Add files using upload-large-folder tool

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

@@ -0,0 +1,31 @@
+# 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_blenderbot_small import *
+    from .modeling_blenderbot_small import *
+    from .modeling_flax_blenderbot_small import *
+    from .modeling_tf_blenderbot_small import *
+    from .tokenization_blenderbot_small import *
+    from .tokenization_blenderbot_small_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/blenderbot_small/configuration_blenderbot_small.py

@@ -0,0 +1,390 @@
+# coding=utf-8
+# Copyright 2021 The Facebook, Inc. 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.
+"""BlenderbotSmall model configuration"""
+
+from collections import OrderedDict
+from typing import Any, Mapping, Optional
+
+from ... import PreTrainedTokenizer
+from ...configuration_utils import PretrainedConfig
+from ...file_utils import TensorType, is_torch_available
+from ...onnx import OnnxConfig, OnnxConfigWithPast, OnnxSeq2SeqConfigWithPast
+from ...onnx.utils import compute_effective_axis_dimension
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class BlenderbotSmallConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`BlenderbotSmallModel`]. It is used to instantiate
+    an BlenderbotSmall 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 BlenderbotSmall
+    [facebook/blenderbot_small-90M](https://huggingface.co/facebook/blenderbot_small-90M) 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 50265):
+            Vocabulary size of the BlenderbotSmall model. Defines the number of different tokens that can be
+            represented by the `inputs_ids` passed when calling [`BlenderbotSmallModel`] or [`TFBlenderbotSmallModel`].
+        d_model (`int`, *optional*, defaults to 512):
+            Dimensionality of the layers and the pooler layer.
+        encoder_layers (`int`, *optional*, defaults to 8):
+            Number of encoder layers.
+        decoder_layers (`int`, *optional*, defaults to 8):
+            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 2048):
+            Dimensionality of the "intermediate" (often named feed-forward) layer in decoder.
+        encoder_ffn_dim (`int`, *optional*, defaults to 2048):
+            Dimensionality of the "intermediate" (often named feed-forward) layer in decoder.
+        activation_function (`str` or `function`, *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.
+        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.
+        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).
+        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.
+        scale_embedding (`bool`, *optional*, defaults to `False`):
+            Scale embeddings by diving by sqrt(d_model).
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models)
+        forced_eos_token_id (`int`, *optional*, defaults to 2):
+            The id of the token to force as the last generated token when `max_length` is reached. Usually set to
+            `eos_token_id`.
+
+    Example:
+
+    ```python
+    >>> from transformers import BlenderbotSmallConfig, BlenderbotSmallModel
+
+    >>> # Initializing a BlenderbotSmall facebook/blenderbot_small-90M style configuration
+    >>> configuration = BlenderbotSmallConfig()
+
+    >>> # Initializing a model (with random weights) from the facebook/blenderbot_small-90M style configuration
+    >>> model = BlenderbotSmallModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "blenderbot-small"
+    keys_to_ignore_at_inference = ["past_key_values"]
+    attribute_map = {"num_attention_heads": "encoder_attention_heads", "hidden_size": "d_model"}
+
+    def __init__(
+        self,
+        vocab_size=50265,
+        max_position_embeddings=512,
+        encoder_layers=8,
+        encoder_ffn_dim=2048,
+        encoder_attention_heads=16,
+        decoder_layers=8,
+        decoder_ffn_dim=2048,
+        decoder_attention_heads=16,
+        encoder_layerdrop=0.0,
+        decoder_layerdrop=0.0,
+        use_cache=True,
+        is_encoder_decoder=True,
+        activation_function="gelu",
+        d_model=512,
+        dropout=0.1,
+        attention_dropout=0.0,
+        activation_dropout=0.0,
+        init_std=0.02,
+        decoder_start_token_id=1,
+        scale_embedding=False,
+        pad_token_id=0,
+        bos_token_id=1,
+        eos_token_id=2,
+        forced_eos_token_id=2,
+        **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.use_cache = use_cache
+        self.num_hidden_layers = encoder_layers
+        self.scale_embedding = scale_embedding  # scale factor will be sqrt(d_model) if True
+
+        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,
+            forced_eos_token_id=forced_eos_token_id,
+            **kwargs,
+        )
+
+
+# Copied from transformers.models.bart.configuration_bart.BartOnnxConfig
+class BlenderbotSmallOnnxConfig(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__ = ["BlenderbotSmallConfig", "BlenderbotSmallOnnxConfig"]

lib/python3.12/site-packages/transformers/models/blenderbot_small/modeling_blenderbot_small.py

@@ -0,0 +1,1511 @@
+# coding=utf-8
+# Copyright 2021 The Facebook, Inc. 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.
+"""PyTorch BlenderbotSmall model."""
+
+import copy
+import math
+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 ...generation import GenerationMixin
+from ...modeling_attn_mask_utils import _prepare_4d_attention_mask, _prepare_4d_causal_attention_mask
+from ...modeling_outputs import (
+    BaseModelOutput,
+    BaseModelOutputWithPastAndCrossAttentions,
+    CausalLMOutputWithCrossAttentions,
+    Seq2SeqLMOutput,
+    Seq2SeqModelOutput,
+)
+from ...modeling_utils import PreTrainedModel
+from ...utils import (
+    add_end_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    logging,
+    replace_return_docstrings,
+)
+from .configuration_blenderbot_small import BlenderbotSmallConfig
+
+
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = "BlenderbotSmallConfig"
+
+
+# Copied from transformers.models.bart.modeling_bart.shift_tokens_right
+def shift_tokens_right(input_ids: torch.Tensor, pad_token_id: int, decoder_start_token_id: int):
+    """
+    Shift input ids one token to the right.
+    """
+    shifted_input_ids = input_ids.new_zeros(input_ids.shape)
+    shifted_input_ids[:, 1:] = input_ids[:, :-1].clone()
+    shifted_input_ids[:, 0] = decoder_start_token_id
+
+    if pad_token_id is None:
+        raise ValueError("self.model.config.pad_token_id has to be defined.")
+    # replace possible -100 values in labels by `pad_token_id`
+    shifted_input_ids.masked_fill_(shifted_input_ids == -100, pad_token_id)
+
+    return shifted_input_ids
+
+
+# Copied from transformers.models.blenderbot.modeling_blenderbot.BlenderbotLearnedPositionalEmbedding with Blenderbot->BlenderbotSmall
+class BlenderbotSmallLearnedPositionalEmbedding(nn.Embedding):
+    """
+    This module learns positional embeddings up to a fixed maximum size.
+    """
+
+    def __init__(self, num_embeddings: int, embedding_dim: int):
+        super().__init__(num_embeddings, embedding_dim)
+
+    def forward(self, input_ids_shape: torch.Size, past_key_values_length: int = 0):
+        """`input_ids_shape` is expected to be [bsz x seqlen]."""
+        bsz, seq_len = input_ids_shape[:2]
+        positions = torch.arange(
+            past_key_values_length, past_key_values_length + seq_len, dtype=torch.long, device=self.weight.device
+        )
+        return super().forward(positions)
+
+
+# Copied from transformers.models.bart.modeling_bart.BartAttention with Bart->BlenderbotSmall
+class BlenderbotSmallAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(
+        self,
+        embed_dim: int,
+        num_heads: int,
+        dropout: float = 0.0,
+        is_decoder: bool = False,
+        bias: bool = True,
+        is_causal: bool = False,
+        config: Optional[BlenderbotSmallConfig] = None,
+    ):
+        super().__init__()
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.dropout = dropout
+        self.head_dim = embed_dim // num_heads
+        self.config = config
+
+        if (self.head_dim * num_heads) != self.embed_dim:
+            raise ValueError(
+                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim}"
+                f" and `num_heads`: {num_heads})."
+            )
+        self.scaling = self.head_dim**-0.5
+        self.is_decoder = is_decoder
+        self.is_causal = is_causal
+
+        self.k_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+        self.v_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+        self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+        self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+
+    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,
+        key_value_states: Optional[torch.Tensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        layer_head_mask: Optional[torch.Tensor] = None,
+        output_attentions: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        """Input shape: Batch x Time x Channel"""
+
+        # if key_value_states are provided this layer is used as a cross-attention layer
+        # for the decoder
+        is_cross_attention = key_value_states is not None
+
+        bsz, tgt_len, _ = hidden_states.size()
+
+        # get query proj
+        query_states = self.q_proj(hidden_states) * self.scaling
+        # get key, value proj
+        # `past_key_value[0].shape[2] == key_value_states.shape[1]`
+        # is checking that the `sequence_length` of the `past_key_value` is the same as
+        # the provided `key_value_states` to support prefix tuning
+        if (
+            is_cross_attention
+            and past_key_value is not None
+            and past_key_value[0].shape[2] == key_value_states.shape[1]
+        ):
+            # reuse k,v, cross_attentions
+            key_states = past_key_value[0]
+            value_states = past_key_value[1]
+        elif is_cross_attention:
+            # cross_attentions
+            key_states = self._shape(self.k_proj(key_value_states), -1, bsz)
+            value_states = self._shape(self.v_proj(key_value_states), -1, bsz)
+        elif past_key_value is not None:
+            # reuse k, v, self_attention
+            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
+            key_states = torch.cat([past_key_value[0], key_states], dim=2)
+            value_states = torch.cat([past_key_value[1], value_states], dim=2)
+        else:
+            # self_attention
+            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
+
+        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_states, value_states)
+
+        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.reshape(*proj_shape)
+        value_states = value_states.reshape(*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()}"
+            )
+
+        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 layer_head_mask is not None:
+            if layer_head_mask.size() != (self.num_heads,):
+                raise ValueError(
+                    f"Head mask for a single layer should be of size {(self.num_heads,)}, but is"
+                    f" {layer_head_mask.size()}"
+                )
+            attn_weights = layer_head_mask.view(1, -1, 1, 1) * attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
+            attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
+
+        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 be 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)
+
+        # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
+        # partitioned across GPUs when using tensor-parallelism.
+        attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)
+
+        attn_output = self.out_proj(attn_output)
+
+        return attn_output, attn_weights_reshaped, past_key_value
+
+
+# Copied from transformers.models.bart.modeling_bart.BartEncoderLayer with Bart->BlenderbotSmall, BART->BLENDERBOT_SMALL
+class BlenderbotSmallEncoderLayer(nn.Module):
+    def __init__(self, config: BlenderbotSmallConfig):
+        super().__init__()
+        self.embed_dim = config.d_model
+
+        self.self_attn = BLENDERBOT_SMALL_ATTENTION_CLASSES[config._attn_implementation](
+            embed_dim=self.embed_dim,
+            num_heads=config.encoder_attention_heads,
+            dropout=config.attention_dropout,
+            config=config,
+        )
+        self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
+        self.dropout = config.dropout
+        self.activation_fn = ACT2FN[config.activation_function]
+        self.activation_dropout = config.activation_dropout
+        self.fc1 = nn.Linear(self.embed_dim, config.encoder_ffn_dim)
+        self.fc2 = nn.Linear(config.encoder_ffn_dim, self.embed_dim)
+        self.final_layer_norm = nn.LayerNorm(self.embed_dim)
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        attention_mask: torch.FloatTensor,
+        layer_head_mask: torch.FloatTensor,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.FloatTensor, Optional[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.
+            layer_head_mask (`torch.FloatTensor`): mask for attention heads in a given layer of size
+                `(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, attn_weights, _ = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            layer_head_mask=layer_head_mask,
+            output_attentions=output_attentions,
+        )
+        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+        hidden_states = residual + hidden_states
+        hidden_states = self.self_attn_layer_norm(hidden_states)
+
+        residual = hidden_states
+        hidden_states = self.activation_fn(self.fc1(hidden_states))
+        hidden_states = nn.functional.dropout(hidden_states, p=self.activation_dropout, training=self.training)
+        hidden_states = self.fc2(hidden_states)
+        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+        hidden_states = residual + hidden_states
+        hidden_states = self.final_layer_norm(hidden_states)
+
+        if hidden_states.dtype == torch.float16 and (
+            torch.isinf(hidden_states).any() or torch.isnan(hidden_states).any()
+        ):
+            clamp_value = torch.finfo(hidden_states.dtype).max - 1000
+            hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (attn_weights,)
+
+        return outputs
+
+
+# TODO: Implement attention with SDPA for TimeSeriesTransformer.
+BLENDERBOT_SMALL_ATTENTION_CLASSES = {
+    "eager": BlenderbotSmallAttention,
+}
+
+
+# Copied from transformers.models.bart.modeling_bart.BartDecoderLayer with Bart->BlenderbotSmall, BART->BLENDERBOT_SMALL
+class BlenderbotSmallDecoderLayer(nn.Module):
+    def __init__(self, config: BlenderbotSmallConfig):
+        super().__init__()
+        self.embed_dim = config.d_model
+
+        self.self_attn = BLENDERBOT_SMALL_ATTENTION_CLASSES[config._attn_implementation](
+            embed_dim=self.embed_dim,
+            num_heads=config.decoder_attention_heads,
+            dropout=config.attention_dropout,
+            is_decoder=True,
+            is_causal=True,
+            config=config,
+        )
+        self.dropout = config.dropout
+        self.activation_fn = ACT2FN[config.activation_function]
+        self.activation_dropout = config.activation_dropout
+
+        self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
+        self.encoder_attn = BLENDERBOT_SMALL_ATTENTION_CLASSES[config._attn_implementation](
+            self.embed_dim,
+            config.decoder_attention_heads,
+            dropout=config.attention_dropout,
+            is_decoder=True,
+            config=config,
+        )
+        self.encoder_attn_layer_norm = nn.LayerNorm(self.embed_dim)
+        self.fc1 = nn.Linear(self.embed_dim, config.decoder_ffn_dim)
+        self.fc2 = nn.Linear(config.decoder_ffn_dim, self.embed_dim)
+        self.final_layer_norm = nn.LayerNorm(self.embed_dim)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        layer_head_mask: Optional[torch.Tensor] = None,
+        cross_attn_layer_head_mask: Optional[torch.Tensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = True,
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, 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.
+            encoder_hidden_states (`torch.FloatTensor`):
+                cross attention input to the layer of shape `(batch, seq_len, embed_dim)`
+            encoder_attention_mask (`torch.FloatTensor`): encoder attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            layer_head_mask (`torch.FloatTensor`): mask for attention heads in a given layer of size
+                `(encoder_attention_heads,)`.
+            cross_attn_layer_head_mask (`torch.FloatTensor`): mask for cross-attention heads in a given layer of
+                size `(decoder_attention_heads,)`.
+            past_key_value (`Tuple(torch.FloatTensor)`): cached past key and value projection states
+            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
+
+        # Self Attention
+        # 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
+        # add present self-attn cache to positions 1,2 of present_key_value tuple
+        hidden_states, self_attn_weights, present_key_value = self.self_attn(
+            hidden_states=hidden_states,
+            past_key_value=self_attn_past_key_value,
+            attention_mask=attention_mask,
+            layer_head_mask=layer_head_mask,
+            output_attentions=output_attentions,
+        )
+        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+        hidden_states = residual + hidden_states
+        hidden_states = self.self_attn_layer_norm(hidden_states)
+
+        # Cross-Attention Block
+        cross_attn_present_key_value = None
+        cross_attn_weights = None
+        if encoder_hidden_states is not None:
+            residual = hidden_states
+
+            # cross_attn cached key/values tuple is at positions 3,4 of present_key_value tuple
+            cross_attn_past_key_value = past_key_value[-2:] if past_key_value is not None else None
+            hidden_states, cross_attn_weights, cross_attn_present_key_value = self.encoder_attn(
+                hidden_states=hidden_states,
+                key_value_states=encoder_hidden_states,
+                attention_mask=encoder_attention_mask,
+                layer_head_mask=cross_attn_layer_head_mask,
+                past_key_value=cross_attn_past_key_value,
+                output_attentions=output_attentions,
+            )
+            hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+            hidden_states = residual + hidden_states
+            hidden_states = self.encoder_attn_layer_norm(hidden_states)
+
+            # add cross-attn to positions 3,4 of present_key_value tuple
+            present_key_value = present_key_value + cross_attn_present_key_value
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.activation_fn(self.fc1(hidden_states))
+        hidden_states = nn.functional.dropout(hidden_states, p=self.activation_dropout, training=self.training)
+        hidden_states = self.fc2(hidden_states)
+        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+        hidden_states = residual + hidden_states
+        hidden_states = self.final_layer_norm(hidden_states)
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights, cross_attn_weights)
+
+        if use_cache:
+            outputs += (present_key_value,)
+
+        return outputs
+
+
+class BlenderbotSmallPreTrainedModel(PreTrainedModel):
+    config_class = BlenderbotSmallConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+
+    def _init_weights(self, module):
+        std = self.config.init_std
+        if isinstance(module, nn.Linear):
+            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_()
+
+    @property
+    def dummy_inputs(self):
+        pad_token = self.config.pad_token_id
+        input_ids = torch.tensor([[0, 6, 10, 4, 2], [0, 8, 12, 2, pad_token]], device=self.device)
+        dummy_inputs = {
+            "attention_mask": input_ids.ne(pad_token),
+            "input_ids": input_ids,
+            "decoder_input_ids": input_ids,
+        }
+        return dummy_inputs
+
+
+BLENDERBOT_SMALL_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 ([`BlenderbotSmallConfig`]):
+            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.
+"""
+
+BLENDERBOT_SMALL_GENERATION_EXAMPLE = r"""
+    Conversation example:
+
+    ```python
+    >>> from transformers import AutoTokenizer, BlenderbotSmallForConditionalGeneration
+
+    >>> mname = "facebook/blenderbot_small-90M"
+    >>> model = BlenderbotSmallForConditionalGeneration.from_pretrained(mname)
+    >>> tokenizer = AutoTokenizer.from_pretrained(mname)
+    >>> UTTERANCE = "My friends are cool but they eat too many carbs."
+    >>> print("Human: ", UTTERANCE)
+    Human:  My friends are cool but they eat too many carbs.
+
+    >>> inputs = tokenizer([UTTERANCE], return_tensors="pt")
+    >>> reply_ids = model.generate(**inputs)
+    >>> print("Bot: ", tokenizer.batch_decode(reply_ids, skip_special_tokens=True)[0])
+    Bot:  what kind of carbs do they eat? i don't know much about carbs.
+
+    >>> REPLY = "I'm not sure"
+    >>> print("Human: ", REPLY)
+    Human: I'm not sure
+
+    >>> NEXT_UTTERANCE = (
+    ...     "My friends are cool but they eat too many carbs.__end__ __start__what kind of carbs do they eat? "
+    ...     "i don't know much about carbs__end__ "
+    ...     "__start__ I'm not sure."
+    ... )
+    >>> inputs = tokenizer([NEXT_UTTERANCE], return_tensors="pt")
+    >>> next_reply_ids = model.generate(**inputs)
+    >>> print("Bot: ", tokenizer.batch_decode(next_reply_ids, skip_special_tokens=True)[0])
+    Bot:  they eat a lot of carbs. carbs are high in fat, protein, and fats.
+    ```
+"""
+
+BLENDERBOT_SMALL_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)
+        decoder_input_ids (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*):
+            Indices of decoder input sequence tokens in the vocabulary.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are decoder input IDs?](../glossary#decoder-input-ids)
+
+            BlenderbotSmall uses the `bos_token_id` as the starting token for `decoder_input_ids` generation. If
+            `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
+            `past_key_values`).
+        decoder_attention_mask (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*):
+            Default behavior: generate a tensor that ignores pad tokens in `decoder_input_ids`. Causal mask will also
+            be used by default.
+        head_mask (`torch.Tensor` of shape `(encoder_layers, encoder_attention_heads)`, *optional*):
+            Mask to nullify selected heads of the attention modules in the encoder. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        decoder_head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
+            Mask to nullify selected heads of the attention modules in the decoder. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        cross_attn_head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
+            Mask to nullify selected heads of the cross-attention modules in the decoder. Mask values selected in `[0,
+            1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        encoder_outputs (`tuple(tuple(torch.FloatTensor)`, *optional*):
+            Tuple consists of (`last_hidden_state`, *optional*: `hidden_states`, *optional*: `attentions`)
+            `last_hidden_state` of shape `(batch_size, sequence_length, hidden_size)`, *optional*) is a sequence of
+            hidden-states at the output of the last layer of the encoder. Used in the cross-attention of the decoder.
+        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.
+        decoder_inputs_embeds (`torch.FloatTensor` of shape `(batch_size, target_sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `decoder_input_ids` you can choose to directly pass an embedded
+            representation. If `past_key_values` is used, optionally only the last `decoder_inputs_embeds` have to be
+            input (see `past_key_values`). This is useful if you want more control over how to convert
+            `decoder_input_ids` indices into associated vectors than the model's internal embedding lookup matrix.
+
+            If `decoder_input_ids` and `decoder_inputs_embeds` are both unset, `decoder_inputs_embeds` takes the value
+            of `inputs_embeds`.
+        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.
+"""
+
+
+class BlenderbotSmallEncoder(BlenderbotSmallPreTrainedModel):
+    """
+    Transformer encoder consisting of *config.encoder_layers* self attention layers. Each layer is a
+    [`BlenderbotSmallEncoderLayer`].
+
+    Args:
+        config: BlenderbotSmallConfig
+        embed_tokens (nn.Embedding): output embedding
+    """
+
+    def __init__(self, config: BlenderbotSmallConfig, embed_tokens: Optional[nn.Embedding] = None):
+        super().__init__(config)
+
+        self.dropout = config.dropout
+        self.layerdrop = config.encoder_layerdrop
+
+        embed_dim = config.d_model
+        self.padding_idx = config.pad_token_id
+        self.max_source_positions = config.max_position_embeddings
+        self.embed_scale = math.sqrt(embed_dim) if config.scale_embedding else 1.0
+
+        if embed_tokens is not None:
+            self.embed_tokens = embed_tokens
+        else:
+            self.embed_tokens = nn.Embedding(config.vocab_size, embed_dim, self.padding_idx)
+
+        self.embed_positions = BlenderbotSmallLearnedPositionalEmbedding(
+            config.max_position_embeddings,
+            embed_dim,
+        )
+        self.layers = nn.ModuleList([BlenderbotSmallEncoderLayer(config) for _ in range(config.encoder_layers)])
+        self.layernorm_embedding = nn.LayerNorm(embed_dim)
+
+        self.gradient_checkpointing = False
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        head_mask=None,
+        inputs_embeds=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        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)
+            head_mask (`torch.Tensor` of shape `(encoder_layers, encoder_attention_heads)`, *optional*):
+                Mask to nullify selected heads of the 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 `(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.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            output_hidden_states (`bool`, *optional*):
+                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
+                for more detail.
+            return_dict (`bool`, *optional*):
+                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # retrieve input_ids and inputs_embeds
+        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:
+            self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask)
+            input_shape = input_ids.size()
+            input_ids = input_ids.view(-1, input_shape[-1])
+        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 inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids) * self.embed_scale
+
+        embed_pos = self.embed_positions(input_shape)
+
+        hidden_states = inputs_embeds + embed_pos
+        hidden_states = self.layernorm_embedding(hidden_states)
+        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+
+        # expand attention_mask
+        if attention_mask is not None:
+            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+            attention_mask = _prepare_4d_attention_mask(attention_mask, inputs_embeds.dtype)
+
+        encoder_states = () if output_hidden_states else None
+        all_attentions = () if output_attentions else None
+
+        # check if head_mask has a correct number of layers specified if desired
+        if head_mask is not None:
+            if head_mask.size()[0] != len(self.layers):
+                raise ValueError(
+                    f"The head_mask should be specified for {len(self.layers)} layers, but it is for"
+                    f" {head_mask.size()[0]}."
+                )
+        for idx, encoder_layer in enumerate(self.layers):
+            if output_hidden_states:
+                encoder_states = encoder_states + (hidden_states,)
+            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
+            to_drop = False
+            if self.training:
+                dropout_probability = torch.rand([])
+                if dropout_probability < self.layerdrop:  # skip the layer
+                    to_drop = True
+
+            if to_drop:
+                layer_outputs = (None, None)
+            else:
+                if self.gradient_checkpointing and self.training:
+                    layer_outputs = self._gradient_checkpointing_func(
+                        encoder_layer.__call__,
+                        hidden_states,
+                        attention_mask,
+                        (head_mask[idx] if head_mask is not None else None),
+                        output_attentions,
+                    )
+                else:
+                    layer_outputs = encoder_layer(
+                        hidden_states,
+                        attention_mask,
+                        layer_head_mask=(head_mask[idx] if head_mask is not None else None),
+                        output_attentions=output_attentions,
+                    )
+
+                hidden_states = layer_outputs[0]
+
+            if output_attentions:
+                all_attentions = all_attentions + (layer_outputs[1],)
+
+        if output_hidden_states:
+            encoder_states = encoder_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None)
+        return BaseModelOutput(
+            last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions
+        )
+
+
+class BlenderbotSmallDecoder(BlenderbotSmallPreTrainedModel):
+    """
+    Transformer decoder consisting of *config.decoder_layers* layers. Each layer is a [`BlenderbotSmallDecoderLayer`]
+
+    Args:
+        config: BlenderbotSmallConfig
+        embed_tokens (nn.Embedding): output embedding
+    """
+
+    def __init__(self, config: BlenderbotSmallConfig, embed_tokens: Optional[nn.Embedding] = None):
+        super().__init__(config)
+        self.dropout = config.dropout
+        self.layerdrop = config.decoder_layerdrop
+        self.padding_idx = config.pad_token_id
+        self.max_target_positions = config.max_position_embeddings
+        self.embed_scale = math.sqrt(config.d_model) if config.scale_embedding else 1.0
+
+        if embed_tokens is not None:
+            self.embed_tokens = embed_tokens
+        else:
+            self.embed_tokens = nn.Embedding(config.vocab_size, config.d_model, self.padding_idx)
+
+        self.embed_positions = BlenderbotSmallLearnedPositionalEmbedding(
+            config.max_position_embeddings,
+            config.d_model,
+        )
+        self.layers = nn.ModuleList([BlenderbotSmallDecoderLayer(config) for _ in range(config.decoder_layers)])
+        self.layernorm_embedding = nn.LayerNorm(config.d_model)
+
+        self.gradient_checkpointing = False
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        head_mask=None,
+        cross_attn_head_mask=None,
+        past_key_values=None,
+        inputs_embeds=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        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)
+            encoder_hidden_states (`torch.FloatTensor` of shape `(batch_size, encoder_sequence_length, hidden_size)`, *optional*):
+                Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention
+                of the decoder.
+            encoder_attention_mask (`torch.LongTensor` of shape `(batch_size, encoder_sequence_length)`, *optional*):
+                Mask to avoid performing cross-attention on padding tokens indices of encoder input_ids. 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)
+            head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
+                Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:
+
+                - 1 indicates the head is **not masked**,
+                - 0 indicates the head is **masked**.
+
+            cross_attn_head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
+                Mask to nullify selected heads of the cross-attention modules in the decoder to avoid performing
+                cross-attention on hidden heads. Mask values selected in `[0, 1]`:
+
+                - 1 indicates the head is **not masked**,
+                - 0 indicates the head is **masked**.
+
+            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.
+            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
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # retrieve input_ids and inputs_embeds
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+            input_ids = input_ids.view(-1, input_shape[-1])
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+        else:
+            raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
+
+        # 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 inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids) * self.embed_scale
+
+        attention_mask = _prepare_4d_causal_attention_mask(
+            attention_mask, input_shape, inputs_embeds, past_key_values_length
+        )
+
+        # expand encoder attention mask
+        if encoder_hidden_states is not None and encoder_attention_mask is not None:
+            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+            encoder_attention_mask = _prepare_4d_attention_mask(
+                encoder_attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]
+            )
+
+        # embed positions
+        positions = self.embed_positions(input_shape, past_key_values_length)
+
+        # BlenderbotSmall applies layer norm on hidden_states
+        inputs_embeds = self.layernorm_embedding(inputs_embeds)
+        hidden_states = inputs_embeds + positions
+
+        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                logger.warning_once(
+                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                )
+                use_cache = False
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        all_cross_attentions = () if (output_attentions and encoder_hidden_states is not None) else None
+        next_decoder_cache = () if use_cache else None
+
+        # check if head_mask/cross_attn_head_mask has a correct number of layers specified if desired
+        for attn_mask, mask_name in zip([head_mask, cross_attn_head_mask], ["head_mask", "cross_attn_head_mask"]):
+            if attn_mask is not None:
+                if attn_mask.size()[0] != len(self.layers):
+                    raise ValueError(
+                        f"The `{mask_name}` should be specified for {len(self.layers)} layers, but it is for"
+                        f" {head_mask.size()[0]}."
+                    )
+        for idx, decoder_layer in enumerate(self.layers):
+            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+            if self.training:
+                dropout_probability = torch.rand([])
+                if dropout_probability < self.layerdrop:
+                    continue
+
+            past_key_value = past_key_values[idx] if past_key_values is not None else None
+
+            if self.gradient_checkpointing and self.training:
+                layer_outputs = self._gradient_checkpointing_func(
+                    decoder_layer.__call__,
+                    hidden_states,
+                    attention_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    head_mask[idx] if head_mask is not None else None,
+                    cross_attn_head_mask[idx] if cross_attn_head_mask is not None else None,
+                    None,
+                    output_attentions,
+                    use_cache,
+                )
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    encoder_hidden_states=encoder_hidden_states,
+                    encoder_attention_mask=encoder_attention_mask,
+                    layer_head_mask=(head_mask[idx] if head_mask is not None else None),
+                    cross_attn_layer_head_mask=(
+                        cross_attn_head_mask[idx] if cross_attn_head_mask is not None else None
+                    ),
+                    past_key_value=past_key_value,
+                    output_attentions=output_attentions,
+                    use_cache=use_cache,
+                )
+            hidden_states = layer_outputs[0]
+
+            if use_cache:
+                next_decoder_cache += (layer_outputs[3 if output_attentions else 1],)
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+                if encoder_hidden_states is not None:
+                    all_cross_attentions += (layer_outputs[2],)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        next_cache = next_decoder_cache if use_cache else None
+        if not return_dict:
+            return tuple(
+                v
+                for v in [hidden_states, next_cache, all_hidden_states, all_self_attns, all_cross_attentions]
+                if v is not None
+            )
+        return BaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+            cross_attentions=all_cross_attentions,
+        )
+
+
+@add_start_docstrings(
+    "The bare BlenderbotSmall Model outputting raw hidden-states without any specific head on top.",
+    BLENDERBOT_SMALL_START_DOCSTRING,
+)
+class BlenderbotSmallModel(BlenderbotSmallPreTrainedModel):
+    _tied_weights_keys = ["decoder.embed_tokens.weight", "encoder.embed_tokens.weight"]
+
+    def __init__(self, config: BlenderbotSmallConfig):
+        super().__init__(config)
+
+        padding_idx, vocab_size = config.pad_token_id, config.vocab_size
+        self.shared = nn.Embedding(vocab_size, config.d_model, padding_idx)
+
+        self.encoder = BlenderbotSmallEncoder(config, self.shared)
+        self.decoder = BlenderbotSmallDecoder(config, self.shared)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.shared
+
+    def set_input_embeddings(self, value):
+        self.shared = value
+        self.encoder.embed_tokens = self.shared
+        self.decoder.embed_tokens = self.shared
+
+    def get_encoder(self):
+        return self.encoder
+
+    def get_decoder(self):
+        return self.decoder
+
+    @add_start_docstrings_to_model_forward(BLENDERBOT_SMALL_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=Seq2SeqModelOutput, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        decoder_input_ids: Optional[torch.LongTensor] = None,
+        decoder_attention_mask: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        decoder_head_mask: Optional[torch.Tensor] = None,
+        cross_attn_head_mask: Optional[torch.Tensor] = None,
+        encoder_outputs: Optional[Union[Tuple, BaseModelOutput]] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        decoder_inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple[torch.FloatTensor], Seq2SeqModelOutput]:
+        r"""
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer, BlenderbotSmallModel
+
+        >>> model = BlenderbotSmallModel.from_pretrained("facebook/blenderbot_small-90M")
+        >>> tokenizer = AutoTokenizer.from_pretrained("facebook/blenderbot_small-90M")
+
+        >>> inputs = tokenizer("Studies have been shown that owning a dog is good for you", return_tensors="pt")
+        >>> decoder_inputs = tokenizer("Studies show that", return_tensors="pt")  # Batch size 1
+        >>> outputs = model(input_ids=inputs.input_ids, decoder_input_ids=decoder_inputs.input_ids)
+
+        >>> last_hidden_states = outputs.last_hidden_state
+        >>> list(last_hidden_states.shape)
+        [1, 3, 512]
+        ```"""
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if encoder_outputs is None:
+            encoder_outputs = self.encoder(
+                input_ids=input_ids,
+                attention_mask=attention_mask,
+                head_mask=head_mask,
+                inputs_embeds=inputs_embeds,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+            )
+        # If the user passed a tuple for encoder_outputs, we wrap it in a BaseModelOutput when return_dict=True
+        elif return_dict and not isinstance(encoder_outputs, BaseModelOutput):
+            encoder_outputs = BaseModelOutput(
+                last_hidden_state=encoder_outputs[0],
+                hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None,
+                attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None,
+            )
+
+        # decoder outputs consists of (dec_features, past_key_value, dec_hidden, dec_attn)
+        decoder_outputs = self.decoder(
+            input_ids=decoder_input_ids,
+            attention_mask=decoder_attention_mask,
+            encoder_hidden_states=encoder_outputs[0],
+            encoder_attention_mask=attention_mask,
+            head_mask=decoder_head_mask,
+            cross_attn_head_mask=cross_attn_head_mask,
+            past_key_values=past_key_values,
+            inputs_embeds=decoder_inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        if not return_dict:
+            return decoder_outputs + encoder_outputs
+
+        return Seq2SeqModelOutput(
+            last_hidden_state=decoder_outputs.last_hidden_state,
+            past_key_values=decoder_outputs.past_key_values,
+            decoder_hidden_states=decoder_outputs.hidden_states,
+            decoder_attentions=decoder_outputs.attentions,
+            cross_attentions=decoder_outputs.cross_attentions,
+            encoder_last_hidden_state=encoder_outputs.last_hidden_state,
+            encoder_hidden_states=encoder_outputs.hidden_states,
+            encoder_attentions=encoder_outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    "The BlenderbotSmall Model with a language modeling head. Can be used for summarization.",
+    BLENDERBOT_SMALL_START_DOCSTRING,
+)
+class BlenderbotSmallForConditionalGeneration(BlenderbotSmallPreTrainedModel, GenerationMixin):
+    base_model_prefix = "model"
+    _keys_to_ignore_on_load_missing = ["final_logits_bias"]
+    _tied_weights_keys = ["decoder.embed_tokens.weight", "encoder.embed_tokens.weight", "lm_head.weight"]
+
+    def __init__(self, config: BlenderbotSmallConfig):
+        super().__init__(config)
+        self.model = BlenderbotSmallModel(config)
+        self.register_buffer("final_logits_bias", torch.zeros((1, self.model.shared.num_embeddings)))
+        self.lm_head = nn.Linear(config.d_model, self.model.shared.num_embeddings, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_encoder(self):
+        return self.model.get_encoder()
+
+    def get_decoder(self):
+        return self.model.get_decoder()
+
+    def resize_token_embeddings(
+        self, new_num_tokens: int, pad_to_multiple_of: Optional[int] = None, mean_resizing: bool = True
+    ) -> nn.Embedding:
+        new_embeddings = super().resize_token_embeddings(new_num_tokens, pad_to_multiple_of, mean_resizing)
+        self._resize_final_logits_bias(new_embeddings.weight.shape[0])
+        return new_embeddings
+
+    def _resize_final_logits_bias(self, new_num_tokens: int) -> None:
+        old_num_tokens = self.final_logits_bias.shape[-1]
+        if new_num_tokens <= old_num_tokens:
+            new_bias = self.final_logits_bias[:, :new_num_tokens]
+        else:
+            extra_bias = torch.zeros((1, new_num_tokens - old_num_tokens), device=self.final_logits_bias.device)
+            new_bias = torch.cat([self.final_logits_bias, extra_bias], dim=1)
+        self.register_buffer("final_logits_bias", new_bias)
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    @add_start_docstrings_to_model_forward(BLENDERBOT_SMALL_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=Seq2SeqLMOutput, config_class=_CONFIG_FOR_DOC)
+    @add_end_docstrings(BLENDERBOT_SMALL_GENERATION_EXAMPLE)
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        decoder_input_ids: Optional[torch.LongTensor] = None,
+        decoder_attention_mask: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        decoder_head_mask: Optional[torch.Tensor] = None,
+        cross_attn_head_mask: Optional[torch.Tensor] = None,
+        encoder_outputs: Optional[Union[Tuple, BaseModelOutput]] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        decoder_inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple[torch.FloatTensor], Seq2SeqLMOutput]:
+        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]`.
+
+        Returns:
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if labels is not None:
+            if use_cache:
+                logger.warning("The `use_cache` argument is changed to `False` since `labels` is provided.")
+            use_cache = False
+            if decoder_input_ids is None and decoder_inputs_embeds is None:
+                decoder_input_ids = shift_tokens_right(
+                    labels, self.config.pad_token_id, self.config.decoder_start_token_id
+                )
+
+        outputs = self.model(
+            input_ids,
+            attention_mask=attention_mask,
+            decoder_input_ids=decoder_input_ids,
+            encoder_outputs=encoder_outputs,
+            decoder_attention_mask=decoder_attention_mask,
+            head_mask=head_mask,
+            decoder_head_mask=decoder_head_mask,
+            cross_attn_head_mask=cross_attn_head_mask,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            decoder_inputs_embeds=decoder_inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        lm_logits = self.lm_head(outputs[0]) + self.final_logits_bias
+
+        masked_lm_loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            masked_lm_loss = loss_fct(lm_logits.view(-1, self.config.vocab_size), labels.view(-1))
+
+        if not return_dict:
+            output = (lm_logits,) + outputs[1:]
+            return ((masked_lm_loss,) + output) if masked_lm_loss is not None else output
+
+        return Seq2SeqLMOutput(
+            loss=masked_lm_loss,
+            logits=lm_logits,
+            past_key_values=outputs.past_key_values,
+            decoder_hidden_states=outputs.decoder_hidden_states,
+            decoder_attentions=outputs.decoder_attentions,
+            cross_attentions=outputs.cross_attentions,
+            encoder_last_hidden_state=outputs.encoder_last_hidden_state,
+            encoder_hidden_states=outputs.encoder_hidden_states,
+            encoder_attentions=outputs.encoder_attentions,
+        )
+
+    @staticmethod
+    def _reorder_cache(past_key_values, beam_idx):
+        reordered_past = ()
+        for layer_past in past_key_values:
+            # cached cross_attention states don't have to be reordered -> they are always the same
+            reordered_past += (
+                tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past[:2])
+                + layer_past[2:],
+            )
+        return reordered_past
+
+
+# Copied from transformers.models.bart.modeling_bart.BartDecoderWrapper with Bart->BlenderbotSmall
+class BlenderbotSmallDecoderWrapper(BlenderbotSmallPreTrainedModel):
+    """
+    This wrapper class is a helper class to correctly load pretrained checkpoints when the causal language model is
+    used in combination with the [`EncoderDecoderModel`] framework.
+    """
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.decoder = BlenderbotSmallDecoder(config)
+
+    def forward(self, *args, **kwargs):
+        return self.decoder(*args, **kwargs)
+
+
+# Copied from transformers.models.bart.modeling_bart.BartForCausalLM with Bart->BlenderbotSmall, facebook/bart-base->facebook/blenderbot_small-90M
+class BlenderbotSmallForCausalLM(BlenderbotSmallPreTrainedModel, GenerationMixin):
+    _tied_weights_keys = ["lm_head.weight"]
+
+    def __init__(self, config):
+        config = copy.deepcopy(config)
+        config.is_decoder = True
+        config.is_encoder_decoder = False
+        super().__init__(config)
+        self.model = BlenderbotSmallDecoderWrapper(config)
+
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.decoder.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.decoder.embed_tokens = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def set_decoder(self, decoder):
+        self.model.decoder = decoder
+
+    def get_decoder(self):
+        return self.model.decoder
+
+    @replace_return_docstrings(output_type=CausalLMOutputWithCrossAttentions, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        cross_attn_head_mask: Optional[torch.Tensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, CausalLMOutputWithCrossAttentions]:
+        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)
+            encoder_hidden_states  (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+                Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention
+                if the model is configured as a decoder.
+            encoder_attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used
+                in the cross-attention if the model is configured as a decoder. Mask values selected in `[0, 1]`:
+            head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
+                Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:
+
+                - 1 indicates the head is **not masked**,
+                - 0 indicates the head is **masked**.
+
+            cross_attn_head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
+                Mask to nullify selected heads of the cross-attention modules. Mask values selected in `[0, 1]`:
+
+                - 1 indicates the head is **not masked**,
+                - 0 indicates the head is **masked**.
+
+            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)`. The two additional
+                tensors are only required when the model is used as a decoder in a Sequence to Sequence model.
+
+                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)`.
+            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]`.
+            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`).
+
+                - 1 for tokens that are **not masked**,
+                - 0 for tokens that are **masked**.
+            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.
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer, BlenderbotSmallForCausalLM
+
+        >>> tokenizer = AutoTokenizer.from_pretrained("facebook/blenderbot_small-90M")
+        >>> model = BlenderbotSmallForCausalLM.from_pretrained("facebook/blenderbot_small-90M", add_cross_attention=False)
+        >>> assert model.config.is_decoder, f"{model.__class__} has to be configured as a decoder."
+        >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
+        >>> outputs = model(**inputs)
+
+        >>> logits = outputs.logits
+        >>> expected_shape = [1, inputs.input_ids.shape[-1], model.config.vocab_size]
+        >>> list(logits.shape) == expected_shape
+        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
+
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        outputs = self.model.decoder(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            head_mask=head_mask,
+            cross_attn_head_mask=cross_attn_head_mask,
+            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,
+        )
+
+        logits = self.lm_head(outputs[0])
+
+        loss = None
+        if labels is not None:
+            labels = labels.to(logits.device)
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(logits.view(-1, self.config.vocab_size), labels.view(-1))
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return CausalLMOutputWithCrossAttentions(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            cross_attentions=outputs.cross_attentions,
+        )
+
+    @staticmethod
+    def _reorder_cache(past_key_values, beam_idx):
+        reordered_past = ()
+        for layer_past in past_key_values:
+            reordered_past += (
+                tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past),
+            )
+        return reordered_past
+
+
+__all__ = [
+    "BlenderbotSmallForCausalLM",
+    "BlenderbotSmallForConditionalGeneration",
+    "BlenderbotSmallModel",
+    "BlenderbotSmallPreTrainedModel",
+]

