmbeddings"]
+
+    def __init__(self, config: CLIPConfig):
+        super().__init__(config)
+
+        if not isinstance(config.text_config, CLIPTextConfig):
+            raise TypeError(
+                "config.text_config is expected to be of type CLIPTextConfig but is of type"
+                f" {type(config.text_config)}."
+            )
+
+        if not isinstance(config.vision_config, CLIPVisionConfig):
+            raise TypeError(
+                "config.vision_config is expected to be of type CLIPVisionConfig but is of type"
+                f" {type(config.vision_config)}."
+            )
+
+        text_config = config.text_config
+        vision_config = config.vision_config
+
+        self.projection_dim = config.projection_dim
+        self.text_embed_dim = text_config.hidden_size
+        self.vision_embed_dim = vision_config.hidden_size
+
+        text_model = CLIPTextModel._from_config(text_config)
+        self.text_model = text_model.text_model
+
+        vision_model = CLIPVisionModel._from_config(vision_config)
+        self.vision_model = vision_model.vision_model
+
+        self.visual_projection = nn.Linear(self.vision_embed_dim, self.projection_dim, bias=False)
+        self.text_projection = nn.Linear(self.text_embed_dim, self.projection_dim, bias=False)
+        self.logit_scale = nn.Parameter(torch.tensor(self.config.logit_scale_init_value))
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(CLIP_TEXT_INPUTS_DOCSTRING)
+    def get_text_features(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+    ) -> torch.FloatTensor:
+        r"""
+        Returns:
+            text_features (`torch.FloatTensor` of shape `(batch_size, output_dim`): The text embeddings obtained by
+            applying the projection layer to the pooled output of [`CLIPTextModel`].
+
+        Examples:
+
+        ```python
+        >>> from transformers import AutoTokenizer, CLIPModel
+
+        >>> model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32")
+        >>> tokenizer = AutoTokenizer.from_pretrained("openai/clip-vit-base-patch32")
+
+        >>> inputs = tokenizer(["a photo of a cat", "a photo of a dog"], padding=True, return_tensors="pt")
+        >>> text_features = model.get_text_features(**inputs)
+        ```"""
+        # Use CLIP model's config for some fields (if specified) instead of those of vision & text components.
+        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
+        )
+
+        text_outputs: BaseModelOutputWithPooling = self.text_model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+        )
+
+        pooled_output = text_outputs.pooler_output
+        text_features = self.text_projection(pooled_output)
+
+        return text_features
+
+    @add_start_docstrings_to_model_forward(CLIP_VISION_INPUTS_DOCSTRING)
+    def get_image_features(
+        self,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        interpolate_pos_encoding: bool = False,
+    ) -> torch.FloatTensor:
+        r"""
+        Returns:
+            image_features (`torch.FloatTensor` of shape `(batch_size, output_dim`): The image embeddings obtained by
+            applying the projection layer to the pooled output of [`CLIPVisionModel`].
+
+        Examples:
+
+        ```python
+        >>> from PIL import Image
+        >>> import requests
+        >>> from transformers import AutoProcessor, CLIPModel
+
+        >>> model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32")
+        >>> processor = AutoProcessor.from_pretrained("openai/clip-vit-base-patch32")
+
+        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> image = Image.open(requests.get(url, stream=True).raw)
+
+        >>> inputs = processor(images=image, return_tensors="pt")
+
+        >>> image_features = model.get_image_features(**inputs)
+        ```"""
+        # Use CLIP model's config for some fields (if specified) instead of those of vision & text components.
+        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
+        )
+
+        vision_outputs: BaseModelOutputWithPooling = self.vision_model(
+            pixel_values=pixel_values,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            interpolate_pos_encoding=interpolate_pos_encoding,
+        )
+
+        pooled_output = vision_outputs.pooler_output
+        image_features = self.visual_projection(pooled_output)
+
+        return image_features
+
+    @can_return_tuple
+    @add_start_docstrings_to_model_forward(CLIP_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=CLIPOutput, config_class=CLIPConfig)
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        return_loss: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        interpolate_pos_encoding: bool = False,
+    ) -> CLIPOutput:
+        r"""
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from PIL import Image
+        >>> import requests
+        >>> from transformers import AutoProcessor, CLIPModel
+
+        >>> model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32")
+        >>> processor = AutoProcessor.from_pretrained("openai/clip-vit-base-patch32")
+
+        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> image = Image.open(requests.get(url, stream=True).raw)
+
+        >>> inputs = processor(
+        ...     text=["a photo of a cat", "a photo of a dog"], images=image, return_tensors="pt", padding=True
+        ... )
+
+        >>> outputs = model(**inputs)
+        >>> logits_per_image = outputs.logits_per_image  # this is the image-text similarity score
+        >>> probs = logits_per_image.softmax(dim=1)  # we can take the softmax to get the label probabilities
+        ```"""
+        # Use CLIP model's config for some fields (if specified) instead of those of vision & text components.
+        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
+        )
+
+        vision_outputs: BaseModelOutputWithPooling = self.vision_model(
+            pixel_values=pixel_values,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            interpolate_pos_encoding=interpolate_pos_encoding,
+        )
+
+        text_outputs: BaseModelOutputWithPooling = self.text_model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+        )
+
+        image_embeds = vision_outputs.pooler_output
+        image_embeds = self.visual_projection(image_embeds)
+
+        text_embeds = text_outputs.pooler_output
+        text_embeds = self.text_projection(text_embeds)
+
+        # normalized features
+        image_embeds = image_embeds / _get_vector_norm(image_embeds)
+        text_embeds = text_embeds / _get_vector_norm(text_embeds)
+
+        # cosine similarity as logits
+        logits_per_text = torch.matmul(text_embeds, image_embeds.t().to(text_embeds.device))
+        logits_per_text = logits_per_text * self.logit_scale.exp().to(text_embeds.device)
+
+        logits_per_image = logits_per_text.t()
+
+        loss = None
+        if return_loss:
+            loss = clip_loss(logits_per_text)
+
+        return CLIPOutput(
+            loss=loss,
+            logits_per_image=logits_per_image,
+            logits_per_text=logits_per_text,
+            text_embeds=text_embeds,
+            image_embeds=image_embeds,
+            text_model_output=text_outputs,
+            vision_model_output=vision_outputs,
+        )
+
+
+@add_start_docstrings(
+    """
+    CLIP Text Model with a projection layer on top (a linear layer on top of the pooled output).
+    """,
+    CLIP_START_DOCSTRING,
+)
+class CLIPTextModelWithProjection(CLIPPreTrainedModel):
+    config_class = CLIPTextConfig
+
+    _no_split_modules = ["CLIPTextEmbeddings", "CLIPEncoderLayer"]
+
+    def __init__(self, config: CLIPTextConfig):
+        super().__init__(config)
+
+        text_model = CLIPTextModel._from_config(config)
+        self.text_model = text_model.text_model
+
+        self.text_projection = nn.Linear(config.hidden_size, config.projection_dim, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self) -> nn.Module:
+        return self.text_model.embeddings.token_embedding
+
+    def set_input_embeddings(self, value):
+        self.text_model.embeddings.token_embedding = value
+
+    @can_return_tuple
+    @add_start_docstrings_to_model_forward(CLIP_TEXT_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=CLIPTextModelOutput, config_class=CLIPTextConfig)
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+    ) -> CLIPTextModelOutput:
+        r"""
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from transformers import AutoTokenizer, CLIPTextModelWithProjection
+
+        >>> model = CLIPTextModelWithProjection.from_pretrained("openai/clip-vit-base-patch32")
+        >>> tokenizer = AutoTokenizer.from_pretrained("openai/clip-vit-base-patch32")
+
+        >>> inputs = tokenizer(["a photo of a cat", "a photo of a dog"], padding=True, return_tensors="pt")
+
+        >>> outputs = model(**inputs)
+        >>> text_embeds = outputs.text_embeds
+        ```"""
+
+        text_outputs: BaseModelOutputWithPooling = self.text_model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+        )
+        pooled_output = text_outputs.pooler_output
+        text_embeds = self.text_projection(pooled_output)
+
+        return CLIPTextModelOutput(
+            text_embeds=text_embeds,
+            last_hidden_state=text_outputs.last_hidden_state,
+            hidden_states=text_outputs.hidden_states,
+            attentions=text_outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    CLIP Vision Model with a projection layer on top (a linear layer on top of the pooled output).
+    """,
+    CLIP_START_DOCSTRING,
+)
+class CLIPVisionModelWithProjection(CLIPPreTrainedModel):
+    config_class = CLIPVisionConfig
+    main_input_name = "pixel_values"
+
+    def __init__(self, config: CLIPVisionConfig):
+        super().__init__(config)
+
+        vision_model = CLIPVisionModel._from_config(config)
+        self.vision_model = vision_model.vision_model
+
+        self.visual_projection = nn.Linear(config.hidden_size, config.projection_dim, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self) -> nn.Module:
+        return self.vision_model.embeddings.patch_embedding
+
+    @can_return_tuple
+    @add_start_docstrings_to_model_forward(CLIP_VISION_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=CLIPVisionModelOutput, config_class=CLIPVisionConfig)
+    def forward(
+        self,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        interpolate_pos_encoding: bool = False,
+    ) -> CLIPVisionModelOutput:
+        r"""
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from PIL import Image
+        >>> import requests
+        >>> from transformers import AutoProcessor, CLIPVisionModelWithProjection
+
+        >>> model = CLIPVisionModelWithProjection.from_pretrained("openai/clip-vit-base-patch32")
+        >>> processor = AutoProcessor.from_pretrained("openai/clip-vit-base-patch32")
+
+        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> image = Image.open(requests.get(url, stream=True).raw)
+
+        >>> inputs = processor(images=image, return_tensors="pt")
+
+        >>> outputs = model(**inputs)
+        >>> image_embeds = outputs.image_embeds
+        ```"""
+
+        vision_outputs: BaseModelOutputWithPooling = self.vision_model(
+            pixel_values=pixel_values,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            interpolate_pos_encoding=interpolate_pos_encoding,
+        )
+        pooled_output = vision_outputs.pooler_output
+        image_embeds = self.visual_projection(pooled_output)
+
+        return CLIPVisionModelOutput(
+            image_embeds=image_embeds,
+            last_hidden_state=vision_outputs.last_hidden_state,
+            hidden_states=vision_outputs.hidden_states,
+            attentions=vision_outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    CLIP vision encoder with an image classification head on top (a linear layer on top of the pooled final hidden states of
+    the patch tokens) e.g. for ImageNet.
+    """,
+    CLIP_START_DOCSTRING,
+)
+class CLIPForImageClassification(CLIPPreTrainedModel):
+    main_input_name = "pixel_values"
+
+    def __init__(self, config: CLIPConfig) -> None:
+        super().__init__(config)
+
+        self.num_labels = config.num_labels
+        vision_model = CLIPVisionModel._from_config(config.vision_config)
+        self.vision_model = vision_model.vision_model
+
+        # Classifier head
+        self.classifier = (
+            nn.Linear(config.vision_config.hidden_size, config.num_labels) if config.num_labels > 0 else nn.Identity()
+        )
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @can_return_tuple
+    @add_start_docstrings_to_model_forward(CLIP_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        checkpoint=_IMAGE_CLASS_CHECKPOINT,
+        output_type=ImageClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+        expected_output=_IMAGE_CLASS_EXPECTED_OUTPUT,
+    )
+    def forward(
+        self,
+        pixel_values: Optional[torch.Tensor] = None,
+        labels: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+    ) -> ImageClassifierOutput:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the image classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        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
+        )
+
+        outputs: BaseModelOutputWithPooling = self.vision_model(
+            pixel_values,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+        )
+
+        sequence_output = outputs.last_hidden_state
+
+        # average pool the patch tokens
+        sequence_output = torch.mean(sequence_output[:, 1:, :], dim=1)
+        # apply classifier
+        logits = self.classifier(sequence_output)
+
+        loss = None
+        if labels is not None:
+            # move labels to correct device to enable model parallelism
+            labels = labels.to(logits.device)
+            if self.config.problem_type is None:
+                if self.num_labels == 1:
+                    self.config.problem_type = "regression"
+                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
+                    self.config.problem_type = "single_label_classification"
+                else:
+                    self.config.problem_type = "multi_label_classification"
+
+            if self.config.problem_type == "regression":
+                loss_fct = MSELoss()
+                if self.num_labels == 1:
+                    loss = loss_fct(logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(logits, labels)
+            elif self.config.problem_type == "single_label_classification":
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            elif self.config.problem_type == "multi_label_classification":
+                loss_fct = BCEWithLogitsLoss()
+                loss = loss_fct(logits, labels)
+
+        return ImageClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+__all__ = [
+    "CLIPModel",
+    "CLIPPreTrainedModel",
+    "CLIPTextModel",
+    "CLIPTextModelWithProjection",
+    "CLIPVisionModel",
+    "CLIPVisionModelWithProjection",
+    "CLIPForImageClassification",
+]

lib/python3.12/site-packages/transformers/models/clip/modeling_flax_clip.py

@@ -0,0 +1,1306 @@
+# coding=utf-8
+# Copyright 2021 The OpenAI Team Authors, The Google Flax Team Authors and 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.
+
+from typing import Any, Optional, Tuple, Union
+
+import flax
+import flax.linen as nn
+import jax
+import jax.numpy as jnp
+from flax.core.frozen_dict import FrozenDict, freeze, unfreeze
+from flax.linen import combine_masks, make_causal_mask
+from flax.linen.attention import dot_product_attention_weights
+from flax.traverse_util import flatten_dict, unflatten_dict
+from jax import lax
+
+from ...modeling_flax_outputs import FlaxBaseModelOutput, FlaxBaseModelOutputWithPooling
+from ...modeling_flax_utils import (
+    ACT2FN,
+    FlaxPreTrainedModel,
+    append_replace_return_docstrings,
+    overwrite_call_docstring,
+)
+from ...utils import ModelOutput, add_start_docstrings, logging
+from .configuration_clip import CLIPConfig, CLIPTextConfig, CLIPVisionConfig
+
+
+logger = logging.get_logger(__name__)
+
+CLIP_START_DOCSTRING = r"""
+
+    This model inherits from [`FlaxPreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading, saving and converting weights from PyTorch models)
+
+    This model is also a
+    [flax.linen.Module](https://flax.readthedocs.io/en/latest/api_reference/flax.linen/module.html) subclass. Use it as
+    a regular Flax linen Module and refer to the Flax documentation for all matter related to general usage and
+    behavior.
+
+    Finally, this model supports inherent JAX features such as:
+
+    - [Just-In-Time (JIT) compilation](https://jax.readthedocs.io/en/latest/jax.html#just-in-time-compilation-jit)
+    - [Automatic Differentiation](https://jax.readthedocs.io/en/latest/jax.html#automatic-differentiation)
+    - [Vectorization](https://jax.readthedocs.io/en/latest/jax.html#vectorization-vmap)
+    - [Parallelization](https://jax.readthedocs.io/en/latest/jax.html#parallelization-pmap)
+
+    Parameters:
+        config ([`CLIPConfig`]): 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 [`~FlaxPreTrainedModel.from_pretrained`] method to load the model weights.
+        dtype (`jax.numpy.dtype`, *optional*, defaults to `jax.numpy.float32`):
+            The data type of the computation. Can be one of `jax.numpy.float32`, `jax.numpy.float16` (on GPUs) and
+            `jax.numpy.bfloat16` (on TPUs).
+
+            This can be used to enable mixed-precision training or half-precision inference on GPUs or TPUs. If
+            specified all the computation will be performed with the given `dtype`.
+
+            **Note that this only specifies the dtype of the computation and does not influence the dtype of model
+            parameters.**
+
+            If you wish to change the dtype of the model parameters, see [`~FlaxPreTrainedModel.to_fp16`] and
+            [`~FlaxPreTrainedModel.to_bf16`].
+"""
+
+CLIP_TEXT_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`numpy.ndarray` 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 (`numpy.ndarray` 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)
+        position_ids (`numpy.ndarray` 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.max_position_embeddings - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        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.
+"""
+
+CLIP_VISION_INPUTS_DOCSTRING = r"""
+    Args:
+        pixel_values (`numpy.ndarray` of shape `(batch_size, num_channels, height, width)`):
+            Pixel values. Padding will be ignored by default should you provide it. Pixel values can be obtained using
+            [`AutoImageProcessor`]. See [`CLIPImageProcessor.__call__`] for details.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+CLIP_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`numpy.ndarray` 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 (`numpy.ndarray` 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)
+        position_ids (`numpy.ndarray` 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.max_position_embeddings - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        pixel_values (`numpy.ndarray` of shape `(batch_size, num_channels, height, width)`):
+            Pixel values. Padding will be ignored by default should you provide it. Pixel values can be obtained using
+            [`AutoImageProcessor`]. See [`CLIPImageProcessor.__call__`] for details.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+@flax.struct.dataclass
+class FlaxCLIPTextModelOutput(ModelOutput):
+    """
+    Base class for text model's outputs that also contains a pooling of the last hidden states.