lib/python3.12/site-packages/transformers/models/blenderbot_small/modeling_flax_blenderbot_small.py

@@ -0,0 +1,1528 @@
+# coding=utf-8
+# Copyright 2021 The Facebook, Inc. 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.
+"""Flax BlenderbotSmall model."""
+
+import math
+import random
+from functools import partial
+from typing import Callable, Optional, Tuple
+
+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 jax.random import PRNGKey
+
+from ...modeling_flax_outputs import (
+    FlaxBaseModelOutput,
+    FlaxBaseModelOutputWithPastAndCrossAttentions,
+    FlaxCausalLMOutputWithCrossAttentions,
+    FlaxSeq2SeqLMOutput,
+    FlaxSeq2SeqModelOutput,
+)
+from ...modeling_flax_utils import (
+    ACT2FN,
+    FlaxPreTrainedModel,
+    append_call_sample_docstring,
+    append_replace_return_docstrings,
+    overwrite_call_docstring,
+)
+from ...utils import add_start_docstrings, logging, replace_return_docstrings
+from .configuration_blenderbot_small import BlenderbotSmallConfig
+
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "facebook/blenderbot_small-90M"
+_CONFIG_FOR_DOC = "BlenderbotSmallConfig"
+
+BLENDERBOT_SMALL_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 or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a Flax Linen
+    [flax.nn.Module](https://flax.readthedocs.io/en/latest/_autosummary/flax.nn.module.html) subclass. Use it as a
+    regular Flax 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 ([`BlenderbotSmallConfig`]): 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`].
+"""
+
+BLENDERBOT_SMALL_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`jnp.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 (`jnp.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)
+        decoder_input_ids (`jnp.ndarray` of shape `(batch_size, target_sequence_length)`, *optional*):
+            Indices of decoder input sequence tokens in the vocabulary.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are decoder input IDs?](../glossary#decoder-input-ids)
+
+            For translation and summarization training, `decoder_input_ids` should be provided. If no
+            `decoder_input_ids` is provided, the model will create this tensor by shifting the `input_ids` to the right
+            for denoising pre-training following the paper.
+        decoder_attention_mask (`jnp.ndarray` of shape `(batch_size, target_sequence_length)`, *optional*):
+            Default behavior: generate a tensor that ignores pad tokens in `decoder_input_ids`. Causal mask will also
+            be used by default.
+
+            If you want to change padding behavior, you should modify to your needs. See diagram 1 in [the
+            paper](https://arxiv.org/abs/1910.13461) for more information on the default strategy.
+        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]`.
+        decoder_position_ids (`numpy.ndarray` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each decoder input sequence tokens in the position embeddings. Selected in the
+            range `[0, config.max_position_embeddings - 1]`.
+        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.
+"""
+
+
+BLENDERBOT_SMALL_ENCODE_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`jnp.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 (`jnp.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]`.
+        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.
+"""
+
+BLENDERBOT_SMALL_DECODE_INPUTS_DOCSTRING = r"""
+    Args:
+        decoder_input_ids (`jnp.ndarray` of shape `(batch_size, target_sequence_length)`):
+            Indices of decoder input sequence tokens in the vocabulary.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are decoder input IDs?](../glossary#decoder-input-ids)
+
+            For translation and summarization training, `decoder_input_ids` should be provided. If no
+            `decoder_input_ids` is provided, the model will create this tensor by shifting the `input_ids` to the right
+            for denoising pre-training following the paper.
+        encoder_outputs (`tuple(tuple(jnp.ndarray)`):
+            Tuple consists of (`last_hidden_state`, *optional*: `hidden_states`, *optional*: `attentions`)
+            `last_hidden_state` of shape `(batch_size, sequence_length, hidden_size)`, *optional*) is a sequence of
+            hidden-states at the output of the last layer of the encoder. Used in the cross-attention of the decoder.
+        encoder_attention_mask (`jnp.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)
+        decoder_attention_mask (`jnp.ndarray` of shape `(batch_size, target_sequence_length)`, *optional*):
+            Default behavior: generate a tensor that ignores pad tokens in `decoder_input_ids`. Causal mask will also
+            be used by default.
+
+            If you want to change padding behavior, you should modify to your needs. See diagram 1 in [the
+            paper](https://arxiv.org/abs/1910.13461) for more information on the default strategy.
+        decoder_position_ids (`numpy.ndarray` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each decoder input sequence tokens in the position embeddings. Selected in the
+            range `[0, config.max_position_embeddings - 1]`.
+        past_key_values (`Dict[str, np.ndarray]`, *optional*, returned by `init_cache` or when passing previous `past_key_values`):
+            Dictionary of pre-computed hidden-states (key and values in the attention blocks) that can be used for fast
+            auto-regressive decoding. Pre-computed key and value hidden-states are of shape *[batch_size, max_length]*.
+        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.
+"""
+
+
+# Copied from transformers.models.bart.modeling_flax_bart.shift_tokens_right
+def shift_tokens_right(input_ids: jnp.ndarray, pad_token_id: int, decoder_start_token_id: int) -> jnp.ndarray:
+    """
+    Shift input ids one token to the right.
+    """
+    shifted_input_ids = jnp.zeros_like(input_ids)
+    shifted_input_ids = shifted_input_ids.at[:, 1:].set(input_ids[:, :-1])
+    shifted_input_ids = shifted_input_ids.at[:, 0].set(decoder_start_token_id)
+
+    shifted_input_ids = jnp.where(shifted_input_ids == -100, pad_token_id, shifted_input_ids)
+    return shifted_input_ids
+
+
+# Copied from transformers.models.bart.modeling_flax_bart.FlaxBartAttention with Bart->BlenderbotSmall
+class FlaxBlenderbotSmallAttention(nn.Module):
+    config: BlenderbotSmallConfig
+    embed_dim: int
+    num_heads: int
+    dropout: float = 0.0
+    causal: bool = False
+    bias: bool = True
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
+
+    def setup(self) -> None:
+        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}"
+                f" and `num_heads`: {self.num_heads})."
+            )
+
+        dense = partial(
+            nn.Dense,
+            self.embed_dim,
+            use_bias=self.bias,
+            dtype=self.dtype,
+            kernel_init=jax.nn.initializers.normal(self.config.init_std),
+        )
+
+        self.q_proj, self.k_proj, self.v_proj = dense(), dense(), dense()
+        self.out_proj = dense()
+
+        self.dropout_layer = nn.Dropout(rate=self.dropout)
+
+        if self.causal:
+            self.causal_mask = make_causal_mask(
+                jnp.ones((1, self.config.max_position_embeddings), dtype="bool"), dtype="bool"
+            )
+
+    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,))
+
+    @nn.compact
+    def _concatenate_to_cache(self, key, value, query, attention_mask):
+        """
+        This function takes projected key, value states from a single input token and concatenates the states to cached
+        states from previous steps. This function is slightly adapted from the official Flax repository:
+        https://github.com/google/flax/blob/491ce18759622506588784b4fca0e4bf05f8c8cd/flax/linen/attention.py#L252
+        """
+        # detect if we're initializing by absence of existing cache data.
+        is_initialized = self.has_variable("cache", "cached_key")
+        cached_key = self.variable("cache", "cached_key", jnp.zeros, key.shape, key.dtype)
+        cached_value = self.variable("cache", "cached_value", jnp.zeros, value.shape, value.dtype)
+        cache_index = self.variable("cache", "cache_index", lambda: jnp.array(0, dtype=jnp.int32))
+
+        if is_initialized:
+            *batch_dims, max_length, num_heads, depth_per_head = cached_key.value.shape
+            # update key, value caches with our new 1d spatial slices
+            cur_index = cache_index.value
+            indices = (0,) * len(batch_dims) + (cur_index, 0, 0)
+            key = lax.dynamic_update_slice(cached_key.value, key, indices)
+            value = lax.dynamic_update_slice(cached_value.value, value, indices)
+            cached_key.value = key
+            cached_value.value = value
+            num_updated_cache_vectors = query.shape[1]
+            cache_index.value = cache_index.value + num_updated_cache_vectors
+            # causal mask for cached decoder self-attention: our single query position should only attend to those key positions that have already been generated and cached, not the remaining zero elements.
+            pad_mask = jnp.broadcast_to(
+                jnp.arange(max_length) < cur_index + num_updated_cache_vectors,
+                tuple(batch_dims) + (1, num_updated_cache_vectors, max_length),
+            )
+            attention_mask = combine_masks(pad_mask, attention_mask)
+        return key, value, attention_mask
+
+    def __call__(
+        self,
+        hidden_states: jnp.ndarray,
+        key_value_states: Optional[jnp.ndarray] = None,
+        attention_mask: Optional[jnp.ndarray] = None,
+        init_cache: bool = False,
+        deterministic: bool = True,
+    ) -> Tuple[jnp.ndarray]:
+        """Input shape: Batch x Time x Channel"""
+
+        # if key_value_states are provided this layer is used as a cross-attention layer
+        # for the decoder
+        is_cross_attention = key_value_states is not None
+        batch_size = hidden_states.shape[0]
+
+        # get query proj
+        query_states = self.q_proj(hidden_states)
+        # get key, value proj
+        if is_cross_attention:
+            # cross_attentions
+            key_states = self.k_proj(key_value_states)
+            value_states = self.v_proj(key_value_states)
+        else:
+            # self_attention
+            key_states = self.k_proj(hidden_states)
+            value_states = self.v_proj(hidden_states)
+
+        query_states = self._split_heads(query_states)
+        key_states = self._split_heads(key_states)
+        value_states = self._split_heads(value_states)
+
+        # handle cache prepare causal attention mask
+        if self.causal:
+            query_length, key_length = query_states.shape[1], key_states.shape[1]
+            if self.has_variable("cache", "cached_key"):
+                mask_shift = self.variables["cache"]["cache_index"]
+                max_decoder_length = self.variables["cache"]["cached_key"].shape[1]
+                causal_mask = lax.dynamic_slice(
+                    self.causal_mask, (0, 0, mask_shift, 0), (1, 1, query_length, max_decoder_length)
+                )
+            else:
+                causal_mask = self.causal_mask[:, :, :query_length, :key_length]
+            causal_mask = jnp.broadcast_to(causal_mask, (batch_size,) + causal_mask.shape[1:])
+
+        # combine masks if needed
+        if attention_mask is not None and self.causal:
+            attention_mask = jnp.broadcast_to(jnp.expand_dims(attention_mask, axis=(-3, -2)), causal_mask.shape)
+            attention_mask = combine_masks(attention_mask, causal_mask)
+        elif self.causal:
+            attention_mask = causal_mask
+        elif attention_mask is not None:
+            attention_mask = jnp.expand_dims(attention_mask, axis=(-3, -2))
+
+        # During fast autoregressive decoding, we feed one position at a time,
+        # and cache the keys and values step by step.
+        if self.causal and (self.has_variable("cache", "cached_key") or init_cache):
+            key_states, value_states, attention_mask = self._concatenate_to_cache(
+                key_states, value_states, query_states, attention_mask
+            )
+
+        # Convert the boolean attention mask to an attention bias.
+        if attention_mask is not None:
+            # attention mask in the form of attention bias
+            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_states,
+            key_states,
+            bias=attention_bias,
+            dropout_rng=dropout_rng,
+            dropout_rate=self.dropout,
+            broadcast_dropout=True,
+            deterministic=deterministic,
+            dtype=self.dtype,
+            precision=None,
+        )
+
+        attn_output = jnp.einsum("...hqk,...khd->...qhd", attn_weights, value_states)
+        attn_output = self._merge_heads(attn_output)
+        attn_output = self.out_proj(attn_output)
+
+        return attn_output, attn_weights
+
+
+# Copied from transformers.models.bart.modeling_flax_bart.FlaxBartEncoderLayer with Bart->BlenderbotSmall
+class FlaxBlenderbotSmallEncoderLayer(nn.Module):
+    config: BlenderbotSmallConfig
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self) -> None:
+        self.embed_dim = self.config.d_model
+        self.self_attn = FlaxBlenderbotSmallAttention(
+            config=self.config,
+            embed_dim=self.embed_dim,
+            num_heads=self.config.encoder_attention_heads,
+            dropout=self.config.attention_dropout,
+            dtype=self.dtype,
+        )
+        self.self_attn_layer_norm = nn.LayerNorm(dtype=self.dtype, epsilon=1e-05)
+        self.dropout_layer = nn.Dropout(rate=self.config.dropout)
+        self.activation_fn = ACT2FN[self.config.activation_function]
+        self.activation_dropout_layer = nn.Dropout(rate=self.config.activation_dropout)
+        self.fc1 = nn.Dense(
+            self.config.encoder_ffn_dim,
+            dtype=self.dtype,
+            kernel_init=jax.nn.initializers.normal(self.config.init_std),
+        )
+        self.fc2 = nn.Dense(
+            self.embed_dim, dtype=self.dtype, kernel_init=jax.nn.initializers.normal(self.config.init_std)
+        )
+        self.final_layer_norm = nn.LayerNorm(dtype=self.dtype, epsilon=1e-05)
+
+    def __call__(
+        self,
+        hidden_states: jnp.ndarray,
+        attention_mask: jnp.ndarray,
+        output_attentions: bool = True,
+        deterministic: bool = True,
+    ) -> Tuple[jnp.ndarray]:
+        residual = hidden_states
+        hidden_states, attn_weights = self.self_attn(hidden_states=hidden_states, attention_mask=attention_mask)
+
+        hidden_states = self.dropout_layer(hidden_states, deterministic=deterministic)
+        hidden_states = residual + hidden_states
+        hidden_states = self.self_attn_layer_norm(hidden_states)
+
+        residual = hidden_states
+        hidden_states = self.activation_fn(self.fc1(hidden_states))
+        hidden_states = self.activation_dropout_layer(hidden_states, deterministic=deterministic)
+        hidden_states = self.fc2(hidden_states)
+        hidden_states = self.dropout_layer(hidden_states, deterministic=deterministic)
+        hidden_states = residual + hidden_states
+        hidden_states = self.final_layer_norm(hidden_states)
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (attn_weights,)
+
+        return outputs
+
+
+# Copied from transformers.models.bart.modeling_flax_bart.FlaxBartEncoderLayerCollection with Bart->BlenderbotSmall
+class FlaxBlenderbotSmallEncoderLayerCollection(nn.Module):
+    config: BlenderbotSmallConfig
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
+
+    def setup(self):
+        self.layers = [
+            FlaxBlenderbotSmallEncoderLayer(self.config, name=str(i), dtype=self.dtype)
+            for i in range(self.config.encoder_layers)
+        ]
+        self.layerdrop = self.config.encoder_layerdrop
+
+    def __call__(
+        self,
+        hidden_states,
+        attention_mask,
+        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 encoder_layer in self.layers:
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
+            dropout_probability = random.uniform(0, 1)
+            if not deterministic and (dropout_probability < self.layerdrop):  # skip the layer
+                layer_outputs = (None, None)
+            else:
+                layer_outputs = encoder_layer(
+                    hidden_states,
+                    attention_mask,
+                    output_attentions,
+                    deterministic,
+                )
+            hidden_states = layer_outputs[0]
+            if output_attentions:
+                all_attentions = all_attentions + (layer_outputs[1],)
+
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        outputs = (hidden_states, all_hidden_states, all_attentions)
+
+        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
+        )
+
+
+# Copied from transformers.models.bart.modeling_flax_bart.FlaxBartDecoderLayer with Bart->BlenderbotSmall
+class FlaxBlenderbotSmallDecoderLayer(nn.Module):
+    config: BlenderbotSmallConfig
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self) -> None:
+        self.embed_dim = self.config.d_model
+        self.self_attn = FlaxBlenderbotSmallAttention(
+            config=self.config,
+            embed_dim=self.embed_dim,
+            num_heads=self.config.decoder_attention_heads,
+            dropout=self.config.attention_dropout,
+            causal=True,
+            dtype=self.dtype,
+        )
+        self.dropout_layer = nn.Dropout(rate=self.config.dropout)
+        self.activation_fn = ACT2FN[self.config.activation_function]
+        self.activation_dropout_layer = nn.Dropout(rate=self.config.activation_dropout)
+
+        self.self_attn_layer_norm = nn.LayerNorm(dtype=self.dtype, epsilon=1e-05)
+        self.encoder_attn = FlaxBlenderbotSmallAttention(
+            config=self.config,
+            embed_dim=self.embed_dim,
+            num_heads=self.config.decoder_attention_heads,
+            dropout=self.config.attention_dropout,
+            dtype=self.dtype,
+        )
+        self.encoder_attn_layer_norm = nn.LayerNorm(dtype=self.dtype, epsilon=1e-05)
+        self.fc1 = nn.Dense(
+            self.config.decoder_ffn_dim,
+            dtype=self.dtype,
+            kernel_init=jax.nn.initializers.normal(self.config.init_std),
+        )
+        self.fc2 = nn.Dense(
+            self.embed_dim, dtype=self.dtype, kernel_init=jax.nn.initializers.normal(self.config.init_std)
+        )
+        self.final_layer_norm = nn.LayerNorm(dtype=self.dtype, epsilon=1e-05)
+
+    def __call__(
+        self,
+        hidden_states: jnp.ndarray,
+        attention_mask: jnp.ndarray,
+        encoder_hidden_states: Optional[jnp.ndarray] = None,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        init_cache: bool = False,
+        output_attentions: bool = True,
+        deterministic: bool = True,
+    ) -> Tuple[jnp.ndarray]:
+        residual = hidden_states
+
+        # Self Attention
+        hidden_states, self_attn_weights = self.self_attn(
+            hidden_states=hidden_states, attention_mask=attention_mask, init_cache=init_cache
+        )
+        hidden_states = self.dropout_layer(hidden_states, deterministic=deterministic)
+        hidden_states = residual + hidden_states
+        hidden_states = self.self_attn_layer_norm(hidden_states)
+
+        # Cross-Attention Block
+        cross_attn_weights = None
+        if encoder_hidden_states is not None:
+            residual = hidden_states
+
+            hidden_states, cross_attn_weights = self.encoder_attn(
+                hidden_states=hidden_states,
+                key_value_states=encoder_hidden_states,
+                attention_mask=encoder_attention_mask,
+            )
+            hidden_states = self.dropout_layer(hidden_states, deterministic=deterministic)
+            hidden_states = residual + hidden_states
+            hidden_states = self.encoder_attn_layer_norm(hidden_states)
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.activation_fn(self.fc1(hidden_states))
+        hidden_states = self.activation_dropout_layer(hidden_states, deterministic=deterministic)
+        hidden_states = self.fc2(hidden_states)
+        hidden_states = self.dropout_layer(hidden_states, deterministic=deterministic)
+        hidden_states = residual + hidden_states
+        hidden_states = self.final_layer_norm(hidden_states)
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights, cross_attn_weights)
+
+        return outputs
+
+
+# Copied from transformers.models.bart.modeling_flax_bart.FlaxBartDecoderLayerCollection with Bart->BlenderbotSmall
+class FlaxBlenderbotSmallDecoderLayerCollection(nn.Module):
+    config: BlenderbotSmallConfig
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
+
+    def setup(self):
+        self.layers = [
+            FlaxBlenderbotSmallDecoderLayer(self.config, name=str(i), dtype=self.dtype)
+            for i in range(self.config.decoder_layers)
+        ]
+        self.layerdrop = self.config.decoder_layerdrop
+
+    def __call__(
+        self,
+        hidden_states,
+        attention_mask,
+        encoder_hidden_states: Optional[jnp.ndarray] = None,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        deterministic: bool = True,
+        init_cache: bool = False,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ):
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        all_cross_attentions = () if (output_attentions and encoder_hidden_states is not None) else None
+
+        for decoder_layer in self.layers:
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+                # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
+            dropout_probability = random.uniform(0, 1)
+            if not deterministic and (dropout_probability < self.layerdrop):
+                layer_outputs = (None, None, None)
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    encoder_hidden_states=encoder_hidden_states,
+                    encoder_attention_mask=encoder_attention_mask,
+                    init_cache=init_cache,
+                    output_attentions=output_attentions,
+                    deterministic=deterministic,
+                )
+
+            hidden_states = layer_outputs[0]
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+                if encoder_hidden_states is not None:
+                    all_cross_attentions += (layer_outputs[2],)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        outputs = [hidden_states, all_hidden_states, all_self_attns, all_cross_attentions]
+
+        if not return_dict:
+            return tuple(v for v in outputs if v is not None)
+
+        return FlaxBaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+            cross_attentions=all_cross_attentions,
+        )
+
+
+class FlaxBlenderbotSmallEncoder(nn.Module):
+    config: BlenderbotSmallConfig
+    embed_tokens: nn.Embed
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
+
+    def setup(self):
+        self.dropout_layer = nn.Dropout(rate=self.config.dropout)
+
+        embed_dim = self.config.d_model
+        self.padding_idx = self.config.pad_token_id
+        self.max_source_positions = self.config.max_position_embeddings
+        self.embed_scale = math.sqrt(embed_dim) if self.config.scale_embedding else 1.0
+
+        self.embed_positions = nn.Embed(
+            self.config.max_position_embeddings,
+            embed_dim,
+            embedding_init=jax.nn.initializers.normal(self.config.init_std),
+        )
+        self.layers = FlaxBlenderbotSmallEncoderLayerCollection(self.config, self.dtype)
+        self.layernorm_embedding = nn.LayerNorm(dtype=self.dtype, epsilon=1e-05)
+
+    def __call__(
+        self,
+        input_ids,
+        attention_mask,
+        position_ids,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+        deterministic: bool = True,
+    ):
+        input_shape = input_ids.shape
+        input_ids = input_ids.reshape(-1, input_shape[-1])
+
+        inputs_embeds = self.embed_tokens(input_ids) * self.embed_scale
+
+        embed_pos = self.embed_positions(position_ids)
+
+        hidden_states = inputs_embeds + embed_pos
+        hidden_states = self.layernorm_embedding(hidden_states)
+        hidden_states = self.dropout_layer(hidden_states, deterministic=deterministic)
+
+        outputs = self.layers(
+            hidden_states,
+            attention_mask,
+            deterministic=deterministic,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        if not return_dict:
+            return outputs
+
+        return FlaxBaseModelOutput(
+            last_hidden_state=outputs.last_hidden_state,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+class FlaxBlenderbotSmallDecoder(nn.Module):
+    config: BlenderbotSmallConfig
+    embed_tokens: nn.Embed
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
+
+    def setup(self):
+        self.dropout_layer = nn.Dropout(rate=self.config.dropout)
+
+        embed_dim = self.config.d_model
+        self.padding_idx = self.config.pad_token_id
+        self.max_target_positions = self.config.max_position_embeddings
+        self.embed_scale = math.sqrt(self.config.d_model) if self.config.scale_embedding else 1.0
+
+        self.embed_positions = nn.Embed(
+            self.config.max_position_embeddings,
+            embed_dim,
+            embedding_init=jax.nn.initializers.normal(self.config.init_std),
+        )
+
+        self.layers = FlaxBlenderbotSmallDecoderLayerCollection(self.config, self.dtype)
+        self.layernorm_embedding = nn.LayerNorm(dtype=self.dtype, epsilon=1e-05)
+
+    def __call__(
+        self,
+        input_ids,
+        attention_mask,
+        position_ids,
+        encoder_hidden_states: Optional[jnp.ndarray] = None,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        init_cache: bool = False,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+        deterministic: bool = True,
+    ):
+        input_shape = input_ids.shape
+        input_ids = input_ids.reshape(-1, input_shape[-1])
+
+        inputs_embeds = self.embed_tokens(input_ids) * self.embed_scale
+
+        # embed positions
+        positions = self.embed_positions(position_ids)
+
+        # BlenderbotSmall applies layer norm on inputs_embeds in decoder
+        inputs_embeds = self.layernorm_embedding(inputs_embeds)
+        hidden_states = inputs_embeds + positions
+
+        hidden_states = self.dropout_layer(hidden_states, deterministic=deterministic)
+
+        outputs = self.layers(
+            hidden_states,
+            attention_mask,
+            encoder_hidden_states,
+            encoder_attention_mask,
+            deterministic=deterministic,
+            init_cache=init_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        if not return_dict:
+            return outputs
+
+        return FlaxBaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=outputs.last_hidden_state,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            cross_attentions=outputs.cross_attentions,
+        )
+
+
+# Copied from transformers.models.bart.modeling_flax_bart.FlaxBartModule with Bart->BlenderbotSmall
+class FlaxBlenderbotSmallModule(nn.Module):
+    config: BlenderbotSmallConfig
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
+
+    def setup(self):
+        self.shared = nn.Embed(
+            self.config.vocab_size,
+            self.config.d_model,
+            embedding_init=jax.nn.initializers.normal(self.config.init_std),
+            dtype=self.dtype,
+        )
+
+        self.encoder = FlaxBlenderbotSmallEncoder(self.config, dtype=self.dtype, embed_tokens=self.shared)
+        self.decoder = FlaxBlenderbotSmallDecoder(self.config, dtype=self.dtype, embed_tokens=self.shared)
+
+    def _get_encoder_module(self):
+        return self.encoder
+
+    def _get_decoder_module(self):
+        return self.decoder
+
+    def __call__(
+        self,
+        input_ids,
+        attention_mask,
+        decoder_input_ids,
+        decoder_attention_mask,
+        position_ids,
+        decoder_position_ids,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+        deterministic: bool = True,
+    ):
+        encoder_outputs = self.encoder(
+            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,
+            deterministic=deterministic,
+        )
+
+        decoder_outputs = self.decoder(
+            input_ids=decoder_input_ids,
+            attention_mask=decoder_attention_mask,
+            position_ids=decoder_position_ids,
+            encoder_hidden_states=encoder_outputs[0],
+            encoder_attention_mask=attention_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            deterministic=deterministic,
+        )
+
+        if not return_dict:
+            return decoder_outputs + encoder_outputs
+
+        return FlaxSeq2SeqModelOutput(
+            last_hidden_state=decoder_outputs.last_hidden_state,
+            decoder_hidden_states=decoder_outputs.hidden_states,
+            decoder_attentions=decoder_outputs.attentions,
+            cross_attentions=decoder_outputs.cross_attentions,
+            encoder_last_hidden_state=encoder_outputs.last_hidden_state,
+            encoder_hidden_states=encoder_outputs.hidden_states,
+            encoder_attentions=encoder_outputs.attentions,
+        )
+
+
+class FlaxBlenderbotSmallPreTrainedModel(FlaxPreTrainedModel):
+    config_class = BlenderbotSmallConfig
+    base_model_prefix: str = "model"
+    module_class: nn.Module = None
+
+    def __init__(
+        self,
+        config: BlenderbotSmallConfig,
+        input_shape: Tuple[int] = (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 tensors
+        input_ids = jnp.zeros(input_shape, dtype="i4")
+        # make sure initialization pass will work for FlaxBlenderbotSmallForSequenceClassificationModule
+        input_ids = input_ids.at[(..., -1)].set(self.config.eos_token_id)
+        attention_mask = jnp.ones_like(input_ids)
+        decoder_input_ids = input_ids
+        decoder_attention_mask = jnp.ones_like(input_ids)
+
+        batch_size, sequence_length = input_ids.shape
+        position_ids = jnp.broadcast_to(jnp.arange(sequence_length)[None, :], (batch_size, sequence_length))
+        decoder_position_ids = jnp.broadcast_to(jnp.arange(sequence_length)[None, :], (batch_size, sequence_length))
+
+        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,
+            decoder_input_ids,
+            decoder_attention_mask,
+            position_ids,
+            decoder_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 init_cache(self, batch_size, max_length, encoder_outputs):
+        r"""
+        Args:
+            batch_size (`int`):
+                batch_size used for fast auto-regressive decoding. Defines the batch size of the initialized cache.
+            max_length (`int`):
+                maximum possible length for auto-regressive decoding. Defines the sequence length of the initialized
+                cache.
+            encoder_outputs (`Union[FlaxBaseModelOutput, tuple(tuple(jnp.ndarray)]`):
+                `encoder_outputs` consists of (`last_hidden_state`, *optional*: `hidden_states`, *optional*:
+                `attentions`). `last_hidden_state` of shape `(batch_size, sequence_length, hidden_size)`, *optional*)
+                is a sequence of hidden-states at the output of the last layer of the encoder. Used in the
+                cross-attention of the decoder.
+        """
+        # init input variables to retrieve cache
+        decoder_input_ids = jnp.ones((batch_size, max_length), dtype="i4")
+        decoder_attention_mask = jnp.ones_like(decoder_input_ids)
+        decoder_position_ids = jnp.broadcast_to(
+            jnp.arange(jnp.atleast_2d(decoder_input_ids).shape[-1]), decoder_input_ids.shape
+        )
+
+        def _decoder_forward(module, decoder_input_ids, decoder_attention_mask, decoder_position_ids, **kwargs):
+            decoder_module = module._get_decoder_module()
+            return decoder_module(
+                decoder_input_ids,
+                decoder_attention_mask,
+                decoder_position_ids,
+                **kwargs,
+            )
+
+        init_variables = self.module.init(
+            jax.random.PRNGKey(0),
+            decoder_input_ids=decoder_input_ids,
+            decoder_attention_mask=decoder_attention_mask,
+            decoder_position_ids=decoder_position_ids,
+            encoder_hidden_states=encoder_outputs[0],
+            init_cache=True,
+            method=_decoder_forward,  # we only need to call the decoder to init the cache
+        )
+        return unfreeze(init_variables["cache"])
+
+    @add_start_docstrings(BLENDERBOT_SMALL_ENCODE_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=FlaxBaseModelOutput, config_class=BlenderbotSmallConfig)
+    def encode(
+        self,
+        input_ids: jnp.ndarray,
+        attention_mask: Optional[jnp.ndarray] = None,
+        position_ids: Optional[jnp.ndarray] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        train: bool = False,
+        params: dict = None,
+        dropout_rng: PRNGKey = None,
+    ):
+        r"""
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer, FlaxBlenderbotSmallForConditionalGeneration
+
+        >>> model = FlaxBlenderbotSmallForConditionalGeneration.from_pretrained("facebook/blenderbot_small-90M")
+        >>> tokenizer = AutoTokenizer.from_pretrained("facebook/blenderbot_small-90M")
+
+        >>> text = "My friends are cool but they eat too many carbs."
+        >>> inputs = tokenizer(text, max_length=1024, return_tensors="np")
+        >>> encoder_outputs = model.encode(**inputs)
+        ```"""
+        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 attention_mask is None:
+            attention_mask = jnp.ones_like(input_ids)
+        if position_ids is None:
+            batch_size, sequence_length = input_ids.shape
+            position_ids = jnp.broadcast_to(jnp.arange(sequence_length)[None, :], (batch_size, sequence_length))
+
+        # Handle any PRNG if needed
+        rngs = {}
+        if dropout_rng is not None:
+            rngs["dropout"] = dropout_rng
+
+        def _encoder_forward(module, input_ids, attention_mask, position_ids, **kwargs):
+            encode_module = module._get_encoder_module()
+            return encode_module(input_ids, attention_mask, position_ids, **kwargs)
+
+        return self.module.apply(
+            {"params": params or self.params},
+            input_ids=jnp.array(input_ids, dtype="i4"),
+            attention_mask=jnp.array(attention_mask, dtype="i4"),
+            position_ids=jnp.array(position_ids, dtype="i4"),
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            deterministic=not train,
+            rngs=rngs,
+            method=_encoder_forward,
+        )
+
+    @add_start_docstrings(BLENDERBOT_SMALL_DECODE_INPUTS_DOCSTRING)
+    @replace_return_docstrings(
+        output_type=FlaxBaseModelOutputWithPastAndCrossAttentions, config_class=BlenderbotSmallConfig
+    )
+    def decode(
+        self,
+        decoder_input_ids,
+        encoder_outputs,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        decoder_attention_mask: Optional[jnp.ndarray] = None,
+        decoder_position_ids: Optional[jnp.ndarray] = None,
+        past_key_values: dict = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        train: bool = False,
+        params: dict = None,
+        dropout_rng: PRNGKey = None,
+    ):
+        r"""
+        Returns:
+
+        Example:
+
+        ```python
+        >>> import jax.numpy as jnp
+        >>> from transformers import AutoTokenizer, FlaxBlenderbotSmallForConditionalGeneration
+
+        >>> model = FlaxBlenderbotSmallForConditionalGeneration.from_pretrained("facebook/blenderbot_small-90M")
+        >>> tokenizer = AutoTokenizer.from_pretrained("facebook/blenderbot_small-90M")
+
+        >>> text = "My friends are cool but they eat too many carbs."
+        >>> inputs = tokenizer(text, max_length=1024, return_tensors="np")
+        >>> encoder_outputs = model.encode(**inputs)
+
+        >>> decoder_start_token_id = model.config.decoder_start_token_id
+        >>> decoder_input_ids = jnp.ones((inputs.input_ids.shape[0], 1), dtype="i4") * decoder_start_token_id
+
+        >>> outputs = model.decode(decoder_input_ids, encoder_outputs)
+        >>> last_decoder_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.return_dict
+
+        encoder_hidden_states = encoder_outputs[0]
+        if encoder_attention_mask is None:
+            batch_size, sequence_length = encoder_hidden_states.shape[:2]
+            encoder_attention_mask = jnp.ones((batch_size, sequence_length))
+
+        batch_size, sequence_length = decoder_input_ids.shape
+        if decoder_attention_mask is None:
+            decoder_attention_mask = jnp.ones((batch_size, sequence_length))
+
+        if decoder_position_ids is None:
+            if past_key_values is not None:
+                raise ValueError("Make sure to provide `decoder_position_ids` when passing `past_key_values`.")
+
+            decoder_position_ids = jnp.broadcast_to(
+                jnp.arange(sequence_length)[None, :], (batch_size, sequence_length)
+            )
+
+        # Handle any PRNG if needed
+        rngs = {}
+        if dropout_rng is not None:
+            rngs["dropout"] = dropout_rng
+
+        inputs = {"params": params or self.params}
+
+        # if past_key_values are passed then cache is already initialized a private flag init_cache has to be
+        # passed down to ensure cache is used. It has to be made sure that cache is marked as mutable so that
+        # it can be changed by FlaxBlenderbotSmallAttention module
+        if past_key_values:
+            inputs["cache"] = past_key_values
+            mutable = ["cache"]
+        else:
+            mutable = False
+
+        def _decoder_forward(module, decoder_input_ids, decoder_attention_mask, decoder_position_ids, **kwargs):
+            decoder_module = module._get_decoder_module()
+            return decoder_module(
+                decoder_input_ids,
+                decoder_attention_mask,
+                decoder_position_ids,
+                **kwargs,
+            )
+
+        outputs = self.module.apply(
+            inputs,
+            decoder_input_ids=jnp.array(decoder_input_ids, dtype="i4"),
+            decoder_attention_mask=jnp.array(decoder_attention_mask, dtype="i4"),
+            decoder_position_ids=jnp.array(decoder_position_ids, dtype="i4"),
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=jnp.array(encoder_attention_mask, dtype="i4"),
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            deterministic=not train,
+            rngs=rngs,
+            mutable=mutable,
+            method=_decoder_forward,
+        )
+
+        # add updated cache to model output
+        if past_key_values is not None and return_dict:
+            outputs, past = outputs
+            outputs["past_key_values"] = unfreeze(past["cache"])
+            return outputs
+        elif past_key_values is not None and not return_dict:
+            outputs, past = outputs
+            outputs = outputs[:1] + (unfreeze(past["cache"]),) + outputs[1:]
+
+        return outputs
+
+    def __call__(
+        self,
+        input_ids: jnp.ndarray,
+        attention_mask: Optional[jnp.ndarray] = None,
+        decoder_input_ids: Optional[jnp.ndarray] = None,
+        decoder_attention_mask: Optional[jnp.ndarray] = None,
+        position_ids: Optional[jnp.ndarray] = None,
+        decoder_position_ids: Optional[jnp.ndarray] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        train: bool = False,
+        params: dict = None,
+        dropout_rng: PRNGKey = 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
+
+        # prepare encoder inputs
+        if attention_mask is None:
+            attention_mask = jnp.ones_like(input_ids)
+        if position_ids is None:
+            batch_size, sequence_length = input_ids.shape
+            position_ids = jnp.broadcast_to(jnp.arange(sequence_length)[None, :], (batch_size, sequence_length))
+
+        # prepare decoder inputs
+        if decoder_input_ids is None:
+            decoder_input_ids = shift_tokens_right(
+                input_ids, self.config.pad_token_id, decoder_start_token_id=self.config.decoder_start_token_id
+            )
+        if decoder_attention_mask is None:
+            decoder_attention_mask = jnp.ones_like(decoder_input_ids)
+        if decoder_position_ids is None:
+            batch_size, sequence_length = decoder_input_ids.shape
+            decoder_position_ids = jnp.broadcast_to(
+                jnp.arange(sequence_length)[None, :], (batch_size, sequence_length)
+            )
+
+        # Handle any PRNG if needed
+        rngs = {"dropout": dropout_rng} if dropout_rng is not None else {}
+
+        return self.module.apply(
+            {"params": params or self.params},
+            input_ids=jnp.array(input_ids, dtype="i4"),
+            attention_mask=jnp.array(attention_mask, dtype="i4"),
+            position_ids=jnp.array(position_ids, dtype="i4"),
+            decoder_input_ids=jnp.array(decoder_input_ids, dtype="i4"),
+            decoder_attention_mask=jnp.array(decoder_attention_mask, dtype="i4"),
+            decoder_position_ids=jnp.array(decoder_position_ids, dtype="i4"),
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            deterministic=not train,
+            rngs=rngs,
+        )
+
+
+@add_start_docstrings(
+    "The bare BlenderbotSmall Model transformer outputting raw hidden-states without any specific head on top.",
+    BLENDERBOT_SMALL_START_DOCSTRING,
+)
+class FlaxBlenderbotSmallModel(FlaxBlenderbotSmallPreTrainedModel):
+    config: BlenderbotSmallConfig
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
+    module_class = FlaxBlenderbotSmallModule
+
+
+append_call_sample_docstring(FlaxBlenderbotSmallModel, _CHECKPOINT_FOR_DOC, FlaxSeq2SeqModelOutput, _CONFIG_FOR_DOC)
+
+
+# Copied from transformers.models.bart.modeling_flax_bart.FlaxBartForConditionalGenerationModule with Bart->BlenderbotSmall
+class FlaxBlenderbotSmallForConditionalGenerationModule(nn.Module):
+    config: BlenderbotSmallConfig
+    dtype: jnp.dtype = jnp.float32
+    bias_init: Callable[..., jnp.ndarray] = jax.nn.initializers.zeros
+
+    def setup(self):
+        self.model = FlaxBlenderbotSmallModule(config=self.config, dtype=self.dtype)
+        self.lm_head = nn.Dense(
+            self.model.shared.num_embeddings,
+            use_bias=False,
+            dtype=self.dtype,
+            kernel_init=jax.nn.initializers.normal(self.config.init_std),
+        )
+        self.final_logits_bias = self.param("final_logits_bias", self.bias_init, (1, self.model.shared.num_embeddings))
+
+    def _get_encoder_module(self):
+        return self.model.encoder
+
+    def _get_decoder_module(self):
+        return self.model.decoder
+
+    def __call__(
+        self,
+        input_ids,
+        attention_mask,
+        decoder_input_ids,
+        decoder_attention_mask,
+        position_ids,
+        decoder_position_ids,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+        deterministic: bool = True,
+    ):
+        outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            decoder_input_ids=decoder_input_ids,
+            decoder_attention_mask=decoder_attention_mask,
+            position_ids=position_ids,
+            decoder_position_ids=decoder_position_ids,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            deterministic=deterministic,
+        )
+
+        hidden_states = outputs[0]
+
+        if self.config.tie_word_embeddings:
+            shared_embedding = self.model.variables["params"]["shared"]["embedding"]
+            lm_logits = self.lm_head.apply({"params": {"kernel": shared_embedding.T}}, hidden_states)
+        else:
+            lm_logits = self.lm_head(hidden_states)
+
+        lm_logits += jax.lax.stop_gradient(self.final_logits_bias.astype(self.dtype))
+
+        if not return_dict:
+            output = (lm_logits,) + outputs[1:]
+            return output
+
+        return FlaxSeq2SeqLMOutput(
+            logits=lm_logits,
+            decoder_hidden_states=outputs.decoder_hidden_states,
+            decoder_attentions=outputs.decoder_attentions,
+            cross_attentions=outputs.cross_attentions,
+            encoder_last_hidden_state=outputs.encoder_last_hidden_state,
+            encoder_hidden_states=outputs.encoder_hidden_states,
+            encoder_attentions=outputs.encoder_attentions,
+        )
+
+
+@add_start_docstrings(
+    "The BLENDERBOT_SMALL Model with a language modeling head. Can be used for summarization.",
+    BLENDERBOT_SMALL_START_DOCSTRING,
+)
+class FlaxBlenderbotSmallForConditionalGeneration(FlaxBlenderbotSmallPreTrainedModel):
+    module_class = FlaxBlenderbotSmallForConditionalGenerationModule
+    dtype: jnp.dtype = jnp.float32
+
+    @add_start_docstrings(BLENDERBOT_SMALL_DECODE_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=FlaxCausalLMOutputWithCrossAttentions, config_class=BlenderbotSmallConfig)
+    def decode(
+        self,
+        decoder_input_ids,
+        encoder_outputs,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        decoder_attention_mask: Optional[jnp.ndarray] = None,
+        decoder_position_ids: Optional[jnp.ndarray] = None,
+        past_key_values: dict = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        deterministic: bool = True,
+        params: dict = None,
+        dropout_rng: PRNGKey = None,
+    ):
+        r"""
+        Returns:
+
+        Example:
+
+        ```python
+        >>> import jax.numpy as jnp
+        >>> from transformers import AutoTokenizer, FlaxBlenderbotSmallForConditionalGeneration
+
+        >>> model = FlaxBlenderbotSmallForConditionalGeneration.from_pretrained("facebook/blenderbot_small-90M")
+        >>> tokenizer = AutoTokenizer.from_pretrained("facebook/blenderbot_small-90M")
+
+        >>> text = "My friends are cool but they eat too many carbs."
+        >>> inputs = tokenizer(text, max_length=1024, return_tensors="np")
+        >>> encoder_outputs = model.encode(**inputs)
+
+        >>> decoder_start_token_id = model.config.decoder_start_token_id
+        >>> decoder_input_ids = jnp.ones((inputs.input_ids.shape[0], 1), dtype="i4") * decoder_start_token_id
+
+        >>> outputs = model.decode(decoder_input_ids, encoder_outputs)
+        >>> logits = outputs.logits
+        ```"""
+        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
+
+        encoder_hidden_states = encoder_outputs[0]
+        if encoder_attention_mask is None:
+            batch_size, sequence_length = encoder_hidden_states.shape[:2]
+            encoder_attention_mask = jnp.ones((batch_size, sequence_length))
+
+        batch_size, sequence_length = decoder_input_ids.shape
+        if decoder_attention_mask is None:
+            decoder_attention_mask = jnp.ones((batch_size, sequence_length))
+
+        if decoder_position_ids is None:
+            if past_key_values is not None:
+                raise ValueError("Make sure to provide `decoder_position_ids` when passing `past_key_values`.")
+
+            decoder_position_ids = jnp.broadcast_to(
+                jnp.arange(sequence_length)[None, :], (batch_size, sequence_length)
+            )
+
+        # Handle any PRNG if needed
+        rngs = {}
+        if dropout_rng is not None:
+            rngs["dropout"] = dropout_rng
+
+        inputs = {"params": params or self.params}
+
+        # if past_key_values are passed then cache is already initialized a private flag init_cache has to be
+        # passed down to ensure cache is used. It has to be made sure that cache is marked as mutable so that
+        # it can be changed by FlaxBlenderbotSmallAttention module
+        if past_key_values:
+            inputs["cache"] = past_key_values
+            mutable = ["cache"]
+        else:
+            mutable = False
+
+        def _decoder_forward(module, decoder_input_ids, decoder_attention_mask, decoder_position_ids, **kwargs):
+            decoder_module = module._get_decoder_module()
+            outputs = decoder_module(
+                decoder_input_ids,
+                decoder_attention_mask,
+                decoder_position_ids,
+                **kwargs,
+            )
+            hidden_states = outputs[0]
+
+            if self.config.tie_word_embeddings:
+                shared_embedding = module.model.variables["params"]["shared"]["embedding"]
+                lm_logits = module.lm_head.apply({"params": {"kernel": shared_embedding.T}}, hidden_states)
+            else:
+                lm_logits = module.lm_head(hidden_states)
+
+            lm_logits += module.final_logits_bias.astype(self.dtype)
+            return lm_logits, outputs
+
+        outputs = self.module.apply(
+            inputs,
+            decoder_input_ids=jnp.array(decoder_input_ids, dtype="i4"),
+            decoder_attention_mask=jnp.array(decoder_attention_mask, dtype="i4"),
+            decoder_position_ids=jnp.array(decoder_position_ids, dtype="i4"),
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=jnp.array(encoder_attention_mask, dtype="i4"),
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            deterministic=deterministic,
+            rngs=rngs,
+            mutable=mutable,
+            method=_decoder_forward,
+        )
+
+        if past_key_values is None:
+            lm_logits, decoder_outputs = outputs
+        else:
+            (lm_logits, decoder_outputs), past = outputs
+
+        if return_dict:
+            outputs = FlaxCausalLMOutputWithCrossAttentions(
+                logits=lm_logits,
+                hidden_states=decoder_outputs.hidden_states,
+                attentions=decoder_outputs.attentions,
+                cross_attentions=decoder_outputs.cross_attentions,
+            )
+        else:
+            outputs = (lm_logits,) + decoder_outputs[1:]
+
+        # add updated cache to model output
+        if past_key_values is not None and return_dict:
+            outputs["past_key_values"] = unfreeze(past["cache"])
+            return outputs
+        elif past_key_values is not None and not return_dict:
+            outputs = outputs[:1] + (unfreeze(past["cache"]),) + outputs[1:]
+
+        return outputs
+
+    def prepare_inputs_for_generation(
+        self,
+        decoder_input_ids,
+        max_length,
+        attention_mask: Optional[jax.Array] = None,
+        decoder_attention_mask: Optional[jax.Array] = None,
+        encoder_outputs=None,
+        **kwargs,
+    ):
+        # initializing the cache
+        batch_size, seq_length = decoder_input_ids.shape
+
+        past_key_values = self.init_cache(batch_size, max_length, encoder_outputs)
+        # Note that usually one would have to put 0's in the attention_mask for x > input_ids.shape[-1] and x < cache_length.
+        # But since the decoder uses a causal mask, those positions are masked anyways.
+        # Thus we can create a single static attention_mask here, which is more efficient for compilation
+        extended_attention_mask = jnp.ones((batch_size, max_length), dtype="i4")
+        if decoder_attention_mask is not None:
+            position_ids = decoder_attention_mask.cumsum(axis=-1) - 1
+            extended_attention_mask = lax.dynamic_update_slice(extended_attention_mask, decoder_attention_mask, (0, 0))
+        else:
+            position_ids = jnp.broadcast_to(jnp.arange(seq_length, dtype="i4")[None, :], (batch_size, seq_length))
+
+        return {
+            "past_key_values": past_key_values,
+            "encoder_outputs": encoder_outputs,
+            "encoder_attention_mask": attention_mask,
+            "decoder_attention_mask": extended_attention_mask,
+            "decoder_position_ids": position_ids,
+        }
+
+    def update_inputs_for_generation(self, model_outputs, model_kwargs):
+        model_kwargs["past_key_values"] = model_outputs.past_key_values
+        model_kwargs["decoder_position_ids"] = model_kwargs["decoder_position_ids"][:, -1:] + 1
+        return model_kwargs
+
+
+FLAX_BLENDERBOT_SMALL_CONDITIONAL_GENERATION_DOCSTRING = """
+    Returns:
+
+    Summarization example:
+
+    ```py
+    >>> from transformers import AutoTokenizer, FlaxBlenderbotSmallForConditionalGeneration
+
+    >>> model = FlaxBlenderbotSmallForConditionalGeneration.from_pretrained("facebook/blenderbot_small-90M")
+    >>> tokenizer = AutoTokenizer.from_pretrained("facebook/blenderbot_small-90M")
+
+    >>> ARTICLE_TO_SUMMARIZE = "My friends are cool but they eat too many carbs."
+    >>> inputs = tokenizer([ARTICLE_TO_SUMMARIZE], max_length=1024, return_tensors="np")
+
+    >>> # Generate Summary
+    >>> summary_ids = model.generate(inputs["input_ids"]).sequences
+    >>> print(tokenizer.batch_decode(summary_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False))
+    ```
+
+    Mask filling example:
+
+    ```py
+    >>> from transformers import AutoTokenizer, FlaxBlenderbotSmallForConditionalGeneration
+
+    >>> tokenizer = AutoTokenizer.from_pretrained("facebook/blenderbot_small-90M")
+    >>> TXT = "My friends are <mask> but they eat too many carbs."
+
+    >>> model = FlaxBlenderbotSmallForConditionalGeneration.from_pretrained("facebook/blenderbot_small-90M")
+    >>> input_ids = tokenizer([TXT], return_tensors="np")["input_ids"]
+    >>> logits = model(input_ids).logits
+
+    >>> masked_index = (input_ids[0] == tokenizer.mask_token_id).nonzero().item()
+    >>> probs = jax.nn.softmax(logits[0, masked_index], axis=0)
+    >>> values, predictions = jax.lax.top_k(probs)
+
+    >>> tokenizer.decode(predictions).split()
+    ```
+"""
+
+overwrite_call_docstring(
+    FlaxBlenderbotSmallForConditionalGeneration,
+    BLENDERBOT_SMALL_INPUTS_DOCSTRING + FLAX_BLENDERBOT_SMALL_CONDITIONAL_GENERATION_DOCSTRING,
+)
+append_replace_return_docstrings(
+    FlaxBlenderbotSmallForConditionalGeneration, output_type=FlaxSeq2SeqLMOutput, config_class=_CONFIG_FOR_DOC
+)
+
+
+__all__ = [
+    "FlaxBlenderbotSmallForConditionalGeneration",
+    "FlaxBlenderbotSmallModel",
+    "FlaxBlenderbotSmallPreTrainedModel",
+]

lib/python3.12/site-packages/transformers/models/blenderbot_small/modeling_tf_blenderbot_small.py

@@ -0,0 +1,1528 @@
+# coding=utf-8
+# Copyright 2021 The Facebook, Inc 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.
+"""TF 2.0 BlenderbotSmall model."""
+
+from __future__ import annotations
+
+import random
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import tensorflow as tf
+
+from ...activations_tf import get_tf_activation
+from ...modeling_tf_outputs import (
+    TFBaseModelOutput,
+    TFBaseModelOutputWithPastAndCrossAttentions,
+    TFSeq2SeqLMOutput,
+    TFSeq2SeqModelOutput,
+)
+
+# Public API
+from ...modeling_tf_utils import (
+    TFCausalLanguageModelingLoss,
+    TFPreTrainedModel,
+    keras,
+    keras_serializable,
+    unpack_inputs,
+)
+from ...tf_utils import check_embeddings_within_bounds, shape_list, stable_softmax
+from ...utils import (
+    add_code_sample_docstrings,
+    add_end_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    logging,
+    replace_return_docstrings,
+)
+from .configuration_blenderbot_small import BlenderbotSmallConfig
+
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "facebook/blenderbot_small-90M"
+_CONFIG_FOR_DOC = "BlenderbotSmallConfig"
+
+
+LARGE_NEGATIVE = -1e8
+
+
+# Copied from transformers.models.bart.modeling_tf_bart.shift_tokens_right
+def shift_tokens_right(input_ids: tf.Tensor, pad_token_id: int, decoder_start_token_id: int):
+    pad_token_id = tf.cast(pad_token_id, input_ids.dtype)
+    decoder_start_token_id = tf.cast(decoder_start_token_id, input_ids.dtype)
+    start_tokens = tf.fill(
+        (shape_list(input_ids)[0], 1), tf.convert_to_tensor(decoder_start_token_id, input_ids.dtype)
+    )
+    shifted_input_ids = tf.concat([start_tokens, input_ids[:, :-1]], -1)
+    # replace possible -100 values in labels by `pad_token_id`
+    shifted_input_ids = tf.where(
+        shifted_input_ids == -100,
+        tf.fill(shape_list(shifted_input_ids), tf.convert_to_tensor(pad_token_id, input_ids.dtype)),
+        shifted_input_ids,
+    )
+
+    # "Verify that `labels` has only positive values and -100"
+    assert_gte0 = tf.debugging.assert_greater_equal(shifted_input_ids, tf.constant(0, dtype=input_ids.dtype))
+
+    # Make sure the assertion op is called by wrapping the result in an identity no-op
+    with tf.control_dependencies([assert_gte0]):
+        shifted_input_ids = tf.identity(shifted_input_ids)
+
+    return shifted_input_ids
+
+
+# Copied from transformers.models.bart.modeling_tf_bart._make_causal_mask
+def _make_causal_mask(input_ids_shape: tf.TensorShape, past_key_values_length: int = 0):
+    """
+    Make causal mask used for bi-directional self-attention.
+    """
+    bsz = input_ids_shape[0]
+    tgt_len = input_ids_shape[1]
+    mask = tf.ones((tgt_len, tgt_len)) * LARGE_NEGATIVE
+    mask_cond = tf.range(shape_list(mask)[-1])
+
+    mask = tf.where(mask_cond < tf.reshape(mask_cond + 1, (shape_list(mask)[-1], 1)), 0.0, mask)
+
+    if past_key_values_length > 0:
+        mask = tf.concat([tf.zeros((tgt_len, past_key_values_length)), mask], axis=-1)
+
+    return tf.tile(mask[None, None, :, :], (bsz, 1, 1, 1))
+
+
+# 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
+
+
+# Copied from transformers.models.blenderbot.modeling_tf_blenderbot.TFBlenderbotLearnedPositionalEmbedding with Blenderbot->BlenderbotSmall
+class TFBlenderbotSmallLearnedPositionalEmbedding(keras.layers.Embedding):
+    """
+    This module learns positional embeddings up to a fixed maximum size.
+    """
+
+    def __init__(self, num_embeddings: int, embedding_dim: int, **kwargs):
+        super().__init__(num_embeddings, embedding_dim, **kwargs)
+
+    def call(
+        self, input_shape: tf.TensorShape, past_key_values_length: int = 0, position_ids: tf.Tensor | None = None
+    ):
+        """Input is expected to be of size [bsz x seqlen]."""
+        if position_ids is None:
+            seq_len = input_shape[1]
+            position_ids = tf.range(seq_len, delta=1, name="range")
+            position_ids += past_key_values_length
+
+        return super().call(tf.cast(position_ids, dtype=tf.int32))
+
+
+# Copied from transformers.models.bart.modeling_tf_bart.TFBartAttention with Bart->BlenderbotSmall
+class TFBlenderbotSmallAttention(keras.layers.Layer):
+    """Multi-headed attention from "Attention Is All You Need"""
+
+    def __init__(
+        self,
+        embed_dim: int,
+        num_heads: int,
+        dropout: float = 0.0,
+        is_decoder: bool = False,
+        bias: bool = True,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.embed_dim = embed_dim
+
+        self.num_heads = num_heads
+        self.dropout = keras.layers.Dropout(dropout)
+        self.head_dim = embed_dim // num_heads
+        if (self.head_dim * num_heads) != self.embed_dim:
+            raise ValueError(
+                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim}"
+                f" and `num_heads`: {num_heads})."
+            )
+        self.scaling = self.head_dim**-0.5
+        self.is_decoder = is_decoder
+
+        self.k_proj = keras.layers.Dense(embed_dim, use_bias=bias, name="k_proj")
+        self.q_proj = keras.layers.Dense(embed_dim, use_bias=bias, name="q_proj")
+        self.v_proj = keras.layers.Dense(embed_dim, use_bias=bias, name="v_proj")
+        self.out_proj = keras.layers.Dense(embed_dim, use_bias=bias, name="out_proj")
+
+    def _shape(self, tensor: tf.Tensor, seq_len: int, bsz: int):
+        return tf.transpose(tf.reshape(tensor, (bsz, seq_len, self.num_heads, self.head_dim)), (0, 2, 1, 3))
+
+    def call(
+        self,
+        hidden_states: tf.Tensor,
+        key_value_states: tf.Tensor | None = None,
+        past_key_value: Tuple[Tuple[tf.Tensor]] | None = None,
+        attention_mask: tf.Tensor | None = None,
+        layer_head_mask: tf.Tensor | None = None,
+        training: Optional[bool] = False,
+    ) -> Tuple[tf.Tensor, tf.Tensor | None]:
+        """Input shape: Batch x Time x Channel"""
+
+        # if key_value_states are provided this layer is used as a cross-attention layer
+        # for the decoder
+        is_cross_attention = key_value_states is not None
+        bsz, tgt_len, embed_dim = shape_list(hidden_states)
+
+        # get query proj
+        query_states = self.q_proj(hidden_states) * self.scaling
+        # get key, value proj
+        if is_cross_attention and past_key_value is not None:
+            # reuse k,v, cross_attentions
+            key_states = past_key_value[0]
+            value_states = past_key_value[1]
+        elif is_cross_attention:
+            # cross_attentions
+            key_states = self._shape(self.k_proj(key_value_states), -1, bsz)
+            value_states = self._shape(self.v_proj(key_value_states), -1, bsz)
+        elif past_key_value is not None:
+            # reuse k, v, self_attention
+            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
+            key_states = tf.concat([past_key_value[0], key_states], axis=2)
+            value_states = tf.concat([past_key_value[1], value_states], axis=2)
+        else:
+            # self_attention
+            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
+
+        if self.is_decoder:
+            # if cross_attention save Tuple(tf.Tensor, tf.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(tf.Tensor, tf.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_states, value_states)
+
+        proj_shape = (bsz * self.num_heads, -1, self.head_dim)
+        query_states = tf.reshape(self._shape(query_states, tgt_len, bsz), proj_shape)
+        key_states = tf.reshape(key_states, proj_shape)
+        value_states = tf.reshape(value_states, proj_shape)
+
+        src_len = shape_list(key_states)[1]
+        attn_weights = tf.matmul(query_states, key_states, transpose_b=True)
+
+        tf.debugging.assert_equal(
+            shape_list(attn_weights),
+            [bsz * self.num_heads, tgt_len, src_len],
+            message=(
+                f"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is"
+                f" {shape_list(attn_weights)}"
+            ),
+        )
+
+        if attention_mask is not None:
+            tf.debugging.assert_equal(
+                shape_list(attention_mask),
+                [bsz, 1, tgt_len, src_len],
+                message=(
+                    f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is"
+                    f" {shape_list(attention_mask)}"
+                ),
+            )
+
+            attention_mask = tf.cast(attention_mask, dtype=attn_weights.dtype)
+            attn_weights = tf.reshape(attn_weights, (bsz, self.num_heads, tgt_len, src_len)) + attention_mask
+            attn_weights = tf.reshape(attn_weights, (bsz * self.num_heads, tgt_len, src_len))
+
+        attn_weights = stable_softmax(attn_weights, axis=-1)
+
+        if layer_head_mask is not None:
+            tf.debugging.assert_equal(
+                shape_list(layer_head_mask),
+                [self.num_heads],
+                message=(
+                    f"Head mask for a single layer should be of size {(self.num_heads)}, but is"
+                    f" {shape_list(layer_head_mask)}"
+                ),
+            )
+
+            attn_weights = tf.reshape(layer_head_mask, (1, -1, 1, 1)) * tf.reshape(
+                attn_weights, (bsz, self.num_heads, tgt_len, src_len)
+            )
+            attn_weights = tf.reshape(attn_weights, (bsz * self.num_heads, tgt_len, src_len))
+
+        attn_probs = self.dropout(attn_weights, training=training)
+        attn_output = tf.matmul(attn_probs, value_states)
+
+        tf.debugging.assert_equal(
+            shape_list(attn_output),
+            [bsz * self.num_heads, tgt_len, self.head_dim],
+            message=(
+                f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
+                f" {shape_list(attn_output)}"
+            ),
+        )
+
+        attn_output = tf.transpose(
+            tf.reshape(attn_output, (bsz, self.num_heads, tgt_len, self.head_dim)), (0, 2, 1, 3)
+        )
+        attn_output = tf.reshape(attn_output, (bsz, tgt_len, embed_dim))
+
+        attn_output = self.out_proj(attn_output)
+        attn_weights: tf.Tensor = tf.reshape(attn_weights, (bsz, self.num_heads, tgt_len, src_len))
+
+        return attn_output, attn_weights, past_key_value
+
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        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, "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, "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])
+
+
+# Copied from transformers.models.bart.modeling_tf_bart.TFBartEncoderLayer with Bart->BlenderbotSmall
+class TFBlenderbotSmallEncoderLayer(keras.layers.Layer):
+    def __init__(self, config: BlenderbotSmallConfig, **kwargs):
+        super().__init__(**kwargs)
+        self.embed_dim = config.d_model
+        self.self_attn = TFBlenderbotSmallAttention(
+            self.embed_dim, config.encoder_attention_heads, dropout=config.attention_dropout, name="self_attn"
+        )
+        self.self_attn_layer_norm = keras.layers.LayerNormalization(epsilon=1e-5, name="self_attn_layer_norm")
+        self.dropout = keras.layers.Dropout(config.dropout)
+        self.activation_fn = get_tf_activation(config.activation_function)
+        self.activation_dropout = keras.layers.Dropout(config.activation_dropout)
+        self.fc1 = keras.layers.Dense(config.encoder_ffn_dim, name="fc1")
+        self.fc2 = keras.layers.Dense(self.embed_dim, name="fc2")
+        self.final_layer_norm = keras.layers.LayerNormalization(epsilon=1e-5, name="final_layer_norm")
+        self.config = config
+
+    def call(
+        self,
+        hidden_states: tf.Tensor,
+        attention_mask: np.ndarray | tf.Tensor | None,
+        layer_head_mask: tf.Tensor | None,
+        training: Optional[bool] = False,
+    ) -> 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.
+            layer_head_mask (`tf.Tensor`): mask for attention heads in a given layer of size
+                `(encoder_attention_heads,)`
+        """
+        residual = hidden_states
+        hidden_states, self_attn_weights, _ = self.self_attn(
+            hidden_states=hidden_states, attention_mask=attention_mask, layer_head_mask=layer_head_mask
+        )
+
+        tf.debugging.assert_equal(
+            shape_list(hidden_states),
+            shape_list(residual),
+            message=f"Self attn modified the shape of query {shape_list(residual)} to {shape_list(hidden_states)}",
+        )
+
+        hidden_states = self.dropout(hidden_states, training=training)
+        hidden_states = residual + hidden_states
+        hidden_states = self.self_attn_layer_norm(hidden_states)
+
+        residual = hidden_states
+        hidden_states = self.activation_fn(self.fc1(hidden_states))
+        hidden_states = self.activation_dropout(hidden_states, training=training)
+        hidden_states = self.fc2(hidden_states)
+        hidden_states = self.dropout(hidden_states, training=training)
+        hidden_states = residual + hidden_states
+        hidden_states = self.final_layer_norm(hidden_states)
+
+        return hidden_states, self_attn_weights
+
+    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, "self_attn_layer_norm", None) is not None:
+            with tf.name_scope(self.self_attn_layer_norm.name):
+                self.self_attn_layer_norm.build([None, None, self.embed_dim])
+        if getattr(self, "fc1", None) is not None:
+            with tf.name_scope(self.fc1.name):
+                self.fc1.build([None, None, self.embed_dim])
+        if getattr(self, "fc2", None) is not None:
+            with tf.name_scope(self.fc2.name):
+                self.fc2.build([None, None, self.config.encoder_ffn_dim])
+        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])
+
+
+# Copied from transformers.models.bart.modeling_tf_bart.TFBartDecoderLayer with Bart->BlenderbotSmall
+class TFBlenderbotSmallDecoderLayer(keras.layers.Layer):
+    def __init__(self, config: BlenderbotSmallConfig, **kwargs):
+        super().__init__(**kwargs)
+        self.embed_dim = config.d_model
+        self.self_attn = TFBlenderbotSmallAttention(
+            embed_dim=self.embed_dim,
+            num_heads=config.decoder_attention_heads,
+            dropout=config.attention_dropout,
+            name="self_attn",
+            is_decoder=True,
+        )
+        self.dropout = keras.layers.Dropout(config.dropout)
+        self.activation_fn = get_tf_activation(config.activation_function)
+        self.activation_dropout = keras.layers.Dropout(config.activation_dropout)
+
+        self.self_attn_layer_norm = keras.layers.LayerNormalization(epsilon=1e-5, name="self_attn_layer_norm")
+        self.encoder_attn = TFBlenderbotSmallAttention(
+            self.embed_dim,
+            config.decoder_attention_heads,
+            dropout=config.attention_dropout,
+            name="encoder_attn",
+            is_decoder=True,
+        )
+        self.encoder_attn_layer_norm = keras.layers.LayerNormalization(epsilon=1e-5, name="encoder_attn_layer_norm")
+        self.fc1 = keras.layers.Dense(config.decoder_ffn_dim, name="fc1")
+        self.fc2 = keras.layers.Dense(self.embed_dim, name="fc2")
+        self.final_layer_norm = keras.layers.LayerNormalization(epsilon=1e-5, name="final_layer_norm")
+        self.config = config
+
+    def call(
+        self,
+        hidden_states: tf.Tensor,
+        attention_mask: np.ndarray | tf.Tensor | None = None,
+        encoder_hidden_states: np.ndarray | tf.Tensor | None = None,
+        encoder_attention_mask: np.ndarray | tf.Tensor | None = None,
+        layer_head_mask: tf.Tensor | None = None,
+        cross_attn_layer_head_mask: tf.Tensor | None = None,
+        past_key_value: Optional[Tuple[Tuple[Union[np.ndarray, tf.Tensor]]]] = None,
+        training: Optional[bool] = False,
+    ) -> Tuple[tf.Tensor, tf.Tensor, Tuple[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.
+            encoder_hidden_states (`tf.Tensor`):
+                cross attention input to the layer of shape `(batch, seq_len, embed_dim)`
+            encoder_attention_mask (`tf.Tensor`): encoder attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            layer_head_mask (`tf.Tensor`): mask for attention heads in a given layer of size
+                `(decoder_attention_heads,)`
+            cross_attn_layer_head_mask (`tf.Tensor`): mask for heads of the cross-attention module.
+                `(decoder_attention_heads,)`
+            past_key_value (`Tuple(tf.Tensor)`): cached past key and value projection states
+        """
+        residual = hidden_states
+
+        # Self Attention
+        # 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
+        # add present self-attn cache to positions 1,2 of present_key_value tuple
+        hidden_states, self_attn_weights, present_key_value = self.self_attn(
+            hidden_states=hidden_states,
+            past_key_value=self_attn_past_key_value,
+            attention_mask=attention_mask,
+            layer_head_mask=layer_head_mask,
+        )
+        hidden_states = self.dropout(hidden_states, training=training)
+        hidden_states = residual + hidden_states
+        hidden_states = self.self_attn_layer_norm(hidden_states)
+
+        # Cross-Attention Block
+        cross_attn_present_key_value = None
+        cross_attn_weights = None
+        if encoder_hidden_states is not None:
+            residual = hidden_states
+
+            # cross_attn cached key/values tuple is at positions 3,4 of present_key_value tuple
+            cross_attn_past_key_value = past_key_value[-2:] if past_key_value is not None else None
+            hidden_states, cross_attn_weights, cross_attn_present_key_value = self.encoder_attn(
+                hidden_states=hidden_states,
+                key_value_states=encoder_hidden_states,
+                attention_mask=encoder_attention_mask,
+                layer_head_mask=cross_attn_layer_head_mask,
+                past_key_value=cross_attn_past_key_value,
+            )
+            hidden_states = self.dropout(hidden_states, training=training)
+            hidden_states = residual + hidden_states
+            hidden_states = self.encoder_attn_layer_norm(hidden_states)
+
+            # add cross-attn to positions 3,4 of present_key_value tuple
+            present_key_value = present_key_value + cross_attn_present_key_value
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.activation_fn(self.fc1(hidden_states))
+        hidden_states = self.activation_dropout(hidden_states, training=training)
+        hidden_states = self.fc2(hidden_states)
+        hidden_states = self.dropout(hidden_states, training=training)
+        hidden_states = residual + hidden_states
+        hidden_states = self.final_layer_norm(hidden_states)
+
+        return (
+            hidden_states,
+            self_attn_weights,
+            cross_attn_weights,
+            present_key_value,
+        )
+
+    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, "self_attn_layer_norm", None) is not None:
+            with tf.name_scope(self.self_attn_layer_norm.name):
+                self.self_attn_layer_norm.build([None, None, self.embed_dim])
+        if getattr(self, "encoder_attn", None) is not None:
+            with tf.name_scope(self.encoder_attn.name):
+                self.encoder_attn.build(None)
+        if getattr(self, "encoder_attn_layer_norm", None) is not None:
+            with tf.name_scope(self.encoder_attn_layer_norm.name):
+                self.encoder_attn_layer_norm.build([None, None, self.embed_dim])
+        if getattr(self, "fc1", None) is not None:
+            with tf.name_scope(self.fc1.name):
+                self.fc1.build([None, None, self.embed_dim])
+        if getattr(self, "fc2", None) is not None:
+            with tf.name_scope(self.fc2.name):
+                self.fc2.build([None, None, self.config.decoder_ffn_dim])
+        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])
+
+
+class TFBlenderbotSmallPreTrainedModel(TFPreTrainedModel):
+    config_class = BlenderbotSmallConfig
+    base_model_prefix = "model"
+
+
+BLENDERBOT_SMALL_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 ([`BlenderbotSmallConfig`]): 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.
+"""
+
+BLENDERBOT_SMALL_GENERATION_EXAMPLE = r"""
+    Conversation example::
+
+    ```py
+    >>> from transformers import AutoTokenizer, TFBlenderbotSmallForConditionalGeneration
+
+    >>> mname = "facebook/blenderbot_small-90M"
+    >>> model = BlenderbotSmallForConditionalGeneration.from_pretrained(mname)
+    >>> tokenizer = AutoTokenizer.from_pretrained(mname)
+
+    >>> UTTERANCE = "My friends are cool but they eat too many carbs."
+    >>> print("Human: ", UTTERANCE)
+    >>> inputs = tokenizer([UTTERANCE], return_tensors="tf")
+
+    >>> reply_ids = model.generate(**inputs)
+    >>> print("Bot: ", tokenizer.batch_decode(reply_ids, skip_special_tokens=True)[0])
+    what kind of carbs do they eat? i don't know much about carbs.
+
+    >>> REPLY = "I'm not sure"
+    >>> print("Human: ", REPLY)
+    >>> NEXT_UTTERANCE = (
+    ...     "My friends are cool but they eat too many carbs.</s> "
+    ...     "<s>what kind of carbs do they eat? i don't know much about carbs.</s> "
+    ...     "<s>I'm not sure."
+    ... )
+
+    >>> inputs = tokenizer([NEXT_UTTERANCE], return_tensors="tf")
+    >>> inputs.pop("token_type_ids")
+    >>> next_reply_ids = model.generate(**inputs)
+    >>> print("Bot: ", tokenizer.batch_decode(next_reply_ids, skip_special_tokens=True)[0])
+    ```
+"""
+
+BLENDERBOT_SMALL_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`tf.Tensor` of shape `({0})`):
+            Indices of input sequence tokens in the vocabulary.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`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)
+        decoder_input_ids (`tf.Tensor` of shape `(batch_size, target_sequence_length)`, *optional*):
+            Indices of decoder input sequence tokens in the vocabulary.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are decoder input IDs?](../glossary#decoder-input-ids)
+
+            BlenderbotSmall uses the `bos_token_id` as the starting token for `decoder_input_ids` generation. If
+            `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
+            `past_key_values`).
+        decoder_attention_mask (`tf.Tensor` of shape `(batch_size, target_sequence_length)`, *optional*):
+            will be made by default and ignore pad tokens. It is not recommended to set this for most use cases.
+        decoder_position_ids (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each decoder input sequence tokens in the position embeddings. Selected in the
+            range `[0, config.max_position_embeddings - 1]`.
+        head_mask (`tf.Tensor` of shape `(encoder_layers, encoder_attention_heads)`, *optional*):
+            Mask to nullify selected heads of the attention modules in the encoder. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        decoder_head_mask (`tf.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
+            Mask to nullify selected heads of the attention modules in the decoder. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        cross_attn_head_mask (`tf.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
+            Mask to nullify selected heads of the cross-attention modules. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        encoder_outputs (`tf.FloatTensor`, *optional*):
+            hidden states at the output of the last layer of the encoder. Used in the cross-attention of the decoder.
+            of shape `(batch_size, sequence_length, hidden_size)` is a sequence of
+        past_key_values (`Tuple[Tuple[tf.Tensor]]` of length `config.n_layers`)
+            contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
+            If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
+            don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
+            `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+            `past_key_values`). Set to `False` during training, `True` during generation
+        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).
+"""
+
+
+@keras_serializable
+class TFBlenderbotSmallEncoder(keras.layers.Layer):
+    config_class = BlenderbotSmallConfig
+    """
+    Transformer encoder consisting of *config.encoder_layers* self attention layers. Each layer is a
+    [`TFBlenderbotSmallEncoderLayer`].
+
+    Args:
+        config: BlenderbotSmallConfig
+    """
+
+    def __init__(self, config: BlenderbotSmallConfig, embed_tokens: Optional[keras.layers.Embedding] = None, **kwargs):
+        super().__init__(**kwargs)
+        self.config = config
+        self.dropout = keras.layers.Dropout(config.dropout)
+        self.layerdrop = config.encoder_layerdrop
+        self.padding_idx = config.pad_token_id
+        self.max_source_positions = config.max_position_embeddings
+        self.embed_scale = tf.math.sqrt(float(config.d_model)) if config.scale_embedding else 1.0
+
+        self.embed_tokens = embed_tokens
+        self.embed_positions = TFBlenderbotSmallLearnedPositionalEmbedding(
+            config.max_position_embeddings,
+            config.d_model,
+            name="embed_positions",
+        )
+        self.layers = [TFBlenderbotSmallEncoderLayer(config, name=f"layers.{i}") for i in range(config.encoder_layers)]
+        self.layernorm_embedding = keras.layers.LayerNormalization(epsilon=1e-5, name="layernorm_embedding")
+        self.embed_dim = config.d_model
+
+    def get_embed_tokens(self):
+        return self.embed_tokens
+
+    def set_embed_tokens(self, embed_tokens):
+        self.embed_tokens = embed_tokens
+
+    @unpack_inputs
+    def call(
+        self,
+        input_ids=None,
+        inputs_embeds=None,
+        attention_mask=None,
+        head_mask=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+        training=False,
+    ):
+        """
+        Args:
+            input_ids (`tf.Tensor` 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 (`tf.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)
+            head_mask (`tf.Tensor` of shape `(encoder_layers, encoder_attention_heads)`, `optional):
+                Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:
+
+                - 1 indicates the head is **not masked**,
+                - 0 indicates the head is **masked**.
+
+            inputs_embeds (`tf.Tensor` 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.
+            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).
+        """
+        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 = shape_list(input_ids)
+        elif inputs_embeds is not None:
+            input_shape = shape_list(inputs_embeds)[:-1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        if inputs_embeds is None:
+            check_embeddings_within_bounds(input_ids, self.embed_tokens.input_dim)
+            inputs_embeds = self.embed_tokens(input_ids) * self.embed_scale
+
+        embed_pos = self.embed_positions(input_shape)
+        hidden_states = inputs_embeds + embed_pos
+        hidden_states = self.layernorm_embedding(hidden_states)
+        hidden_states = self.dropout(hidden_states, training=training)
+
+        # check attention mask and invert
+        if attention_mask is not None:
+            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+            attention_mask = _expand_mask(attention_mask)
+        else:
+            attention_mask = None
+
+        encoder_states = () if output_hidden_states else None
+        all_attentions = () if output_attentions else None
+
+        # check if head_mask has a correct number of layers specified if desired
+        if head_mask is not None:
+            tf.debugging.assert_equal(
+                shape_list(head_mask)[0],
+                len(self.layers),
+                message=(
+                    f"The head_mask should be specified for {len(self.layers)} layers, but it is for"
+                    f" {shape_list(head_mask)[0]}."
+                ),
+            )
+
+        # encoder layers
+        for idx, encoder_layer in enumerate(self.layers):