+
+    Args:
+        text_embeds (`jnp.ndarray` of shape `(batch_size, output_dim`):
+            The text embeddings obtained by applying the projection layer to the pooled output of
+            [`FlaxCLIPTextModel`].
+        last_hidden_state (`jnp.ndarray` of shape `(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the model.
+        hidden_states (`tuple(jnp.ndarray)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `jnp.ndarray` (one for the output of the embeddings + 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 initial embedding outputs.
+        attentions (`tuple(jnp.ndarray)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
+            Tuple of `jnp.ndarray` (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.
+    """
+
+    text_embeds: jnp.ndarray = None
+    last_hidden_state: jnp.ndarray = None
+    hidden_states: Optional[Tuple[jnp.ndarray, ...]] = None
+    attentions: Optional[Tuple[jnp.ndarray, ...]] = None
+
+
+@flax.struct.dataclass
+class FlaxCLIPOutput(ModelOutput):
+    """
+    Args:
+        logits_per_image:(`jnp.ndarray` of shape `(image_batch_size, text_batch_size)`):
+            The scaled dot product scores between `image_embeds` and `text_embeds`. This represents the image-text
+            similarity scores.
+        logits_per_text:(`jnp.ndarray` of shape `(text_batch_size, image_batch_size)`):
+            The scaled dot product scores between `text_embeds` and `image_embeds`. This represents the text-image
+            similarity scores.
+        text_embeds(`jnp.ndarray` of shape `(batch_size, output_dim`):
+            The text embeddings obtained by applying the projection layer to the pooled output of
+            [`FlaxCLIPTextModel`].
+        image_embeds(`jnp.ndarray` of shape `(batch_size, output_dim`):
+            The image embeddings obtained by applying the projection layer to the pooled output of
+            [`FlaxCLIPVisionModel`].
+        text_model_output(`FlaxBaseModelOutputWithPooling`):
+            The output of the [`FlaxCLIPTextModel`].
+        vision_model_output(`FlaxBaseModelOutputWithPooling`):
+            The output of the [`FlaxCLIPVisionModel`].
+    """
+
+    logits_per_image: jnp.ndarray = None
+    logits_per_text: jnp.ndarray = None
+    text_embeds: jnp.ndarray = None
+    image_embeds: jnp.ndarray = None
+    text_model_output: FlaxBaseModelOutputWithPooling = None
+    vision_model_output: FlaxBaseModelOutputWithPooling = None
+
+    def to_tuple(self) -> Tuple[Any]:
+        return tuple(
+            self[k] if k not in ["text_model_output", "vision_model_output"] else getattr(self, k).to_tuple()
+            for k in self.keys()
+        )
+
+
+class FlaxCLIPVisionEmbeddings(nn.Module):
+    config: CLIPVisionConfig
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        embed_dim = self.config.hidden_size
+        image_size = self.config.image_size
+        patch_size = self.config.patch_size
+
+        self.class_embedding = self.param("class_embedding", jax.nn.initializers.normal(stddev=0.02), (embed_dim,))
+
+        self.patch_embedding = nn.Conv(
+            embed_dim,
+            kernel_size=(patch_size, patch_size),
+            strides=(patch_size, patch_size),
+            padding="VALID",
+            use_bias=False,
+            dtype=self.dtype,
+            kernel_init=jax.nn.initializers.normal(),
+        )
+
+        self.num_patches = (image_size // patch_size) ** 2
+        num_positions = self.num_patches + 1
+        self.position_embedding = nn.Embed(num_positions, embed_dim, embedding_init=jax.nn.initializers.normal())
+        self.position_ids = jnp.expand_dims(jnp.arange(0, num_positions, dtype="i4"), axis=0)
+
+    def __call__(self, pixel_values):
+        patch_embeds = self.patch_embedding(pixel_values)
+        batch_size, height, width, channels = patch_embeds.shape
+        patch_embeds = jnp.reshape(patch_embeds, (batch_size, height * width, channels))
+
+        class_embeds = jnp.expand_dims(self.class_embedding, axis=(0, 1))
+        class_embeds = jnp.tile(class_embeds, (batch_size, 1, 1))
+        embeddings = jnp.concatenate([class_embeds, patch_embeds], axis=1)
+        embeddings = embeddings + self.position_embedding(self.position_ids)
+        return embeddings
+
+
+class FlaxCLIPTextEmbeddings(nn.Module):
+    config: CLIPTextConfig
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        embed_dim = self.config.hidden_size
+
+        self.token_embedding = nn.Embed(self.config.vocab_size, embed_dim, embedding_init=jax.nn.initializers.normal())
+        self.position_embedding = nn.Embed(
+            self.config.max_position_embeddings, embed_dim, embedding_init=jax.nn.initializers.normal()
+        )
+        self.position_ids = jnp.expand_dims(
+            jnp.arange(0, self.config.max_position_embeddings, dtype="i4"), axis=(0, 1)
+        )
+
+    def __call__(self, input_ids, position_ids):
+        input_embeds = self.token_embedding(input_ids.astype("i4"))
+        position_embeds = self.position_embedding(position_ids.astype("i4"))
+
+        embeddings = input_embeds + position_embeds
+        return embeddings
+
+
+class FlaxCLIPAttention(nn.Module):
+    config: Union[CLIPTextConfig, CLIPVisionConfig]
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        self.embed_dim = self.config.hidden_size
+        self.num_heads = self.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 = self.config.attention_dropout
+
+        self.k_proj = nn.Dense(self.embed_dim, dtype=self.dtype, kernel_init=jax.nn.initializers.normal(0.01))
+        self.v_proj = nn.Dense(self.embed_dim, dtype=self.dtype, kernel_init=jax.nn.initializers.normal(0.01))
+        self.q_proj = nn.Dense(self.embed_dim, dtype=self.dtype, kernel_init=jax.nn.initializers.normal(0.01))
+        self.out_proj = nn.Dense(self.embed_dim, dtype=self.dtype, kernel_init=jax.nn.initializers.normal(0.01))
+
+        self.causal = isinstance(self.config, CLIPTextConfig)
+        if self.causal:
+            self.causal_mask = make_causal_mask(jnp.ones((1, self.config.max_position_embeddings), dtype="i4"))
+
+    def _split_heads(self, hidden_states):
+        return hidden_states.reshape(hidden_states.shape[:2] + (self.num_heads, self.head_dim))
+
+    def _merge_heads(self, hidden_states):
+        return hidden_states.reshape(hidden_states.shape[:2] + (self.embed_dim,))
+
+    def __call__(
+        self,
+        hidden_states,
+        attention_mask=None,
+        deterministic: bool = True,
+        output_attentions: bool = False,
+    ):
+        query = self.q_proj(hidden_states)
+        key = self.k_proj(hidden_states)
+        value = self.v_proj(hidden_states)
+
+        query = self._split_heads(query)
+        key = self._split_heads(key)
+        value = self._split_heads(value)
+
+        causal_attention_mask = None
+        if self.causal:
+            query_length, key_length = query.shape[1], key.shape[1]
+            causal_attention_mask = self.causal_mask[:, :, key_length - query_length : key_length, :key_length]
+
+        if attention_mask is not None and causal_attention_mask is not None:
+            attention_mask = jnp.expand_dims(attention_mask, axis=(-3, -2))
+            attention_mask = combine_masks(attention_mask, causal_attention_mask, dtype="i4")
+        elif causal_attention_mask is not None:
+            attention_mask = causal_attention_mask
+        elif attention_mask is not None:
+            attention_mask = jnp.expand_dims(attention_mask, axis=(-3, -2))
+
+        if attention_mask is not None:
+            attention_bias = lax.select(
+                attention_mask > 0,
+                jnp.full(attention_mask.shape, 0.0).astype(self.dtype),
+                jnp.full(attention_mask.shape, jnp.finfo(self.dtype).min).astype(self.dtype),
+            )
+        else:
+            attention_bias = None
+
+        dropout_rng = None
+        if not deterministic and self.dropout > 0.0:
+            dropout_rng = self.make_rng("dropout")
+
+        attn_weights = dot_product_attention_weights(
+            query,
+            key,
+            bias=attention_bias,
+            dropout_rng=dropout_rng,
+            dropout_rate=self.dropout,
+            deterministic=deterministic,
+            dtype=self.dtype,
+            precision=None,
+        )
+
+        attn_output = jnp.einsum("...hqk,...khd->...qhd", attn_weights, value)
+        attn_output = self._merge_heads(attn_output)
+        attn_output = self.out_proj(attn_output)
+
+        outputs = (attn_output, attn_weights) if output_attentions else (attn_output,)
+        return outputs
+
+
+class FlaxCLIPMLP(nn.Module):
+    config: Union[CLIPTextConfig, CLIPVisionConfig]
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        self.activation_fn = ACT2FN[self.config.hidden_act]
+        self.fc1 = nn.Dense(
+            self.config.intermediate_size,
+            dtype=self.dtype,
+            kernel_init=jax.nn.initializers.normal(0.01),
+        )
+        self.fc2 = nn.Dense(self.config.hidden_size, dtype=self.dtype, kernel_init=jax.nn.initializers.normal(0.01))
+
+    def __call__(self, hidden_states):
+        hidden_states = self.fc1(hidden_states)
+        hidden_states = self.activation_fn(hidden_states)
+        hidden_states = self.fc2(hidden_states)
+        return hidden_states
+
+
+class FlaxCLIPEncoderLayer(nn.Module):
+    config: Union[CLIPTextConfig, CLIPVisionConfig]
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        self.self_attn = FlaxCLIPAttention(self.config, dtype=self.dtype)
+        self.layer_norm1 = nn.LayerNorm(epsilon=self.config.layer_norm_eps, dtype=self.dtype)
+        self.mlp = FlaxCLIPMLP(self.config, dtype=self.dtype)
+        self.layer_norm2 = nn.LayerNorm(epsilon=self.config.layer_norm_eps, dtype=self.dtype)
+
+    def __call__(
+        self,
+        hidden_states,
+        attention_mask,
+        deterministic: bool = True,
+        output_attentions: bool = False,
+    ):
+        residual = hidden_states
+
+        hidden_states = self.layer_norm1(hidden_states)
+        attn_outputs = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            deterministic=deterministic,
+            output_attentions=output_attentions,
+        )
+        hidden_states = attn_outputs[0]
+        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_outputs[1:]
+
+        return outputs
+
+
+class FlaxCLIPLayerCollection(nn.Module):
+    config: Union[CLIPTextConfig, CLIPVisionConfig]
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        self.layers = [
+            FlaxCLIPEncoderLayer(self.config, name=str(i), dtype=self.dtype)
+            for i in range(self.config.num_hidden_layers)
+        ]
+
+    def __call__(
+        self,
+        hidden_states,
+        attention_mask=None,
+        deterministic: bool = True,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ):
+        all_attentions = () if output_attentions else None
+        all_hidden_states = () if output_hidden_states else None
+
+        for layer in self.layers:
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            layer_outputs = layer(
+                hidden_states, attention_mask, deterministic=deterministic, output_attentions=output_attentions
+            )
+            hidden_states = layer_outputs[0]
+
+            if output_attentions:
+                all_attentions += (layer_outputs[1],)
+
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        outputs = (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in outputs if v is not None)
+
+        return FlaxBaseModelOutput(
+            last_hidden_state=hidden_states, hidden_states=all_hidden_states, attentions=all_attentions
+        )
+
+
+class FlaxCLIPEncoder(nn.Module):
+    config: Union[CLIPTextConfig, CLIPVisionConfig]
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        self.layers = FlaxCLIPLayerCollection(self.config, dtype=self.dtype)
+
+    def __call__(
+        self,
+        inputs_embeds,
+        attention_mask=None,
+        deterministic: bool = True,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ):
+        return self.layers(
+            hidden_states=inputs_embeds,
+            attention_mask=attention_mask,
+            deterministic=deterministic,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+
+class FlaxCLIPTextTransformer(nn.Module):
+    config: CLIPTextConfig
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        self.embeddings = FlaxCLIPTextEmbeddings(self.config, dtype=self.dtype)
+        self.encoder = FlaxCLIPEncoder(self.config, dtype=self.dtype)
+        self.final_layer_norm = nn.LayerNorm(epsilon=self.config.layer_norm_eps, dtype=self.dtype)
+
+        # For `pooled_output` computation
+        self.eos_token_id = self.config.eos_token_id
+
+    def __call__(
+        self,
+        input_ids,
+        attention_mask,
+        position_ids,
+        deterministic: bool = True,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ):
+        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
+
+        hidden_states = self.embeddings(input_ids=input_ids, position_ids=position_ids)
+
+        encoder_outputs = self.encoder(
+            inputs_embeds=hidden_states,
+            attention_mask=attention_mask,
+            deterministic=deterministic,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        last_hidden_state = encoder_outputs[0]
+        last_hidden_state = self.final_layer_norm(last_hidden_state)
+
+        if self.eos_token_id == 2:
+            # The `eos_token_id` was incorrect before PR #24773: Let's keep what have been done here.