+            if output_hidden_states:
+                encoder_states = encoder_states + (hidden_states,)
+            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
+            dropout_probability = random.uniform(0, 1)
+            if training and (dropout_probability < self.layerdrop):  # skip the layer
+                continue
+
+            hidden_states, attn = encoder_layer(
+                hidden_states,
+                attention_mask,
+                head_mask[idx] if head_mask is not None else None,
+            )
+
+            if output_attentions:
+                all_attentions += (attn,)
+
+        if output_hidden_states:
+            encoder_states = encoder_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None)
+        return TFBaseModelOutput(
+            last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions
+        )
+
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "embed_positions", None) is not None:
+            with tf.name_scope(self.embed_positions.name):
+                self.embed_positions.build(None)
+        if getattr(self, "layernorm_embedding", None) is not None:
+            with tf.name_scope(self.layernorm_embedding.name):
+                self.layernorm_embedding.build([None, None, self.embed_dim])
+        if getattr(self, "layers", None) is not None:
+            for layer in self.layers:
+                with tf.name_scope(layer.name):
+                    layer.build(None)
+
+
+@keras_serializable
+class TFBlenderbotSmallDecoder(keras.layers.Layer):
+    config_class = BlenderbotSmallConfig
+    """
+    Transformer decoder consisting of *config.decoder_layers* layers. Each layer is a [`TFBlenderbotSmallDecoderLayer`]
+
+    Args:
+        config: BlenderbotSmallConfig
+        embed_tokens: output embedding
+    """
+
+    def __init__(self, config: BlenderbotSmallConfig, embed_tokens: Optional[keras.layers.Embedding] = None, **kwargs):
+        super().__init__(**kwargs)
+        self.config = config
+        self.padding_idx = config.pad_token_id
+        self.embed_tokens = embed_tokens
+        self.layerdrop = config.decoder_layerdrop
+        self.embed_positions = TFBlenderbotSmallLearnedPositionalEmbedding(
+            config.max_position_embeddings,
+            config.d_model,
+            name="embed_positions",
+        )
+        self.embed_scale = tf.math.sqrt(float(config.d_model)) if config.scale_embedding else 1.0
+        self.layers = [TFBlenderbotSmallDecoderLayer(config, name=f"layers.{i}") for i in range(config.decoder_layers)]
+        self.layernorm_embedding = keras.layers.LayerNormalization(epsilon=1e-5, name="layernorm_embedding")
+
+        self.dropout = keras.layers.Dropout(config.dropout)
+
+    def get_embed_tokens(self):
+        return self.embed_tokens
+
+    def set_embed_tokens(self, embed_tokens):
+        self.embed_tokens = embed_tokens
+
+    @unpack_inputs
+    def call(
+        self,
+        input_ids=None,
+        inputs_embeds=None,
+        attention_mask=None,
+        position_ids=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        head_mask=None,
+        cross_attn_head_mask=None,
+        past_key_values=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+        training=False,
+    ):
+        r"""
+        Args:
+            input_ids (`tf.Tensor` 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 (`tf.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 (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Indices of positions of each decoder input sequence tokens in the position embeddings. Selected in the
+                range `[0, config.max_position_embeddings - 1]`.
+            encoder_hidden_states (`tf.Tensor` of shape `(batch_size, encoder_sequence_length, hidden_size)`, *optional*):
+                Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention
+                of the decoder.
+            encoder_attention_mask (`tf.Tensor` of shape `(batch_size, encoder_sequence_length)`, *optional*):
+                Mask to avoid performing cross-attention on padding tokens indices of encoder input_ids. 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)
+            head_mask (`tf.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
+                Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:
+
+                - 1 indicates the head is **not masked**,
+                - 0 indicates the head is **masked**.
+
+            cross_attn_head_mask (`tf.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
+                Mask to nullify selected heads of the cross-attention modules. Mask values selected in `[0, 1]`:
+
+                - 1 indicates the head is **not masked**,
+                - 0 indicates the head is **masked**.
+
+            past_key_values (`Tuple[Tuple[tf.Tensor]]` of length `config.n_layers` with each tuple having 2 tuples each of which has 2 tensors of shape `(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
+                Contains precomputed key and value hidden-states of the attention blocks. Can be used to speed up
+                decoding.
+
+                If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those
+                that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of
+                all `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+            inputs_embeds (`tf.Tensor` 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.
+            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).
+        """
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = shape_list(input_ids)
+        elif inputs_embeds is not None:
+            input_shape = shape_list(inputs_embeds)[:-1]
+        else:
+            raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
+
+        past_key_values_length = shape_list(past_key_values[0][0])[2] if past_key_values is not None else 0
+
+        if inputs_embeds is None:
+            check_embeddings_within_bounds(input_ids, self.embed_tokens.input_dim)
+            inputs_embeds = self.embed_tokens(input_ids) * self.embed_scale
+
+        # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+        if input_shape[-1] > 1:
+            combined_attention_mask = _make_causal_mask(input_shape, past_key_values_length=past_key_values_length)
+        else:
+            combined_attention_mask = _expand_mask(
+                tf.ones((input_shape[0], input_shape[1] + past_key_values_length)), tgt_len=input_shape[-1]
+            )
+
+        if attention_mask is not None:
+            combined_attention_mask = combined_attention_mask + _expand_mask(attention_mask, tgt_len=input_shape[-1])
+
+        if encoder_hidden_states is not None and encoder_attention_mask is not None:
+            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+            encoder_attention_mask = _expand_mask(encoder_attention_mask, tgt_len=input_shape[-1])
+
+        # embed positions
+        if position_ids is None:
+            positions = self.embed_positions(input_shape, past_key_values_length)
+        else:
+            positions = self.embed_positions(input_shape, position_ids=position_ids)
+
+        hidden_states = self.layernorm_embedding(inputs_embeds) + positions
+        hidden_states = self.dropout(hidden_states, training=training)
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        all_cross_attns = () if (output_attentions and encoder_hidden_states is not None) else None
+        present_key_values = () if use_cache else None
+
+        # check if head_mask and cross_attn_head_mask have a correct number of layers specified if desired
+        for attn_mask_name, attn_mask in [("head_mask", head_mask), ("cross_attn_head_mask", cross_attn_head_mask)]:
+            if attn_mask is not None:
+                tf.debugging.assert_equal(
+                    shape_list(attn_mask)[0],
+                    len(self.layers),
+                    message=(
+                        f"The {attn_mask_name} should be specified for {len(self.layers)} layers, but it is for"
+                        f" {shape_list(attn_mask)[0]}."
+                    ),
+                )
+
+        for idx, decoder_layer in enumerate(self.layers):
+            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+            dropout_probability = random.uniform(0, 1)
+
+            if training and (dropout_probability < self.layerdrop):
+                continue
+
+            past_key_value = past_key_values[idx] if past_key_values is not None else None
+
+            hidden_states, layer_self_attn, layer_cross_attn, present_key_value = decoder_layer(
+                hidden_states,
+                attention_mask=combined_attention_mask,
+                encoder_hidden_states=encoder_hidden_states,
+                encoder_attention_mask=encoder_attention_mask,
+                layer_head_mask=head_mask[idx] if head_mask is not None else None,
+                cross_attn_layer_head_mask=cross_attn_head_mask[idx] if cross_attn_head_mask is not None else None,
+                past_key_value=past_key_value,
+            )
+
+            if use_cache:
+                present_key_values += (present_key_value,)
+
+            if output_attentions:
+                all_self_attns += (layer_self_attn,)
+
+                if encoder_hidden_states is not None:
+                    all_cross_attns += (layer_cross_attn,)
+
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        if not return_dict:
+            return hidden_states, present_key_values, all_hidden_states, all_self_attns, all_cross_attns
+        else:
+            return TFBaseModelOutputWithPastAndCrossAttentions(
+                last_hidden_state=hidden_states,
+                past_key_values=present_key_values,
+                hidden_states=all_hidden_states,
+                attentions=all_self_attns,
+                cross_attentions=all_cross_attns,
+            )
+
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "embed_positions", None) is not None:
+            with tf.name_scope(self.embed_positions.name):
+                self.embed_positions.build(None)
+        if getattr(self, "layernorm_embedding", None) is not None:
+            with tf.name_scope(self.layernorm_embedding.name):
+                self.layernorm_embedding.build([None, None, self.config.d_model])
+        if getattr(self, "layers", None) is not None:
+            for layer in self.layers:
+                with tf.name_scope(layer.name):
+                    layer.build(None)
+
+
+@keras_serializable
+class TFBlenderbotSmallMainLayer(keras.layers.Layer):
+    config_class = BlenderbotSmallConfig
+
+    def __init__(self, config: BlenderbotSmallConfig, **kwargs):
+        super().__init__(**kwargs)
+
+        self.config = config
+        self.shared = keras.layers.Embedding(
+            input_dim=config.vocab_size,
+            output_dim=config.d_model,
+            embeddings_initializer=keras.initializers.TruncatedNormal(stddev=self.config.init_std),
+            name="model.shared",
+        )
+        # Additional attribute to specify the expected name scope of the layer (for loading/storing weights)
+        self.shared.load_weight_prefix = "model.shared"
+
+        self.encoder = TFBlenderbotSmallEncoder(config, self.shared, name="encoder")
+        self.decoder = TFBlenderbotSmallDecoder(config, self.shared, name="decoder")
+
+    def get_input_embeddings(self):
+        return self.shared
+
+    def set_input_embeddings(self, new_embeddings):
+        self.shared = new_embeddings
+        self.encoder.embed_tokens = self.shared
+        self.decoder.embed_tokens = self.shared
+
+    @unpack_inputs
+    def call(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        decoder_input_ids=None,
+        decoder_attention_mask=None,
+        decoder_position_ids=None,
+        head_mask=None,
+        decoder_head_mask=None,
+        cross_attn_head_mask=None,
+        encoder_outputs: Optional[Union[Tuple, TFBaseModelOutput]] = None,
+        past_key_values=None,
+        inputs_embeds=None,
+        decoder_inputs_embeds=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+        training=False,
+        **kwargs,
+    ):
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+
+        if encoder_outputs is None:
+            encoder_outputs = self.encoder(
+                input_ids=input_ids,
+                attention_mask=attention_mask,
+                head_mask=head_mask,
+                inputs_embeds=inputs_embeds,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+                training=training,
+            )
+        # If the user passed a tuple for encoder_outputs, we wrap it in a TFBaseModelOutput when return_dict=True
+        elif return_dict and not isinstance(encoder_outputs, TFBaseModelOutput):
+            encoder_outputs = TFBaseModelOutput(
+                last_hidden_state=encoder_outputs[0],
+                hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None,
+                attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None,
+            )
+        # If the user passed a TFBaseModelOutput for encoder_outputs, we wrap it in a tuple when return_dict=False
+        elif not return_dict and not isinstance(encoder_outputs, tuple):
+            encoder_outputs = encoder_outputs.to_tuple()
+
+        decoder_outputs = self.decoder(
+            decoder_input_ids,
+            attention_mask=decoder_attention_mask,
+            position_ids=decoder_position_ids,
+            encoder_hidden_states=encoder_outputs[0],
+            encoder_attention_mask=attention_mask,
+            head_mask=decoder_head_mask,
+            cross_attn_head_mask=cross_attn_head_mask,
+            past_key_values=past_key_values,
+            inputs_embeds=decoder_inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            training=training,
+        )
+
+        if not return_dict:
+            return decoder_outputs + encoder_outputs
+
+        return TFSeq2SeqModelOutput(
+            last_hidden_state=decoder_outputs.last_hidden_state,
+            past_key_values=decoder_outputs.past_key_values,
+            decoder_hidden_states=decoder_outputs.hidden_states,
+            decoder_attentions=decoder_outputs.attentions,
+            cross_attentions=decoder_outputs.cross_attentions,
+            encoder_last_hidden_state=encoder_outputs.last_hidden_state,
+            encoder_hidden_states=encoder_outputs.hidden_states,
+            encoder_attentions=encoder_outputs.attentions,
+        )
+
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        # The shared/tied weights expect to be in the model base namespace
+        # Adding "/" to the end (not the start!) of a tf.name_scope puts it in the root namespace rather than
+        # the current one.
+        with tf.name_scope(self.shared.load_weight_prefix + "/" + self.shared.name + "/"):
+            self.shared.build(None)
+        if getattr(self, "encoder", None) is not None:
+            with tf.name_scope(self.encoder.name):
+                self.encoder.build(None)
+        if getattr(self, "decoder", None) is not None:
+            with tf.name_scope(self.decoder.name):
+                self.decoder.build(None)
+
+
+@add_start_docstrings(
+    "The bare BLENDERBOT_SMALL Model outputting raw hidden-states without any specific head on top.",
+    BLENDERBOT_SMALL_START_DOCSTRING,
+)
+class TFBlenderbotSmallModel(TFBlenderbotSmallPreTrainedModel):
+    def __init__(self, config: BlenderbotSmallConfig, *inputs, **kwargs):
+        super().__init__(config, *inputs, **kwargs)
+
+        self.model = TFBlenderbotSmallMainLayer(config, name="model")
+
+    def get_encoder(self):
+        return self.model.encoder
+
+    def get_decoder(self):
+        return self.model.decoder
+
+    @unpack_inputs
+    @add_start_docstrings_to_model_forward(BLENDERBOT_SMALL_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=TFSeq2SeqModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def call(
+        self,
+        input_ids: tf.Tensor | None = None,
+        attention_mask: tf.Tensor | None = None,
+        decoder_input_ids: tf.Tensor | None = None,
+        decoder_attention_mask: tf.Tensor | None = None,
+        decoder_position_ids: tf.Tensor | None = None,
+        head_mask: tf.Tensor | None = None,
+        decoder_head_mask: tf.Tensor | None = None,
+        cross_attn_head_mask: tf.Tensor | None = None,
+        encoder_outputs: Optional[Union[Tuple, TFBaseModelOutput]] = None,
+        past_key_values: List[tf.Tensor] | None = None,
+        inputs_embeds: tf.Tensor | None = None,
+        decoder_inputs_embeds: tf.Tensor | None = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        training: Optional[bool] = False,
+        **kwargs,
+    ) -> Union[Tuple[tf.Tensor], TFSeq2SeqModelOutput]:
+        outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            decoder_input_ids=decoder_input_ids,
+            decoder_attention_mask=decoder_attention_mask,
+            decoder_position_ids=decoder_position_ids,
+            head_mask=head_mask,
+            decoder_head_mask=decoder_head_mask,
+            cross_attn_head_mask=cross_attn_head_mask,
+            encoder_outputs=encoder_outputs,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            decoder_inputs_embeds=decoder_inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            training=training,
+        )
+
+        return outputs
+
+    # Copied from transformers.models.bart.modeling_tf_bart.TFBartModel.serving_output
+    def serving_output(self, output):
+        pkv = tf.tuple(output.past_key_values)[1] if self.config.use_cache else None
+        dec_hs = tf.convert_to_tensor(output.decoder_hidden_states) if self.config.output_hidden_states else None
+        dec_attns = tf.convert_to_tensor(output.decoder_attentions) if self.config.output_attentions else None
+        cross_attns = tf.convert_to_tensor(output.cross_attentions) if self.config.output_attentions else None
+        enc_hs = tf.convert_to_tensor(output.encoder_hidden_states) if self.config.output_hidden_states else None
+        enc_attns = tf.convert_to_tensor(output.encoder_attentions) if self.config.output_attentions else None
+
+        return TFSeq2SeqModelOutput(
+            last_hidden_state=output.last_hidden_state,
+            past_key_values=pkv,
+            decoder_hidden_states=dec_hs,
+            decoder_attentions=dec_attns,
+            cross_attentions=cross_attns,
+            encoder_last_hidden_state=output.encoder_last_hidden_state,
+            encoder_hidden_states=enc_hs,
+            encoder_attentions=enc_attns,
+        )
+
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "model", None) is not None:
+            with tf.name_scope(self.model.name):
+                self.model.build(None)
+
+
+# Copied from transformers.models.bart.modeling_tf_bart.BiasLayer
+class BiasLayer(keras.layers.Layer):
+    """
+    Bias as a layer. It is used for serialization purposes: `keras.Model.save_weights` stores on a per-layer basis,
+    so all weights have to be registered in a layer.
+    """
+
+    def __init__(self, shape, initializer, trainable, name, **kwargs):
+        super().__init__(name=name, **kwargs)
+        # Note: the name of this variable will NOT be scoped when serialized, i.e. it will not be in the format of
+        # "outer_layer/inner_layer/.../name:0". Instead, it will be "name:0". For further details, see:
+        # https://github.com/huggingface/transformers/pull/18833#issuecomment-1233090214
+        self.bias = self.add_weight(name=name, shape=shape, initializer=initializer, trainable=trainable)
+
+    def call(self, x):
+        return x + self.bias
+
+
+@add_start_docstrings(
+    "The BLENDERBOT_SMALL Model with a language modeling head. Can be used for summarization.",
+    BLENDERBOT_SMALL_START_DOCSTRING,
+)
+class TFBlenderbotSmallForConditionalGeneration(TFBlenderbotSmallPreTrainedModel, TFCausalLanguageModelingLoss):
+    _keys_to_ignore_on_load_unexpected = [
+        r"model.encoder.embed_tokens.weight",
+        r"model.decoder.embed_tokens.weight",
+    ]
+
+    def __init__(self, config, *inputs, **kwargs):
+        super().__init__(config, *inputs, **kwargs)
+        self.model = TFBlenderbotSmallMainLayer(config, name="model")
+        self.use_cache = config.use_cache
+        # final_bias_logits is registered as a buffer in pytorch, so not trainable for the sake of consistency.
+        self.bias_layer = BiasLayer(
+            name="final_logits_bias", shape=[1, config.vocab_size], initializer="zeros", trainable=False
+        )
+
+    def get_decoder(self):
+        return self.model.decoder
+
+    def get_encoder(self):
+        return self.model.encoder
+
+    def get_output_embeddings(self):
+        return self.get_input_embeddings()
+
+    def set_output_embeddings(self, value):
+        self.set_input_embeddings(value)
+
+    def get_bias(self):
+        return {"final_logits_bias": self.bias_layer.bias}
+
+    def set_bias(self, value):
+        # Replaces the existing layers containing bias for correct (de)serialization.
+        vocab_size = value["final_logits_bias"].shape[-1]
+        self.bias_layer = BiasLayer(
+            name="final_logits_bias", shape=[1, vocab_size], initializer="zeros", trainable=False
+        )
+        self.bias_layer.bias.assign(value["final_logits_bias"])
+
+    @unpack_inputs
+    @add_start_docstrings_to_model_forward(BLENDERBOT_SMALL_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=TFSeq2SeqLMOutput, config_class=_CONFIG_FOR_DOC)
+    @add_end_docstrings(BLENDERBOT_SMALL_GENERATION_EXAMPLE)
+    def call(
+        self,
+        input_ids: tf.Tensor | None = None,
+        attention_mask: tf.Tensor | None = None,
+        decoder_input_ids: tf.Tensor | None = None,
+        decoder_attention_mask: tf.Tensor | None = None,
+        decoder_position_ids: tf.Tensor | None = None,
+        head_mask: tf.Tensor | None = None,
+        decoder_head_mask: tf.Tensor | None = None,
+        cross_attn_head_mask: tf.Tensor | None = None,
+        encoder_outputs: Optional[TFBaseModelOutput] = None,
+        past_key_values: List[tf.Tensor] | None = None,
+        inputs_embeds: tf.Tensor | None = None,
+        decoder_inputs_embeds: tf.Tensor | None = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        labels: tf.Tensor | None = None,
+        training: Optional[bool] = False,
+    ) -> Union[Tuple[tf.Tensor], TFSeq2SeqLMOutput]:
+        r"""
+        labels (`tf.tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        Returns:
+
+        """
+
+        if labels is not None:
+            labels = tf.where(
+                labels == self.config.pad_token_id,
+                tf.cast(tf.fill(shape_list(labels), -100), labels.dtype),
+                labels,
+            )
+            use_cache = False
+            if decoder_input_ids is None and decoder_inputs_embeds is None:
+                decoder_input_ids = shift_tokens_right(
+                    labels, self.config.pad_token_id, self.config.decoder_start_token_id
+                )
+
+        outputs = self.model(
+            input_ids,
+            attention_mask=attention_mask,
+            decoder_input_ids=decoder_input_ids,
+            decoder_attention_mask=decoder_attention_mask,
+            decoder_position_ids=decoder_position_ids,
+            head_mask=head_mask,
+            decoder_head_mask=decoder_head_mask,
+            cross_attn_head_mask=cross_attn_head_mask,
+            encoder_outputs=encoder_outputs,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            decoder_inputs_embeds=decoder_inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            training=training,
+        )
+        lm_logits = tf.matmul(outputs[0], self.model.shared.weights, transpose_b=True)
+        lm_logits = self.bias_layer(lm_logits)
+        masked_lm_loss = None if labels is None else self.hf_compute_loss(labels, lm_logits)
+
+        if not return_dict:
+            output = (lm_logits,) + outputs[1:]
+            return ((masked_lm_loss,) + output) if masked_lm_loss is not None else output
+        return TFSeq2SeqLMOutput(
+            loss=masked_lm_loss,
+            logits=lm_logits,
+            past_key_values=outputs.past_key_values,  # index 1 of d outputs
+            decoder_hidden_states=outputs.decoder_hidden_states,  # index 2 of d outputs
+            decoder_attentions=outputs.decoder_attentions,  # index 3 of d outputs
+            cross_attentions=outputs.cross_attentions,  # index 4 of d outputs
+            encoder_last_hidden_state=outputs.encoder_last_hidden_state,  # index 0 of encoder outputs
+            encoder_hidden_states=outputs.encoder_hidden_states,  # 1 of e out
+            encoder_attentions=outputs.encoder_attentions,  # 2 of e out
+        )
+
+    # Copied from transformers.models.bart.modeling_tf_bart.TFBartForConditionalGeneration.serving_output
+    def serving_output(self, output):
+        pkv = tf.tuple(output.past_key_values)[1] if self.config.use_cache else None
+        dec_hs = tf.convert_to_tensor(output.decoder_hidden_states) if self.config.output_hidden_states else None
+        dec_attns = tf.convert_to_tensor(output.decoder_attentions) if self.config.output_attentions else None
+        cross_attns = tf.convert_to_tensor(output.cross_attentions) if self.config.output_attentions else None
+        enc_hs = tf.convert_to_tensor(output.encoder_hidden_states) if self.config.output_hidden_states else None
+        enc_attns = tf.convert_to_tensor(output.encoder_attentions) if self.config.output_attentions else None
+
+        return TFSeq2SeqLMOutput(
+            logits=output.logits,
+            past_key_values=pkv,
+            decoder_hidden_states=dec_hs,
+            decoder_attentions=dec_attns,
+            cross_attentions=cross_attns,
+            encoder_last_hidden_state=output.encoder_last_hidden_state,
+            encoder_hidden_states=enc_hs,
+            encoder_attentions=enc_attns,
+        )
+
+    # Copied from transformers.models.bart.modeling_tf_bart.TFBartForConditionalGeneration.prepare_inputs_for_generation
+    def prepare_inputs_for_generation(
+        self,
+        decoder_input_ids,
+        past_key_values=None,
+        attention_mask=None,
+        decoder_attention_mask=None,
+        head_mask=None,
+        decoder_head_mask=None,
+        cross_attn_head_mask=None,
+        use_cache=None,
+        encoder_outputs=None,
+        **kwargs,
+    ):
+        # cut decoder_input_ids if past_key_values is used
+        if past_key_values is not None:
+            decoder_input_ids = decoder_input_ids[:, -1:]
+
+        if decoder_attention_mask is not None:  # xla
+            decoder_position_ids = tf.math.cumsum(decoder_attention_mask, axis=-1, exclusive=True)[:, -1:]
+        elif past_key_values is not None:  # no xla + past_key_values
+            decoder_position_ids = past_key_values[0][0].shape[2]
+        else:  # no xla + no past_key_values
+            decoder_position_ids = tf.range(decoder_input_ids.shape[1])
+
+        return {
+            "input_ids": None,  # encoder_outputs is defined. input_ids not needed
+            "encoder_outputs": encoder_outputs,
+            "past_key_values": past_key_values,
+            "decoder_input_ids": decoder_input_ids,
+            "attention_mask": attention_mask,
+            "decoder_attention_mask": decoder_attention_mask,
+            "decoder_position_ids": decoder_position_ids,
+            "head_mask": head_mask,
+            "decoder_head_mask": decoder_head_mask,
+            "cross_attn_head_mask": cross_attn_head_mask,
+            "use_cache": use_cache,  # change this to avoid caching (presumably for debugging)
+        }
+
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "model", None) is not None:
+            with tf.name_scope(self.model.name):
+                self.model.build(None)
+        if getattr(self, "bias_layer", None) is not None:
+            with tf.name_scope(self.bias_layer.name):
+                self.bias_layer.build(None)
+
+
+__all__ = ["TFBlenderbotSmallForConditionalGeneration", "TFBlenderbotSmallModel", "TFBlenderbotSmallPreTrainedModel"]

lib/python3.12/site-packages/transformers/models/blenderbot_small/tokenization_blenderbot_small.py

@@ -0,0 +1,222 @@
+# coding=utf-8
+# Copyright 2021 The Facebook Inc. 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.
+"""Tokenization class for BlenderbotSmall."""
+
+import json
+import os
+from typing import Dict, List, Optional, Tuple
+
+import regex as re
+
+from ...tokenization_utils import PreTrainedTokenizer
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+VOCAB_FILES_NAMES = {
+    "vocab_file": "vocab.json",
+    "merges_file": "merges.txt",
+    "tokenizer_config_file": "tokenizer_config.json",
+}
+
+
+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
+
+    pairs = set(pairs)
+    return pairs
+
+
+class BlenderbotSmallTokenizer(PreTrainedTokenizer):
+    """
+    Constructs a Blenderbot-90M tokenizer based on BPE (Byte-Pair-Encoding)
+
+    This tokenizer inherits from [`PreTrainedTokenizer`] which contains most of the main methods. Users should refer to
+    the superclass for more information regarding methods.
+
+    Args:
+        vocab_file (`str`):
+            File containing the vocabulary.
+        merges_file (`str`):
+            Path to the merges file.
+        bos_token (`str`, *optional*, defaults to `"__start__"`):
+            The beginning of sentence token.
+        eos_token (`str`, *optional*, defaults to `"__end__"`):
+            The end of sentence token.
+        unk_token (`str`, *optional*, defaults to `"__unk__"`):
+            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.
+        pad_token (`str`, *optional*, defaults to `"__null__"`):
+            The token used for padding, for example when batching sequences of different lengths.
+        kwargs (*optional*):
+            Additional keyword arguments passed along to [`PreTrainedTokenizer`]
+    """
+
+    vocab_files_names = VOCAB_FILES_NAMES
+    model_input_names = ["input_ids", "attention_mask"]
+
+    def __init__(
+        self,
+        vocab_file,
+        merges_file,
+        bos_token="__start__",
+        eos_token="__end__",
+        unk_token="__unk__",
+        pad_token="__null__",
+        **kwargs,
+    ):
+        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()}
+        with open(merges_file, encoding="utf-8") as merges_handle:
+            merges = merges_handle.read().split("\n")[1:-1]
+        merges = [tuple(merge.split()) for merge in merges]
+        self.bpe_ranks = dict(zip(merges, range(len(merges))))
+        self.cache = {}
+        super().__init__(unk_token=unk_token, bos_token=bos_token, eos_token=eos_token, pad_token=pad_token, **kwargs)
+
+    @property
+    def vocab_size(self) -> int:
+        return len(self.encoder)
+
+    def get_vocab(self) -> Dict:
+        return dict(self.encoder, **self.added_tokens_encoder)
+
+    def bpe(self, token: str) -> str:
+        if token in self.cache:
+            return self.cache[token]
+        token = re.sub("([.,!?()])", r" \1", token)
+        token = re.sub("(')", r" \1 ", token)
+        token = re.sub(r"\s{2,}", " ", token)
+        if "\n" in token:
+            token = token.replace("\n", " __newln__")
+
+        tokens = token.split(" ")
+        words = []
+        for token in tokens:
+            if not len(token):
+                continue
+
+            token = token.lower()
+            word = tuple(token)
+            word = tuple(list(word[:-1]) + [word[-1] + "</w>"])
+            pairs = get_pairs(word)
+
+            if not pairs:
+                words.append(token)
+                continue
+
+            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)
+                        new_word.extend(word[i:j])
+                        i = j
+                    except ValueError:
+                        new_word.extend(word[i:])
+                        break
+
+                    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)
+            word = word[:-4]
+
+            self.cache[token] = word
+            words.append(word)
+        return " ".join(words)
+
+    def _tokenize(self, text: str) -> List[str]:
+        """Split a string into tokens using BPE."""
+        split_tokens = []
+
+        words = re.findall(r"\S+\n?", text)
+
+        for token in words:
+            split_tokens.extend(list(self.bpe(token).split(" ")))
+        return split_tokens
+
+    def _convert_token_to_id(self, token: str) -> int:
+        """Converts a token to an id using the vocab."""
+        token = token.lower()
+        return self.encoder.get(token, self.encoder.get(self.unk_token))
+
+    def _convert_id_to_token(self, index: int) -> str:
+        """Converts an index (integer) in a token (str) using the vocab."""
+        return self.decoder.get(index, self.unk_token)
+
+    def convert_tokens_to_string(self, tokens: List[str]) -> str:
+        """Converts a sequence of tokens in a single string."""
+        out_string = " ".join(tokens).replace("@@ ", "").strip()
+        return out_string
+
+    def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
+        if not os.path.isdir(save_directory):
+            logger.error(f"Vocabulary path ({save_directory}) should be a 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(
+                        f"Saving vocabulary to {merge_file}: BPE merge indices are not consecutive."
+                        " Please check that the tokenizer is not corrupted!"
+                    )
+                    index = token_index
+                writer.write(" ".join(bpe_tokens) + "\n")
+                index += 1
+
+        return vocab_file, merge_file
+
+
+__all__ = ["BlenderbotSmallTokenizer"]

lib/python3.12/site-packages/transformers/models/blenderbot_small/tokenization_blenderbot_small_fast.py

@@ -0,0 +1,103 @@
+# coding=utf-8
+# Copyright 2021, The Facebook, Inc. 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.
+"""Fast tokenization class for BlenderbotSmall."""
+
+from typing import List, Optional
+
+from tokenizers import ByteLevelBPETokenizer
+
+from ...tokenization_utils_fast import PreTrainedTokenizerFast
+from ...utils import logging
+from .tokenization_blenderbot_small import BlenderbotSmallTokenizer
+
+
+logger = logging.get_logger(__name__)
+
+VOCAB_FILES_NAMES = {
+    "vocab_file": "vocab.json",
+    "merges_file": "merges.txt",
+    "tokenizer_config_file": "tokenizer_config.json",
+}
+
+
+class BlenderbotSmallTokenizerFast(PreTrainedTokenizerFast):
+    """
+    Construct a "fast" BlenderbotSmall tokenizer (backed by HuggingFace's *tokenizers* library).
+
+    Args:
+        vocab_file (`str`):
+            Path to the vocabulary file.
+    """
+
+    vocab_files_names = VOCAB_FILES_NAMES
+    slow_tokenizer_class = BlenderbotSmallTokenizer
+
+    def __init__(
+        self,
+        vocab_file=None,
+        merges_file=None,
+        unk_token="<|endoftext|>",
+        bos_token="<|endoftext|>",
+        eos_token="<|endoftext|>",
+        add_prefix_space=False,
+        trim_offsets=True,
+        **kwargs,
+    ):
+        super().__init__(
+            ByteLevelBPETokenizer(
+                vocab=vocab_file,
+                merges=merges_file,
+                add_prefix_space=add_prefix_space,
+                trim_offsets=trim_offsets,
+            ),
+            bos_token=bos_token,
+            eos_token=eos_token,
+            unk_token=unk_token,
+            **kwargs,
+        )
+        self.add_prefix_space = add_prefix_space
+
+    def build_inputs_with_special_tokens(self, token_ids_0, token_ids_1=None):
+        output = [self.bos_token_id] + token_ids_0 + [self.eos_token_id]
+        if token_ids_1 is None:
+            return output
+
+        return output + [self.eos_token_id] + token_ids_1 + [self.eos_token_id]
+
+    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. BlenderbotSmall
+        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.
+        """
+        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 + sep + token_ids_1 + sep) * [0]
+
+
+__all__ = ["BlenderbotSmallTokenizerFast"]

lib/python3.12/site-packages/transformers/models/byt5/__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_byt5 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/byt5/tokenization_byt5.py

@@ -0,0 +1,236 @@
+# coding=utf-8
+# Copyright 2021 T5 Authors and 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 class for model ByT5."""
+
+import warnings
+from typing import List, Optional, Tuple
+
+from ...tokenization_utils import AddedToken, PreTrainedTokenizer
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class ByT5Tokenizer(PreTrainedTokenizer):
+    """
+    Construct a ByT5 tokenizer. ByT5 simply uses raw bytes utf-8 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:
+        eos_token (`str`, *optional*, defaults to `"</s>"`):
+            The end of sequence token.
+
+            <Tip>
+
+            When building a sequence using special tokens, this is not the token that is used for the end of sequence.
+            The token used is the `sep_token`.
+
+            </Tip>
+
+        unk_token (`str`, *optional*, defaults to `"<unk>"`):
+            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.
+        pad_token (`str`, *optional*, defaults to `"<pad>"`):
+            The token used for padding, for example when batching sequences of different lengths.
+        extra_ids (`int`, *optional*, defaults to 125):
+            Add a number of extra ids added to the end of the vocabulary for use as sentinels. These tokens are
+            accessible as "<extra_id_{%d}>" where "{%d}" is a number between 0 and extra_ids-1. Extra tokens are
+            indexed from the end of the vocabulary up to beginning ("<extra_id_0>" is the last token in the vocabulary
+            like in ByT5 preprocessing see
+            [here](https://github.com/google-research/text-to-text-transfer-transformer/blob/9fd7b14a769417be33bc6c850f9598764913c833/t5/data/preprocessors.py#L2117)).
+        additional_special_tokens (`List[str]`, *optional*):
+            Additional special tokens used by the tokenizer.
+    """
+
+    model_input_names = ["input_ids", "attention_mask"]
+
+    def __init__(
+        self,
+        eos_token="</s>",
+        unk_token="<unk>",
+        pad_token="<pad>",
+        extra_ids=125,
+        additional_special_tokens=None,
+        **kwargs,
+    ) -> None:
+        # Add extra_ids to the special token list
+        if extra_ids > 0 and additional_special_tokens is None:
+            additional_special_tokens = [f"<extra_id_{i}>" for i in range(extra_ids)]
+        elif extra_ids > 0 and additional_special_tokens is not None and len(additional_special_tokens) > 0:
+            # Check that we have the right number of extra_id special tokens
+            extra_tokens = len(set(filter(lambda x: bool("extra_id" in str(x)), additional_special_tokens)))
+            if extra_tokens != extra_ids:
+                raise ValueError(
+                    f"Both extra_ids ({extra_ids}) and additional_special_tokens ({additional_special_tokens}) are"
+                    " provided to ByT5Tokenizer. In this case the additional_special_tokens must include the"
+                    " extra_ids tokens"
+                )
+
+        pad_token = AddedToken(pad_token, lstrip=True, rstrip=True) if isinstance(pad_token, str) else pad_token
+        # we force left and right stripping for backward compatibility. The byt5tests depend on this.
+        eos_token = AddedToken(eos_token, lstrip=True, rstrip=True) if isinstance(eos_token, str) else eos_token
+        unk_token = AddedToken(unk_token, lstrip=True, rstrip=True) if isinstance(unk_token, str) else unk_token
+        # unk token needs to be in the vocab with correct index
+        self._added_tokens_decoder = {0: pad_token, 1: eos_token, 2: unk_token}
+        self.offset = len(self._added_tokens_decoder)
+        self._utf_vocab_size = 2**8  # utf is 8 bits
+        super().__init__(
+            eos_token=eos_token,
+            unk_token=unk_token,
+            pad_token=pad_token,
+            extra_ids=0,
+            additional_special_tokens=additional_special_tokens,  # TODO extra ids are not used :sweatywmile:
+            **kwargs,
+        )
+
+    @property
+    def vocab_size(self):
+        return self._utf_vocab_size
+
+    def get_vocab(self):
+        vocab = {self.convert_ids_to_tokens(i): i for i in range(self.vocab_size + self.offset)}
+        vocab.update(self.added_tokens_encoder)
+        return vocab
+
+    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
+            )
+
+        # normal case: some special tokens
+        if token_ids_1 is None:
+            return ([0] * len(token_ids_0)) + [1]
+        return ([0] * len(token_ids_0)) + [1] + ([0] * len(token_ids_1)) + [1]
+
+    def _add_eos_if_not_present(self, token_ids: List[int]) -> List[int]:
+        """Do not add eos again if user already added it."""
+        if len(token_ids) > 0 and token_ids[-1] == self.eos_token_id:
+            warnings.warn(
+                f"This sequence already has {self.eos_token}. In future versions this behavior may lead to duplicated"
+                " eos tokens being added."
+            )
+            return token_ids
+        else:
+            return token_ids + [self.eos_token_id]
+
+    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. ByT5 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.
+        """
+        eos = [self.eos_token_id]
+
+        if token_ids_1 is None:
+            return len(token_ids_0 + eos) * [0]
+        return len(token_ids_0 + eos + token_ids_1 + eos) * [0]
+
+    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 sequence has the following format:
+
+        - single sequence: `X </s>`
+        - pair of sequences: `A </s> B </s>`
+
+        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.
+        """
+        token_ids_0 = self._add_eos_if_not_present(token_ids_0)
+        if token_ids_1 is None:
+            return token_ids_0
+        else:
+            token_ids_1 = self._add_eos_if_not_present(token_ids_1)
+            return token_ids_0 + token_ids_1
+
+    def _tokenize(self, text: str) -> List[str]:
+        """Take as input a string and return a list of strings (tokens) for words/sub-words"""
+        tokens = [chr(i) for i in text.encode("utf-8")]
+        return tokens
+
+    def _convert_token_to_id(self, token):
+        """Converts a token (str) in an id using the vocab."""
+
+        if len(token) != 1:
+            token_id = None
+        else:
+            token_id = ord(token) + self.offset
+
+        return token_id
+
+    def _convert_id_to_token(self, index):
+        """Converts an index (integer) in a token (str) using the vocab."""
+        token = chr(index - self.offset)
+        return token
+
+    def convert_tokens_to_string(self, tokens):
+        """Converts a sequence of tokens (string) in a single string."""
+        bstring = b""
+        for token in tokens:
+            if token in self.added_tokens_decoder:
+                tok_string = self.added_tokens_decoder[token].encode("utf-8")
+            elif token in self.added_tokens_encoder:
+                tok_string = token.encode("utf-8")
+            else:
+                tok_string = bytes([ord(token)])
+            bstring += tok_string
+        string = bstring.decode("utf-8", errors="ignore")
+        return string
+
+    # ByT5Tokenizer has no vocab file
+    def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
+        return ()
+
+
+__all__ = ["ByT5Tokenizer"]

lib/python3.12/site-packages/transformers/models/clipseg/__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_clipseg import *
+    from .modeling_clipseg import *
+    from .processing_clipseg 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/clipseg/configuration_clipseg.py

@@ -0,0 +1,396 @@
+# 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.
+"""CLIPSeg model configuration"""
+
+from ...configuration_utils import PretrainedConfig
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class CLIPSegTextConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`CLIPSegModel`]. It is used to instantiate an
+    CLIPSeg 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 CLIPSeg
+    [CIDAS/clipseg-rd64](https://huggingface.co/CIDAS/clipseg-rd64) 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 CLIPSeg text model. Defines the number of different tokens that can be represented
+            by the `inputs_ids` passed when calling [`CLIPSegModel`].
+        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.
+        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 CLIPSegTextConfig, CLIPSegTextModel
+
+    >>> # Initializing a CLIPSegTextConfig with CIDAS/clipseg-rd64 style configuration
+    >>> configuration = CLIPSegTextConfig()
+
+    >>> # Initializing a CLIPSegTextModel (with random weights) from the CIDAS/clipseg-rd64 style configuration
+    >>> model = CLIPSegTextModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "clipseg_text_model"
+    base_config_key = "text_config"
+
+    def __init__(
+        self,
+        vocab_size=49408,
+        hidden_size=512,
+        intermediate_size=2048,
+        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,
+        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.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 CLIPSegVisionConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`CLIPSegModel`]. It is used to instantiate an
+    CLIPSeg 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 CLIPSeg
+    [CIDAS/clipseg-rd64](https://huggingface.co/CIDAS/clipseg-rd64) 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.
+        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 CLIPSegVisionConfig, CLIPSegVisionModel
+
+    >>> # Initializing a CLIPSegVisionConfig with CIDAS/clipseg-rd64 style configuration
+    >>> configuration = CLIPSegVisionConfig()
+
+    >>> # Initializing a CLIPSegVisionModel (with random weights) from the CIDAS/clipseg-rd64 style configuration
+    >>> model = CLIPSegVisionModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "clipseg_vision_model"
+    base_config_key = "vision_config"
+
+    def __init__(
+        self,
+        hidden_size=768,
+        intermediate_size=3072,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        num_channels=3,
+        image_size=224,
+        patch_size=32,
+        hidden_act="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.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 CLIPSegConfig(PretrainedConfig):
+    r"""
+    [`CLIPSegConfig`] is the configuration class to store the configuration of a [`CLIPSegModel`]. It is used to
+    instantiate a CLIPSeg 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 CLIPSeg
+    [CIDAS/clipseg-rd64](https://huggingface.co/CIDAS/clipseg-rd64) 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 [`CLIPSegTextConfig`].
+        vision_config (`dict`, *optional*):
+            Dictionary of configuration options used to initialize [`CLIPSegVisionConfig`].
+        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 CLIPSeg implementation.
+        extract_layers (`List[int]`, *optional*, defaults to `[3, 6, 9]`):
+            Layers to extract when forwarding the query image through the frozen visual backbone of CLIP.
+        reduce_dim (`int`, *optional*, defaults to 64):
+            Dimensionality to reduce the CLIP vision embedding.
+        decoder_num_attention_heads (`int`, *optional*, defaults to 4):
+            Number of attention heads in the decoder of CLIPSeg.
+        decoder_attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        decoder_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.
+        decoder_intermediate_size (`int`, *optional*, defaults to 2048):
+            Dimensionality of the "intermediate" (i.e., feed-forward) layers in the Transformer decoder.
+        conditional_layer (`int`, *optional*, defaults to 0):
+            The layer to use of the Transformer encoder whose activations will be combined with the condition
+            embeddings using FiLM (Feature-wise Linear Modulation). If 0, the last layer is used.
+        use_complex_transposed_convolution (`bool`, *optional*, defaults to `False`):
+            Whether to use a more complex transposed convolution in the decoder, enabling more fine-grained
+            segmentation.
+        kwargs (*optional*):
+            Dictionary of keyword arguments.
+
+    Example:
+
+    ```python
+    >>> from transformers import CLIPSegConfig, CLIPSegModel
+
+    >>> # Initializing a CLIPSegConfig with CIDAS/clipseg-rd64 style configuration
+    >>> configuration = CLIPSegConfig()
+
+    >>> # Initializing a CLIPSegModel (with random weights) from the CIDAS/clipseg-rd64 style configuration
+    >>> model = CLIPSegModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+
+    >>> # We can also initialize a CLIPSegConfig from a CLIPSegTextConfig and a CLIPSegVisionConfig
+
+    >>> # Initializing a CLIPSegText and CLIPSegVision configuration
+    >>> config_text = CLIPSegTextConfig()
+    >>> config_vision = CLIPSegVisionConfig()
+
+    >>> config = CLIPSegConfig.from_text_vision_configs(config_text, config_vision)
+    ```"""
+
+    model_type = "clipseg"
+    sub_configs = {"text_config": CLIPSegTextConfig, "vision_config": CLIPSegVisionConfig}
+
+    def __init__(
+        self,
+        text_config=None,
+        vision_config=None,
+        projection_dim=512,
+        logit_scale_init_value=2.6592,
+        extract_layers=[3, 6, 9],
+        reduce_dim=64,
+        decoder_num_attention_heads=4,
+        decoder_attention_dropout=0.0,
+        decoder_hidden_act="quick_gelu",
+        decoder_intermediate_size=2048,
+        conditional_layer=0,
+        use_complex_transposed_convolution=False,
+        **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 = CLIPSegTextConfig(**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 `CLIPSegTextConfig`. 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 = CLIPSegVisionConfig(**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 `CLIPSegVisionConfig`. "
+                            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 `CLIPSegTextConfig` with default values.")
+
+        if vision_config is None:
+            vision_config = {}
+            logger.info("`vision_config` is `None`. initializing the `CLIPSegVisionConfig` with default values.")
+
+        self.text_config = CLIPSegTextConfig(**text_config)
+        self.vision_config = CLIPSegVisionConfig(**vision_config)
+
+        self.projection_dim = projection_dim
+        self.logit_scale_init_value = logit_scale_init_value
+        self.extract_layers = extract_layers
+        self.reduce_dim = reduce_dim
+        self.decoder_num_attention_heads = decoder_num_attention_heads
+        self.decoder_attention_dropout = decoder_attention_dropout
+        self.decoder_hidden_act = decoder_hidden_act
+        self.decoder_intermediate_size = decoder_intermediate_size
+        self.conditional_layer = conditional_layer
+        self.initializer_factor = 1.0
+        self.use_complex_transposed_convolution = use_complex_transposed_convolution
+
+    @classmethod
+    def from_text_vision_configs(cls, text_config: CLIPSegTextConfig, vision_config: CLIPSegVisionConfig, **kwargs):
+        r"""
+        Instantiate a [`CLIPSegConfig`] (or a derived class) from clipseg text model configuration and clipseg vision
+        model configuration.
+
+        Returns:
+            [`CLIPSegConfig`]: An instance of a configuration object
+        """
+
+        return cls(text_config=text_config.to_dict(), vision_config=vision_config.to_dict(), **kwargs)
+
+
+__all__ = ["CLIPSegConfig", "CLIPSegTextConfig", "CLIPSegVisionConfig"]