+            # A CLIP model with such `eos_token_id` in the config can't work correctly with extra new tokens added
+            # ------------------------------------------------------------
+            # text_embeds.shape = [batch_size, sequence_length, transformer.width]
+            # take features from the EOS embedding (eos_token_id is the highest number in each sequence)
+            pooled_output = last_hidden_state[jnp.arange(last_hidden_state.shape[0]), input_ids.argmax(axis=-1)]
+        else:
+            # (no need to cast from bool to int after comparing to `eos_token_id`)
+            pooled_output = last_hidden_state[
+                jnp.arange(last_hidden_state.shape[0]), (input_ids == self.eos_token_id).argmax(axis=-1)
+            ]
+
+        if not return_dict:
+            return (last_hidden_state, pooled_output) + encoder_outputs[1:]
+
+        return FlaxBaseModelOutputWithPooling(
+            last_hidden_state=last_hidden_state,
+            pooler_output=pooled_output,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )
+
+
+class FlaxCLIPVisionTransformer(nn.Module):
+    config: CLIPVisionConfig
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        self.embeddings = FlaxCLIPVisionEmbeddings(self.config, dtype=self.dtype)
+        self.pre_layrnorm = nn.LayerNorm(epsilon=self.config.layer_norm_eps, dtype=self.dtype)
+        self.encoder = FlaxCLIPEncoder(self.config, dtype=self.dtype)
+        self.post_layernorm = nn.LayerNorm(epsilon=self.config.layer_norm_eps, dtype=self.dtype)
+
+    def __call__(
+        self,
+        pixel_values=None,
+        deterministic: bool = True,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict: bool = True,
+    ):
+        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
+
+        hidden_states = self.embeddings(pixel_values)
+        hidden_states = self.pre_layrnorm(hidden_states)
+
+        encoder_outputs = self.encoder(
+            inputs_embeds=hidden_states,
+            deterministic=deterministic,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        last_hidden_state = encoder_outputs[0]
+        pooled_output = last_hidden_state[:, 0, :]
+        pooled_output = self.post_layernorm(pooled_output)
+
+        if not return_dict:
+            return (last_hidden_state, pooled_output) + encoder_outputs[1:]
+
+        return FlaxBaseModelOutputWithPooling(
+            last_hidden_state=last_hidden_state,
+            pooler_output=pooled_output,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )
+
+
+class FlaxCLIPTextPreTrainedModel(FlaxPreTrainedModel):
+    config_class = CLIPTextConfig
+    module_class: nn.Module = None
+
+    def __init__(
+        self,
+        config: CLIPTextConfig,
+        input_shape=(1, 1),
+        seed: int = 0,
+        dtype: jnp.dtype = jnp.float32,
+        _do_init: bool = True,
+        **kwargs,
+    ):
+        module = self.module_class(config=config, dtype=dtype, **kwargs)
+        super().__init__(config, module, input_shape=input_shape, seed=seed, dtype=dtype, _do_init=_do_init)
+
+    def init_weights(self, rng: jax.random.PRNGKey, input_shape: Tuple, params: FrozenDict = None) -> FrozenDict:
+        # init input tensor
+        input_ids = jnp.zeros(input_shape, dtype="i4")
+        position_ids = jnp.broadcast_to(jnp.arange(jnp.atleast_2d(input_ids).shape[-1]), input_shape)
+        attention_mask = jnp.ones_like(input_ids)
+
+        params_rng, dropout_rng = jax.random.split(rng)
+        rngs = {"params": params_rng, "dropout": dropout_rng}
+
+        random_params = self.module.init(rngs, input_ids, attention_mask, position_ids)["params"]
+
+        if params is not None:
+            random_params = flatten_dict(unfreeze(random_params))
+            params = flatten_dict(unfreeze(params))
+            for missing_key in self._missing_keys:
+                params[missing_key] = random_params[missing_key]
+            self._missing_keys = set()
+            return freeze(unflatten_dict(params))
+        else:
+            return random_params
+
+    def __call__(
+        self,
+        input_ids,
+        attention_mask=None,
+        position_ids=None,
+        params: dict = None,
+        dropout_rng: jax.random.PRNGKey = None,
+        train: bool = False,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ):
+        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.return_dict
+
+        if position_ids is None:
+            position_ids = jnp.broadcast_to(jnp.arange(jnp.atleast_2d(input_ids).shape[-1]), input_ids.shape)
+
+        if attention_mask is None:
+            attention_mask = jnp.ones_like(input_ids)
+
+        # Handle any PRNG if needed
+        rngs = {}
+        if dropout_rng is not None:
+            rngs["dropout"] = dropout_rng
+
+        return self.module.apply(
+            {"params": params or self.params},
+            jnp.array(input_ids, dtype="i4"),
+            jnp.array(attention_mask, dtype="i4"),
+            jnp.array(position_ids, dtype="i4"),
+            not train,
+            output_attentions,
+            output_hidden_states,
+            return_dict,
+            rngs=rngs,
+        )
+
+
+class FlaxCLIPVisionPreTrainedModel(FlaxPreTrainedModel):
+    config_class = CLIPVisionConfig
+    main_input_name = "pixel_values"
+    module_class: nn.Module = None
+
+    def __init__(
+        self,
+        config: CLIPVisionConfig,
+        input_shape: Optional[Tuple] = None,
+        seed: int = 0,
+        dtype: jnp.dtype = jnp.float32,
+        _do_init: bool = True,
+        **kwargs,
+    ):
+        if input_shape is None:
+            input_shape = (1, config.image_size, config.image_size, 3)
+        module = self.module_class(config=config, dtype=dtype, **kwargs)
+        super().__init__(config, module, input_shape=input_shape, seed=seed, dtype=dtype, _do_init=_do_init)
+
+    def init_weights(self, rng: jax.random.PRNGKey, input_shape: Tuple, params: FrozenDict = None) -> FrozenDict:
+        # init input tensor
+        pixel_values = jax.random.normal(rng, input_shape)
+
+        params_rng, dropout_rng = jax.random.split(rng)
+        rngs = {"params": params_rng, "dropout": dropout_rng}
+
+        random_params = self.module.init(rngs, pixel_values)["params"]
+
+        if params is not None:
+            random_params = flatten_dict(unfreeze(random_params))
+            params = flatten_dict(unfreeze(params))
+            for missing_key in self._missing_keys:
+                params[missing_key] = random_params[missing_key]
+            self._missing_keys = set()
+            return freeze(unflatten_dict(params))
+        else:
+            return random_params
+
+    def __call__(
+        self,
+        pixel_values,
+        params: dict = None,
+        dropout_rng: jax.random.PRNGKey = None,
+        train: bool = False,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ):
+        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.return_dict
+
+        pixel_values = jnp.transpose(pixel_values, (0, 2, 3, 1))
+
+        # Handle any PRNG if needed
+        rngs = {}
+        if dropout_rng is not None:
+            rngs["dropout"] = dropout_rng
+
+        return self.module.apply(
+            {"params": params or self.params},
+            jnp.array(pixel_values, dtype=jnp.float32),
+            not train,
+            output_attentions,
+            output_hidden_states,
+            return_dict,
+            rngs=rngs,
+        )
+
+
+class FlaxCLIPPreTrainedModel(FlaxPreTrainedModel):
+    config_class = CLIPConfig
+    module_class: nn.Module = None
+
+    def __init__(
+        self,
+        config: CLIPConfig,
+        input_shape: Optional[Tuple] = None,
+        seed: int = 0,
+        dtype: jnp.dtype = jnp.float32,
+        _do_init: bool = True,
+        **kwargs,
+    ):
+        if input_shape is None:
+            input_shape = ((1, 1), (1, config.vision_config.image_size, config.vision_config.image_size, 3))
+        module = self.module_class(config=config, dtype=dtype, **kwargs)
+        super().__init__(config, module, input_shape=input_shape, seed=seed, dtype=dtype, _do_init=_do_init)
+
+    def init_weights(self, rng: jax.random.PRNGKey, input_shape: Tuple, params: FrozenDict = None) -> FrozenDict:
+        # init input tensor
+        input_ids = jnp.zeros(input_shape[0], dtype="i4")
+        position_ids = jnp.broadcast_to(jnp.arange(jnp.atleast_2d(input_ids).shape[-1]), input_shape[0])
+        attention_mask = jnp.ones_like(input_ids)
+
+        pixel_values = jax.random.normal(rng, input_shape[1])
+
+        params_rng, dropout_rng = jax.random.split(rng)
+        rngs = {"params": params_rng, "dropout": dropout_rng}
+
+        random_params = self.module.init(rngs, input_ids, pixel_values, attention_mask, position_ids)["params"]
+
+        if params is not None:
+            random_params = flatten_dict(unfreeze(random_params))
+            params = flatten_dict(unfreeze(params))
+            for missing_key in self._missing_keys:
+                params[missing_key] = random_params[missing_key]
+            self._missing_keys = set()
+            return freeze(unflatten_dict(params))
+        else:
+            return random_params
+
+    def __call__(
+        self,
+        input_ids,
+        pixel_values,
+        attention_mask=None,
+        position_ids=None,
+        params: dict = None,
+        dropout_rng: jax.random.PRNGKey = None,
+        train: bool = False,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ):
+        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.return_dict
+
+        if position_ids is None:
+            position_ids = jnp.broadcast_to(jnp.arange(jnp.atleast_2d(input_ids).shape[-1]), input_ids.shape)
+
+        if attention_mask is None:
+            attention_mask = jnp.ones_like(input_ids)
+
+        pixel_values = jnp.transpose(pixel_values, (0, 2, 3, 1))
+
+        # Handle any PRNG if needed
+        rngs = {}
+        if dropout_rng is not None:
+            rngs["dropout"] = dropout_rng
+
+        return self.module.apply(
+            {"params": params or self.params},
+            jnp.array(input_ids, dtype="i4"),
+            jnp.array(pixel_values, dtype=jnp.float32),
+            jnp.array(attention_mask, dtype="i4"),
+            jnp.array(position_ids, dtype="i4"),
+            not train,
+            output_attentions,
+            output_hidden_states,
+            return_dict,
+            rngs=rngs,
+        )
+
+    def get_text_features(
+        self,
+        input_ids,
+        attention_mask=None,
+        position_ids=None,
+        params: dict = None,
+        dropout_rng: jax.random.PRNGKey = None,
+        train=False,
+    ):
+        r"""
+        Args:
+            input_ids (`numpy.ndarray` 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)
+
+        Returns:
+            text_features (`jnp.ndarray` of shape `(batch_size, output_dim`): The text embeddings obtained by applying
+            the projection layer to the pooled output of [`FlaxCLIPTextModel`].
+
+        Examples:
+
+        ```python
+        >>> from transformers import AutoTokenizer, FlaxCLIPModel
+
+        >>> model = FlaxCLIPModel.from_pretrained("openai/clip-vit-base-patch32")
+        >>> tokenizer = AutoTokenizer.from_pretrained("openai/clip-vit-base-patch32")
+
+        >>> inputs = tokenizer(["a photo of a cat", "a photo of a dog"], padding=True, return_tensors="np")
+        >>> text_features = model.get_text_features(**inputs)
+        ```"""
+        if position_ids is None:
+            position_ids = jnp.broadcast_to(jnp.arange(jnp.atleast_2d(input_ids).shape[-1]), input_ids.shape)
+
+        if attention_mask is None:
+            attention_mask = jnp.ones_like(input_ids)
+
+        # Handle any PRNG if needed
+        rngs = {}
+        if dropout_rng is not None:
+            rngs["dropout"] = dropout_rng
+
+        def _get_features(module, input_ids, attention_mask, position_ids, deterministic):
+            text_outputs = module.text_model(
+                input_ids=input_ids,
+                attention_mask=attention_mask,
+                position_ids=position_ids,
+                deterministic=deterministic,
+            )
+            pooled_output = text_outputs[1]
+            text_features = module.text_projection(pooled_output)
+            return text_features
+
+        return self.module.apply(
+            {"params": params or self.params},
+            jnp.array(input_ids, dtype="i4"),
+            jnp.array(attention_mask, dtype="i4"),
+            jnp.array(position_ids, dtype="i4"),
+            not train,
+            method=_get_features,
+            rngs=rngs,
+        )
+
+    def get_image_features(
+        self, pixel_values, params: dict = None, dropout_rng: jax.random.PRNGKey = None, train=False
+    ):
+        r"""
+        Args:
+            pixel_values (`numpy.ndarray` of shape `(batch_size, num_channels, height, width)`):
+                Pixel values. Padding will be ignored by default should you provide it. Pixel values can be obtained
+                using [`AutoImageProcessor`]. See [`CLIPImageProcessor.__call__`] for details.
+
+        Returns:
+            image_features (`jnp.ndarray` of shape `(batch_size, output_dim`): The image embeddings obtained by
+            applying the projection layer to the pooled output of [`FlaxCLIPVisionModel`]
+
+        Examples:
+
+        ```python
+        >>> from PIL import Image
+        >>> import requests
+        >>> from transformers import AutoProcessor, FlaxCLIPModel
+
+        >>> model = FlaxCLIPModel.from_pretrained("openai/clip-vit-base-patch32")
+        >>> processor = AutoProcessor.from_pretrained("openai/clip-vit-base-patch32")
+
+        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> image = Image.open(requests.get(url, stream=True).raw)
+
+        >>> inputs = processor(images=image, return_tensors="np")
+
+        >>> image_features = model.get_image_features(**inputs)
+        ```"""
+        pixel_values = jnp.transpose(pixel_values, (0, 2, 3, 1))
+
+        # Handle any PRNG if needed
+        rngs = {}
+        if dropout_rng is not None:
+            rngs["dropout"] = dropout_rng
+
+        def _get_features(module, pixel_values, deterministic):
+            vision_outputs = module.vision_model(pixel_values=pixel_values, deterministic=deterministic)
+            pooled_output = vision_outputs[1]  # pooled_output
+            image_features = module.visual_projection(pooled_output)
+            return image_features
+
+        return self.module.apply(
+            {"params": params or self.params},
+            jnp.array(pixel_values, dtype=jnp.float32),
+            not train,
+            method=_get_features,
+            rngs=rngs,
+        )
+
+
+class FlaxCLIPTextModule(nn.Module):
+    config: CLIPTextConfig
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        self.text_model = FlaxCLIPTextTransformer(self.config, dtype=self.dtype)
+
+    def __call__(
+        self,
+        input_ids,
+        attention_mask,
+        position_ids,
+        deterministic: bool = True,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ):
+        return self.text_model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            deterministic=deterministic,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+
+class FlaxCLIPTextModel(FlaxCLIPTextPreTrainedModel):
+    module_class = FlaxCLIPTextModule
+
+
+FLAX_CLIP_TEXT_MODEL_DOCSTRING = """
+    Returns:
+
+    Example:
+
+    ```python
+    >>> from transformers import AutoTokenizer, FlaxCLIPTextModel
+
+    >>> model = FlaxCLIPTextModel.from_pretrained("openai/clip-vit-base-patch32")
+    >>> tokenizer = AutoTokenizer.from_pretrained("openai/clip-vit-base-patch32")
+
+    >>> inputs = tokenizer(["a photo of a cat", "a photo of a dog"], padding=True, return_tensors="np")
+
+    >>> outputs = model(**inputs)
+    >>> last_hidden_state = outputs.last_hidden_state
+    >>> pooler_output = outputs.pooler_output  # pooled (EOS token) states
+    ```
+"""
+
+overwrite_call_docstring(FlaxCLIPTextModel, CLIP_TEXT_INPUTS_DOCSTRING + FLAX_CLIP_TEXT_MODEL_DOCSTRING)
+append_replace_return_docstrings(
+    FlaxCLIPTextModel, output_type=FlaxBaseModelOutputWithPooling, config_class=CLIPTextConfig
+)
+
+
+class FlaxCLIPTextModelWithProjectionModule(nn.Module):
+    config: CLIPTextConfig
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        self.text_model = FlaxCLIPTextTransformer(self.config, dtype=self.dtype)
+        self.text_projection = nn.Dense(self.config.projection_dim, use_bias=False, dtype=self.dtype)
+
+    def __call__(
+        self,
+        input_ids,
+        attention_mask,
+        position_ids,
+        deterministic: bool = True,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ):
+        text_outputs = self.text_model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            deterministic=deterministic,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        pooled_output = text_outputs[1]
+        text_embeds = self.text_projection(pooled_output)
+
+        if not return_dict:
+            return (text_embeds, text_outputs[0]) + text_outputs[2:]
+
+        return FlaxCLIPTextModelOutput(
+            text_embeds=text_embeds,
+            last_hidden_state=text_outputs.last_hidden_state,
+            hidden_states=text_outputs.hidden_states,
+            attentions=text_outputs.attentions,
+        )
+
+
+class FlaxCLIPTextModelWithProjection(FlaxCLIPTextPreTrainedModel):
+    module_class = FlaxCLIPTextModelWithProjectionModule
+
+
+FLAX_CLIP_TEXT_MODEL_WITH_PROJECTION_DOCSTRING = """
+    Returns:
+
+    Example:
+
+    ```python
+    >>> from transformers import AutoTokenizer, FlaxCLIPTextModelWithProjection
+
+    >>> model = FlaxCLIPTextModelWithProjection.from_pretrained("openai/clip-vit-base-patch32")
+    >>> tokenizer = AutoTokenizer.from_pretrained("openai/clip-vit-base-patch32")
+
+    >>> inputs = tokenizer(["a photo of a cat", "a photo of a dog"], padding=True, return_tensors="np")
+
+    >>> outputs = model(**inputs)
+    >>> text_embeds = outputs.text_embeds
+    ```
+"""
+
+overwrite_call_docstring(
+    FlaxCLIPTextModelWithProjection, CLIP_TEXT_INPUTS_DOCSTRING + FLAX_CLIP_TEXT_MODEL_WITH_PROJECTION_DOCSTRING
+)
+append_replace_return_docstrings(
+    FlaxCLIPTextModelWithProjection, output_type=FlaxCLIPTextModelOutput, config_class=CLIPTextConfig
+)
+
+
+class FlaxCLIPVisionModule(nn.Module):
+    config: CLIPVisionConfig
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        self.vision_model = FlaxCLIPVisionTransformer(self.config, dtype=self.dtype)
+
+    def __call__(
+        self,
+        pixel_values,
+        deterministic: bool = True,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ):
+        return self.vision_model(
+            pixel_values=pixel_values,
+            deterministic=deterministic,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+
+class FlaxCLIPVisionModel(FlaxCLIPVisionPreTrainedModel):
+    module_class = FlaxCLIPVisionModule
+
+
+FLAX_CLIP_VISION_MODEL_DOCSTRING = """
+    Returns:
+
+    Example:
+
+    ```python
+    >>> from PIL import Image
+    >>> import requests
+    >>> from transformers import AutoProcessor, FlaxCLIPVisionModel
+
+    >>> model = FlaxCLIPVisionModel.from_pretrained("openai/clip-vit-base-patch32")
+    >>> processor = AutoProcessor.from_pretrained("openai/clip-vit-base-patch32")
+
+    >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+    >>> image = Image.open(requests.get(url, stream=True).raw)
+
+    >>> inputs = processor(images=image, return_tensors="np")
+
+    >>> outputs = model(**inputs)
+    >>> last_hidden_state = outputs.last_hidden_state
+    >>> pooler_output = outputs.pooler_output  # pooled CLS states
+    ```
+"""
+
+overwrite_call_docstring(FlaxCLIPVisionModel, CLIP_VISION_INPUTS_DOCSTRING + FLAX_CLIP_VISION_MODEL_DOCSTRING)
+append_replace_return_docstrings(
+    FlaxCLIPVisionModel, output_type=FlaxBaseModelOutputWithPooling, config_class=CLIPVisionConfig
+)
+
+
+class FlaxCLIPModule(nn.Module):
+    config: CLIPConfig
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        text_config = self.config.text_config
+        vision_config = self.config.vision_config
+
+        self.projection_dim = self.config.projection_dim
+        self.text_embed_dim = text_config.hidden_size
+        self.vision_embed_dim = vision_config.hidden_size
+
+        self.text_model = FlaxCLIPTextTransformer(text_config, dtype=self.dtype)
+        self.vision_model = FlaxCLIPVisionTransformer(vision_config, dtype=self.dtype)
+
+        self.visual_projection = nn.Dense(
+            self.projection_dim,
+            dtype=self.dtype,
+            kernel_init=jax.nn.initializers.normal(0.02),
+            use_bias=False,
+        )
+        self.text_projection = nn.Dense(
+            self.projection_dim,
+            dtype=self.dtype,
+            kernel_init=jax.nn.initializers.normal(0.02),
+            use_bias=False,
+        )
+
+        self.logit_scale = self.param(
+            "logit_scale", lambda _, shape: jnp.ones(shape) * self.config.logit_scale_init_value, []
+        )
+
+    def __call__(
+        self,
+        input_ids=None,
+        pixel_values=None,
+        attention_mask=None,
+        position_ids=None,
+        deterministic: bool = True,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        return_dict = return_dict if return_dict is not None else self.config.return_dict
+
+        vision_outputs = self.vision_model(
+            pixel_values=pixel_values,
+            deterministic=deterministic,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        text_outputs = self.text_model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            deterministic=deterministic,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        image_embeds = vision_outputs[1]
+        image_embeds = self.visual_projection(image_embeds)
+
+        text_embeds = text_outputs[1]
+        text_embeds = self.text_projection(text_embeds)
+
+        # normalized features
+        image_embeds = image_embeds / jnp.linalg.norm(image_embeds, axis=-1, keepdims=True)
+        text_embeds = text_embeds / jnp.linalg.norm(text_embeds, axis=-1, keepdims=True)
+
+        # cosine similarity as logits
+        logit_scale = jnp.exp(self.logit_scale)
+        logits_per_text = jnp.matmul(text_embeds, image_embeds.T) * logit_scale
+        logits_per_image = logits_per_text.T
+
+        if not return_dict:
+            return (logits_per_image, logits_per_text, text_embeds, image_embeds, text_outputs, vision_outputs)
+
+        return FlaxCLIPOutput(
+            logits_per_image=logits_per_image,
+            logits_per_text=logits_per_text,
+            text_embeds=text_embeds,
+            image_embeds=image_embeds,
+            text_model_output=text_outputs,
+            vision_model_output=vision_outputs,
+        )
+
+
+@add_start_docstrings(CLIP_START_DOCSTRING)
+class FlaxCLIPModel(FlaxCLIPPreTrainedModel):
+    module_class = FlaxCLIPModule
+
+
+FLAX_CLIP_MODEL_DOCSTRING = """
+    Returns:
+
+    Example:
+
+    ```python
+    >>> import jax
+    >>> from PIL import Image
+    >>> import requests
+    >>> from transformers import AutoProcessor, FlaxCLIPModel
+
+    >>> model = FlaxCLIPModel.from_pretrained("openai/clip-vit-base-patch32")
+    >>> processor = AutoProcessor.from_pretrained("openai/clip-vit-base-patch32")
+
+    >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+    >>> image = Image.open(requests.get(url, stream=True).raw)
+
+    >>> inputs = processor(
+    ...     text=["a photo of a cat", "a photo of a dog"], images=image, return_tensors="np", padding=True
+    ... )
+
+    >>> outputs = model(**inputs)
+    >>> logits_per_image = outputs.logits_per_image  # this is the image-text similarity score
+    >>> probs = jax.nn.softmax(logits_per_image, axis=1)  # we can take the softmax to get the label probabilities
+    ```
+"""
+
+overwrite_call_docstring(FlaxCLIPModel, CLIP_INPUTS_DOCSTRING + FLAX_CLIP_MODEL_DOCSTRING)
+append_replace_return_docstrings(FlaxCLIPModel, output_type=FlaxCLIPOutput, config_class=CLIPConfig)
+
+
+__all__ = [
+    "FlaxCLIPModel",
+    "FlaxCLIPPreTrainedModel",
+    "FlaxCLIPTextModel",
+    "FlaxCLIPTextPreTrainedModel",
+    "FlaxCLIPTextModelWithProjection",
+    "FlaxCLIPVisionModel",
+    "FlaxCLIPVisionPreTrainedModel",
+]

lib/python3.12/site-packages/transformers/models/clip/modeling_tf_clip.py

@@ -0,0 +1,1460 @@
+# coding=utf-8
+# Copyright 2021 The OpenAI Team Authors and 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.