lib/python3.12/site-packages/transformers/models/clipseg/modeling_clipseg.py

@@ -0,0 +1,1522 @@
+# coding=utf-8
+# Copyright 2022 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 CLIPSeg model."""
+
+import copy
+import math
+from dataclasses import dataclass
+from typing import Any, Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+
+from ...activations import ACT2FN
+from ...modeling_attn_mask_utils import _create_4d_causal_attention_mask, _prepare_4d_attention_mask
+from ...modeling_outputs import BaseModelOutput, BaseModelOutputWithPooling
+from ...modeling_utils import PreTrainedModel
+from ...utils import (
+    ModelOutput,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    logging,
+    replace_return_docstrings,
+    torch_int,
+)
+from .configuration_clipseg import CLIPSegConfig, CLIPSegTextConfig, CLIPSegVisionConfig
+
+
+logger = logging.get_logger(__name__)
+
+
+_CHECKPOINT_FOR_DOC = "CIDAS/clipseg-rd64-refined"
+
+
+# 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: torch.Tensor) -> torch.Tensor:
+    return nn.functional.cross_entropy(logits, torch.arange(len(logits), device=logits.device))
+
+
+# Copied from transformers.models.clip.modeling_clip.clip_loss with clip->clipseg
+def clipseg_loss(similarity: torch.Tensor) -> torch.Tensor:
+    caption_loss = contrastive_loss(similarity)
+    image_loss = contrastive_loss(similarity.t())
+    return (caption_loss + image_loss) / 2.0
+
+
+@dataclass
+# Copied from transformers.models.clip.modeling_clip.CLIPOutput with CLIP->CLIPSeg
+class CLIPSegOutput(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 [`CLIPSegTextModel`].
+        image_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim`):
+            The image embeddings obtained by applying the projection layer to the pooled output of [`CLIPSegVisionModel`].
+        text_model_output (`BaseModelOutputWithPooling`):
+            The output of the [`CLIPSegTextModel`].
+        vision_model_output (`BaseModelOutputWithPooling`):
+            The output of the [`CLIPSegVisionModel`].
+    """
+
+    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()
+        )
+
+
+@dataclass
+class CLIPSegDecoderOutput(ModelOutput):
+    """
+    Args:
+        logits (`torch.FloatTensor` of shape `(batch_size, height, width)`):
+            Classification scores for each pixel.
+        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)`.
+        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.
+    """
+
+    logits: Optional[torch.FloatTensor] = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+@dataclass
+class CLIPSegImageSegmentationOutput(ModelOutput):
+    """
+    Args:
+        loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `return_loss` is `True`):
+            Contrastive loss for image-text similarity.
+        ...
+        vision_model_output (`BaseModelOutputWithPooling`):
+            The output of the [`CLIPSegVisionModel`].
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    logits: Optional[torch.FloatTensor] = None
+    conditional_embeddings: Optional[torch.FloatTensor] = None
+    pooled_output: Optional[torch.FloatTensor] = None
+    vision_model_output: BaseModelOutputWithPooling = None
+    decoder_output: CLIPSegDecoderOutput = None
+
+    def to_tuple(self) -> Tuple[Any]:
+        return tuple(
+            self[k] if k not in ["vision_model_output", "decoder_output"] else getattr(self, k).to_tuple()
+            for k in self.keys()
+        )
+
+
+class CLIPSegVisionEmbeddings(nn.Module):
+    # Copied from transformers.models.clip.modeling_clip.CLIPVisionEmbeddings.__init__ with CLIP->CLIPSeg
+    def __init__(self, config: CLIPSegVisionConfig):
+        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=True) -> 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})."
+            )
+        patch_embeds = self.patch_embedding(pixel_values)  # 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
+
+
+# Copied from transformers.models.clip.modeling_clip.CLIPTextEmbeddings with CLIP->CLIPSeg
+class CLIPSegTextEmbeddings(nn.Module):
+    def __init__(self, config: CLIPSegTextConfig):
+        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
+
+
+# Copied from transformers.models.clip.modeling_clip.CLIPAttention with CLIP->CLIPSeg
+class CLIPSegAttention(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
+
+
+# Copied from transformers.models.clip.modeling_clip.CLIPMLP with CLIP->CLIPSeg
+class CLIPSegMLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.activation_fn = ACT2FN[config.hidden_act]
+        self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
+        self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.fc1(hidden_states)
+        hidden_states = self.activation_fn(hidden_states)
+        hidden_states = self.fc2(hidden_states)
+        return hidden_states
+
+
+# Copied from transformers.models.altclip.modeling_altclip.AltCLIPEncoderLayer with AltCLIP->CLIPSeg
+class CLIPSegEncoderLayer(nn.Module):
+    def __init__(self, config: CLIPSegConfig):
+        super().__init__()
+        self.embed_dim = config.hidden_size
+        self.self_attn = CLIPSegAttention(config)
+        self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
+        self.mlp = CLIPSegMLP(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 CLIPSegPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = CLIPSegConfig
+    base_model_prefix = "clip"
+    supports_gradient_checkpointing = True
+
+    def _init_weights(self, module):
+        """Initialize the weights"""
+        factor = self.config.initializer_factor
+        if isinstance(module, CLIPSegTextEmbeddings):
+            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, CLIPSegVisionEmbeddings):
+            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, CLIPSegAttention):
+            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, CLIPSegMLP):
+            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, CLIPSegModel):
+            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,
+            )
+
+        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_()
+
+
+CLIPSEG_START_DOCSTRING = r"""
+    This model is 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 ([`CLIPSegConfig`]): 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.
+"""
+
+CLIPSEG_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.
+"""
+
+CLIPSEG_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 to `True`):
+            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.
+"""
+
+CLIPSEG_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 to `True`):
+            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.
+"""
+
+
+# Copied from transformers.models.altclip.modeling_altclip.AltCLIPEncoder with AltCLIP->CLIPSeg
+class CLIPSegEncoder(nn.Module):
+    """
+    Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a
+    [`CLIPSegEncoderLayer`].
+
+    Args:
+        config: CLIPSegConfig
+    """
+
+    def __init__(self, config: CLIPSegConfig):
+        super().__init__()
+        self.config = config
+        self.layers = nn.ModuleList([CLIPSegEncoderLayer(config) for _ in range(config.num_hidden_layers)])
+        self.gradient_checkpointing = False
+
+    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,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutput]:
+        r"""
+        Args:
+            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
+                Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
+                This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+                than the model's internal embedding lookup matrix.
+            attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+                - 1 for tokens that are **not masked**,
+                - 0 for tokens that are **masked**.
+
+                [What are attention masks?](../glossary#attention-mask)
+            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
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        encoder_states = () if output_hidden_states else None
+        all_attentions = () if output_attentions else None
+
+        hidden_states = inputs_embeds
+        for idx, encoder_layer in enumerate(self.layers):
+            if output_hidden_states:
+                encoder_states = encoder_states + (hidden_states,)
+            if self.gradient_checkpointing and self.training:
+                layer_outputs = self._gradient_checkpointing_func(
+                    encoder_layer.__call__,
+                    hidden_states,
+                    attention_mask,
+                    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,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None)
+        return BaseModelOutput(
+            last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions
+        )
+
+
+class CLIPSegTextTransformer(nn.Module):
+    def __init__(self, config: CLIPSegTextConfig):
+        super().__init__()
+        self.config = config
+        embed_dim = config.hidden_size
+        self.embeddings = CLIPSegTextEmbeddings(config)
+        self.encoder = CLIPSegEncoder(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
+
+    @add_start_docstrings_to_model_forward(CLIPSEG_TEXT_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=CLIPSegTextConfig)
+    # Adapted from transformers.models.clip.modeling_clip.CLIPTextTransformer.forward with clip->clipseg, CLIP->CLIPSeg
+    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,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPooling]:
+        r"""
+        Returns:
+
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if 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)
+
+        # CLIPSeg's text model uses causal mask, prepare it here.
+        # https://github.com/openai/CLIPSeg/blob/cfcffb90e69f37bf2ff1e988237a0fbe41f33c04/clipseg/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:
+            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+            attention_mask = _prepare_4d_attention_mask(attention_mask, hidden_states.dtype)
+
+        encoder_outputs = 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,
+            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 CLIPSeg 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),
+            ]
+
+        if not return_dict:
+            return (last_hidden_state, pooled_output) + encoder_outputs[1:]
+
+        return BaseModelOutputWithPooling(
+            last_hidden_state=last_hidden_state,
+            pooler_output=pooled_output,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )
+
+
+class CLIPSegTextModel(CLIPSegPreTrainedModel):
+    config_class = CLIPSegTextConfig
+
+    _no_split_modules = ["CLIPSegTextEmbeddings", "CLIPSegEncoderLayer"]
+
+    def __init__(self, config: CLIPSegTextConfig):
+        super().__init__(config)
+        self.text_model = CLIPSegTextTransformer(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
+
+    @add_start_docstrings_to_model_forward(CLIPSEG_TEXT_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=CLIPSegTextConfig)
+    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,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPooling]:
+        r"""
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from transformers import AutoTokenizer, CLIPSegTextModel
+
+        >>> tokenizer = AutoTokenizer.from_pretrained("CIDAS/clipseg-rd64-refined")
+        >>> model = CLIPSegTextModel.from_pretrained("CIDAS/clipseg-rd64-refined")
+
+        >>> 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,
+            return_dict=return_dict,
+        )
+
+
+class CLIPSegVisionTransformer(nn.Module):
+    # Copied from transformers.models.altclip.modeling_altclip.AltCLIPVisionTransformer.__init__ with AltCLIP->CLIPSeg
+    def __init__(self, config: CLIPSegVisionConfig):
+        super().__init__()
+        self.config = config
+        embed_dim = config.hidden_size
+
+        self.embeddings = CLIPSegVisionEmbeddings(config)
+        self.pre_layrnorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
+        self.encoder = CLIPSegEncoder(config)
+        self.post_layernorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
+
+    @add_start_docstrings_to_model_forward(CLIPSEG_VISION_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=CLIPSegVisionConfig)
+    def forward(
+        self,
+        pixel_values: Optional[torch.FloatTensor],
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        interpolate_pos_encoding: Optional[bool] = True,
+    ) -> Union[Tuple, BaseModelOutputWithPooling]:
+        r"""
+        Returns:
+
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        hidden_states = self.embeddings(pixel_values, interpolate_pos_encoding=interpolate_pos_encoding)
+        hidden_states = self.pre_layrnorm(hidden_states)
+
+        encoder_outputs = self.encoder(
+            inputs_embeds=hidden_states,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        last_hidden_state = encoder_outputs[0]
+        pooled_output = last_hidden_state[:, 0, :]
+        pooled_output = self.post_layernorm(pooled_output)
+
+        if not return_dict:
+            return (last_hidden_state, pooled_output) + encoder_outputs[1:]
+
+        return BaseModelOutputWithPooling(
+            last_hidden_state=last_hidden_state,
+            pooler_output=pooled_output,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )
+
+
+class CLIPSegVisionModel(CLIPSegPreTrainedModel):
+    config_class = CLIPSegVisionConfig
+    main_input_name = "pixel_values"
+
+    def __init__(self, config: CLIPSegVisionConfig):
+        super().__init__(config)
+        self.vision_model = CLIPSegVisionTransformer(config)
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self) -> nn.Module:
+        return self.vision_model.embeddings.patch_embedding
+
+    @add_start_docstrings_to_model_forward(CLIPSEG_VISION_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=CLIPSegVisionConfig)
+    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] = True,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPooling]:
+        r"""
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from PIL import Image
+        >>> import requests
+        >>> from transformers import AutoProcessor, CLIPSegVisionModel
+
+        >>> processor = AutoProcessor.from_pretrained("CIDAS/clipseg-rd64-refined")
+        >>> model = CLIPSegVisionModel.from_pretrained("CIDAS/clipseg-rd64-refined")
+
+        >>> 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,
+            return_dict=return_dict,
+        )
+
+
+@add_start_docstrings(CLIPSEG_START_DOCSTRING)
+class CLIPSegModel(CLIPSegPreTrainedModel):
+    config_class = CLIPSegConfig
+
+    def __init__(self, config: CLIPSegConfig):
+        super().__init__(config)
+
+        if not isinstance(config.text_config, CLIPSegTextConfig):
+            raise TypeError(
+                "config.text_config is expected to be of type CLIPSegTextConfig but is of type"
+                f" {type(config.text_config)}."
+            )
+
+        if not isinstance(config.vision_config, CLIPSegVisionConfig):
+            raise TypeError(
+                "config.vision_config is expected to be of type CLIPSegVisionConfig 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
+
+        self.text_model = CLIPSegTextTransformer(text_config)
+        self.vision_model = CLIPSegVisionTransformer(vision_config)
+
+        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(CLIPSEG_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,
+        return_dict: 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 [`CLIPSegTextModel`].
+
+        Examples:
+
+        ```python
+        >>> from transformers import AutoTokenizer, CLIPSegModel
+
+        >>> tokenizer = AutoTokenizer.from_pretrained("CIDAS/clipseg-rd64-refined")
+        >>> model = CLIPSegModel.from_pretrained("CIDAS/clipseg-rd64-refined")
+
+        >>> 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 CLIPSEG 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
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        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,
+        )
+
+        pooled_output = text_outputs[1]
+        text_features = self.text_projection(pooled_output)
+
+        return text_features
+
+    @add_start_docstrings_to_model_forward(CLIPSEG_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 = True,
+        return_dict: Optional[bool] = None,
+    ) -> 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 [`CLIPSegVisionModel`].
+
+        Examples:
+
+        ```python
+        >>> from PIL import Image
+        >>> import requests
+        >>> from transformers import AutoProcessor, CLIPSegModel
+
+        >>> processor = AutoProcessor.from_pretrained("CIDAS/clipseg-rd64-refined")
+        >>> model = CLIPSegModel.from_pretrained("CIDAS/clipseg-rd64-refined")
+
+        >>> 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 CLIPSEG 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
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        vision_outputs = self.vision_model(
+            pixel_values=pixel_values,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            interpolate_pos_encoding=interpolate_pos_encoding,
+            return_dict=return_dict,
+        )
+
+        pooled_output = vision_outputs[1]  # pooled_output
+        image_features = self.visual_projection(pooled_output)
+
+        return image_features
+
+    @add_start_docstrings_to_model_forward(CLIPSEG_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=CLIPSegOutput, config_class=CLIPSegConfig)
+    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 = True,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, CLIPSegOutput]:
+        r"""
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from PIL import Image
+        >>> import requests
+        >>> from transformers import AutoProcessor, CLIPSegModel
+
+        >>> processor = AutoProcessor.from_pretrained("CIDAS/clipseg-rd64-refined")
+        >>> model = CLIPSegModel.from_pretrained("CIDAS/clipseg-rd64-refined")
+
+        >>> 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 CLIPSEG 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
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        vision_outputs = self.vision_model(
+            pixel_values=pixel_values,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            interpolate_pos_encoding=interpolate_pos_encoding,
+            return_dict=return_dict,
+        )
+
+        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,
+        )
+
+        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 / image_embeds.norm(p=2, dim=-1, keepdim=True)
+        text_embeds = text_embeds / text_embeds.norm(p=2, dim=-1, keepdim=True)
+
+        # cosine similarity as logits
+        logit_scale = self.logit_scale.exp()
+        logits_per_text = torch.matmul(text_embeds, image_embeds.t()) * logit_scale
+        logits_per_image = logits_per_text.t()
+
+        loss = None
+        if return_loss:
+            loss = clipseg_loss(logits_per_text)
+
+        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 CLIPSegOutput(
+            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 CLIPSegDecoderLayer(nn.Module):
+    """
+    CLIPSeg decoder layer, which is identical to `CLIPSegEncoderLayer`, except that normalization is applied after
+    self-attention/MLP, rather than before.
+    """
+
+    # Copied from transformers.models.altclip.modeling_altclip.AltCLIPEncoderLayer.__init__ with AltCLIP->CLIPSeg
+    def __init__(self, config: CLIPSegConfig):
+        super().__init__()
+        self.embed_dim = config.hidden_size
+        self.self_attn = CLIPSegAttention(config)
+        self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
+        self.mlp = CLIPSegMLP(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, 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
+        hidden_states = self.layer_norm1(hidden_states)
+
+        residual = hidden_states
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+        hidden_states = self.layer_norm2(hidden_states)
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (attn_weights,)
+
+        return outputs
+
+
+class CLIPSegDecoder(CLIPSegPreTrainedModel):
+    def __init__(self, config: CLIPSegConfig):
+        super().__init__(config)
+
+        self.conditional_layer = config.conditional_layer
+
+        self.film_mul = nn.Linear(config.projection_dim, config.reduce_dim)
+        self.film_add = nn.Linear(config.projection_dim, config.reduce_dim)
+
+        if config.use_complex_transposed_convolution:
+            transposed_kernels = (config.vision_config.patch_size // 4, config.vision_config.patch_size // 4)
+
+            self.transposed_convolution = nn.Sequential(
+                nn.Conv2d(config.reduce_dim, config.reduce_dim, kernel_size=3, padding=1),
+                nn.ReLU(),
+                nn.ConvTranspose2d(
+                    config.reduce_dim,
+                    config.reduce_dim // 2,
+                    kernel_size=transposed_kernels[0],
+                    stride=transposed_kernels[0],
+                ),
+                nn.ReLU(),
+                nn.ConvTranspose2d(
+                    config.reduce_dim // 2, 1, kernel_size=transposed_kernels[1], stride=transposed_kernels[1]
+                ),
+            )
+        else:
+            self.transposed_convolution = nn.ConvTranspose2d(
+                config.reduce_dim, 1, config.vision_config.patch_size, stride=config.vision_config.patch_size
+            )
+
+        depth = len(config.extract_layers)
+        self.reduces = nn.ModuleList(
+            [nn.Linear(config.vision_config.hidden_size, config.reduce_dim) for _ in range(depth)]
+        )
+
+        decoder_config = copy.deepcopy(config.vision_config)
+        decoder_config.hidden_size = config.reduce_dim
+        decoder_config.num_attention_heads = config.decoder_num_attention_heads
+        decoder_config.intermediate_size = config.decoder_intermediate_size
+        decoder_config.hidden_act = "relu"
+        self.layers = nn.ModuleList([CLIPSegDecoderLayer(decoder_config) for _ in range(len(config.extract_layers))])
+
+    def forward(
+        self,
+        hidden_states: Tuple[torch.Tensor],
+        conditional_embeddings: torch.Tensor,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = True,
+    ):
+        all_hidden_states = () if output_hidden_states else None
+        all_attentions = () if output_attentions else None
+
+        activations = hidden_states[::-1]
+
+        output = None
+        for i, (activation, layer, reduce) in enumerate(zip(activations, self.layers, self.reduces)):
+            if output is not None:
+                output = reduce(activation) + output
+            else:
+                output = reduce(activation)
+
+            if i == self.conditional_layer:
+                output = self.film_mul(conditional_embeddings) * output.permute(1, 0, 2) + self.film_add(
+                    conditional_embeddings
+                )
+                output = output.permute(1, 0, 2)
+
+            layer_outputs = layer(
+                output, attention_mask=None, causal_attention_mask=None, output_attentions=output_attentions
+            )
+
+            output = layer_outputs[0]
+
+            if output_hidden_states:
+                all_hidden_states += (output,)
+
+            if output_attentions:
+                all_attentions += (layer_outputs[1],)
+
+        output = output[:, 1:, :].permute(0, 2, 1)  # remove cls token and reshape to [batch_size, reduce_dim, seq_len]
+
+        size = int(math.sqrt(output.shape[2]))
+
+        batch_size = conditional_embeddings.shape[0]
+        output = output.view(batch_size, output.shape[1], size, size)
+
+        logits = self.transposed_convolution(output).squeeze(1)
+
+        if not return_dict:
+            return tuple(v for v in [logits, all_hidden_states, all_attentions] if v is not None)
+
+        return CLIPSegDecoderOutput(
+            logits=logits,
+            hidden_states=all_hidden_states,
+            attentions=all_attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    CLIPSeg model with a Transformer-based decoder on top for zero-shot and one-shot image segmentation.
+    """,
+    CLIPSEG_START_DOCSTRING,
+)
+class CLIPSegForImageSegmentation(CLIPSegPreTrainedModel):
+    config_class = CLIPSegConfig
+
+    def __init__(self, config: CLIPSegConfig):
+        super().__init__(config)
+
+        self.config = config
+
+        self.clip = CLIPSegModel(config)
+        self.extract_layers = config.extract_layers
+
+        self.decoder = CLIPSegDecoder(config)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_conditional_embeddings(
+        self,
+        batch_size: Optional[int] = None,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        conditional_pixel_values: Optional[torch.Tensor] = None,
+    ):
+        if input_ids is not None:
+            # compute conditional embeddings from texts
+            if len(input_ids) != batch_size:
+                raise ValueError("Make sure to pass as many prompt texts as there are query images")
+            with torch.no_grad():
+                conditional_embeddings = self.clip.get_text_features(
+                    input_ids, attention_mask=attention_mask, position_ids=position_ids
+                )
+        elif conditional_pixel_values is not None:
+            # compute conditional embeddings from images
+            if len(conditional_pixel_values) != batch_size:
+                raise ValueError("Make sure to pass as many prompt images as there are query images")
+            with torch.no_grad():
+                conditional_embeddings = self.clip.get_image_features(conditional_pixel_values)
+        else:
+            raise ValueError(
+                "Invalid conditional, should be either provided as `input_ids` or `conditional_pixel_values`"
+            )
+
+        return conditional_embeddings
+
+    @add_start_docstrings_to_model_forward(CLIPSEG_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=CLIPSegImageSegmentationOutput, config_class=CLIPSegTextConfig)
+    def forward(
+        self,
+        input_ids: Optional[torch.FloatTensor] = None,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        conditional_pixel_values: Optional[torch.FloatTensor] = None,
+        conditional_embeddings: Optional[torch.FloatTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        interpolate_pos_encoding: bool = True,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, CLIPSegOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence 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).
+
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from transformers import AutoProcessor, CLIPSegForImageSegmentation
+        >>> from PIL import Image
+        >>> import requests
+
+        >>> processor = AutoProcessor.from_pretrained("CIDAS/clipseg-rd64-refined")
+        >>> model = CLIPSegForImageSegmentation.from_pretrained("CIDAS/clipseg-rd64-refined")
+
+        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> image = Image.open(requests.get(url, stream=True).raw)
+        >>> texts = ["a cat", "a remote", "a blanket"]
+        >>> inputs = processor(text=texts, images=[image] * len(texts), padding=True, return_tensors="pt")
+
+        >>> outputs = model(**inputs)
+
+        >>> logits = outputs.logits
+        >>> print(logits.shape)
+        torch.Size([3, 352, 352])
+        ```"""
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # step 1: forward the query images through the frozen CLIP vision encoder
+        with torch.no_grad():
+            vision_outputs = self.clip.vision_model(
+                pixel_values=pixel_values,
+                output_attentions=output_attentions,
+                output_hidden_states=True,  # we need the intermediate hidden states
+                interpolate_pos_encoding=interpolate_pos_encoding,
+                return_dict=return_dict,
+            )
+            pooled_output = self.clip.visual_projection(vision_outputs[1])
+
+            hidden_states = vision_outputs.hidden_states if return_dict else vision_outputs[2]
+            # we add +1 here as the hidden states also include the initial embeddings
+            activations = [hidden_states[i + 1] for i in self.extract_layers]
+
+            # update vision_outputs
+            if return_dict:
+                vision_outputs = BaseModelOutputWithPooling(
+                    last_hidden_state=vision_outputs.last_hidden_state,
+                    pooler_output=vision_outputs.pooler_output,
+                    hidden_states=vision_outputs.hidden_states if output_hidden_states else None,
+                    attentions=vision_outputs.attentions,
+                )
+            else:
+                vision_outputs = (
+                    vision_outputs[:2] + vision_outputs[3:] if not output_hidden_states else vision_outputs
+                )
+
+        # step 2: compute conditional embeddings, either from text, images or an own provided embedding
+        if conditional_embeddings is None:
+            conditional_embeddings = self.get_conditional_embeddings(
+                batch_size=pixel_values.shape[0],
+                input_ids=input_ids,
+                attention_mask=attention_mask,
+                position_ids=position_ids,
+                conditional_pixel_values=conditional_pixel_values,
+            )
+        else:
+            if conditional_embeddings.shape[0] != pixel_values.shape[0]:
+                raise ValueError(
+                    "Make sure to pass as many conditional embeddings as there are query images in the batch"
+                )
+            if conditional_embeddings.shape[1] != self.config.projection_dim:
+                raise ValueError(
+                    "Make sure that the feature dimension of the conditional embeddings matches"
+                    " `config.projection_dim`."
+                )
+
+        # step 3: forward both the pooled output and the activations through the lightweight decoder to predict masks
+        decoder_outputs = self.decoder(
+            activations,
+            conditional_embeddings,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        logits = decoder_outputs.logits if return_dict else decoder_outputs[0]
+
+        loss = None
+        if labels is not None:
+            # move labels to the correct device to enable PP
+            labels = labels.to(logits.device)
+            loss_fn = nn.BCEWithLogitsLoss()
+            loss = loss_fn(logits, labels)
+
+        if not return_dict:
+            output = (logits, conditional_embeddings, pooled_output, vision_outputs, decoder_outputs)
+            return ((loss,) + output) if loss is not None else output
+
+        return CLIPSegImageSegmentationOutput(
+            loss=loss,
+            logits=logits,
+            conditional_embeddings=conditional_embeddings,
+            pooled_output=pooled_output,
+            vision_model_output=vision_outputs,
+            decoder_output=decoder_outputs,
+        )
+
+
+__all__ = [
+    "CLIPSegModel",
+    "CLIPSegPreTrainedModel",
+    "CLIPSegTextModel",
+    "CLIPSegVisionModel",
+    "CLIPSegForImageSegmentation",
+]

lib/python3.12/site-packages/transformers/models/clipseg/processing_clipseg.py

@@ -0,0 +1,164 @@
+# coding=utf-8
+# Copyright 2022 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 CLIPSeg
+"""
+
+import warnings
+
+from ...processing_utils import ProcessorMixin
+from ...tokenization_utils_base import BatchEncoding
+
+
+class CLIPSegProcessor(ProcessorMixin):
+    r"""
+    Constructs a CLIPSeg processor which wraps a CLIPSeg image processor and a CLIP tokenizer into a single processor.
+
+    [`CLIPSegProcessor`] offers all the functionalities of [`ViTImageProcessor`] and [`CLIPTokenizerFast`]. See the
+    [`~CLIPSegProcessor.__call__`] and [`~CLIPSegProcessor.decode`] for more information.
+
+    Args:
+        image_processor ([`ViTImageProcessor`], *optional*):
+            The image processor is a required input.
+        tokenizer ([`CLIPTokenizerFast`], *optional*):
+            The tokenizer is a required input.
+    """
+
+    attributes = ["image_processor", "tokenizer"]
+    image_processor_class = ("ViTImageProcessor", "ViTImageProcessorFast")
+    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, visual_prompt=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
+        ViTImageProcessor's [`~ViTImageProcessor.__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.
+            visual_prompt (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, `List[PIL.Image.Image]`, `List[np.ndarray]`, `List[torch.Tensor]`):
+                The visual prompt image or batch of images to be prepared. Each visual prompt image can be a PIL image,
+                NumPy array or PyTorch tensor. In case of a NumPy array/PyTorch tensor, each image should be of shape
+                (C, H, W), where C is a number of channels, H and W are image height and width.
+
+            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`.
+        """
+        if text is None and visual_prompt is None and images is None:
+            raise ValueError("You have to specify either text, visual prompt or images.")
+
+        if text is not None and visual_prompt is not None:
+            raise ValueError("You have to specify exactly one type of prompt. Either text or visual prompt.")
+
+        if text is not None:
+            encoding = self.tokenizer(text, return_tensors=return_tensors, **kwargs)
+
+        if visual_prompt is not None:
+            prompt_features = self.image_processor(visual_prompt, return_tensors=return_tensors, **kwargs)
+
+        if images is not None:
+            image_features = self.image_processor(images, return_tensors=return_tensors, **kwargs)
+
+        if visual_prompt is not None and images is not None:
+            encoding = {
+                "pixel_values": image_features.pixel_values,
+                "conditional_pixel_values": prompt_features.pixel_values,
+            }
+            return encoding
+        elif 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
+        elif visual_prompt is not None:
+            encoding = {
+                "conditional_pixel_values": prompt_features.pixel_values,
+            }
+            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 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__ = ["CLIPSegProcessor"]

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

@@ -0,0 +1,29 @@
+# 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_codegen import *
+    from .modeling_codegen import *
+    from .tokenization_codegen import *
+    from .tokenization_codegen_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/codegen/configuration_codegen.py

@@ -0,0 +1,230 @@
+# coding=utf-8
+# Copyright 2022 Salesforce authors, The EleutherAI, and HuggingFace Teams. 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.
+"""CodeGen model configuration"""
+
+from collections import OrderedDict
+from typing import Any, List, Mapping, Optional
+
+from ... import PreTrainedTokenizer, TensorType, is_torch_available
+from ...configuration_utils import PretrainedConfig
+from ...onnx import OnnxConfigWithPast, PatchingSpec
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class CodeGenConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`CodeGenModel`]. It is used to instantiate a
+    CodeGen 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 CodeGen
+    [Salesforce/codegen-2B-mono](https://huggingface.co/Salesforce/codegen-2B-mono) 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 50400):
+            Vocabulary size of the CodeGen model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`CodeGenModel`].
+        n_positions (`int`, *optional*, defaults to 2048):
+            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).
+        n_ctx (`int`, *optional*, defaults to 2048):
+            This attribute is used in `CodeGenModel.__init__` without any real effect.
+        n_embd (`int`, *optional*, defaults to 4096):
+            Dimensionality of the embeddings and hidden states.
+        n_layer (`int`, *optional*, defaults to 28):
+            Number of hidden layers in the Transformer encoder.
+        n_head (`int`, *optional*, defaults to 16):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        rotary_dim (`int`, *optional*, defaults to 64):
+            Number of dimensions in the embedding that Rotary Position Embedding is applied to.
+        n_inner (`int`, *optional*):
+            Dimensionality of the inner feed-forward layers. `None` will set it to 4 times n_embd
+        activation_function (`str`, *optional*, defaults to `"gelu_new"`):
+            Activation function, to be selected in the list `["relu", "silu", "gelu", "tanh", "gelu_new"]`.
+        resid_pdrop (`float`, *optional*, defaults to 0.0):
+            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
+        embd_pdrop (`int`, *optional*, defaults to 0.0):
+            The dropout ratio for the embeddings.
+        attn_pdrop (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention.
+        layer_norm_epsilon (`float`, *optional*, defaults to 1e-05):
+            The epsilon to use in the layer normalization layers.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models).
+        bos_token_id (`int`, *optional*, defaults to 50256):
+            Beginning of stream token id.
+        eos_token_id (`int`, *optional*, defaults to 50256):
+            End of stream token id.
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether the model's input and output word embeddings should be tied. Note that this is only relevant if the
+            model has a output word embedding layer.
+
+    Example:
+
+    ```python
+    >>> from transformers import CodeGenConfig, CodeGenModel
+
+    >>> # Initializing a CodeGen 6B configuration
+    >>> configuration = CodeGenConfig()
+
+    >>> # Initializing a model (with random weights) from the configuration
+    >>> model = CodeGenModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "codegen"
+    attribute_map = {
+        "max_position_embeddings": "n_positions",
+        "hidden_size": "n_embd",
+        "num_attention_heads": "n_head",
+        "num_hidden_layers": "n_layer",
+    }
+
+    def __init__(
+        self,
+        vocab_size=50400,
+        n_positions=2048,
+        n_ctx=2048,
+        n_embd=4096,
+        n_layer=28,
+        n_head=16,
+        rotary_dim=64,
+        n_inner=None,
+        activation_function="gelu_new",
+        resid_pdrop=0.0,
+        embd_pdrop=0.0,
+        attn_pdrop=0.0,
+        layer_norm_epsilon=1e-5,
+        initializer_range=0.02,
+        use_cache=True,
+        bos_token_id=50256,
+        eos_token_id=50256,
+        tie_word_embeddings=False,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.n_ctx = n_ctx
+        self.n_positions = n_positions
+        self.n_embd = n_embd
+        self.n_layer = n_layer
+        self.n_head = n_head
+        self.n_inner = n_inner
+        self.rotary_dim = rotary_dim
+        self.activation_function = activation_function
+        self.resid_pdrop = resid_pdrop
+        self.embd_pdrop = embd_pdrop
+        self.attn_pdrop = attn_pdrop
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.initializer_range = initializer_range
+        self.use_cache = use_cache
+
+        self.bos_token_id = bos_token_id
+        self.eos_token_id = eos_token_id
+
+        super().__init__(
+            bos_token_id=bos_token_id, eos_token_id=eos_token_id, tie_word_embeddings=tie_word_embeddings, **kwargs
+        )
+
+
+# Copied from transformers.models.gpt2.configuration_gpt2.GPT2OnnxConfig
+class CodeGenOnnxConfig(OnnxConfigWithPast):
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        task: str = "default",
+        patching_specs: List[PatchingSpec] = None,
+        use_past: bool = False,
+    ):
+        super().__init__(config, task=task, patching_specs=patching_specs, use_past=use_past)
+        if not getattr(self._config, "pad_token_id", None):
+            # TODO: how to do that better?
+            self._config.pad_token_id = 0
+
+    @property
+    def inputs(self) -> Mapping[str, Mapping[int, str]]:
+        common_inputs = OrderedDict({"input_ids": {0: "batch", 1: "sequence"}})
+        if self.use_past:
+            self.fill_with_past_key_values_(common_inputs, direction="inputs")
+            common_inputs["attention_mask"] = {0: "batch", 1: "past_sequence + sequence"}
+        else:
+            common_inputs["attention_mask"] = {0: "batch", 1: "sequence"}
+
+        return common_inputs
+
+    @property
+    def num_layers(self) -> int:
+        return self._config.n_layer
+
+    @property
+    def num_attention_heads(self) -> int:
+        return self._config.n_head
+
+    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]:
+        common_inputs = super(OnnxConfigWithPast, self).generate_dummy_inputs(
+            tokenizer, batch_size=batch_size, seq_length=seq_length, is_pair=is_pair, framework=framework
+        )
+
+        # We need to order the input in the way they appears in the forward()
+        ordered_inputs = OrderedDict({"input_ids": common_inputs["input_ids"]})
+
+        # Need to add the past_keys
+        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
+                past_shape = (
+                    batch,
+                    self.num_attention_heads,
+                    past_key_values_length,
+                    self._config.hidden_size // self.num_attention_heads,
+                )
+                ordered_inputs["past_key_values"] = [
+                    (torch.zeros(past_shape), torch.zeros(past_shape)) for _ in range(self.num_layers)
+                ]
+
+        ordered_inputs["attention_mask"] = common_inputs["attention_mask"]
+        if self.use_past:
+            mask_dtype = ordered_inputs["attention_mask"].dtype
+            ordered_inputs["attention_mask"] = torch.cat(
+                [ordered_inputs["attention_mask"], torch.ones(batch, past_key_values_length, dtype=mask_dtype)], dim=1
+            )
+
+        return ordered_inputs
+
+    @property
+    def default_onnx_opset(self) -> int:
+        return 13
+
+
+__all__ = ["CodeGenConfig", "CodeGenOnnxConfig"]