+"""TF 2.0 CLIP model."""
+
+from __future__ import annotations
+
+import math
+from dataclasses import dataclass
+from typing import Any, Optional, Tuple, Union
+
+import numpy as np
+import tensorflow as tf
+
+from ...activations_tf import get_tf_activation
+from ...modeling_tf_outputs import TFBaseModelOutput, TFBaseModelOutputWithPooling
+
+# Public API
+from ...modeling_tf_utils import (
+    TFModelInputType,
+    TFPreTrainedModel,
+    get_initializer,
+    keras,
+    keras_serializable,
+    unpack_inputs,
+)
+from ...tf_utils import check_embeddings_within_bounds, shape_list, stable_softmax
+from ...utils import (
+    ModelOutput,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    logging,
+    replace_return_docstrings,
+)
+from .configuration_clip import CLIPConfig, CLIPTextConfig, CLIPVisionConfig
+
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "openai/clip-vit-base-patch32"
+
+
+LARGE_NEGATIVE = -1e8
+
+
+# Copied from transformers.models.bart.modeling_tf_bart._expand_mask
+def _expand_mask(mask: tf.Tensor, tgt_len: Optional[int] = None):
+    """
+    Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
+    """
+    src_len = shape_list(mask)[1]
+    tgt_len = tgt_len if tgt_len is not None else src_len
+    one_cst = tf.constant(1.0)
+    mask = tf.cast(mask, dtype=one_cst.dtype)
+    expanded_mask = tf.tile(mask[:, None, None, :], (1, 1, tgt_len, 1))
+
+    return (one_cst - expanded_mask) * LARGE_NEGATIVE
+
+
+# contrastive loss function, adapted from
+# https://sachinruk.github.io/blog/pytorch/pytorch%20lightning/loss%20function/gpu/2021/03/07/CLIP.html
+def contrastive_loss(logits: tf.Tensor) -> tf.Tensor:
+    return tf.math.reduce_mean(
+        keras.metrics.sparse_categorical_crossentropy(
+            y_true=tf.range(shape_list(logits)[0]), y_pred=logits, from_logits=True
+        )
+    )
+
+
+def clip_loss(similarity: tf.Tensor) -> tf.Tensor:
+    caption_loss = contrastive_loss(similarity)
+    image_loss = contrastive_loss(tf.transpose(similarity))
+    return (caption_loss + image_loss) / 2.0
+
+
+@dataclass
+class TFCLIPOutput(ModelOutput):
+    """
+    Args:
+        loss (`tf.Tensor` of shape `(1,)`, *optional*, returned when `return_loss` is `True`):
+            Contrastive loss for image-text similarity.
+        logits_per_image:(`tf.Tensor` of shape `(image_batch_size, text_batch_size)`):
+            The scaled dot product scores between `image_embeds` and `text_embeds`. This represents the image-text
+            similarity scores.
+        logits_per_text:(`tf.Tensor` of shape `(text_batch_size, image_batch_size)`):
+            The scaled dot product scores between `text_embeds` and `image_embeds`. This represents the text-image
+            similarity scores.
+        text_embeds(`tf.Tensor` of shape `(batch_size, output_dim`):
+            The text embeddings obtained by applying the projection layer to the pooled output of [`TFCLIPTextModel`].
+        image_embeds(`tf.Tensor` of shape `(batch_size, output_dim`):
+            The image embeddings obtained by applying the projection layer to the pooled output of
+            [`TFCLIPVisionModel`].
+        text_model_output([`~modeling_tf_utils.TFBaseModelOutputWithPooling`]):
+            The output of the [`TFCLIPTextModel`].
+        vision_model_output([`~modeling_tf_utils.TFBaseModelOutputWithPooling`]):
+            The output of the [`TFCLIPVisionModel`].
+    """
+
+    loss: tf.Tensor | None = None
+    logits_per_image: Optional[tf.Tensor] = None
+    logits_per_text: Optional[tf.Tensor] = None
+    text_embeds: Optional[tf.Tensor] = None
+    image_embeds: Optional[tf.Tensor] = None
+    text_model_output: TFBaseModelOutputWithPooling = None
+    vision_model_output: TFBaseModelOutputWithPooling = None
+
+    def to_tuple(self) -> Tuple[Any]:
+        return tuple(
+            self[k] if k not in ["text_model_output", "vision_model_output"] else getattr(self, k).to_tuple()
+            for k in self.keys()
+        )
+
+
+class TFCLIPVisionEmbeddings(keras.layers.Layer):
+    def __init__(self, config: CLIPVisionConfig, **kwargs):
+        super().__init__(**kwargs)
+
+        self.embed_dim = config.hidden_size
+        self.image_size = config.image_size
+        self.patch_size = config.patch_size
+
+        self.num_patches = (self.image_size // self.patch_size) ** 2
+        self.num_positions = self.num_patches + 1
+
+        self.config = config
+
+        self.patch_embedding = keras.layers.Conv2D(
+            filters=self.embed_dim,
+            kernel_size=self.patch_size,
+            strides=self.patch_size,
+            padding="valid",
+            data_format="channels_last",
+            use_bias=False,
+            kernel_initializer=get_initializer(self.config.initializer_range * self.config.initializer_factor),
+            name="patch_embedding",
+        )
+
+    def build(self, input_shape: tf.TensorShape = None):
+        factor = self.config.initializer_factor
+
+        self.class_embedding = self.add_weight(
+            shape=(self.embed_dim,),
+            initializer=get_initializer(self.embed_dim**-0.5 * factor),
+            trainable=True,
+            name="class_embedding",
+        )
+
+        with tf.name_scope("position_embedding"):
+            self.position_embedding = self.add_weight(
+                shape=(self.num_positions, self.embed_dim),
+                initializer=get_initializer(self.config.initializer_range * factor),
+                trainable=True,
+                name="embeddings",
+            )
+
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "patch_embedding", None) is not None:
+            with tf.name_scope(self.patch_embedding.name):
+                self.patch_embedding.build([None, None, None, self.config.num_channels])
+
+    def call(self, pixel_values: tf.Tensor) -> tf.Tensor:
+        """`pixel_values` is expected to be of NCHW format."""
+
+        batch_size, num_channels, height, width = shape_list(pixel_values)
+
+        # When running on CPU, `tf.nn.conv2d` doesn't support `NCHW` format.
+        # So change the input format from `NCHW` to `NHWC`.
+        # shape = (batch_size, in_height, in_width, in_channels=num_channels)
+        pixel_values = tf.transpose(pixel_values, perm=(0, 2, 3, 1))
+
+        patch_embeds = self.patch_embedding(pixel_values)
+
+        # Change the 2D spatial dimensions to a single temporal dimension.
+        # shape = (batch_size, num_patches, out_channels=embed_dim)
+        patch_embeds = tf.reshape(tensor=patch_embeds, shape=(batch_size, self.num_patches, -1))
+
+        # add the [CLS] token to the embedded patch tokens
+        class_embeds = tf.broadcast_to(self.class_embedding, shape=(batch_size, 1, self.embed_dim))
+        embeddings = tf.concat((class_embeds, patch_embeds), axis=1)
+
+        embeddings = embeddings + self.position_embedding
+
+        return embeddings
+
+
+class TFCLIPTextEmbeddings(keras.layers.Layer):
+    def __init__(self, config: CLIPTextConfig, **kwargs):
+        super().__init__(**kwargs)
+
+        self.embed_dim = config.hidden_size
+
+        self.config = config
+
+    def build(self, input_shape: tf.TensorShape = None):
+        with tf.name_scope("token_embedding"):
+            self.weight = self.add_weight(
+                shape=(self.config.vocab_size, self.embed_dim),
+                initializer=get_initializer(self.config.initializer_factor * self.config.initializer_range),
+                trainable=True,
+                name="weight",
+            )
+
+        with tf.name_scope("position_embedding"):
+            self.position_embedding = self.add_weight(
+                shape=(self.config.max_position_embeddings, self.embed_dim),
+                initializer=get_initializer(self.config.initializer_factor * self.config.initializer_range),
+                trainable=True,
+                name="embeddings",
+            )
+
+        super().build(input_shape)
+
+    def call(
+        self,
+        input_ids: Optional[tf.Tensor] = None,
+        position_ids: Optional[tf.Tensor] = None,
+        inputs_embeds: Optional[tf.Tensor] = None,
+    ) -> tf.Tensor:
+        """
+        Applies embedding based on inputs tensor.
+
+        Returns:
+            final_embeddings (`tf.Tensor`): output embedding tensor.
+        """
+        if input_ids is None and inputs_embeds is None:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        if inputs_embeds is None:
+            check_embeddings_within_bounds(input_ids, self.config.vocab_size)
+            inputs_embeds = tf.gather(params=self.weight, indices=input_ids)
+
+        input_shape = shape_list(inputs_embeds)[:-1]
+
+        if position_ids is None:
+            position_ids = tf.expand_dims(tf.range(start=0, limit=input_shape[-1]), axis=0)
+
+        position_embeds = tf.gather(params=self.position_embedding, indices=position_ids)
+        position_embeds = tf.tile(input=position_embeds, multiples=(input_shape[0], 1, 1))
+        final_embeddings = inputs_embeds + position_embeds
+
+        return final_embeddings
+
+
+class TFCLIPAttention(keras.layers.Layer):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config: CLIPConfig, **kwargs):
+        super().__init__(**kwargs)
+
+        self.embed_dim = config.hidden_size
+        self.num_attention_heads = config.num_attention_heads
+        self.attention_head_size = self.embed_dim // self.num_attention_heads
+        if self.attention_head_size * self.num_attention_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_attention_heads})."
+            )
+
+        factor = config.initializer_factor
+        in_proj_std = (self.embed_dim**-0.5) * ((2 * config.num_hidden_layers) ** -0.5) * factor
+        out_proj_std = (self.embed_dim**-0.5) * factor
+
+        self.sqrt_att_head_size = math.sqrt(self.attention_head_size)
+
+        self.q_proj = keras.layers.Dense(
+            units=self.embed_dim, kernel_initializer=get_initializer(in_proj_std), name="q_proj"
+        )
+        self.k_proj = keras.layers.Dense(
+            units=self.embed_dim, kernel_initializer=get_initializer(in_proj_std), name="k_proj"
+        )
+        self.v_proj = keras.layers.Dense(
+            units=self.embed_dim, kernel_initializer=get_initializer(in_proj_std), name="v_proj"
+        )
+
+        self.dropout = keras.layers.Dropout(rate=config.attention_dropout)
+
+        self.out_proj = keras.layers.Dense(
+            units=self.embed_dim, kernel_initializer=get_initializer(out_proj_std), name="out_proj"
+        )
+
+    # copied from transformers.models.bert.modeling_tf_bert.TFBertSelfAttention.transpose_for_scores
+    def transpose_for_scores(self, tensor: tf.Tensor, batch_size: int) -> tf.Tensor:
+        # Reshape from [batch_size, seq_length, all_head_size] to [batch_size, seq_length, num_attention_heads, attention_head_size]
+        tensor = tf.reshape(tensor=tensor, shape=(batch_size, -1, self.num_attention_heads, self.attention_head_size))
+
+        # Transpose the tensor from [batch_size, seq_length, num_attention_heads, attention_head_size] to [batch_size, num_attention_heads, seq_length, attention_head_size]
+        return tf.transpose(tensor, perm=[0, 2, 1, 3])
+
+    def call(
+        self,
+        hidden_states: tf.Tensor,
+        attention_mask: tf.Tensor,
+        causal_attention_mask: tf.Tensor,
+        output_attentions: bool,
+        training: bool = False,
+    ) -> Tuple[tf.Tensor]:
+        """Input shape: Batch x Time x Channel"""
+
+        batch_size = shape_list(hidden_states)[0]
+        mixed_query_layer = self.q_proj(inputs=hidden_states)
+        mixed_key_layer = self.k_proj(inputs=hidden_states)
+        mixed_value_layer = self.v_proj(inputs=hidden_states)
+        query_layer = self.transpose_for_scores(mixed_query_layer, batch_size)
+        key_layer = self.transpose_for_scores(mixed_key_layer, batch_size)
+        value_layer = self.transpose_for_scores(mixed_value_layer, batch_size)
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        # (batch size, num_heads, seq_len_q, seq_len_k)
+        attention_scores = tf.matmul(query_layer, key_layer, transpose_b=True)
+        dk = tf.cast(self.sqrt_att_head_size, dtype=attention_scores.dtype)
+        attention_scores = tf.divide(attention_scores, dk)
+
+        # apply the causal_attention_mask first
+        if causal_attention_mask is not None:
+            # Apply the causal attention mask (precomputed for all layers in TFCLIPModel call() function)
+            attention_scores = tf.add(attention_scores, causal_attention_mask)
+
+        if attention_mask is not None:
+            # Apply the attention mask (precomputed for all layers in TFCLIPModel call() function)
+            attention_scores = tf.add(attention_scores, attention_mask)
+
+        # Normalize the attention scores to probabilities.