lib/python3.12/site-packages/transformers/models/codegen/modeling_codegen.py

@@ -0,0 +1,835 @@
+# coding=utf-8
+# Copyright 2022 Salesforce authors, The EleutherAI, and HuggingFace Teams. 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 CodeGen model."""
+
+from typing import Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+
+from ...activations import ACT2FN
+from ...cache_utils import Cache, DynamicCache, StaticCache
+from ...generation import GenerationMixin
+from ...modeling_attn_mask_utils import AttentionMaskConverter
+from ...modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
+from ...modeling_utils import PreTrainedModel
+from ...utils import (
+    add_code_sample_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    is_torch_flex_attn_available,
+    logging,
+)
+from .configuration_codegen import CodeGenConfig
+
+
+if is_torch_flex_attn_available():
+    from torch.nn.attention.flex_attention import BlockMask
+
+    from ...integrations.flex_attention import make_flex_block_causal_mask
+
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "Salesforce/codegen-2B-mono"
+_CONFIG_FOR_DOC = "CodeGenConfig"
+
+
+# Copied from transformers.models.gptj.modeling_gptj.create_sinusoidal_positions
+def create_sinusoidal_positions(num_pos: int, dim: int) -> torch.Tensor:
+    inv_freq = 1.0 / (10000 ** (torch.arange(0, dim, 2, dtype=torch.int64) / dim))
+    sinusoid_inp = torch.einsum("i , j -> i j", torch.arange(num_pos, dtype=torch.int64).float(), inv_freq).float()
+    return torch.cat((torch.sin(sinusoid_inp), torch.cos(sinusoid_inp)), dim=1)
+
+
+# Copied from transformers.models.gptj.modeling_gptj.rotate_every_two
+def rotate_every_two(x: torch.Tensor) -> torch.Tensor:
+    x1 = x[:, :, :, ::2]
+    x2 = x[:, :, :, 1::2]
+    x = torch.stack((-x2, x1), dim=-1)
+    return x.flatten(-2)  # in einsum notation: rearrange(x, '... d j -> ... (d j)')
+
+
+# Copied from transformers.models.gptj.modeling_gptj.apply_rotary_pos_emb
+def apply_rotary_pos_emb(tensor: torch.Tensor, sin: torch.Tensor, cos: torch.Tensor) -> torch.Tensor:
+    sin = torch.repeat_interleave(sin[:, :, None, :], 2, 3)
+    cos = torch.repeat_interleave(cos[:, :, None, :], 2, 3)
+    return (tensor * cos) + (rotate_every_two(tensor) * sin)
+
+
+class CodeGenAttention(nn.Module):
+    def __init__(self, config, layer_idx=None):
+        super().__init__()
+
+        max_positions = config.max_position_embeddings
+        self.attn_dropout = nn.Dropout(config.attn_pdrop)
+        self.resid_dropout = nn.Dropout(config.resid_pdrop)
+        self.layer_idx = layer_idx
+        if layer_idx is None:
+            logger.warning_once(
+                f"Instantiating {self.__class__.__name__} without passing a `layer_idx` is not recommended and will "
+                "lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` "
+                "when creating this class."
+            )
+
+        self.embed_dim = config.hidden_size
+        self.num_attention_heads = config.num_attention_heads
+        self.head_dim = self.embed_dim // self.num_attention_heads
+        if self.head_dim * self.num_attention_heads != self.embed_dim:
+            raise ValueError(
+                f"embed_dim must be divisible by num_attention_heads (got `embed_dim`: {self.embed_dim} and"
+                f" `num_attention_heads`: {self.num_attention_heads})."
+            )
+        self.scale_attn = torch.sqrt(torch.tensor(self.head_dim, dtype=torch.float32)).to(torch.get_default_dtype())
+        self.qkv_proj = nn.Linear(self.embed_dim, self.embed_dim * 3, bias=False)
+
+        self.out_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=False)
+        self.rotary_dim = config.rotary_dim
+        pos_embd_dim = self.rotary_dim or self.embed_dim
+        self.embed_positions = create_sinusoidal_positions(max_positions, pos_embd_dim)
+
+    def _split_heads(self, x, n_head, dim_head, mp_num):
+        reshaped = x.reshape(x.shape[:-1] + (n_head // mp_num, dim_head))
+        reshaped = reshaped.reshape(x.shape[:-2] + (-1,) + reshaped.shape[-1:])
+        return reshaped
+
+    def _merge_heads(self, tensor, num_attention_heads, attn_head_size):
+        """
+        Merges attn_head_size dim and num_attn_heads dim into n_ctx
+        """
+        if len(tensor.shape) == 5:
+            tensor = tensor.permute(0, 1, 3, 2, 4).contiguous()
+        elif len(tensor.shape) == 4:
+            tensor = tensor.permute(0, 2, 1, 3).contiguous()
+        else:
+            raise ValueError(f"Input tensor rank should be one of [4, 5], but is: {len(tensor.shape)}")
+        new_shape = tensor.size()[:-2] + (num_attention_heads * attn_head_size,)
+        return tensor.view(new_shape)
+
+    def _attn(
+        self,
+        query,
+        key,
+        value,
+        attention_mask=None,
+        head_mask=None,
+    ):
+        # Keep the attention weights computation in fp32 to avoid overflow issues
+        query = query.to(torch.float32)
+        key = key.to(torch.float32)
+
+        attn_weights = torch.matmul(query, key.transpose(-1, -2))
+
+        if attention_mask is not None:
+            causal_mask = attention_mask[:, :, :, : key.shape[-2]]
+            attn_weights += causal_mask
+
+        attn_weights = attn_weights / self.scale_attn
+        attn_weights = nn.Softmax(dim=-1)(attn_weights)
+        attn_weights = attn_weights.to(value.dtype)
+        attn_weights = self.attn_dropout(attn_weights)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attn_weights = attn_weights * head_mask
+
+        attn_output = torch.matmul(attn_weights, value)
+
+        return attn_output, attn_weights
+
+    def forward(
+        self,
+        hidden_states: Optional[torch.FloatTensor],
+        layer_past: Optional[Cache] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = False,
+        output_attentions: Optional[bool] = False,
+        cache_position: Optional[torch.LongTensor] = None,
+    ) -> Union[
+        Tuple[torch.Tensor, Tuple[torch.Tensor]],
+        Optional[Tuple[torch.Tensor, Tuple[torch.Tensor], Tuple[torch.Tensor, ...]]],
+    ]:
+        qkv = self.qkv_proj(hidden_states)
+        # TODO(enijkamp): factor out number of logical TPU-v4 cores or make forward pass agnostic
+        mp_num = 4
+        qkv_split = qkv.reshape(qkv.shape[:-1] + (mp_num, -1))
+
+        local_dim = self.head_dim * self.num_attention_heads // mp_num
+        query, value, key = torch.split(qkv_split, local_dim, dim=-1)
+        query = self._split_heads(query, self.num_attention_heads, self.head_dim, mp_num=mp_num)
+        key = self._split_heads(key, self.num_attention_heads, self.head_dim, mp_num=mp_num)
+
+        value = self._split_heads(value, self.num_attention_heads, self.head_dim, mp_num=mp_num)
+        value = value.permute(0, 2, 1, 3)
+
+        embed_positions = self.embed_positions
+        if embed_positions.device != position_ids.device:
+            embed_positions = embed_positions.to(position_ids.device)
+            self.embed_positions = embed_positions
+
+        sincos = embed_positions[position_ids]
+        sin, cos = torch.split(sincos, sincos.shape[-1] // 2, dim=-1)
+
+        if self.rotary_dim is not None:
+            k_rot = key[:, :, :, : self.rotary_dim]
+            k_pass = key[:, :, :, self.rotary_dim :]
+
+            q_rot = query[:, :, :, : self.rotary_dim]
+            q_pass = query[:, :, :, self.rotary_dim :]
+
+            k_rot = apply_rotary_pos_emb(k_rot, sin, cos)
+            q_rot = apply_rotary_pos_emb(q_rot, sin, cos)
+
+            key = torch.cat([k_rot, k_pass], dim=-1)
+            query = torch.cat([q_rot, q_pass], dim=-1)
+        else:
+            key = apply_rotary_pos_emb(key, sin, cos)
+            query = apply_rotary_pos_emb(query, sin, cos)
+
+        key = key.permute(0, 2, 1, 3)
+        query = query.permute(0, 2, 1, 3)
+
+        # Note that this cast is quite ugly, but is not implemented before ROPE as k_rot in the original codebase is always in fp32.
+        # Reference: https://github.com/salesforce/CodeGen/blob/f210c3bb1216c975ad858cd4132c0fdeabf4bfc2/codegen1/jaxformer/hf/codegen/modeling_codegen.py#L38
+        if layer_past is not None:
+            cache_kwargs = {
+                "sin": sin,
+                "cos": cos,
+                "partial_rotation_size": self.rotary_dim,
+                "cache_position": cache_position,
+            }
+            key, value = layer_past.update(key.to(hidden_states.dtype), value, self.layer_idx, cache_kwargs)
+
+        # compute self-attention: V x Softmax(QK^T)
+        attn_output, attn_weights = self._attn(query, key, value, attention_mask, head_mask)
+
+        attn_output = self._merge_heads(attn_output, self.num_attention_heads, self.head_dim)
+        attn_output = self.out_proj(attn_output)
+        attn_output = self.resid_dropout(attn_output)
+
+        outputs = (attn_output, layer_past)
+        if output_attentions:
+            outputs += (attn_weights,)
+
+        return outputs  # a, present, (attentions)
+
+
+# Copied from transformers.models.gptj.modeling_gptj.GPTJMLP with GPTJ->CodeGen
+class CodeGenMLP(nn.Module):
+    def __init__(self, intermediate_size, config):  # in MLP: intermediate_size= 4 * embed_dim
+        super().__init__()
+        embed_dim = config.n_embd
+
+        self.fc_in = nn.Linear(embed_dim, intermediate_size)
+        self.fc_out = nn.Linear(intermediate_size, embed_dim)
+
+        self.act = ACT2FN[config.activation_function]
+        self.dropout = nn.Dropout(config.resid_pdrop)
+
+    def forward(self, hidden_states: Optional[torch.FloatTensor]) -> torch.FloatTensor:
+        hidden_states = self.fc_in(hidden_states)
+        hidden_states = self.act(hidden_states)
+        hidden_states = self.fc_out(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        return hidden_states
+
+
+# Copied from transformers.models.gptj.modeling_gptj.GPTJBlock with GPTJ->CodeGen
+class CodeGenBlock(nn.Module):
+    # Ignore copy
+    def __init__(self, config, layer_idx=None):
+        super().__init__()
+        inner_dim = config.n_inner if config.n_inner is not None else 4 * config.n_embd
+        self.ln_1 = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
+        self.attn = CodeGenAttention(config, layer_idx)
+        self.mlp = CodeGenMLP(inner_dim, config)
+
+    def forward(
+        self,
+        hidden_states: Optional[torch.FloatTensor],
+        layer_past: Optional[Cache] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = False,
+        output_attentions: Optional[bool] = False,
+        cache_position: Optional[torch.LongTensor] = None,
+    ) -> Union[Tuple[torch.Tensor], Optional[Tuple[torch.Tensor, Tuple[torch.FloatTensor, ...]]]]:
+        residual = hidden_states
+        hidden_states = self.ln_1(hidden_states)
+        attn_outputs = self.attn(
+            hidden_states=hidden_states,
+            layer_past=layer_past,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            cache_position=cache_position,
+        )
+        attn_output = attn_outputs[0]  # output_attn: a, present, (attentions)
+        outputs = attn_outputs[1:]
+
+        feed_forward_hidden_states = self.mlp(hidden_states)
+        hidden_states = attn_output + feed_forward_hidden_states + residual
+
+        if use_cache:
+            outputs = (hidden_states,) + outputs
+        else:
+            outputs = (hidden_states,) + outputs[1:]
+
+        return outputs  # hidden_states, present, (attentions)
+
+
+class CodeGenPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = CodeGenConfig
+    base_model_prefix = "transformer"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["CodeGenBlock"]
+    _skip_keys_device_placement = "past_key_values"
+    _supports_cache_class = True
+    _supports_quantized_cache = True
+    _supports_static_cache = True
+
+    def __init__(self, *inputs, **kwargs):
+        super().__init__(*inputs, **kwargs)
+
+    def _init_weights(self, module):
+        """Initialize the weights."""
+        if isinstance(module, (nn.Linear,)):
+            # Slightly different from Mesh Transformer JAX 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)
+
+
+CODEGEN_START_DOCSTRING = r"""
+    This model is a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) sub-class. Use
+    it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and
+    behavior.
+
+    Parameters:
+        config ([`CodeGenConfig`]): 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.
+"""
+
+CODEGEN_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `({0})`):
+            Indices of input sequence tokens in the vocabulary.
+
+            Indices can be obtained using [`AutoProcenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        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)
+        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.n_positions - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        head_mask (`torch.FloatTensor` of shape `(num_attention_heads,)` or `(n_layer, num_attention_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_dim)`, *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.
+        past_key_values (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*):
+            Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
+            blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values`
+            returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`.
+
+            Two formats are allowed:
+            - a [`~cache_utils.Cache`] instance, see our
+            [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache);
+            - 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)`). This is also known as the legacy
+            cache format.
+
+            The model will output the same cache format that is fed as input. If no `past_key_values` are passed, the
+            legacy cache format will be returned.
+
+            If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't
+            have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids`
+            of shape `(batch_size, sequence_length)`.
+        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 bare CodeGen Model transformer outputting raw hidden-states without any specific head on top.",
+    CODEGEN_START_DOCSTRING,
+)
+class CodeGenModel(CodeGenPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.embed_dim = config.n_embd
+        self.vocab_size = config.vocab_size
+        self.wte = nn.Embedding(config.vocab_size, self.embed_dim)
+        self.drop = nn.Dropout(config.embd_pdrop)
+        self.h = nn.ModuleList([CodeGenBlock(config, layer_idx=i) for i in range(config.n_layer)])
+        self.ln_f = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon)
+        self.rotary_dim = min(config.rotary_dim, config.n_ctx // config.num_attention_heads)
+
+        self.gradient_checkpointing = False
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.wte
+
+    def set_input_embeddings(self, new_embeddings):
+        self.wte = new_embeddings
+
+    @add_start_docstrings_to_model_forward(CODEGEN_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=BaseModelOutputWithPast,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Union[Cache, Tuple[Tuple[torch.Tensor]]]] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs,  # NOOP kwargs, for now
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                logger.warning_once(
+                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                )
+                use_cache = False
+
+        if inputs_embeds is None:
+            inputs_embeds = self.wte(input_ids)
+
+        # kept for BC (non `Cache` `past_key_values` inputs)
+        return_legacy_cache = False
+        if use_cache and not isinstance(past_key_values, Cache):
+            return_legacy_cache = True
+            if past_key_values is None:
+                past_key_values = DynamicCache()
+            else:
+                past_key_values = DynamicCache.from_legacy_cache(past_key_values)
+                logger.warning_once(
+                    "We detected that you are passing `past_key_values` as a tuple of tuples. This is deprecated and "
+                    "will be removed in v4.47. Please convert your cache or use an appropriate `Cache` class "
+                    "(https://huggingface.co/docs/transformers/kv_cache#legacy-cache-format)"
+                )
+
+        seq_length = inputs_embeds.shape[1]
+        if cache_position is None:
+            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+            cache_position = torch.arange(past_seen_tokens, past_seen_tokens + seq_length, device=inputs_embeds.device)
+
+        if position_ids is None:
+            position_ids = cache_position.unsqueeze(0)
+
+        causal_mask = self._update_causal_mask(
+            attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions
+        )
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x num_attention_heads x N x N
+        # head_mask has shape n_layer x batch x num_attention_heads x N x N
+        head_mask = self.get_head_mask(head_mask, self.config.n_layer)
+        hidden_states = inputs_embeds
+
+        if token_type_ids is not None:
+            token_type_ids = token_type_ids.view(-1, seq_length)
+            token_type_embeds = self.wte(token_type_ids)
+            hidden_states = hidden_states + token_type_embeds
+
+        hidden_states = self.drop(hidden_states)
+        output_shape = (-1, seq_length, hidden_states.size(-1))
+
+        next_decoder_cache = None
+        all_self_attentions = () if output_attentions else None
+        all_hidden_states = () if output_hidden_states else None
+        for i, block in enumerate(self.h):
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+                outputs = self._gradient_checkpointing_func(
+                    block.__call__,
+                    hidden_states,
+                    None,
+                    causal_mask,
+                    position_ids,
+                    head_mask[i],
+                    use_cache,
+                    output_attentions,
+                    cache_position,
+                )
+            else:
+                outputs = block(
+                    hidden_states=hidden_states,
+                    layer_past=past_key_values,
+                    attention_mask=causal_mask,
+                    position_ids=position_ids,
+                    head_mask=head_mask[i],
+                    use_cache=use_cache,
+                    output_attentions=output_attentions,
+                    cache_position=cache_position,
+                )
+
+            hidden_states = outputs[0]
+            if use_cache is True:
+                next_decoder_cache = outputs[1]
+
+            if output_attentions:
+                all_self_attentions = all_self_attentions + (outputs[2 if use_cache else 1],)
+
+        hidden_states = self.ln_f(hidden_states)
+
+        hidden_states = hidden_states.view(output_shape)
+        # Add last hidden state
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        next_cache = next_decoder_cache if use_cache else None
+        if return_legacy_cache:
+            next_cache = next_cache.to_legacy_cache()
+
+        if not return_dict:
+            return tuple(
+                v for v in [hidden_states, next_cache, all_hidden_states, all_self_attentions] if v is not None
+            )
+
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attentions,
+        )
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaModel._update_causal_mask
+    def _update_causal_mask(
+        self,
+        attention_mask: torch.Tensor,
+        input_tensor: torch.Tensor,
+        cache_position: torch.Tensor,
+        past_key_values: Cache,
+        output_attentions: bool = False,
+    ):
+        if self.config._attn_implementation == "flash_attention_2":
+            if attention_mask is not None and (attention_mask == 0.0).any():
+                return attention_mask
+            return None
+        if self.config._attn_implementation == "flex_attention":
+            if isinstance(attention_mask, torch.Tensor):
+                attention_mask = make_flex_block_causal_mask(attention_mask)
+            if isinstance(attention_mask, BlockMask):
+                return attention_mask
+
+        # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in
+        # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail
+        # to infer the attention mask.
+        past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+        using_static_cache = isinstance(past_key_values, StaticCache)
+
+        # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward
+        if self.config._attn_implementation == "sdpa" and not using_static_cache and not output_attentions:
+            if AttentionMaskConverter._ignore_causal_mask_sdpa(
+                attention_mask,
+                inputs_embeds=input_tensor,
+                past_key_values_length=past_seen_tokens,
+                is_training=self.training,
+            ):
+                return None
+
+        dtype, device = input_tensor.dtype, input_tensor.device
+        sequence_length = input_tensor.shape[1]
+        if using_static_cache:
+            target_length = past_key_values.get_max_cache_shape()
+        else:
+            target_length = (
+                attention_mask.shape[-1]
+                if isinstance(attention_mask, torch.Tensor)
+                else past_seen_tokens + sequence_length + 1
+            )
+
+        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
+        causal_mask = self._prepare_4d_causal_attention_mask_with_cache_position(
+            attention_mask,
+            sequence_length=sequence_length,
+            target_length=target_length,
+            dtype=dtype,
+            device=device,
+            cache_position=cache_position,
+            batch_size=input_tensor.shape[0],
+        )
+
+        if (
+            self.config._attn_implementation == "sdpa"
+            and attention_mask is not None
+            and attention_mask.device.type in ["cuda", "xpu"]
+            and not output_attentions
+        ):
+            # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
+            # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
+            # Details: https://github.com/pytorch/pytorch/issues/110213
+            min_dtype = torch.finfo(dtype).min
+            causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)
+
+        return causal_mask
+
+    @staticmethod
+    # Copied from transformers.models.llama.modeling_llama.LlamaPreTrainedModel._prepare_4d_causal_attention_mask_with_cache_position
+    def _prepare_4d_causal_attention_mask_with_cache_position(
+        attention_mask: torch.Tensor,
+        sequence_length: int,
+        target_length: int,
+        dtype: torch.dtype,
+        device: torch.device,
+        cache_position: torch.Tensor,
+        batch_size: int,
+        **kwargs,
+    ):
+        """
+        Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
+        `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.
+
+        Args:
+            attention_mask (`torch.Tensor`):
+                A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape
+                `(batch_size, 1, query_length, key_value_length)`.
+            sequence_length (`int`):
+                The sequence length being processed.
+            target_length (`int`):
+                The target length: when generating with static cache, the mask should be as long as the static cache,
+                to account for the 0 padding, the part of the cache that is not filled yet.
+            dtype (`torch.dtype`):
+                The dtype to use for the 4D attention mask.
+            device (`torch.device`):
+                The device to place the 4D attention mask on.
+            cache_position (`torch.Tensor`):
+                Indices depicting the position of the input sequence tokens in the sequence.
+            batch_size (`torch.Tensor`):
+                Batch size.
+        """
+        if attention_mask is not None and attention_mask.dim() == 4:
+            # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
+            causal_mask = attention_mask
+        else:
+            min_dtype = torch.finfo(dtype).min
+            causal_mask = torch.full(
+                (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device
+            )
+            if sequence_length != 1:
+                causal_mask = torch.triu(causal_mask, diagonal=1)
+            causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
+            causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
+            if attention_mask is not None:
+                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
+                mask_length = attention_mask.shape[-1]
+                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :].to(
+                    causal_mask.device
+                )
+                padding_mask = padding_mask == 0
+                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
+                    padding_mask, min_dtype
+                )
+
+        return causal_mask
+
+
+@add_start_docstrings(
+    """
+    The CodeGen Model transformer with a language modeling head on top.
+    """,
+    CODEGEN_START_DOCSTRING,
+)
+class CodeGenForCausalLM(CodeGenPreTrainedModel, GenerationMixin):
+    _tied_weights_keys = ["lm_head.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.transformer = CodeGenModel(config)
+        self.lm_head = nn.Linear(config.n_embd, config.vocab_size)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    @add_start_docstrings_to_model_forward(CODEGEN_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=CausalLMOutputWithPast,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Union[Cache, Tuple[Tuple[torch.Tensor]]]] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
+            `labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100`
+            are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]`
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        transformer_outputs = self.transformer(
+            input_ids,
+            past_key_values=past_key_values,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            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,
+        )
+        hidden_states = transformer_outputs[0]
+
+        # make sure sampling in fp16 works correctly and
+        # compute loss in fp32 to match with mesh-tf version
+        # https://github.com/EleutherAI/gpt-neo/blob/89ce74164da2fb16179106f54e2269b5da8db333/models/gpt2/gpt2.py#L179
+        lm_logits = self.lm_head(hidden_states).to(torch.float32)
+
+        loss = None
+        if labels is not None:
+            # move labels to correct device to enable model parallelism
+            labels = labels.to(lm_logits.device)
+            # Flatten the tokens
+            loss = self.loss_function(
+                lm_logits,
+                labels,
+                vocab_size=self.config.vocab_size,
+                **kwargs,
+            )
+
+            loss = loss.to(hidden_states.dtype)
+
+        if not return_dict:
+            output = (lm_logits,) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=lm_logits,
+            past_key_values=transformer_outputs.past_key_values,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
+
+    @staticmethod
+    def _reorder_cache(
+        past_key_values: Tuple[Tuple[torch.Tensor]], beam_idx: torch.Tensor
+    ) -> Tuple[Tuple[torch.Tensor]]:
+        """
+        This function is used to re-order the `past_key_values` cache if [`~PretrainedModel.beam_search`] or
+        [`~PretrainedModel.beam_sample`] is called. This is required to match `past_key_values` with the correct
+        beam_idx at every generation step.
+        """
+        return tuple(
+            tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past)
+            for layer_past in past_key_values
+        )
+
+
+__all__ = ["CodeGenForCausalLM", "CodeGenModel", "CodeGenPreTrainedModel"]

lib/python3.12/site-packages/transformers/models/codegen/tokenization_codegen.py

@@ -0,0 +1,419 @@
+# coding=utf-8
+# Copyright 2022 The Salesforce authors, 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 CodeGen"""
+
+import json
+import os
+from functools import lru_cache
+from typing import TYPE_CHECKING, List, Optional, Tuple, Union
+
+import numpy as np
+import regex as re
+
+from ...utils import is_tf_available, is_torch_available, logging, to_py_obj
+
+
+if TYPE_CHECKING:
+    if is_torch_available():
+        import torch
+    if is_tf_available():
+        import tensorflow as tf
+
+from ...tokenization_utils import AddedToken, PreTrainedTokenizer
+
+
+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
+
+
+class CodeGenTokenizer(PreTrainedTokenizer):
+    """
+    Construct a CodeGen tokenizer. Based on byte-level Byte-Pair-Encoding.
+
+    This tokenizer has been trained to treat spaces like parts of the tokens (a bit like sentencepiece) so a word will
+    be encoded differently whether it is at the beginning of the sentence (without space) or not:
+
+    ```python
+    >>> from transformers import CodeGenTokenizer
+
+    >>> tokenizer = CodeGenTokenizer.from_pretrained("Salesforce/codegen-350M-mono")
+    >>> tokenizer("Hello world")["input_ids"]
+    [15496, 995]
+
+    >>> tokenizer(" Hello world")["input_ids"]
+    [18435, 995]
+    ```
+
+    You can get around that behavior by passing `add_prefix_space=True` when instantiating this tokenizer or when you
+    call it on some text, but since the model was not pretrained this way, it might yield a decrease in performance.
+
+    <Tip>
+
+    When used with `is_split_into_words=True`, this tokenizer will add a space before each word (even the first one).
+
+    </Tip>
+
+    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 `"<|endoftext|>"`):
+            The beginning of sequence token.
+        eos_token (`str`, *optional*, defaults to `"<|endoftext|>"`):
+            The end of sequence token.
+        pad_token (`str`, *optional*):
+            The token used for padding, for example when batching sequences of different lengths.
+        add_prefix_space (`bool`, *optional*, defaults to `False`):
+            Whether or not to add an initial space to the input. This allows to treat the leading word just as any
+            other word. (CodeGen tokenizer detect beginning of words by the preceding space).
+        add_bos_token (`bool`, *optional*, defaults to `False`):
+            Whether to add a beginning of sequence token at the start of sequences.
+        return_token_type_ids (`bool`, *optional*, defaults to `False`):
+            Whether to return token type IDs.
+    """
+
+    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="<|endoftext|>",
+        eos_token="<|endoftext|>",
+        pad_token=None,
+        add_prefix_space=False,
+        add_bos_token=False,
+        return_token_type_ids=False,
+        **kwargs,
+    ):
+        bos_token = AddedToken(bos_token, special=True) if isinstance(bos_token, str) else bos_token
+        eos_token = AddedToken(eos_token, special=True) if isinstance(eos_token, str) else eos_token
+        unk_token = AddedToken(unk_token, special=True) if isinstance(unk_token, str) else unk_token
+        pad_token = AddedToken(pad_token, special=True) if isinstance(pad_token, str) else pad_token
+        self.add_bos_token = add_bos_token
+        self.return_token_type_ids = return_token_type_ids
+        if self.return_token_type_ids:
+            self.model_input_names.append("token_type_ids")
+
+        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().split("\n")[1:-1]
+        bpe_merges = [tuple(merge.split()) for merge in bpe_merges]
+        self.bpe_ranks = dict(zip(bpe_merges, range(len(bpe_merges))))
+        self.cache = {}
+        self.add_prefix_space = add_prefix_space
+
+        # Should have added re.IGNORECASE so BPE merges can happen for capitalized versions of contractions
+        self.pat = re.compile(r"""'s|'t|'re|'ve|'m|'ll|'d| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+""")
+        super().__init__(
+            errors=errors,
+            unk_token=unk_token,
+            bos_token=bos_token,
+            eos_token=eos_token,
+            pad_token=pad_token,
+            add_prefix_space=add_prefix_space,
+            add_bos_token=add_bos_token,
+            return_token_type_ids=return_token_type_ids,
+            **kwargs,
+        )
+
+    @property
+    def vocab_size(self):
+        return len(self.encoder)
+
+    def get_vocab(self):
+        return dict(self.encoder, **self.added_tokens_encoder)
+
+    def bpe(self, token):
+        if token in self.cache:
+            return self.cache[token]
+        word = tuple(token)
+        pairs = get_pairs(word)
+
+        if not pairs:
+            return token
+
+        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 build_inputs_with_special_tokens(self, token_ids_0, token_ids_1=None):
+        if self.add_bos_token:
+            bos_token_ids = [self.bos_token_id]
+        else:
+            bos_token_ids = []
+
+        output = bos_token_ids + token_ids_0
+
+        if token_ids_1 is None:
+            return output
+
+        return output + bos_token_ids + token_ids_1
+
+    def _tokenize(self, text):
+        """Tokenize a string."""
+        bpe_tokens = []
+        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)
+        text = bytearray([self.byte_decoder[c] for c in text]).decode("utf-8", errors=self.errors)
+        return text
+
+    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 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] if self.sep_token_id is not None else []
+        cls = [self.cls_token_id] if self.sep_token_id is not None else []
+        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 not os.path.isdir(save_directory):
+            logger.error(f"Vocabulary path ({save_directory}) should be a 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(
+                        f"Saving vocabulary to {merge_file}: BPE merge indices are not consecutive."
+                        " Please check that the tokenizer is not corrupted!"
+                    )
+                    index = token_index
+                writer.write(" ".join(bpe_tokens) + "\n")
+                index += 1
+
+        return vocab_file, merge_file
+
+    def prepare_for_tokenization(self, text, is_split_into_words=False, **kwargs):
+        add_prefix_space = kwargs.pop("add_prefix_space", self.add_prefix_space)
+        if is_split_into_words or add_prefix_space:
+            text = " " + text
+        return (text, kwargs)
+
+    def decode(
+        self,
+        token_ids: Union[int, List[int], "np.ndarray", "torch.Tensor", "tf.Tensor"],
+        skip_special_tokens: bool = False,
+        clean_up_tokenization_spaces: Optional[bool] = None,
+        truncate_before_pattern: Optional[List[str]] = None,
+        **kwargs,
+    ) -> str:
+        """
+        Converts a sequence of ids in a string, using the tokenizer and vocabulary with options to remove special
+        tokens and clean up tokenization spaces.
+
+        Similar to doing `self.convert_tokens_to_string(self.convert_ids_to_tokens(token_ids))`.
+
+        Args:
+            token_ids (`Union[int, List[int], np.ndarray, torch.Tensor, tf.Tensor]`):
+                List of tokenized input ids. Can be obtained using the `__call__` method.
+            skip_special_tokens (`bool`, *optional*, defaults to `False`):
+                Whether or not to remove special tokens in the decoding.
+            clean_up_tokenization_spaces (`bool`, *optional*):
+                Whether or not to clean up the tokenization spaces. If `None`, will default to
+                `self.clean_up_tokenization_spaces` (available in the `tokenizer_config`).
+            truncate_before_pattern (`List[str]`, *optional*, defaults to `None`):
+                A list of regular expression strings that will be used to truncate the returned string. This can be
+                used to remove extra pieces of code (e.g. truncate if observing a comment symbol "#" at the beginning
+                of a new line). An example pattern could be `["^#", re.escape("<|endoftext|>"), "^'''", "\n\n\n"]`.
+            kwargs (additional keyword arguments, *optional*):
+                Will be passed to the underlying model specific decode method.
+
+        Returns:
+            `str`: The decoded sentence.
+        """
+
+        token_ids = to_py_obj(token_ids)
+
+        decoded_text = super()._decode(
+            token_ids=token_ids,
+            skip_special_tokens=skip_special_tokens,
+            clean_up_tokenization_spaces=clean_up_tokenization_spaces,
+            **kwargs,
+        )
+
+        if truncate_before_pattern is not None and len(truncate_before_pattern) > 0:
+            decoded_text = self.truncate(decoded_text, truncate_before_pattern)
+
+        return decoded_text
+
+    def truncate(self, completion, truncate_before_pattern):
+        def find_re(string, pattern, start_pos):
+            m = pattern.search(string, start_pos)
+            return m.start() if m else -1
+
+        terminals = [re.compile(pattern, re.MULTILINE) for pattern in truncate_before_pattern]
+
+        prints = list(re.finditer("^print", completion, re.MULTILINE))
+
+        if len(prints) > 1:
+            completion = completion[: prints[1].start()]
+
+        defs = list(re.finditer("^def", completion, re.MULTILINE))
+
+        if len(defs) > 1:
+            completion = completion[: defs[1].start()]
+
+        start_pos = 0
+
+        terminals_pos = [
+            pos for pos in [find_re(completion, terminal, start_pos) for terminal in terminals] if pos != -1
+        ]
+
+        if len(terminals_pos) > 0:
+            return completion[: min(terminals_pos)]
+        else:
+            return completion
+
+
+__all__ = ["CodeGenTokenizer"]

lib/python3.12/site-packages/transformers/models/codegen/tokenization_codegen_fast.py

@@ -0,0 +1,265 @@
+# coding=utf-8
+# Copyright 2022 The Salesforce authors, 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."""
+
+import re
+from typing import TYPE_CHECKING, List, Optional, Tuple, Union
+
+import numpy as np
+
+from ...utils import is_tf_available, is_torch_available, logging
+
+
+if TYPE_CHECKING:
+    if is_torch_available():
+        import torch
+    if is_tf_available():
+        import tensorflow as tf
+
+
+from ...tokenization_utils_base import BatchEncoding
+from ...tokenization_utils_fast import PreTrainedTokenizerFast
+from .tokenization_codegen import CodeGenTokenizer
+
+
+logger = logging.get_logger(__name__)
+
+VOCAB_FILES_NAMES = {"vocab_file": "vocab.json", "merges_file": "merges.txt", "tokenizer_file": "tokenizer.json"}
+
+
+class CodeGenTokenizerFast(PreTrainedTokenizerFast):
+    """
+    Construct a "fast" CodeGen tokenizer (backed by HuggingFace's *tokenizers* library). Based on byte-level
+    Byte-Pair-Encoding.
+
+    This tokenizer has been trained to treat spaces like parts of the tokens (a bit like sentencepiece) so a word will
+    be encoded differently whether it is at the beginning of the sentence (without space) or not:
+
+    ```python
+    >>> from transformers import CodeGenTokenizerFast
+
+    >>> tokenizer = CodeGenTokenizerFast.from_pretrained("Salesforce/codegen-350M-mono")
+    >>> tokenizer("Hello world")["input_ids"]
+    [15496, 995]
+
+    >>> tokenizer(" Hello world")["input_ids"]
+    [18435, 995]
+    ```
+
+    You can get around that behavior by passing `add_prefix_space=True` when instantiating this tokenizer, but since
+    the model was not pretrained this way, it might yield a decrease in performance.
+
+    <Tip>
+
+    When used with `is_split_into_words=True`, this tokenizer needs to be instantiated with `add_prefix_space=True`.
+
+    </Tip>
+
+    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*):
+            Path to [tokenizers](https://github.com/huggingface/tokenizers) file (generally has a .json extension) that
+            contains everything needed to load the tokenizer.
+        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 `"<|endoftext|>"`):
+            The beginning of sequence token.
+        eos_token (`str`, *optional*, defaults to `"<|endoftext|>"`):
+            The end of sequence token.
+        add_prefix_space (`bool`, *optional*, defaults to `False`):
+            Whether or not to add an initial space to the input. This allows to treat the leading word just as any
+            other word. (CodeGen tokenizer detect beginning of words by the preceding space).
+        return_token_type_ids (`bool`, *optional*, defaults to `False`):
+            Whether to return token type IDs.
+    """
+
+    vocab_files_names = VOCAB_FILES_NAMES
+    model_input_names = ["input_ids", "attention_mask"]
+    slow_tokenizer_class = CodeGenTokenizer
+
+    def __init__(
+        self,
+        vocab_file=None,
+        merges_file=None,
+        tokenizer_file=None,
+        unk_token="<|endoftext|>",
+        bos_token="<|endoftext|>",
+        eos_token="<|endoftext|>",
+        add_prefix_space=False,
+        return_token_type_ids=False,
+        **kwargs,
+    ):
+        self.return_token_type_ids = return_token_type_ids
+        if self.return_token_type_ids:
+            self.model_input_names.append("token_type_ids")
+
+        super().__init__(
+            vocab_file,
+            merges_file,
+            tokenizer_file=tokenizer_file,
+            unk_token=unk_token,
+            bos_token=bos_token,
+            eos_token=eos_token,
+            add_prefix_space=add_prefix_space,
+            return_token_type_ids=return_token_type_ids,
+            **kwargs,
+        )
+
+        if kwargs.pop("add_bos_token", False):
+            model_id = kwargs.pop("name_or_path", "")
+            raise ValueError(
+                "Currenty GPT2's fast tokenizer does NOT support adding a BOS token. "
+                "Instead you should use GPT2's slow tokenizer class `CodeGenTokenizer` as follows: \n"
+                f"`CodeGenTokenizer.from_pretrained('{model_id}')`\nor\n"
+                f"`AutoTokenizer.from_pretrained('{model_id}', use_fast=False)`\n"
+                "This issue will be fixed soon, see: https://github.com/huggingface/tokenizers/pull/1005."
+                " so that the fast tokenizer works correctly."
+            )
+
+    def _batch_encode_plus(self, *args, **kwargs) -> BatchEncoding:
+        is_split_into_words = kwargs.get("is_split_into_words", False)
+        assert self.add_prefix_space or not is_split_into_words, (
+            f"You need to instantiate {self.__class__.__name__} with add_prefix_space=True "
+            "to use it with pretokenized inputs."
+        )
+
+        return super()._batch_encode_plus(*args, **kwargs)
+
+    def _encode_plus(self, *args, **kwargs) -> BatchEncoding:
+        is_split_into_words = kwargs.get("is_split_into_words", False)
+
+        assert self.add_prefix_space or not is_split_into_words, (
+            f"You need to instantiate {self.__class__.__name__} with add_prefix_space=True "
+            "to use it with pretokenized inputs."
+        )
+
+        return super()._encode_plus(*args, **kwargs)
+
+    # Copied from transformers.models.codegen.tokenization_codegen.CodeGenTokenizer.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 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] if self.sep_token_id is not None else []
+        cls = [self.cls_token_id] if self.sep_token_id is not None else []
+        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]:
+        files = self._tokenizer.model.save(save_directory, name=filename_prefix)
+        return tuple(files)
+
+    def decode(
+        self,
+        token_ids: Union[int, List[int], "np.ndarray", "torch.Tensor", "tf.Tensor"],
+        skip_special_tokens: bool = False,
+        clean_up_tokenization_spaces: Optional[bool] = None,
+        truncate_before_pattern: Optional[List[str]] = None,
+        **kwargs,
+    ) -> str:
+        """
+        Converts a sequence of ids in a string, using the tokenizer and vocabulary with options to remove special
+        tokens and clean up tokenization spaces.
+
+        Similar to doing `self.convert_tokens_to_string(self.convert_ids_to_tokens(token_ids))`.
+
+        Args:
+            token_ids (`Union[int, List[int], np.ndarray, torch.Tensor, tf.Tensor]`):
+                List of tokenized input ids. Can be obtained using the `__call__` method.
+            skip_special_tokens (`bool`, *optional*, defaults to `False`):
+                Whether or not to remove special tokens in the decoding.
+            clean_up_tokenization_spaces (`bool`, *optional*):
+                Whether or not to clean up the tokenization spaces. If `None`, will default to
+                `self.clean_up_tokenization_spaces` (available in the `tokenizer_config`).
+            truncate_before_pattern (`List[str]`, *optional*, defaults to `None`):
+                A list of regular expression strings that will be used to truncate the returned string. This can be
+                used to remove extra pieces of code (e.g. truncate if observing a comment symbol "#" at the beginning
+                of a new line). An example pattern could be `["^#", re.escape("<|endoftext|>"), "^'''", "\n\n\n"]`.
+            kwargs (additional keyword arguments, *optional*):
+                Will be passed to the underlying model specific decode method.
+
+        Returns:
+            `str`: The decoded sentence.
+        """
+
+        decoded_text = super().decode(
+            token_ids=token_ids,
+            skip_special_tokens=skip_special_tokens,
+            clean_up_tokenization_spaces=clean_up_tokenization_spaces,
+            **kwargs,
+        )
+
+        if truncate_before_pattern is not None and len(truncate_before_pattern) > 0:
+            decoded_text = self.truncate(decoded_text, truncate_before_pattern)
+
+        return decoded_text
+
+    def truncate(self, completion, truncate_before_pattern):
+        def find_re(string, pattern, start_pos):
+            m = pattern.search(string, start_pos)
+            return m.start() if m else -1
+
+        terminals = [re.compile(pattern, re.MULTILINE) for pattern in truncate_before_pattern]
+
+        prints = list(re.finditer("^print", completion, re.MULTILINE))
+
+        if len(prints) > 1:
+            completion = completion[: prints[1].start()]
+
+        defs = list(re.finditer("^def", completion, re.MULTILINE))
+
+        if len(defs) > 1:
+            completion = completion[: defs[1].start()]
+
+        start_pos = 0
+
+        terminals_pos = [
+            pos for pos in [find_re(completion, terminal, start_pos) for terminal in terminals] if pos != -1
+        ]
+
+        if len(terminals_pos) > 0:
+            return completion[: min(terminals_pos)]
+        else:
+            return completion
+
+
+__all__ = ["CodeGenTokenizerFast"]