+        _attention_probs = stable_softmax(logits=attention_scores, axis=-1)
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        attention_probs = self.dropout(inputs=_attention_probs, training=training)
+
+        attention_output = tf.matmul(attention_probs, value_layer)
+        attention_output = tf.transpose(attention_output, perm=[0, 2, 1, 3])
+
+        # (batch_size, seq_len_q, embed_dim)
+        attention_output = tf.reshape(tensor=attention_output, shape=(batch_size, -1, self.embed_dim))
+
+        attention_output = self.out_proj(attention_output, training=training)
+        # In TFBert, attention weights are returned after dropout.
+        # However, in CLIP, they are returned before dropout.
+        outputs = (attention_output, _attention_probs) if output_attentions else (attention_output,)
+
+        return outputs
+
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "q_proj", None) is not None:
+            with tf.name_scope(self.q_proj.name):
+                self.q_proj.build([None, None, self.embed_dim])
+        if getattr(self, "k_proj", None) is not None:
+            with tf.name_scope(self.k_proj.name):
+                self.k_proj.build([None, None, self.embed_dim])
+        if getattr(self, "v_proj", None) is not None:
+            with tf.name_scope(self.v_proj.name):
+                self.v_proj.build([None, None, self.embed_dim])
+        if getattr(self, "out_proj", None) is not None:
+            with tf.name_scope(self.out_proj.name):
+                self.out_proj.build([None, None, self.embed_dim])
+
+
+class TFCLIPMLP(keras.layers.Layer):
+    def __init__(self, config: CLIPConfig, **kwargs):
+        super().__init__(**kwargs)
+
+        self.activation_fn = get_tf_activation(config.hidden_act)
+
+        factor = config.initializer_factor
+        in_proj_std = (config.hidden_size**-0.5) * ((2 * config.num_hidden_layers) ** -0.5) * factor
+        fc_std = (2 * config.hidden_size) ** -0.5 * factor
+
+        self.fc1 = keras.layers.Dense(
+            units=config.intermediate_size, kernel_initializer=get_initializer(fc_std), name="fc1"
+        )
+        self.fc2 = keras.layers.Dense(
+            units=config.hidden_size, kernel_initializer=get_initializer(in_proj_std), name="fc2"
+        )
+        self.config = config
+
+    def call(self, hidden_states: tf.Tensor) -> tf.Tensor:
+        hidden_states = self.fc1(inputs=hidden_states)
+        hidden_states = self.activation_fn(hidden_states)
+        hidden_states = self.fc2(inputs=hidden_states)
+        return hidden_states
+
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "fc1", None) is not None:
+            with tf.name_scope(self.fc1.name):
+                self.fc1.build([None, None, self.config.hidden_size])
+        if getattr(self, "fc2", None) is not None:
+            with tf.name_scope(self.fc2.name):
+                self.fc2.build([None, None, self.config.intermediate_size])
+
+
+class TFCLIPEncoderLayer(keras.layers.Layer):
+    def __init__(self, config: CLIPConfig, **kwargs):
+        super().__init__(**kwargs)
+
+        self.embed_dim = config.hidden_size
+        self.self_attn = TFCLIPAttention(config, name="self_attn")
+        self.layer_norm1 = keras.layers.LayerNormalization(epsilon=config.layer_norm_eps, name="layer_norm1")
+        self.mlp = TFCLIPMLP(config, name="mlp")
+        self.layer_norm2 = keras.layers.LayerNormalization(epsilon=config.layer_norm_eps, name="layer_norm2")
+
+    def call(
+        self,
+        hidden_states: tf.Tensor,
+        attention_mask: tf.Tensor,
+        causal_attention_mask: tf.Tensor,
+        output_attentions: bool,
+        training: bool = False,
+    ) -> Tuple[tf.Tensor]:
+        """
+        Args:
+            hidden_states (`tf.Tensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+            attention_mask (`tf.Tensor`): attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            causal_attention_mask (`tf.Tensor`): causal attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            output_attentions (`bool`):
+                Whether or not to return the attentions tensors of all attention layers. See `outputs` under returned
+                tensors for more detail.
+        """
+        residual = hidden_states
+
+        hidden_states = self.layer_norm1(inputs=hidden_states)
+        attention_outputs = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            causal_attention_mask=causal_attention_mask,
+            output_attentions=output_attentions,
+            training=training,
+        )
+        hidden_states = attention_outputs[0]
+        hidden_states = residual + hidden_states
+
+        residual = hidden_states
+        hidden_states = self.layer_norm2(inputs=hidden_states)
+        hidden_states = self.mlp(hidden_states=hidden_states)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,) + attention_outputs[1:]  # add attentions if we output them
+
+        return outputs
+
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "self_attn", None) is not None:
+            with tf.name_scope(self.self_attn.name):
+                self.self_attn.build(None)
+        if getattr(self, "layer_norm1", None) is not None:
+            with tf.name_scope(self.layer_norm1.name):
+                self.layer_norm1.build([None, None, self.embed_dim])
+        if getattr(self, "mlp", None) is not None:
+            with tf.name_scope(self.mlp.name):
+                self.mlp.build(None)
+        if getattr(self, "layer_norm2", None) is not None:
+            with tf.name_scope(self.layer_norm2.name):
+                self.layer_norm2.build([None, None, self.embed_dim])
+
+
+class TFCLIPEncoder(keras.layers.Layer):
+    """
+    Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a
+    [`TFCLIPEncoderLayer`].
+
+    Args:
+        config: CLIPConfig
+    """
+
+    def __init__(self, config: CLIPConfig, **kwargs):
+        super().__init__(**kwargs)
+
+        self.layers = [TFCLIPEncoderLayer(config, name=f"layers_._{i}") for i in range(config.num_hidden_layers)]
+
+    def call(
+        self,
+        hidden_states: tf.Tensor,
+        attention_mask: tf.Tensor,
+        causal_attention_mask: tf.Tensor,
+        output_attentions: bool,
+        output_hidden_states: bool,
+        return_dict: bool,
+        training: bool = False,
+    ) -> Union[TFBaseModelOutput, Tuple[tf.Tensor]]:
+        all_hidden_states = () if output_hidden_states else None
+        all_attentions = () if output_attentions else None
+
+        for i, layer_module in enumerate(self.layers):
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            layer_outputs = layer_module(
+                hidden_states=hidden_states,
+                attention_mask=attention_mask,
+                causal_attention_mask=causal_attention_mask,
+                output_attentions=output_attentions,
+                training=training,
+            )
+            hidden_states = layer_outputs[0]
+
+            if output_attentions:
+                all_attentions = all_attentions + (layer_outputs[1],)
+
+        # Add last layer
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, all_hidden_states, all_attentions] if v is not None)
+
+        return TFBaseModelOutput(
+            last_hidden_state=hidden_states, hidden_states=all_hidden_states, attentions=all_attentions
+        )
+
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "layers", None) is not None:
+            for layer in self.layers:
+                with tf.name_scope(layer.name):
+                    layer.build(None)
+
+
+class TFCLIPTextTransformer(keras.layers.Layer):
+    def __init__(self, config: CLIPTextConfig, **kwargs):
+        super().__init__(**kwargs)
+
+        self.embeddings = TFCLIPTextEmbeddings(config, name="embeddings")
+        self.encoder = TFCLIPEncoder(config, name="encoder")
+        self.final_layer_norm = keras.layers.LayerNormalization(epsilon=config.layer_norm_eps, name="final_layer_norm")
+
+        # For `pooled_output` computation
+        self.eos_token_id = config.eos_token_id
+        self.embed_dim = config.hidden_size
+
+    def call(
+        self,
+        input_ids: TFModelInputType,
+        attention_mask: tf.Tensor,
+        position_ids: tf.Tensor,
+        output_attentions: bool,
+        output_hidden_states: bool,
+        return_dict: bool,
+        training: bool = False,
+    ) -> Union[TFBaseModelOutputWithPooling, Tuple[tf.Tensor]]:
+        input_shape = shape_list(input_ids)
+
+        embedding_output = self.embeddings(input_ids=input_ids, position_ids=position_ids)
+
+        batch_size, seq_length = input_shape
+        # CLIP's text model uses causal mask, prepare it here.
+        # https://github.com/openai/CLIP/blob/cfcffb90e69f37bf2ff1e988237a0fbe41f33c04/clip/model.py#L324
+        causal_attention_mask = self._build_causal_attention_mask(batch_size, seq_length, dtype=embedding_output.dtype)
+
+        # check attention mask and invert
+        # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+        attention_mask = _expand_mask(attention_mask)
+
+        encoder_outputs = self.encoder(
+            hidden_states=embedding_output,
+            attention_mask=attention_mask,
+            causal_attention_mask=causal_attention_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            training=training,
+        )
+
+        sequence_output = encoder_outputs[0]
+        sequence_output = self.final_layer_norm(inputs=sequence_output)
+
+        if self.eos_token_id == 2:
+            # The `eos_token_id` was incorrect before PR #24773: Let's keep what have been done here.
+            # A CLIP model with such `eos_token_id` in the config can't work correctly with extra new tokens added
+            # ------------------------------------------------------------
+            # text_embeds.shape = [batch_size, n_ctx, transformer.width]
+            # take features from the eot embedding (eot_token is the highest number in each sequence)
+            pooled_output = tf.gather_nd(
+                params=sequence_output,
+                indices=tf.stack(
+                    values=(tf.range(input_shape[0], dtype=tf.int64), tf.math.argmax(input_ids, axis=-1)), axis=1
+                ),
+            )
+        else:
+            # The config gets updated `eos_token_id` from PR #24773 (so the use of exta new tokens is possible)
+            pooled_output = tf.gather_nd(
+                params=sequence_output,
+                indices=tf.stack(
+                    values=(
+                        tf.range(input_shape[0], dtype=tf.int64),
+                        tf.math.argmax(tf.cast(input_ids == self.eos_token_id, dtype=tf.int8), axis=-1),
+                    ),
+                    axis=1,
+                ),
+            )
+
+        if not return_dict:
+            return (sequence_output, pooled_output) + encoder_outputs[1:]
+
+        return TFBaseModelOutputWithPooling(
+            last_hidden_state=sequence_output,
+            pooler_output=pooled_output,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )
+
+    def _build_causal_attention_mask(self, batch_size, seq_length, dtype=tf.float32):
+        # It is possible with an unspecified sequence length for seq_length to be
+        # a runtime value, which is unsupported by tf.constant. Per the TensorFlow
+        # docs, tf.fill can handle runtime dynamic shapes:
+        # https://www.tensorflow.org/api_docs/python/tf/fill
+        diag = tf.cast(tf.fill((seq_length,), 0.0), dtype)
+
+        # set an additive 2D attention mask with all places being masked
+        to_mask = tf.cast(tf.fill((seq_length, seq_length), -10000.0), dtype)
+
+        # set diagonal & lower triangular parts to 0 (i.e. the places not to be masked)
+        # TIP: think the 2D matrix as the space of (query_seq, key_seq)
+        to_mask = tf.linalg.band_part(to_mask, 0, -1)
+        # to_mask = tf.linalg.band_part(to_mask, -1, 0)
+        to_mask = tf.linalg.set_diag(to_mask, diagonal=diag)
+
+        return tf.broadcast_to(input=to_mask, shape=(batch_size, 1, seq_length, seq_length))
+
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "embeddings", None) is not None:
+            with tf.name_scope(self.embeddings.name):
+                self.embeddings.build(None)
+        if getattr(self, "encoder", None) is not None:
+            with tf.name_scope(self.encoder.name):
+                self.encoder.build(None)
+        if getattr(self, "final_layer_norm", None) is not None:
+            with tf.name_scope(self.final_layer_norm.name):
+                self.final_layer_norm.build([None, None, self.embed_dim])
+
+
+@keras_serializable
+class TFCLIPTextMainLayer(keras.layers.Layer):
+    config_class = CLIPTextConfig
+
+    def __init__(self, config: CLIPTextConfig, **kwargs):
+        super().__init__(**kwargs)
+        self.config = config
+        self.text_model = TFCLIPTextTransformer(config, name="text_model")
+
+    def get_input_embeddings(self) -> keras.layers.Layer:
+        return self.text_model.embeddings
+
+    def set_input_embeddings(self, value: tf.Variable):
+        self.text_model.embeddings.weight = value
+        self.text_model.embeddings.vocab_size = shape_list(value)[0]
+
+    @unpack_inputs
+    def call(
+        self,
+        input_ids: TFModelInputType | None = None,
+        attention_mask: np.ndarray | tf.Tensor | None = None,
+        position_ids: np.ndarray | tf.Tensor | None = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        training: bool = False,
+    ) -> Union[TFBaseModelOutputWithPooling, Tuple[tf.Tensor]]:
+        if input_ids is None:
+            raise ValueError("You have to specify input_ids")
+
+        input_shape = shape_list(input_ids)
+
+        if attention_mask is None:
+            attention_mask = tf.fill(dims=input_shape, value=1)
+
+        text_model_outputs = self.text_model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            training=training,
+        )
+
+        return text_model_outputs
+
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "text_model", None) is not None:
+            with tf.name_scope(self.text_model.name):
+                self.text_model.build(None)
+
+
+class TFCLIPVisionTransformer(keras.layers.Layer):
+    def __init__(self, config: CLIPVisionConfig, **kwargs):
+        super().__init__(**kwargs)
+
+        self.embeddings = TFCLIPVisionEmbeddings(config, name="embeddings")
+        self.pre_layernorm = keras.layers.LayerNormalization(epsilon=config.layer_norm_eps, name="pre_layrnorm")
+        self.encoder = TFCLIPEncoder(config, name="encoder")
+        self.post_layernorm = keras.layers.LayerNormalization(epsilon=config.layer_norm_eps, name="post_layernorm")
+        self.embed_dim = config.hidden_size
+
+    def call(
+        self,
+        pixel_values: TFModelInputType,
+        output_attentions: bool,
+        output_hidden_states: bool,
+        return_dict: bool,
+        training: bool = False,
+    ) -> Union[TFBaseModelOutputWithPooling, Tuple[tf.Tensor]]:
+        embedding_output = self.embeddings(pixel_values=pixel_values)
+        embedding_output = self.pre_layernorm(inputs=embedding_output)
+
+        encoder_outputs = self.encoder(
+            hidden_states=embedding_output,
+            attention_mask=None,
+            causal_attention_mask=None,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            training=training,
+        )
+
+        sequence_output = encoder_outputs[0]
+        pooled_output = sequence_output[:, 0, :]
+        pooled_output = self.post_layernorm(inputs=pooled_output)
+
+        if not return_dict:
+            return (sequence_output, pooled_output) + encoder_outputs[1:]
+
+        return TFBaseModelOutputWithPooling(
+            last_hidden_state=sequence_output,
+            pooler_output=pooled_output,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )
+
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "embeddings", None) is not None:
+            with tf.name_scope(self.embeddings.name):
+                self.embeddings.build(None)
+        if getattr(self, "pre_layernorm", None) is not None:
+            with tf.name_scope(self.pre_layernorm.name):
+                self.pre_layernorm.build([None, None, self.embed_dim])
+        if getattr(self, "encoder", None) is not None:
+            with tf.name_scope(self.encoder.name):
+                self.encoder.build(None)
+        if getattr(self, "post_layernorm", None) is not None:
+            with tf.name_scope(self.post_layernorm.name):
+                self.post_layernorm.build([None, self.embed_dim])
+
+
+@keras_serializable
+class TFCLIPVisionMainLayer(keras.layers.Layer):
+    config_class = CLIPVisionConfig
+
+    def __init__(self, config: CLIPVisionConfig, **kwargs):
+        super().__init__(**kwargs)
+        self.config = config
+        self.vision_model = TFCLIPVisionTransformer(config, name="vision_model")
+
+    def get_input_embeddings(self) -> keras.layers.Layer:
+        return self.vision_model.embeddings
+
+    @unpack_inputs
+    def call(
+        self,
+        pixel_values: TFModelInputType | None = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        training: bool = False,
+    ) -> Union[TFBaseModelOutputWithPooling, Tuple[tf.Tensor]]:
+        if pixel_values is None:
+            raise ValueError("You have to specify pixel_values")
+
+        vision_model_outputs = self.vision_model(
+            pixel_values=pixel_values,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            training=training,
+        )
+
+        return vision_model_outputs
+
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "vision_model", None) is not None:
+            with tf.name_scope(self.vision_model.name):
+                self.vision_model.build(None)
+
+
+@keras_serializable
+class TFCLIPMainLayer(keras.layers.Layer):
+    config_class = CLIPConfig
+
+    def __init__(self, config: CLIPConfig, **kwargs):
+        super().__init__(**kwargs)
+
+        if not isinstance(config.text_config, CLIPTextConfig):
+            raise TypeError(
+                "config.text_config is expected to be of type CLIPTextConfig but is of type"
+                f" {type(config.text_config)}."
+            )
+
+        if not isinstance(config.vision_config, CLIPVisionConfig):
+            raise TypeError(
+                "config.vision_config is expected to be of type CLIPVisionConfig but is of type"
+                f" {type(config.vision_config)}."
+            )
+
+        self.config = config
+
+        text_config = config.text_config
+        vision_config = config.vision_config
+
+        self.projection_dim = config.projection_dim
+
+        self.text_model = TFCLIPTextTransformer(text_config, name="text_model")
+        self.vision_model = TFCLIPVisionTransformer(vision_config, name="vision_model")
+
+        self.visual_projection = keras.layers.Dense(
+            units=self.projection_dim,
+            kernel_initializer=get_initializer(vision_config.hidden_size**-0.5 * self.config.initializer_factor),
+            use_bias=False,
+            name="visual_projection",
+        )
+
+        self.text_projection = keras.layers.Dense(
+            units=self.projection_dim,
+            kernel_initializer=get_initializer(text_config.hidden_size**-0.5 * self.config.initializer_factor),
+            use_bias=False,
+            name="text_projection",
+        )
+        self.text_embed_dim = text_config.hidden_size
+        self.vision_embed_dim = vision_config.hidden_size
+
+    def build(self, input_shape: tf.TensorShape = None):
+        self.logit_scale = self.add_weight(
+            shape=(1,),
+            initializer=keras.initializers.Constant(self.config.logit_scale_init_value),
+            trainable=True,
+            name="logit_scale",
+        )
+
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "text_model", None) is not None:
+            with tf.name_scope(self.text_model.name):
+                self.text_model.build(None)
+        if getattr(self, "vision_model", None) is not None:
+            with tf.name_scope(self.vision_model.name):
+                self.vision_model.build(None)
+        if getattr(self, "visual_projection", None) is not None:
+            with tf.name_scope(self.visual_projection.name):
+                self.visual_projection.build([None, None, self.vision_embed_dim])
+        if getattr(self, "text_projection", None) is not None:
+            with tf.name_scope(self.text_projection.name):
+                self.text_projection.build([None, None, self.text_embed_dim])
+
+    @unpack_inputs
+    def get_text_features(
+        self,
+        input_ids: TFModelInputType | None = None,
+        attention_mask: np.ndarray | tf.Tensor | None = None,
+        position_ids: np.ndarray | tf.Tensor | None = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        training: bool = False,
+    ) -> tf.Tensor:
+        if input_ids is None:
+            raise ValueError("You have to specify either input_ids")
+
+        input_shape = shape_list(input_ids)
+
+        if attention_mask is None:
+            attention_mask = tf.fill(dims=input_shape, value=1)
+
+        text_outputs = self.text_model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            training=training,
+        )
+
+        pooled_output = text_outputs[1]
+        text_features = self.text_projection(inputs=pooled_output)
+
+        return text_features
+
+    @unpack_inputs
+    def get_image_features(
+        self,
+        pixel_values: TFModelInputType | None = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        training: bool = False,
+    ) -> tf.Tensor:
+        if pixel_values is None:
+            raise ValueError("You have to specify pixel_values")
+
+        vision_outputs = self.vision_model(
+            pixel_values=pixel_values,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            training=training,
+        )
+
+        pooled_output = vision_outputs[1]  # pooled_output
+        image_features = self.visual_projection(inputs=pooled_output)
+
+        return image_features
+
+    @unpack_inputs
+    def call(
+        self,
+        input_ids: TFModelInputType | None = None,
+        pixel_values: TFModelInputType | None = None,
+        attention_mask: np.ndarray | tf.Tensor | None = None,
+        position_ids: np.ndarray | tf.Tensor | None = None,
+        return_loss: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        training: bool = False,
+    ) -> Union[TFCLIPOutput, Tuple[tf.Tensor]]:
+        if input_ids is None:
+            raise ValueError("You have to specify either input_ids")
+        if pixel_values is None:
+            raise ValueError("You have to specify pixel_values")
+
+        input_shape = shape_list(input_ids)
+
+        if attention_mask is None:
+            attention_mask = tf.fill(dims=input_shape, value=1)
+
+        vision_outputs = self.vision_model(
+            pixel_values=pixel_values,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            training=training,
+        )
+
+        text_outputs = self.text_model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            training=training,
+        )
+
+        image_embeds = vision_outputs[1]
+        image_embeds = self.visual_projection(inputs=image_embeds)
+
+        text_embeds = text_outputs[1]
+        text_embeds = self.text_projection(inputs=text_embeds)
+
+        # normalized features
+        image_embeds = image_embeds / tf.norm(tensor=image_embeds, ord="euclidean", axis=-1, keepdims=True)
+        text_embeds = text_embeds / tf.norm(tensor=text_embeds, ord="euclidean", axis=-1, keepdims=True)
+
+        # cosine similarity as logits
+        logit_scale = tf.math.exp(self.logit_scale)
+        logits_per_text = tf.matmul(text_embeds, image_embeds, transpose_b=True) * logit_scale
+        logits_per_image = tf.transpose(logits_per_text)
+
+        loss = None
+        if return_loss:
+            loss = clip_loss(logits_per_text)
+            loss = tf.reshape(loss, (1,))
+
+        if not return_dict:
+            output = (logits_per_image, logits_per_text, text_embeds, image_embeds, text_outputs, vision_outputs)
+            return (loss,) + output if loss is not None else output
+
+        return TFCLIPOutput(
+            loss=loss,
+            logits_per_image=logits_per_image,
+            logits_per_text=logits_per_text,
+            text_embeds=text_embeds,
+            image_embeds=image_embeds,
+            text_model_output=text_outputs,
+            vision_model_output=vision_outputs,
+        )
+
+
+class TFCLIPPreTrainedModel(TFPreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = CLIPConfig
+    base_model_prefix = "clip"
+    _keys_to_ignore_on_load_missing = [r"position_ids"]
+    _keys_to_ignore_on_load_unexpected = [r"position_ids"]
+
+
+CLIP_START_DOCSTRING = r"""
+
+    This model inherits from [`TFPreTrainedModel`]. 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 [keras.Model](https://www.tensorflow.org/api_docs/python/tf/keras/Model) subclass. Use it
+    as a regular TF 2.0 Keras Model and refer to the TF 2.0 documentation for all matter related to general usage and
+    behavior.
+
+    <Tip>
+
+    TensorFlow models and layers in `transformers` accept two formats as input:
+
+    - having all inputs as keyword arguments (like PyTorch models), or
+    - having all inputs as a list, tuple or dict in the first positional argument.
+
+    The reason the second format is supported is that Keras methods prefer this format when passing inputs to models
+    and layers. Because of this support, when using methods like `model.fit()` things should "just work" for you - just
+    pass your inputs and labels in any format that `model.fit()` supports! If, however, you want to use the second
+    format outside of Keras methods like `fit()` and `predict()`, such as when creating your own layers or models with
+    the Keras `Functional` API, there are three possibilities you can use to gather all the input Tensors in the first
+    positional argument:
+
+    - a single Tensor with `input_ids` only and nothing else: `model(input_ids)`
+    - a list of varying length with one or several input Tensors IN THE ORDER given in the docstring:
+    `model([input_ids, attention_mask])` or `model([input_ids, attention_mask, token_type_ids])`
+    - a dictionary with one or several input Tensors associated to the input names given in the docstring:
+    `model({"input_ids": input_ids, "token_type_ids": token_type_ids})`
+
+    Note that when creating models and layers with
+    [subclassing](https://keras.io/guides/making_new_layers_and_models_via_subclassing/) then you don't need to worry
+    about any of this, as you can just pass inputs like you would to any other Python function!
+
+    </Tip>
+
+    Args:
+        config ([`CLIPConfig`]): 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 [`~TFPreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+CLIP_TEXT_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`np.ndarray`, `tf.Tensor`, `List[tf.Tensor]` ``Dict[str, tf.Tensor]` or `Dict[str, np.ndarray]` and each example must have the shape `({0})`):
+            Indices of input sequence tokens in the vocabulary.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.__call__`] and
+            [`PreTrainedTokenizer.encode`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`np.ndarray` or `tf.Tensor` of shape `({0})`, *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)
+        position_ids (`np.ndarray` or `tf.Tensor` of shape `({0})`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.max_position_embeddings - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail. This argument can be used only in eager mode, in graph mode the value in the
+            config will be used instead.
+        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. This argument can be used only in eager mode, in graph mode the value in the config will be
+            used instead.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. This argument can be used in
+            eager mode, in graph mode the value will always be set to True.
+        training (`bool`, *optional*, defaults to `False``):
+            Whether or not to use the model in training mode (some modules like dropout modules have different
+            behaviors between training and evaluation).
+"""
+
+CLIP_VISION_INPUTS_DOCSTRING = r"""
+    Args:
+        pixel_values (`np.ndarray`, `tf.Tensor`, `List[tf.Tensor]` ``Dict[str, tf.Tensor]` or `Dict[str, np.ndarray]` and each example must have the shape `(batch_size, num_channels, height, width)`):
+            Pixel values. Pixel values can be obtained using [`AutoImageProcessor`]. See
+            [`CLIPImageProcessor.__call__`] for details. output_attentions (`bool`, *optional*): Whether or not to
+            return the attentions tensors of all attention layers. See `attentions` under returned tensors for more
+            detail. This argument can be used only in eager mode, in graph mode the value in the config will be used
+            instead.
+        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. This argument can be used only in eager mode, in graph mode the value in the config will be
+            used instead.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. This argument can be used in
+            eager mode, in graph mode the value will always be set to True.
+        training (`bool`, *optional*, defaults to `False``):
+            Whether or not to use the model in training mode (some modules like dropout modules have different
+            behaviors between training and evaluation).
+"""
+
+CLIP_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`np.ndarray`, `tf.Tensor`, `List[tf.Tensor]` ``Dict[str, tf.Tensor]` or `Dict[str, np.ndarray]` and each example must have the shape `({0})`):
+            Indices of input sequence tokens in the vocabulary.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.__call__`] and
+            [`PreTrainedTokenizer.encode`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        pixel_values (`np.ndarray`, `tf.Tensor`, `List[tf.Tensor]` `Dict[str, tf.Tensor]` or `Dict[str, np.ndarray]` and each example must have the shape `(batch_size, num_channels, height, width)`):
+            Pixel values. Pixel values can be obtained using [`AutoImageProcessor`]. See
+            [`CLIPImageProcessor.__call__`] for details.
+        attention_mask (`np.ndarray` or `tf.Tensor` of shape `({0})`, *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)
+        position_ids (`np.ndarray` or `tf.Tensor` of shape `({0})`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.max_position_embeddings - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        return_loss (`bool`, *optional*):
+            Whether or not to return the contrastive loss.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail. This argument can be used only in eager mode, in graph mode the value in the
+            config will be used instead.
+        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. This argument can be used only in eager mode, in graph mode the value in the config will be
+            used instead.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. This argument can be used in
+            eager mode, in graph mode the value will always be set to True.
+        training (`bool`, *optional*, defaults to `False``):
+            Whether or not to use the model in training mode (some modules like dropout modules have different
+            behaviors between training and evaluation).
+"""
+
+
+class TFCLIPTextModel(TFCLIPPreTrainedModel):
+    config_class = CLIPTextConfig
+
+    def __init__(self, config: CLIPTextConfig, *inputs, **kwargs):
+        super().__init__(config, *inputs, **kwargs)
+
+        self.clip = TFCLIPTextMainLayer(config, name="clip")
+
+    @unpack_inputs
+    @add_start_docstrings_to_model_forward(CLIP_TEXT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @replace_return_docstrings(output_type=TFBaseModelOutputWithPooling, config_class=CLIPTextConfig)
+    def call(
+        self,
+        input_ids: TFModelInputType | None = None,
+        attention_mask: np.ndarray | tf.Tensor | None = None,
+        position_ids: np.ndarray | tf.Tensor | None = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        training: Optional[bool] = False,
+    ) -> Union[TFBaseModelOutputWithPooling, Tuple[tf.Tensor]]:
+        r"""
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from transformers import AutoTokenizer, TFCLIPTextModel
+
+        >>> model = TFCLIPTextModel.from_pretrained("openai/clip-vit-base-patch32")
+        >>> tokenizer = AutoTokenizer.from_pretrained("openai/clip-vit-base-patch32")
+
+        >>> inputs = tokenizer(["a photo of a cat", "a photo of a dog"], padding=True, return_tensors="tf")
+
+        >>> outputs = model(**inputs)
+        >>> last_hidden_state = outputs.last_hidden_state
+        >>> pooled_output = outputs.pooler_output  # pooled (EOS token) states
+        ```"""
+
+        outputs = self.clip(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            training=training,
+        )
+
+        return outputs
+
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "clip", None) is not None:
+            with tf.name_scope(self.clip.name):
+                self.clip.build(None)
+
+
+class TFCLIPVisionModel(TFCLIPPreTrainedModel):
+    config_class = CLIPVisionConfig
+    main_input_name = "pixel_values"
+
+    def __init__(self, config: CLIPVisionConfig, *inputs, **kwargs):
+        super().__init__(config, *inputs, **kwargs)
+
+        self.clip = TFCLIPVisionMainLayer(config, name="clip")
+
+    @unpack_inputs
+    @add_start_docstrings_to_model_forward(CLIP_VISION_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=TFBaseModelOutputWithPooling, config_class=CLIPVisionConfig)
+    def call(
+        self,
+        pixel_values: TFModelInputType | None = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        training: Optional[bool] = False,
+    ) -> Union[TFBaseModelOutputWithPooling, Tuple[tf.Tensor]]:
+        r"""
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from PIL import Image
+        >>> import requests
+        >>> from transformers import AutoProcessor, TFCLIPVisionModel
+
+        >>> model = TFCLIPVisionModel.from_pretrained("openai/clip-vit-base-patch32")
+        >>> processor = AutoProcessor.from_pretrained("openai/clip-vit-base-patch32")
+
+        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> image = Image.open(requests.get(url, stream=True).raw)
+
+        >>> inputs = processor(images=image, return_tensors="tf")
+
+        >>> outputs = model(**inputs)
+        >>> last_hidden_state = outputs.last_hidden_state
+        >>> pooled_output = outputs.pooler_output  # pooled CLS states
+        ```"""
+
+        outputs = self.clip(
+            pixel_values=pixel_values,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            training=training,
+        )
+
+        return outputs
+
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "clip", None) is not None:
+            with tf.name_scope(self.clip.name):
+                self.clip.build(None)
+
+
+@add_start_docstrings(CLIP_START_DOCSTRING)
+class TFCLIPModel(TFCLIPPreTrainedModel):
+    config_class = CLIPConfig
+
+    def __init__(self, config: CLIPConfig, *inputs, **kwargs):
+        super().__init__(config, *inputs, **kwargs)
+
+        self.clip = TFCLIPMainLayer(config, name="clip")
+
+    @unpack_inputs
+    @add_start_docstrings_to_model_forward(CLIP_TEXT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    def get_text_features(
+        self,
+        input_ids: TFModelInputType | None = None,
+        attention_mask: np.ndarray | tf.Tensor | None = None,
+        position_ids: np.ndarray | tf.Tensor | None = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        training: bool = False,
+    ) -> tf.Tensor:
+        r"""
+        Returns:
+            text_features (`tf.Tensor` of shape `(batch_size, output_dim`): The text embeddings obtained by applying
+            the projection layer to the pooled output of [`TFCLIPTextModel`].
+
+        Examples:
+
+        ```python
+        >>> from transformers import AutoTokenizer, TFCLIPModel
+
+        >>> model = TFCLIPModel.from_pretrained("openai/clip-vit-base-patch32")
+        >>> tokenizer = AutoTokenizer.from_pretrained("openai/clip-vit-base-patch32")
+
+        >>> inputs = tokenizer(["a photo of a cat", "a photo of a dog"], padding=True, return_tensors="tf")
+        >>> text_features = model.get_text_features(**inputs)
+        ```"""
+
+        text_features = self.clip.get_text_features(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        return text_features
+
+    @unpack_inputs
+    @add_start_docstrings_to_model_forward(CLIP_VISION_INPUTS_DOCSTRING)
+    def get_image_features(
+        self,
+        pixel_values: TFModelInputType | None = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        training: bool = False,
+    ) -> tf.Tensor:
+        r"""
+        Returns:
+            image_features (`tf.Tensor` of shape `(batch_size, output_dim`): The image embeddings obtained by applying
+            the projection layer to the pooled output of [`TFCLIPVisionModel`].
+
+        Examples:
+
+        ```python
+        >>> from PIL import Image
+        >>> import requests
+        >>> from transformers import AutoProcessor, TFCLIPModel
+
+        >>> model = TFCLIPModel.from_pretrained("openai/clip-vit-base-patch32")
+        >>> processor = AutoProcessor.from_pretrained("openai/clip-vit-base-patch32")
+
+        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> image = Image.open(requests.get(url, stream=True).raw)
+
+        >>> inputs = processor(images=image, return_tensors="tf")
+
+        >>> image_features = model.get_image_features(**inputs)
+        ```"""
+
+        image_features = self.clip.get_image_features(
+            pixel_values=pixel_values,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        return image_features
+
+    @unpack_inputs
+    @add_start_docstrings_to_model_forward(CLIP_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @replace_return_docstrings(output_type=TFCLIPOutput, config_class=CLIPConfig)
+    def call(
+        self,
+        input_ids: TFModelInputType | None = None,
+        pixel_values: TFModelInputType | None = None,
+        attention_mask: np.ndarray | tf.Tensor | None = None,
+        position_ids: np.ndarray | tf.Tensor | None = None,
+        return_loss: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        training: bool = False,
+    ) -> Union[TFCLIPOutput, Tuple[tf.Tensor]]:
+        r"""
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> import tensorflow as tf
+        >>> from PIL import Image
+        >>> import requests
+        >>> from transformers import AutoProcessor, TFCLIPModel
+
+        >>> model = TFCLIPModel.from_pretrained("openai/clip-vit-base-patch32")
+        >>> processor = AutoProcessor.from_pretrained("openai/clip-vit-base-patch32")
+
+        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> image = Image.open(requests.get(url, stream=True).raw)
+
+        >>> inputs = processor(
+        ...     text=["a photo of a cat", "a photo of a dog"], images=image, return_tensors="tf", padding=True
+        ... )
+
+        >>> outputs = model(**inputs)
+        >>> logits_per_image = outputs.logits_per_image  # this is the image-text similarity score
+        >>> probs = tf.nn.softmax(logits_per_image, axis=1)  # we can take the softmax to get the label probabilities
+        ```"""
+
+        outputs = self.clip(
+            input_ids=input_ids,
+            pixel_values=pixel_values,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            return_loss=return_loss,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        return outputs
+
+    def serving_output(self, output: TFCLIPOutput) -> TFCLIPOutput:
+        # TODO: As is this currently fails with saved_model=True, because
+        # TensorFlow cannot trace through nested dataclasses. Reference:
+        # https://github.com/huggingface/transformers/pull/16886
+        return output
+
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "clip", None) is not None:
+            with tf.name_scope(self.clip.name):
+                self.clip.build(None)
+
+
+__all__ = ["TFCLIPModel", "TFCLIPPreTrainedModel", "TFCLIPTextModel", "TFCLIPVisionModel"]

lib/python3.12/site-packages/transformers/models/clip/processing_clip.py

@@ -0,0 +1,156 @@
+# coding=utf-8
+# Copyright 2021 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.
+"""
+Image/Text processor class for CLIP
+"""
+
+import warnings
+
+from ...processing_utils import ProcessorMixin
+from ...tokenization_utils_base import BatchEncoding
+
+
+class CLIPProcessor(ProcessorMixin):
+    r"""
+    Constructs a CLIP processor which wraps a CLIP image processor and a CLIP tokenizer into a single processor.
+
+    [`CLIPProcessor`] offers all the functionalities of [`CLIPImageProcessor`] and [`CLIPTokenizerFast`]. See the
+    [`~CLIPProcessor.__call__`] and [`~CLIPProcessor.decode`] for more information.
+
+    Args:
+        image_processor ([`CLIPImageProcessor`], *optional*):
+            The image processor is a required input.
+        tokenizer ([`CLIPTokenizerFast`], *optional*):
+            The tokenizer is a required input.
+    """
+
+    attributes = ["image_processor", "tokenizer"]
+    image_processor_class = ("CLIPImageProcessor", "CLIPImageProcessorFast")
+    tokenizer_class = ("CLIPTokenizer", "CLIPTokenizerFast")
+
+    def __init__(self, image_processor=None, tokenizer=None, **kwargs):
+        feature_extractor = None
+        if "feature_extractor" in kwargs:
+            warnings.warn(
+                "The `feature_extractor` argument is deprecated and will be removed in v5, use `image_processor`"
+                " instead.",
+                FutureWarning,
+            )
+            feature_extractor = kwargs.pop("feature_extractor")
+
+        image_processor = image_processor if image_processor is not None else feature_extractor
+        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`.")
+
+        super().__init__(image_processor, tokenizer)
+
+    def __call__(self, text=None, images=None, return_tensors=None, **kwargs):
+        """
+        Main method to prepare for the model one or several sequences(s) and image(s). This method forwards the `text`
+        and `kwargs` arguments to CLIPTokenizerFast's [`~CLIPTokenizerFast.__call__`] if `text` is not `None` to encode
+        the text. To prepare the image(s), this method forwards the `images` and `kwrags` arguments to
+        CLIPImageProcessor's [`~CLIPImageProcessor.__call__`] if `images` is not `None`. Please refer to the docstring
+        of the above two methods for more information.
+
+        Args:
+            text (`str`, `List[str]`, `List[List[str]]`):
+                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).
+            images (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, `List[PIL.Image.Image]`, `List[np.ndarray]`, `List[torch.Tensor]`):
+                The image or batch of images to be prepared. Each image can be a PIL image, NumPy array or PyTorch
+                tensor. Both channels-first and channels-last formats are supported.
+
+            return_tensors (`str` or [`~utils.TensorType`], *optional*):
+                If set, will return tensors of a particular framework. Acceptable values are:
+
+                - `'tf'`: Return TensorFlow `tf.constant` objects.
+                - `'pt'`: Return PyTorch `torch.Tensor` objects.
+                - `'np'`: Return NumPy `np.ndarray` objects.
+                - `'jax'`: Return JAX `jnp.ndarray` objects.
+
+        Returns:
+            [`BatchEncoding`]: A [`BatchEncoding`] with the following fields:
+
+            - **input_ids** -- List of token ids to be fed to a model. Returned when `text` is not `None`.
+            - **attention_mask** -- List of indices specifying which tokens should be attended to by the model (when
+              `return_attention_mask=True` or if *"attention_mask"* is in `self.model_input_names` and if `text` is not
+              `None`).
+            - **pixel_values** -- Pixel values to be fed to a model. Returned when `images` is not `None`.
+        """
+        tokenizer_kwargs, image_processor_kwargs = {}, {}
+        if kwargs:
+            tokenizer_kwargs = {k: v for k, v in kwargs.items() if k not in self.image_processor._valid_processor_keys}
+            image_processor_kwargs = {
+                k: v for k, v in kwargs.items() if k in self.image_processor._valid_processor_keys
+            }
+
+        if text is None and images is None:
+            raise ValueError("You have to specify either text or images. Both cannot be none.")
+
+        if text is not None:
+            encoding = self.tokenizer(text, return_tensors=return_tensors, **tokenizer_kwargs)
+
+        if images is not None:
+            image_features = self.image_processor(images, return_tensors=return_tensors, **image_processor_kwargs)
+
+        if text is not None and images is not None:
+            encoding["pixel_values"] = image_features.pixel_values
+            return encoding
+        elif text is not None:
+            return encoding
+        else:
+            return BatchEncoding(data=dict(**image_features), tensor_type=return_tensors)
+
+    def batch_decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to CLIPTokenizerFast'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 CLIPTokenizerFast'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))
+
+    @property
+    def feature_extractor_class(self):
+        warnings.warn(
+            "`feature_extractor_class` is deprecated and will be removed in v5. Use `image_processor_class` instead.",
+            FutureWarning,
+        )
+        return self.image_processor_class
+
+    @property
+    def feature_extractor(self):
+        warnings.warn(
+            "`feature_extractor` is deprecated and will be removed in v5. Use `image_processor` instead.",
+            FutureWarning,
+        )
+        return self.image_processor
+
+
+__all__ = ["CLIPProcessor"]

lib/python3.12/site-packages/transformers/models/clip/tokenization_clip.py

@@ -0,0 +1,519 @@
+# coding=utf-8
+# Copyright 2021 The Open AI Team Authors and 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.
+"""Tokenization classes for CLIP."""
+
+import json
+import os
+import unicodedata
+from functools import lru_cache
+from typing import List, Optional, Tuple
+
+import regex as re
+
+from ...tokenization_utils import AddedToken, PreTrainedTokenizer, _is_control, _is_punctuation, _is_whitespace
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+VOCAB_FILES_NAMES = {
+    "vocab_file": "vocab.json",
+    "merges_file": "merges.txt",
+}
+
+
+@lru_cache()
+def bytes_to_unicode():
+    """
+    Returns list of utf-8 byte and a mapping to unicode strings. We specifically avoids mapping to whitespace/control
+    characters the bpe code barfs on.
+
+    The reversible bpe codes work on unicode strings. This means you need a large # of unicode characters in your vocab
+    if you want to avoid UNKs. When you're at something like a 10B token dataset you end up needing around 5K for
+    decent coverage. This is a significant percentage of your normal, say, 32K bpe vocab. To avoid that, we want lookup
+    tables between utf-8 bytes and unicode strings.
+    """
+    bs = (
+        list(range(ord("!"), ord("~") + 1)) + list(range(ord("¡"), ord("¬") + 1)) + list(range(ord("®"), ord("ÿ") + 1))
+    )
+    cs = bs[:]
+    n = 0
+    for b in range(2**8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2**8 + n)
+            n += 1
+    cs = [chr(n) for n in cs]
+    return dict(zip(bs, cs))
+
+
+def get_pairs(word):
+    """
+    Return set of symbol pairs in a word.
+
+    Word is represented as tuple of symbols (symbols being variable-length strings).
+    """
+    pairs = set()
+    prev_char = word[0]
+    for char in word[1:]:
+        pairs.add((prev_char, char))
+        prev_char = char
+    return pairs
+
+
+def whitespace_clean(text):
+    text = re.sub(r"\s+", " ", text)
+    text = text.strip()
+    return text
+
+
+# Copied from transformers.models.bert.tokenization_bert.whitespace_tokenize
+def whitespace_tokenize(text):
+    """Runs basic whitespace cleaning and splitting on a piece of text."""
+    text = text.strip()
+    if not text:
+        return []
+    tokens = text.split()
+    return tokens
+
+
+# Copied from transformers.models.bert.tokenization_bert.BasicTokenizer
+class BasicTokenizer:
+    """
+    Constructs a BasicTokenizer that will run basic tokenization (punctuation splitting, lower casing, etc.).
+
+    Args:
+        do_lower_case (`bool`, *optional*, defaults to `True`):
+            Whether or not to lowercase the input when tokenizing.
+        never_split (`Iterable`, *optional*):
+            Collection of tokens which will never be split during tokenization. Only has an effect when
+            `do_basic_tokenize=True`
+        tokenize_chinese_chars (`bool`, *optional*, defaults to `True`):
+            Whether or not to tokenize Chinese characters.
+
+            This should likely be deactivated for Japanese (see this
+            [issue](https://github.com/huggingface/transformers/issues/328)).
+        strip_accents (`bool`, *optional*):
+            Whether or not to strip all accents. If this option is not specified, then it will be determined by the
+            value for `lowercase` (as in the original BERT).
+        do_split_on_punc (`bool`, *optional*, defaults to `True`):
+            In some instances we want to skip the basic punctuation splitting so that later tokenization can capture
+            the full context of the words, such as contractions.
+    """
+
+    def __init__(
+        self,
+        do_lower_case=True,
+        never_split=None,
+        tokenize_chinese_chars=True,
+        strip_accents=None,
+        do_split_on_punc=True,
+    ):
+        if never_split is None:
+            never_split = []
+        self.do_lower_case = do_lower_case
+        self.never_split = set(never_split)
+        self.tokenize_chinese_chars = tokenize_chinese_chars
+        self.strip_accents = strip_accents
+        self.do_split_on_punc = do_split_on_punc
+
+    def tokenize(self, text, never_split=None):
+        """
+        Basic Tokenization of a piece of text. For sub-word tokenization, see WordPieceTokenizer.
+
+        Args:
+            never_split (`List[str]`, *optional*)
+                Kept for backward compatibility purposes. Now implemented directly at the base class level (see
+                [`PreTrainedTokenizer.tokenize`]) List of token not to split.
+        """
+        # union() returns a new set by concatenating the two sets.
+        never_split = self.never_split.union(set(never_split)) if never_split else self.never_split
+        text = self._clean_text(text)
+
+        # This was added on November 1st, 2018 for the multilingual and Chinese
+        # models. This is also applied to the English models now, but it doesn't
+        # matter since the English models were not trained on any Chinese data
+        # and generally don't have any Chinese data in them (there are Chinese
+        # characters in the vocabulary because Wikipedia does have some Chinese
+        # words in the English Wikipedia.).
+        if self.tokenize_chinese_chars:
+            text = self._tokenize_chinese_chars(text)
+        # prevents treating the same character with different unicode codepoints as different characters
+        unicode_normalized_text = unicodedata.normalize("NFC", text)
+        orig_tokens = whitespace_tokenize(unicode_normalized_text)
+        split_tokens = []
+        for token in orig_tokens:
+            if token not in never_split:
+                if self.do_lower_case:
+                    token = token.lower()
+                    if self.strip_accents is not False:
+                        token = self._run_strip_accents(token)
+                elif self.strip_accents:
+                    token = self._run_strip_accents(token)
+            split_tokens.extend(self._run_split_on_punc(token, never_split))
+
+        output_tokens = whitespace_tokenize(" ".join(split_tokens))
+        return output_tokens
+
+    def _run_strip_accents(self, text):
+        """Strips accents from a piece of text."""
+        text = unicodedata.normalize("NFD", text)
+        output = []
+        for char in text:
+            cat = unicodedata.category(char)
+            if cat == "Mn":
+                continue
+            output.append(char)
+        return "".join(output)
+
+    def _run_split_on_punc(self, text, never_split=None):
+        """Splits punctuation on a piece of text."""
+        if not self.do_split_on_punc or (never_split is not None and text in never_split):
+            return [text]
+        chars = list(text)
+        i = 0
+        start_new_word = True
+        output = []
+        while i < len(chars):
+            char = chars[i]
+            if _is_punctuation(char):
+                output.append([char])
+                start_new_word = True
+            else:
+                if start_new_word:
+                    output.append([])
+                start_new_word = False
+                output[-1].append(char)
+            i += 1
+
+        return ["".join(x) for x in output]
+
+    def _tokenize_chinese_chars(self, text):
+        """Adds whitespace around any CJK character."""
+        output = []
+        for char in text:
+            cp = ord(char)
+            if self._is_chinese_char(cp):
+                output.append(" ")
+                output.append(char)
+                output.append(" ")
+            else:
+                output.append(char)
+        return "".join(output)
+
+    def _is_chinese_char(self, cp):
+        """Checks whether CP is the codepoint of a CJK character."""
+        # This defines a "chinese character" as anything in the CJK Unicode block:
+        #   https://en.wikipedia.org/wiki/CJK_Unified_Ideographs_(Unicode_block)
+        #
+        # Note that the CJK Unicode block is NOT all Japanese and Korean characters,
+        # despite its name. The modern Korean Hangul alphabet is a different block,
+        # as is Japanese Hiragana and Katakana. Those alphabets are used to write
+        # space-separated words, so they are not treated specially and handled
+        # like the all of the other languages.
+        if (
+            (cp >= 0x4E00 and cp <= 0x9FFF)
+            or (cp >= 0x3400 and cp <= 0x4DBF)  #
+            or (cp >= 0x20000 and cp <= 0x2A6DF)  #
+            or (cp >= 0x2A700 and cp <= 0x2B73F)  #
+            or (cp >= 0x2B740 and cp <= 0x2B81F)  #
+            or (cp >= 0x2B820 and cp <= 0x2CEAF)  #
+            or (cp >= 0xF900 and cp <= 0xFAFF)
+            or (cp >= 0x2F800 and cp <= 0x2FA1F)  #
+        ):  #
+            return True
+
+        return False
+
+    def _clean_text(self, text):
+        """Performs invalid character removal and whitespace cleanup on text."""
+        output = []
+        for char in text:
+            cp = ord(char)
+            if cp == 0 or cp == 0xFFFD or _is_control(char):
+                continue
+            if _is_whitespace(char):
+                output.append(" ")
+            else:
+                output.append(char)
+        return "".join(output)
+
+
+class CLIPTokenizer(PreTrainedTokenizer):
+    """
+    Construct a CLIP tokenizer. Based on byte-level Byte-Pair-Encoding.
+
+    This tokenizer inherits from [`PreTrainedTokenizer`] which contains most of the main methods. Users should refer to
+    this superclass for more information regarding those methods.
+
+    Args:
+        vocab_file (`str`):
+            Path to the vocabulary file.
+        merges_file (`str`):
+            Path to the merges file.
+        errors (`str`, *optional*, defaults to `"replace"`):
+            Paradigm to follow when decoding bytes to UTF-8. See
+            [bytes.decode](https://docs.python.org/3/library/stdtypes.html#bytes.decode) for more information.
+        unk_token (`str`, *optional*, defaults to `"<|endoftext|>"`):
+            The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this
+            token instead.
+        bos_token (`str`, *optional*, defaults to `"<|startoftext|>"`):
+            The beginning of sequence token.
+        eos_token (`str`, *optional*, defaults to `"<|endoftext|>"`):
+            The end of sequence token.
+        pad_token (`str`, *optional*, defaults to `"<|endoftext|>"`):
+            The token used for padding, for example when batching sequences of different lengths.
+    """
+
+    vocab_files_names = VOCAB_FILES_NAMES
+    model_input_names = ["input_ids", "attention_mask"]
+
+    def __init__(
+        self,
+        vocab_file,
+        merges_file,
+        errors="replace",
+        unk_token="<|endoftext|>",
+        bos_token="<|startoftext|>",
+        eos_token="<|endoftext|>",
+        pad_token="<|endoftext|>",  # hack to enable padding
+        **kwargs,
+    ):
+        bos_token = AddedToken(bos_token, lstrip=False, rstrip=False) if isinstance(bos_token, str) else bos_token
+        eos_token = AddedToken(eos_token, lstrip=False, rstrip=False) if isinstance(eos_token, str) else eos_token
+        unk_token = AddedToken(unk_token, lstrip=False, rstrip=False) if isinstance(unk_token, str) else unk_token
+        try:
+            import ftfy
+
+            self.fix_text = ftfy.fix_text
+        except ImportError:
+            logger.info("ftfy or spacy is not installed using custom BasicTokenizer instead of ftfy.")
+            self.nlp = BasicTokenizer(strip_accents=False, do_split_on_punc=False)
+            self.fix_text = None
+
+        with open(vocab_file, encoding="utf-8") as vocab_handle:
+            self.encoder = json.load(vocab_handle)
+        self.decoder = {v: k for k, v in self.encoder.items()}
+        self.errors = errors  # how to handle errors in decoding
+        self.byte_encoder = bytes_to_unicode()
+        self.byte_decoder = {v: k for k, v in self.byte_encoder.items()}
+        with open(merges_file, encoding="utf-8") as merges_handle:
+            bpe_merges = merges_handle.read().strip().split("\n")[1 : 49152 - 256 - 2 + 1]
+        bpe_merges = [tuple(merge.split()) for merge in bpe_merges]
+        self.bpe_ranks = dict(zip(bpe_merges, range(len(bpe_merges))))
+        self.cache = {"<|startoftext|>": "<|startoftext|>", "<|endoftext|>": "<|endoftext|>"}
+
+        self.pat = re.compile(
+            r"""<\|startoftext\|>|<\|endoftext\|>|'s|'t|'re|'ve|'m|'ll|'d|[\p{L}]+|[\p{N}]|[^\s\p{L}\p{N}]+""",
+            re.IGNORECASE,
+        )
+
+        super().__init__(
+            errors=errors,
+            unk_token=unk_token,
+            bos_token=bos_token,
+            eos_token=eos_token,
+            pad_token=pad_token,
+            **kwargs,
+        )
+
+    @property
+    def vocab_size(self):
+        return len(self.encoder)
+
+    def get_vocab(self):
+        return dict(self.encoder, **self.added_tokens_encoder)
+
+    def build_inputs_with_special_tokens(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        Build model inputs from a sequence or a pair of sequence for sequence classification tasks by concatenating and
+        adding special tokens. A CLIP sequence has the following format:
+
+        - single sequence: `<|startoftext|> X <|endoftext|>`
+
+        Pairs of sequences are not the expected use case, but they will be handled without a separator.
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs to which the special tokens will be added.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            `List[int]`: List of [input IDs](../glossary#input-ids) with the appropriate special tokens.
+        """
+        bos_token = [self.bos_token_id]
+        eos_token = [self.eos_token_id]
+
+        if token_ids_1 is None:
+            return bos_token + token_ids_0 + eos_token
+        return bos_token + token_ids_0 + eos_token + eos_token + token_ids_1 + eos_token
+
+    def get_special_tokens_mask(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None, already_has_special_tokens: bool = False
+    ) -> List[int]:
+        """
+        Retrieve sequence ids from a token list that has no special tokens added. This method is called when adding
+        special tokens using the tokenizer `prepare_for_model` method.
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+            already_has_special_tokens (`bool`, *optional*, defaults to `False`):
+                Whether or not the token list is already formatted with special tokens for the model.
+
+        Returns:
+            `List[int]`: A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
+        """
+
+        if already_has_special_tokens:
+            return super().get_special_tokens_mask(
+                token_ids_0=token_ids_0, token_ids_1=token_ids_1, already_has_special_tokens=True
+            )
+
+        if token_ids_1 is None:
+            return [1] + ([0] * len(token_ids_0)) + [1]
+        return [1] + ([0] * len(token_ids_0)) + [1] + [1] + ([0] * len(token_ids_1)) + [1]
+
+    def create_token_type_ids_from_sequences(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        Create a mask from the two sequences passed. CLIP does not make use of token type ids, therefore a list of
+        zeros is returned.
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            `List[int]`: List of zeros.
+        """
+        bos_token = [self.bos_token_id]
+        eos_token = [self.eos_token_id]
+
+        if token_ids_1 is None:
+            return len(bos_token + token_ids_0 + eos_token) * [0]
+        return len(bos_token + token_ids_0 + eos_token + eos_token + token_ids_1 + eos_token) * [0]
+
+    def bpe(self, token):
+        if token in self.cache:
+            return self.cache[token]
+        word = tuple(token[:-1]) + (token[-1] + "</w>",)
+        pairs = get_pairs(word)
+
+        if not pairs:
+            return token + "</w>"
+
+        while True:
+            bigram = min(pairs, key=lambda pair: self.bpe_ranks.get(pair, float("inf")))
+            if bigram not in self.bpe_ranks:
+                break
+            first, second = bigram
+            new_word = []
+            i = 0
+            while i < len(word):
+                try:
+                    j = word.index(first, i)
+                except ValueError:
+                    new_word.extend(word[i:])
+                    break
+                else:
+                    new_word.extend(word[i:j])
+                    i = j
+
+                if word[i] == first and i < len(word) - 1 and word[i + 1] == second:
+                    new_word.append(first + second)
+                    i += 2
+                else:
+                    new_word.append(word[i])
+                    i += 1
+            new_word = tuple(new_word)
+            word = new_word
+            if len(word) == 1:
+                break
+            else:
+                pairs = get_pairs(word)
+        word = " ".join(word)
+        self.cache[token] = word
+        return word
+
+    def _tokenize(self, text):
+        """Tokenize a string."""
+        bpe_tokens = []
+        if self.fix_text is None:
+            text = " ".join(self.nlp.tokenize(text))
+        else:
+            text = whitespace_clean(self.fix_text(text)).lower()
+
+        for token in re.findall(self.pat, text):
+            token = "".join(
+                self.byte_encoder[b] for b in token.encode("utf-8")
+            )  # Maps all our bytes to unicode strings, avoiding control tokens of the BPE (spaces in our case)
+            bpe_tokens.extend(bpe_token for bpe_token in self.bpe(token).split(" "))
+        return bpe_tokens
+
+    def _convert_token_to_id(self, token):
+        """Converts a token (str) in an id using the vocab."""
+        return self.encoder.get(token, self.encoder.get(self.unk_token))
+
+    def _convert_id_to_token(self, index):
+        """Converts an index (integer) in a token (str) using the vocab."""
+        return self.decoder.get(index)
+
+    def convert_tokens_to_string(self, tokens):
+        """Converts a sequence of tokens (string) in a single string."""
+        text = "".join(tokens)
+        byte_array = bytearray([self.byte_decoder[c] for c in text])
+        text = byte_array.decode("utf-8", errors=self.errors).replace("</w>", " ").strip()
+        return text
+
+    def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
+        if not os.path.isdir(save_directory):
+            logger.error("Vocabulary path ({}) should be a directory".format(save_directory))
+            return
+        vocab_file = os.path.join(
+            save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]
+        )
+        merge_file = os.path.join(
+            save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["merges_file"]
+        )
+
+        with open(vocab_file, "w", encoding="utf-8") as f:
+            f.write(json.dumps(self.encoder, indent=2, sort_keys=True, ensure_ascii=False) + "\n")
+
+        index = 0
+        with open(merge_file, "w", encoding="utf-8") as writer:
+            writer.write("#version: 0.2\n")
+            for bpe_tokens, token_index in sorted(self.bpe_ranks.items(), key=lambda kv: kv[1]):
+                if index != token_index:
+                    logger.warning(
+                        "Saving vocabulary to {}: BPE merge indices are not consecutive."
+                        " Please check that the tokenizer is not corrupted!".format(merge_file)
+                    )
+                    index = token_index
+                writer.write(" ".join(bpe_tokens) + "\n")
+                index += 1
+
+        return vocab_file, merge_file
+
+
+__all__ = ["CLIPTokenizer"]

lib/python3.12/site-packages/transformers/models/clip/tokenization_clip_fast.py

@@ -0,0 +1,164 @@
+# coding=utf-8
+# Copyright 2021 The Open AI Team Authors and 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.
+"""Tokenization classes for OpenAI GPT."""
+
+from typing import List, Optional, Tuple
+
+from tokenizers import pre_tokenizers
+
+from ...tokenization_utils_fast import PreTrainedTokenizerFast
+from ...utils import logging
+from .tokenization_clip import CLIPTokenizer
+
+
+logger = logging.get_logger(__name__)
+
+VOCAB_FILES_NAMES = {"vocab_file": "vocab.json", "merges_file": "merges.txt", "tokenizer_file": "tokenizer.json"}
+
+
+class CLIPTokenizerFast(PreTrainedTokenizerFast):
+    """
+    Construct a "fast" CLIP tokenizer (backed by HuggingFace's *tokenizers* library). Based on byte-level
+    Byte-Pair-Encoding.
+
+    This tokenizer inherits from [`PreTrainedTokenizerFast`] which contains most of the main methods. Users should
+    refer to this superclass for more information regarding those methods.
+
+    Args:
+        vocab_file (`str`, *optional*):
+            Path to the vocabulary file.
+        merges_file (`str`, *optional*):
+            Path to the merges file.
+        tokenizer_file (`str`, *optional*):
+            The path to a tokenizer file to use instead of the vocab file.
+        unk_token (`str`, *optional*, defaults to `"<|endoftext|>"`):
+            The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this
+            token instead.
+        bos_token (`str`, *optional*, defaults to `"<|startoftext|>"`):
+            The beginning of sequence token.
+        eos_token (`str`, *optional*, defaults to `"<|endoftext|>"`):
+            The end of sequence token.
+        pad_token (`str`, *optional*, defaults to `"<|endoftext|>"`):
+            The token used for padding, for example when batching sequences of different lengths.
+    """
+
+    vocab_files_names = VOCAB_FILES_NAMES
+    model_input_names = ["input_ids", "attention_mask"]
+    slow_tokenizer_class = CLIPTokenizer
+
+    def __init__(
+        self,
+        vocab_file=None,
+        merges_file=None,
+        tokenizer_file=None,
+        unk_token="<|endoftext|>",
+        bos_token="<|startoftext|>",
+        eos_token="<|endoftext|>",
+        pad_token="<|endoftext|>",  # hack to enable padding
+        **kwargs,
+    ):
+        super().__init__(
+            vocab_file,
+            merges_file,
+            tokenizer_file=tokenizer_file,
+            unk_token=unk_token,
+            bos_token=bos_token,
+            eos_token=eos_token,
+            pad_token=pad_token,
+            **kwargs,
+        )
+
+        if not isinstance(self.backend_tokenizer.pre_tokenizer, pre_tokenizers.Sequence):
+            raise ValueError(
+                "The `backend_tokenizer` provided does not match the expected format. The CLIP tokenizer has been"
+                " heavily modified from transformers version 4.17.0. You need to convert the tokenizer you are using"
+                " to be compatible with this version.The easiest way to do so is"
+                ' `CLIPTokenizerFast.from_pretrained("path_to_local_folder_or_hub_repo, from_slow=True)`. If you want'
+                " to use your existing tokenizer, you will have to revert to a version prior to 4.17.0 of"
+                " transformers."
+            )
+        self._wrap_decode_method_backend_tokenizer()
+
+    # Very ugly hack to enable padding to have a correct decoding see https://github.com/huggingface/tokenizers/issues/872
+    def _wrap_decode_method_backend_tokenizer(self):
+        orig_decode_method = self.backend_tokenizer.decode
+
+        ## define this as a local variable to avoid circular reference
+        ## See: https://github.com/huggingface/transformers/issues/30930
+        end_of_word_suffix = self.backend_tokenizer.model.end_of_word_suffix
+
+        def new_decode_method(*args, **kwargs):
+            text = orig_decode_method(*args, **kwargs)
+            text = text.replace(end_of_word_suffix, " ").strip()
+            return text
+
+        self.backend_tokenizer.decode = new_decode_method
+
+    def build_inputs_with_special_tokens(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        Build model inputs from a sequence or a pair of sequence for sequence classification tasks by concatenating and
+        adding special tokens. A CLIP sequence has the following format:
+
+        - single sequence: `<|startoftext|> X <|endoftext|>`
+
+        Pairs of sequences are not the expected use case, but they will be handled without a separator.
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs to which the special tokens will be added.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            `List[int]`: List of [input IDs](../glossary#input-ids) with the appropriate special tokens.
+        """
+        bos_token = [self.bos_token_id]
+        eos_token = [self.eos_token_id]
+
+        if token_ids_1 is None:
+            return bos_token + token_ids_0 + eos_token
+        return bos_token + token_ids_0 + eos_token + eos_token + token_ids_1 + eos_token
+
+    def create_token_type_ids_from_sequences(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        Create a mask from the two sequences passed. CLIP does not make use of token type ids, therefore a list of
+        zeros is returned.
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            `List[int]`: List of zeros.
+        """
+        bos_token = [self.bos_token_id]
+        eos_token = [self.eos_token_id]
+
+        if token_ids_1 is None:
+            return len(bos_token + token_ids_0 + eos_token) * [0]
+        return len(bos_token + token_ids_0 + eos_token + eos_token + token_ids_1 + eos_token) * [0]
+
+    def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
+        files = self._tokenizer.model.save(save_directory, name=filename_prefix)
+        return tuple(files)
+
+
+__all__ = ["CLIPTokenizerFast"]

lib/python3.12/site-packages/transformers/models/colpali/__init__.py

@@ -0,0 +1,28 @@
+# 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 _LazyModule
+from ...utils.import_utils import define_import_structure
+
+
+if TYPE_CHECKING:
+    from .configuration_colpali import *
+    from .modeling_colpali import *
+    from .processing_colpali import *
+else:
+    import sys
+
+    _file = globals()["__file__"]
+    sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)