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

@@ -0,0 +1,29 @@
+# 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_conditional_detr import *
+    from .feature_extraction_conditional_detr import *
+    from .image_processing_conditional_detr import *
+    from .modeling_conditional_detr 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/conditional_detr/configuration_conditional_detr.py

@@ -0,0 +1,278 @@
+# 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.
+"""Conditional DETR model configuration"""
+
+from collections import OrderedDict
+from typing import Mapping
+
+from packaging import version
+
+from ...configuration_utils import PretrainedConfig
+from ...onnx import OnnxConfig
+from ...utils import logging
+from ...utils.backbone_utils import verify_backbone_config_arguments
+from ..auto import CONFIG_MAPPING
+
+
+logger = logging.get_logger(__name__)
+
+
+class ConditionalDetrConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`ConditionalDetrModel`]. It is used to instantiate
+    a Conditional DETR 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 Conditional DETR
+    [microsoft/conditional-detr-resnet-50](https://huggingface.co/microsoft/conditional-detr-resnet-50) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        use_timm_backbone (`bool`, *optional*, defaults to `True`):
+            Whether or not to use the `timm` library for the backbone. If set to `False`, will use the [`AutoBackbone`]
+            API.
+        backbone_config (`PretrainedConfig` or `dict`, *optional*):
+            The configuration of the backbone model. Only used in case `use_timm_backbone` is set to `False` in which
+            case it will default to `ResNetConfig()`.
+        num_channels (`int`, *optional*, defaults to 3):
+            The number of input channels.
+        num_queries (`int`, *optional*, defaults to 100):
+            Number of object queries, i.e. detection slots. This is the maximal number of objects
+            [`ConditionalDetrModel`] can detect in a single image. For COCO, we recommend 100 queries.
+        d_model (`int`, *optional*, defaults to 256):
+            This parameter is a general dimension parameter, defining dimensions for components such as the encoder layer and projection parameters in the decoder layer, among others.
+        encoder_layers (`int`, *optional*, defaults to 6):
+            Number of encoder layers.
+        decoder_layers (`int`, *optional*, defaults to 6):
+            Number of decoder layers.
+        encoder_attention_heads (`int`, *optional*, defaults to 8):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        decoder_attention_heads (`int`, *optional*, defaults to 8):
+            Number of attention heads for each attention layer in the Transformer decoder.
+        decoder_ffn_dim (`int`, *optional*, defaults to 2048):
+            Dimension of the "intermediate" (often named feed-forward) layer in decoder.
+        encoder_ffn_dim (`int`, *optional*, defaults to 2048):
+            Dimension of the "intermediate" (often named feed-forward) layer in decoder.
+        activation_function (`str` or `function`, *optional*, defaults to `"relu"`):
+            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.
+        init_std (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        init_xavier_std (`float`, *optional*, defaults to 1):
+            The scaling factor used for the Xavier initialization gain in the HM Attention map module.
+        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.
+        auxiliary_loss (`bool`, *optional*, defaults to `False`):
+            Whether auxiliary decoding losses (loss at each decoder layer) are to be used.
+        position_embedding_type (`str`, *optional*, defaults to `"sine"`):
+            Type of position embeddings to be used on top of the image features. One of `"sine"` or `"learned"`.
+        backbone (`str`, *optional*, defaults to `"resnet50"`):
+            Name of backbone to use when `backbone_config` is `None`. If `use_pretrained_backbone` is `True`, this
+            will load the corresponding pretrained weights from the timm or transformers library. If `use_pretrained_backbone`
+            is `False`, this loads the backbone's config and uses that to initialize the backbone with random weights.
+        use_pretrained_backbone (`bool`, *optional*, defaults to `True`):
+            Whether to use pretrained weights for the backbone.
+        backbone_kwargs (`dict`, *optional*):
+            Keyword arguments to be passed to AutoBackbone when loading from a checkpoint
+            e.g. `{'out_indices': (0, 1, 2, 3)}`. Cannot be specified if `backbone_config` is set.
+        dilation (`bool`, *optional*, defaults to `False`):
+            Whether to replace stride with dilation in the last convolutional block (DC5). Only supported when
+            `use_timm_backbone` = `True`.
+        class_cost (`float`, *optional*, defaults to 1):
+            Relative weight of the classification error in the Hungarian matching cost.
+        bbox_cost (`float`, *optional*, defaults to 5):
+            Relative weight of the L1 error of the bounding box coordinates in the Hungarian matching cost.
+        giou_cost (`float`, *optional*, defaults to 2):
+            Relative weight of the generalized IoU loss of the bounding box in the Hungarian matching cost.
+        mask_loss_coefficient (`float`, *optional*, defaults to 1):
+            Relative weight of the Focal loss in the panoptic segmentation loss.
+        dice_loss_coefficient (`float`, *optional*, defaults to 1):
+            Relative weight of the DICE/F-1 loss in the panoptic segmentation loss.
+        bbox_loss_coefficient (`float`, *optional*, defaults to 5):
+            Relative weight of the L1 bounding box loss in the object detection loss.
+        giou_loss_coefficient (`float`, *optional*, defaults to 2):
+            Relative weight of the generalized IoU loss in the object detection loss.
+        eos_coefficient (`float`, *optional*, defaults to 0.1):
+            Relative classification weight of the 'no-object' class in the object detection loss.
+        focal_alpha (`float`, *optional*, defaults to 0.25):
+            Alpha parameter in the focal loss.
+
+    Examples:
+
+    ```python
+    >>> from transformers import ConditionalDetrConfig, ConditionalDetrModel
+
+    >>> # Initializing a Conditional DETR microsoft/conditional-detr-resnet-50 style configuration
+    >>> configuration = ConditionalDetrConfig()
+
+    >>> # Initializing a model (with random weights) from the microsoft/conditional-detr-resnet-50 style configuration
+    >>> model = ConditionalDetrModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "conditional_detr"
+    keys_to_ignore_at_inference = ["past_key_values"]
+    attribute_map = {
+        "hidden_size": "d_model",
+        "num_attention_heads": "encoder_attention_heads",
+    }
+
+    def __init__(
+        self,
+        use_timm_backbone=True,
+        backbone_config=None,
+        num_channels=3,
+        num_queries=300,
+        encoder_layers=6,
+        encoder_ffn_dim=2048,
+        encoder_attention_heads=8,
+        decoder_layers=6,
+        decoder_ffn_dim=2048,
+        decoder_attention_heads=8,
+        encoder_layerdrop=0.0,
+        decoder_layerdrop=0.0,
+        is_encoder_decoder=True,
+        activation_function="relu",
+        d_model=256,
+        dropout=0.1,
+        attention_dropout=0.0,
+        activation_dropout=0.0,
+        init_std=0.02,
+        init_xavier_std=1.0,
+        auxiliary_loss=False,
+        position_embedding_type="sine",
+        backbone="resnet50",
+        use_pretrained_backbone=True,
+        backbone_kwargs=None,
+        dilation=False,
+        class_cost=2,
+        bbox_cost=5,
+        giou_cost=2,
+        mask_loss_coefficient=1,
+        dice_loss_coefficient=1,
+        cls_loss_coefficient=2,
+        bbox_loss_coefficient=5,
+        giou_loss_coefficient=2,
+        focal_alpha=0.25,
+        **kwargs,
+    ):
+        # We default to values which were previously hard-coded in the model. This enables configurability of the config
+        # while keeping the default behavior the same.
+        if use_timm_backbone and backbone_kwargs is None:
+            backbone_kwargs = {}
+            if dilation:
+                backbone_kwargs["output_stride"] = 16
+            backbone_kwargs["out_indices"] = [1, 2, 3, 4]
+            backbone_kwargs["in_chans"] = num_channels
+        # Backwards compatibility
+        elif not use_timm_backbone and backbone in (None, "resnet50"):
+            if backbone_config is None:
+                logger.info("`backbone_config` is `None`. Initializing the config with the default `ResNet` backbone.")
+                backbone_config = CONFIG_MAPPING["resnet"](out_features=["stage4"])
+            elif isinstance(backbone_config, dict):
+                backbone_model_type = backbone_config.get("model_type")
+                config_class = CONFIG_MAPPING[backbone_model_type]
+                backbone_config = config_class.from_dict(backbone_config)
+
+        verify_backbone_config_arguments(
+            use_timm_backbone=use_timm_backbone,
+            use_pretrained_backbone=use_pretrained_backbone,
+            backbone=backbone,
+            backbone_config=backbone_config,
+            backbone_kwargs=backbone_kwargs,
+        )
+
+        self.use_timm_backbone = use_timm_backbone
+        self.backbone_config = backbone_config
+        self.num_channels = num_channels
+        self.num_queries = num_queries
+        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.init_xavier_std = init_xavier_std
+        self.encoder_layerdrop = encoder_layerdrop
+        self.decoder_layerdrop = decoder_layerdrop
+        self.num_hidden_layers = encoder_layers
+        self.auxiliary_loss = auxiliary_loss
+        self.position_embedding_type = position_embedding_type
+        self.backbone = backbone
+        self.use_pretrained_backbone = use_pretrained_backbone
+        self.backbone_kwargs = backbone_kwargs
+        self.dilation = dilation
+        # Hungarian matcher
+        self.class_cost = class_cost
+        self.bbox_cost = bbox_cost
+        self.giou_cost = giou_cost
+        # Loss coefficients
+        self.mask_loss_coefficient = mask_loss_coefficient
+        self.dice_loss_coefficient = dice_loss_coefficient
+        self.cls_loss_coefficient = cls_loss_coefficient
+        self.bbox_loss_coefficient = bbox_loss_coefficient
+        self.giou_loss_coefficient = giou_loss_coefficient
+        self.focal_alpha = focal_alpha
+        super().__init__(is_encoder_decoder=is_encoder_decoder, **kwargs)
+
+    @property
+    def num_attention_heads(self) -> int:
+        return self.encoder_attention_heads
+
+    @property
+    def hidden_size(self) -> int:
+        return self.d_model
+
+
+class ConditionalDetrOnnxConfig(OnnxConfig):
+    torch_onnx_minimum_version = version.parse("1.11")
+
+    @property
+    def inputs(self) -> Mapping[str, Mapping[int, str]]:
+        return OrderedDict(
+            [
+                ("pixel_values", {0: "batch", 1: "num_channels", 2: "height", 3: "width"}),
+                ("pixel_mask", {0: "batch"}),
+            ]
+        )
+
+    @property
+    def atol_for_validation(self) -> float:
+        return 1e-5
+
+    @property
+    def default_onnx_opset(self) -> int:
+        return 12
+
+
+__all__ = ["ConditionalDetrConfig", "ConditionalDetrOnnxConfig"]

lib/python3.12/site-packages/transformers/models/conditional_detr/feature_extraction_conditional_detr.py

@@ -0,0 +1,46 @@
+# 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.
+"""Feature extractor class for Conditional DETR."""
+
+import warnings
+
+from ...image_transforms import rgb_to_id as _rgb_to_id
+from ...utils import logging
+from .image_processing_conditional_detr import ConditionalDetrImageProcessor
+
+
+logger = logging.get_logger(__name__)
+
+
+def rgb_to_id(x):
+    warnings.warn(
+        "rgb_to_id has moved and will not be importable from this module from v5. "
+        "Please import from transformers.image_transforms instead.",
+        FutureWarning,
+    )
+    return _rgb_to_id(x)
+
+
+class ConditionalDetrFeatureExtractor(ConditionalDetrImageProcessor):
+    def __init__(self, *args, **kwargs) -> None:
+        warnings.warn(
+            "The class ConditionalDetrFeatureExtractor is deprecated and will be removed in version 5 of Transformers."
+            " Please use ConditionalDetrImageProcessor instead.",
+            FutureWarning,
+        )
+        super().__init__(*args, **kwargs)
+
+
+__all__ = ["ConditionalDetrFeatureExtractor"]

lib/python3.12/site-packages/transformers/models/conditional_detr/image_processing_conditional_detr.py

@@ -0,0 +1,1855 @@
+# 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 Conditional DETR."""
+
+import io
+import pathlib
+from collections import defaultdict
+from typing import Any, Callable, Dict, Iterable, List, Optional, Set, Tuple, Union
+
+import numpy as np
+
+from ...feature_extraction_utils import BatchFeature
+from ...image_processing_utils import BaseImageProcessor, get_size_dict
+from ...image_transforms import (
+    PaddingMode,
+    center_to_corners_format,
+    corners_to_center_format,
+    id_to_rgb,
+    pad,
+    rescale,
+    resize,
+    rgb_to_id,
+    to_channel_dimension_format,
+)
+from ...image_utils import (
+    IMAGENET_DEFAULT_MEAN,
+    IMAGENET_DEFAULT_STD,
+    AnnotationFormat,
+    AnnotationType,
+    ChannelDimension,
+    ImageInput,
+    PILImageResampling,
+    get_image_size,
+    infer_channel_dimension_format,
+    is_scaled_image,
+    make_list_of_images,
+    to_numpy_array,
+    valid_images,
+    validate_annotations,
+    validate_kwargs,
+    validate_preprocess_arguments,
+)
+from ...utils import (
+    TensorType,
+    is_flax_available,
+    is_jax_tensor,
+    is_scipy_available,
+    is_tf_available,
+    is_tf_tensor,
+    is_torch_available,
+    is_torch_tensor,
+    is_vision_available,
+    logging,
+)
+
+
+if is_torch_available():
+    import torch
+    from torch import nn
+
+
+if is_vision_available():
+    import PIL
+
+
+if is_scipy_available():
+    import scipy.special
+    import scipy.stats
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+SUPPORTED_ANNOTATION_FORMATS = (AnnotationFormat.COCO_DETECTION, AnnotationFormat.COCO_PANOPTIC)
+
+
+# Copied from transformers.models.detr.image_processing_detr.get_size_with_aspect_ratio
+def get_size_with_aspect_ratio(image_size, size, max_size=None) -> Tuple[int, int]:
+    """
+    Computes the output image size given the input image size and the desired output size.
+
+    Args:
+        image_size (`Tuple[int, int]`):
+            The input image size.
+        size (`int`):
+            The desired output size.
+        max_size (`int`, *optional*):
+            The maximum allowed output size.
+    """
+    height, width = image_size
+    raw_size = None
+    if max_size is not None:
+        min_original_size = float(min((height, width)))
+        max_original_size = float(max((height, width)))
+        if max_original_size / min_original_size * size > max_size:
+            raw_size = max_size * min_original_size / max_original_size
+            size = int(round(raw_size))
+
+    if (height <= width and height == size) or (width <= height and width == size):
+        oh, ow = height, width
+    elif width < height:
+        ow = size
+        if max_size is not None and raw_size is not None:
+            oh = int(raw_size * height / width)
+        else:
+            oh = int(size * height / width)
+    else:
+        oh = size
+        if max_size is not None and raw_size is not None:
+            ow = int(raw_size * width / height)
+        else:
+            ow = int(size * width / height)
+
+    return (oh, ow)
+
+
+# Copied from transformers.models.detr.image_processing_detr.get_resize_output_image_size
+def get_resize_output_image_size(
+    input_image: np.ndarray,
+    size: Union[int, Tuple[int, int], List[int]],
+    max_size: Optional[int] = None,
+    input_data_format: Optional[Union[str, ChannelDimension]] = None,
+) -> Tuple[int, int]:
+    """
+    Computes the output image size given the input image size and the desired output size. If the desired output size
+    is a tuple or list, the output image size is returned as is. If the desired output size is an integer, the output
+    image size is computed by keeping the aspect ratio of the input image size.
+
+    Args:
+        input_image (`np.ndarray`):
+            The image to resize.
+        size (`int` or `Tuple[int, int]` or `List[int]`):
+            The desired output size.
+        max_size (`int`, *optional*):
+            The maximum allowed output size.
+        input_data_format (`ChannelDimension` or `str`, *optional*):
+            The channel dimension format of the input image. If not provided, it will be inferred from the input image.
+    """
+    image_size = get_image_size(input_image, input_data_format)
+    if isinstance(size, (list, tuple)):
+        return size
+
+    return get_size_with_aspect_ratio(image_size, size, max_size)
+
+
+# Copied from transformers.models.detr.image_processing_detr.get_image_size_for_max_height_width
+def get_image_size_for_max_height_width(
+    input_image: np.ndarray,
+    max_height: int,
+    max_width: int,
+    input_data_format: Optional[Union[str, ChannelDimension]] = None,
+) -> Tuple[int, int]:
+    """
+    Computes the output image size given the input image and the maximum allowed height and width. Keep aspect ratio.
+    Important, even if image_height < max_height and image_width < max_width, the image will be resized
+    to at least one of the edges be equal to max_height or max_width.
+
+    For example:
+        - input_size: (100, 200), max_height: 50, max_width: 50 -> output_size: (25, 50)
+        - input_size: (100, 200), max_height: 200, max_width: 500 -> output_size: (200, 400)
+
+    Args:
+        input_image (`np.ndarray`):
+            The image to resize.
+        max_height (`int`):
+            The maximum allowed height.
+        max_width (`int`):
+            The maximum allowed width.
+        input_data_format (`ChannelDimension` or `str`, *optional*):
+            The channel dimension format of the input image. If not provided, it will be inferred from the input image.
+    """
+    image_size = get_image_size(input_image, input_data_format)
+    height, width = image_size
+    height_scale = max_height / height
+    width_scale = max_width / width
+    min_scale = min(height_scale, width_scale)
+    new_height = int(height * min_scale)
+    new_width = int(width * min_scale)
+    return new_height, new_width
+
+
+# Copied from transformers.models.detr.image_processing_detr.get_numpy_to_framework_fn
+def get_numpy_to_framework_fn(arr) -> Callable:
+    """
+    Returns a function that converts a numpy array to the framework of the input array.
+
+    Args:
+        arr (`np.ndarray`): The array to convert.
+    """
+    if isinstance(arr, np.ndarray):
+        return np.array
+    if is_tf_available() and is_tf_tensor(arr):
+        import tensorflow as tf
+
+        return tf.convert_to_tensor
+    if is_torch_available() and is_torch_tensor(arr):
+        import torch
+
+        return torch.tensor
+    if is_flax_available() and is_jax_tensor(arr):
+        import jax.numpy as jnp
+
+        return jnp.array
+    raise ValueError(f"Cannot convert arrays of type {type(arr)}")
+
+
+# Copied from transformers.models.detr.image_processing_detr.safe_squeeze
+def safe_squeeze(arr: np.ndarray, axis: Optional[int] = None) -> np.ndarray:
+    """
+    Squeezes an array, but only if the axis specified has dim 1.
+    """
+    if axis is None:
+        return arr.squeeze()
+
+    try:
+        return arr.squeeze(axis=axis)
+    except ValueError:
+        return arr
+
+
+# Copied from transformers.models.detr.image_processing_detr.normalize_annotation
+def normalize_annotation(annotation: Dict, image_size: Tuple[int, int]) -> Dict:
+    image_height, image_width = image_size
+    norm_annotation = {}
+    for key, value in annotation.items():
+        if key == "boxes":
+            boxes = value
+            boxes = corners_to_center_format(boxes)
+            boxes /= np.asarray([image_width, image_height, image_width, image_height], dtype=np.float32)
+            norm_annotation[key] = boxes
+        else:
+            norm_annotation[key] = value
+    return norm_annotation
+
+
+# Copied from transformers.models.detr.image_processing_detr.max_across_indices
+def max_across_indices(values: Iterable[Any]) -> List[Any]:
+    """
+    Return the maximum value across all indices of an iterable of values.
+    """
+    return [max(values_i) for values_i in zip(*values)]
+
+
+# Copied from transformers.models.detr.image_processing_detr.get_max_height_width
+def get_max_height_width(
+    images: List[np.ndarray], input_data_format: Optional[Union[str, ChannelDimension]] = None
+) -> List[int]:
+    """
+    Get the maximum height and width across all images in a batch.
+    """
+    if input_data_format is None:
+        input_data_format = infer_channel_dimension_format(images[0])
+
+    if input_data_format == ChannelDimension.FIRST:
+        _, max_height, max_width = max_across_indices([img.shape for img in images])
+    elif input_data_format == ChannelDimension.LAST:
+        max_height, max_width, _ = max_across_indices([img.shape for img in images])
+    else:
+        raise ValueError(f"Invalid channel dimension format: {input_data_format}")
+    return (max_height, max_width)
+
+
+# Copied from transformers.models.detr.image_processing_detr.make_pixel_mask
+def make_pixel_mask(
+    image: np.ndarray, output_size: Tuple[int, int], input_data_format: Optional[Union[str, ChannelDimension]] = None
+) -> np.ndarray:
+    """
+    Make a pixel mask for the image, where 1 indicates a valid pixel and 0 indicates padding.
+
+    Args:
+        image (`np.ndarray`):
+            Image to make the pixel mask for.
+        output_size (`Tuple[int, int]`):
+            Output size of the mask.
+    """
+    input_height, input_width = get_image_size(image, channel_dim=input_data_format)
+    mask = np.zeros(output_size, dtype=np.int64)
+    mask[:input_height, :input_width] = 1
+    return mask
+
+
+# Copied from transformers.models.detr.image_processing_detr.convert_coco_poly_to_mask
+def convert_coco_poly_to_mask(segmentations, height: int, width: int) -> np.ndarray:
+    """
+    Convert a COCO polygon annotation to a mask.
+
+    Args:
+        segmentations (`List[List[float]]`):
+            List of polygons, each polygon represented by a list of x-y coordinates.
+        height (`int`):
+            Height of the mask.
+        width (`int`):
+            Width of the mask.
+    """
+    try:
+        from pycocotools import mask as coco_mask
+    except ImportError:
+        raise ImportError("Pycocotools is not installed in your environment.")
+
+    masks = []
+    for polygons in segmentations:
+        rles = coco_mask.frPyObjects(polygons, height, width)
+        mask = coco_mask.decode(rles)
+        if len(mask.shape) < 3:
+            mask = mask[..., None]
+        mask = np.asarray(mask, dtype=np.uint8)
+        mask = np.any(mask, axis=2)
+        masks.append(mask)
+    if masks:
+        masks = np.stack(masks, axis=0)
+    else:
+        masks = np.zeros((0, height, width), dtype=np.uint8)
+
+    return masks
+
+
+# Copied from transformers.models.detr.image_processing_detr.prepare_coco_detection_annotation with DETR->ConditionalDetr
+def prepare_coco_detection_annotation(
+    image,
+    target,
+    return_segmentation_masks: bool = False,
+    input_data_format: Optional[Union[ChannelDimension, str]] = None,
+):
+    """
+    Convert the target in COCO format into the format expected by ConditionalDetr.
+    """
+    image_height, image_width = get_image_size(image, channel_dim=input_data_format)
+
+    image_id = target["image_id"]
+    image_id = np.asarray([image_id], dtype=np.int64)
+
+    # Get all COCO annotations for the given image.
+    annotations = target["annotations"]
+    annotations = [obj for obj in annotations if "iscrowd" not in obj or obj["iscrowd"] == 0]
+
+    classes = [obj["category_id"] for obj in annotations]
+    classes = np.asarray(classes, dtype=np.int64)
+
+    # for conversion to coco api
+    area = np.asarray([obj["area"] for obj in annotations], dtype=np.float32)
+    iscrowd = np.asarray([obj["iscrowd"] if "iscrowd" in obj else 0 for obj in annotations], dtype=np.int64)
+
+    boxes = [obj["bbox"] for obj in annotations]
+    # guard against no boxes via resizing
+    boxes = np.asarray(boxes, dtype=np.float32).reshape(-1, 4)
+    boxes[:, 2:] += boxes[:, :2]
+    boxes[:, 0::2] = boxes[:, 0::2].clip(min=0, max=image_width)
+    boxes[:, 1::2] = boxes[:, 1::2].clip(min=0, max=image_height)
+
+    keep = (boxes[:, 3] > boxes[:, 1]) & (boxes[:, 2] > boxes[:, 0])
+
+    new_target = {}
+    new_target["image_id"] = image_id
+    new_target["class_labels"] = classes[keep]
+    new_target["boxes"] = boxes[keep]
+    new_target["area"] = area[keep]
+    new_target["iscrowd"] = iscrowd[keep]
+    new_target["orig_size"] = np.asarray([int(image_height), int(image_width)], dtype=np.int64)
+
+    if annotations and "keypoints" in annotations[0]:
+        keypoints = [obj["keypoints"] for obj in annotations]
+        # Converting the filtered keypoints list to a numpy array
+        keypoints = np.asarray(keypoints, dtype=np.float32)
+        # Apply the keep mask here to filter the relevant annotations
+        keypoints = keypoints[keep]
+        num_keypoints = keypoints.shape[0]
+        keypoints = keypoints.reshape((-1, 3)) if num_keypoints else keypoints
+        new_target["keypoints"] = keypoints
+
+    if return_segmentation_masks:
+        segmentation_masks = [obj["segmentation"] for obj in annotations]
+        masks = convert_coco_poly_to_mask(segmentation_masks, image_height, image_width)
+        new_target["masks"] = masks[keep]
+
+    return new_target
+
+
+# Copied from transformers.models.detr.image_processing_detr.masks_to_boxes
+def masks_to_boxes(masks: np.ndarray) -> np.ndarray:
+    """
+    Compute the bounding boxes around the provided panoptic segmentation masks.
+
+    Args:
+        masks: masks in format `[number_masks, height, width]` where N is the number of masks
+
+    Returns:
+        boxes: bounding boxes in format `[number_masks, 4]` in xyxy format
+    """
+    if masks.size == 0:
+        return np.zeros((0, 4))
+
+    h, w = masks.shape[-2:]
+    y = np.arange(0, h, dtype=np.float32)
+    x = np.arange(0, w, dtype=np.float32)
+    # see https://github.com/pytorch/pytorch/issues/50276
+    y, x = np.meshgrid(y, x, indexing="ij")
+
+    x_mask = masks * np.expand_dims(x, axis=0)
+    x_max = x_mask.reshape(x_mask.shape[0], -1).max(-1)
+    x = np.ma.array(x_mask, mask=~(np.array(masks, dtype=bool)))
+    x_min = x.filled(fill_value=1e8)
+    x_min = x_min.reshape(x_min.shape[0], -1).min(-1)
+
+    y_mask = masks * np.expand_dims(y, axis=0)
+    y_max = y_mask.reshape(x_mask.shape[0], -1).max(-1)
+    y = np.ma.array(y_mask, mask=~(np.array(masks, dtype=bool)))
+    y_min = y.filled(fill_value=1e8)
+    y_min = y_min.reshape(y_min.shape[0], -1).min(-1)
+
+    return np.stack([x_min, y_min, x_max, y_max], 1)
+
+
+# Copied from transformers.models.detr.image_processing_detr.prepare_coco_panoptic_annotation with DETR->ConditionalDetr
+def prepare_coco_panoptic_annotation(
+    image: np.ndarray,
+    target: Dict,
+    masks_path: Union[str, pathlib.Path],
+    return_masks: bool = True,
+    input_data_format: Union[ChannelDimension, str] = None,
+) -> Dict:
+    """
+    Prepare a coco panoptic annotation for ConditionalDetr.
+    """
+    image_height, image_width = get_image_size(image, channel_dim=input_data_format)
+    annotation_path = pathlib.Path(masks_path) / target["file_name"]
+
+    new_target = {}
+    new_target["image_id"] = np.asarray([target["image_id"] if "image_id" in target else target["id"]], dtype=np.int64)
+    new_target["size"] = np.asarray([image_height, image_width], dtype=np.int64)
+    new_target["orig_size"] = np.asarray([image_height, image_width], dtype=np.int64)
+
+    if "segments_info" in target:
+        masks = np.asarray(PIL.Image.open(annotation_path), dtype=np.uint32)
+        masks = rgb_to_id(masks)
+
+        ids = np.array([segment_info["id"] for segment_info in target["segments_info"]])
+        masks = masks == ids[:, None, None]
+        masks = masks.astype(np.uint8)
+        if return_masks:
+            new_target["masks"] = masks
+        new_target["boxes"] = masks_to_boxes(masks)
+        new_target["class_labels"] = np.array(
+            [segment_info["category_id"] for segment_info in target["segments_info"]], dtype=np.int64
+        )
+        new_target["iscrowd"] = np.asarray(
+            [segment_info["iscrowd"] for segment_info in target["segments_info"]], dtype=np.int64
+        )
+        new_target["area"] = np.asarray(
+            [segment_info["area"] for segment_info in target["segments_info"]], dtype=np.float32
+        )
+
+    return new_target
+
+
+# Copied from transformers.models.detr.image_processing_detr.get_segmentation_image
+def get_segmentation_image(
+    masks: np.ndarray, input_size: Tuple, target_size: Tuple, stuff_equiv_classes, deduplicate=False
+):
+    h, w = input_size
+    final_h, final_w = target_size
+
+    m_id = scipy.special.softmax(masks.transpose(0, 1), -1)
+
+    if m_id.shape[-1] == 0:
+        # We didn't detect any mask :(
+        m_id = np.zeros((h, w), dtype=np.int64)
+    else:
+        m_id = m_id.argmax(-1).reshape(h, w)
+
+    if deduplicate:
+        # Merge the masks corresponding to the same stuff class
+        for equiv in stuff_equiv_classes.values():
+            for eq_id in equiv:
+                m_id[m_id == eq_id] = equiv[0]
+
+    seg_img = id_to_rgb(m_id)
+    seg_img = resize(seg_img, (final_w, final_h), resample=PILImageResampling.NEAREST)
+    return seg_img
+
+
+# Copied from transformers.models.detr.image_processing_detr.get_mask_area
+def get_mask_area(seg_img: np.ndarray, target_size: Tuple[int, int], n_classes: int) -> np.ndarray:
+    final_h, final_w = target_size
+    np_seg_img = seg_img.astype(np.uint8)
+    np_seg_img = np_seg_img.reshape(final_h, final_w, 3)
+    m_id = rgb_to_id(np_seg_img)
+    area = [(m_id == i).sum() for i in range(n_classes)]
+    return area
+
+
+# Copied from transformers.models.detr.image_processing_detr.score_labels_from_class_probabilities
+def score_labels_from_class_probabilities(logits: np.ndarray) -> Tuple[np.ndarray, np.ndarray]:
+    probs = scipy.special.softmax(logits, axis=-1)
+    labels = probs.argmax(-1, keepdims=True)
+    scores = np.take_along_axis(probs, labels, axis=-1)
+    scores, labels = scores.squeeze(-1), labels.squeeze(-1)
+    return scores, labels
+
+
+# Copied from transformers.models.detr.image_processing_detr.post_process_panoptic_sample with DetrForSegmentation->ConditionalDetrForSegmentation
+def post_process_panoptic_sample(
+    out_logits: np.ndarray,
+    masks: np.ndarray,
+    boxes: np.ndarray,
+    processed_size: Tuple[int, int],
+    target_size: Tuple[int, int],
+    is_thing_map: Dict,
+    threshold=0.85,
+) -> Dict:
+    """
+    Converts the output of [`ConditionalDetrForSegmentation`] into panoptic segmentation predictions for a single sample.
+
+    Args:
+        out_logits (`torch.Tensor`):
+            The logits for this sample.
+        masks (`torch.Tensor`):
+            The predicted segmentation masks for this sample.
+        boxes (`torch.Tensor`):
+            The prediced bounding boxes for this sample. The boxes are in the normalized format `(center_x, center_y,
+            width, height)` and values between `[0, 1]`, relative to the size the image (disregarding padding).
+        processed_size (`Tuple[int, int]`):
+            The processed size of the image `(height, width)`, as returned by the preprocessing step i.e. the size
+            after data augmentation but before batching.
+        target_size (`Tuple[int, int]`):
+            The target size of the image, `(height, width)` corresponding to the requested final size of the
+            prediction.
+        is_thing_map (`Dict`):
+            A dictionary mapping class indices to a boolean value indicating whether the class is a thing or not.
+        threshold (`float`, *optional*, defaults to 0.85):
+            The threshold used to binarize the segmentation masks.
+    """
+    # we filter empty queries and detection below threshold
+    scores, labels = score_labels_from_class_probabilities(out_logits)
+    keep = (labels != out_logits.shape[-1] - 1) & (scores > threshold)
+
+    cur_scores = scores[keep]
+    cur_classes = labels[keep]
+    cur_boxes = center_to_corners_format(boxes[keep])
+
+    if len(cur_boxes) != len(cur_classes):
+        raise ValueError("Not as many boxes as there are classes")
+
+    cur_masks = masks[keep]
+    cur_masks = resize(cur_masks[:, None], processed_size, resample=PILImageResampling.BILINEAR)
+    cur_masks = safe_squeeze(cur_masks, 1)
+    b, h, w = cur_masks.shape
+
+    # It may be that we have several predicted masks for the same stuff class.
+    # In the following, we track the list of masks ids for each stuff class (they are merged later on)
+    cur_masks = cur_masks.reshape(b, -1)
+    stuff_equiv_classes = defaultdict(list)
+    for k, label in enumerate(cur_classes):
+        if not is_thing_map[label]:
+            stuff_equiv_classes[label].append(k)
+
+    seg_img = get_segmentation_image(cur_masks, processed_size, target_size, stuff_equiv_classes, deduplicate=True)
+    area = get_mask_area(cur_masks, processed_size, n_classes=len(cur_scores))
+
+    # We filter out any mask that is too small
+    if cur_classes.size() > 0:
+        # We know filter empty masks as long as we find some
+        filtered_small = np.array([a <= 4 for a in area], dtype=bool)
+        while filtered_small.any():
+            cur_masks = cur_masks[~filtered_small]
+            cur_scores = cur_scores[~filtered_small]
+            cur_classes = cur_classes[~filtered_small]
+            seg_img = get_segmentation_image(cur_masks, (h, w), target_size, stuff_equiv_classes, deduplicate=True)
+            area = get_mask_area(seg_img, target_size, n_classes=len(cur_scores))
+            filtered_small = np.array([a <= 4 for a in area], dtype=bool)
+    else:
+        cur_classes = np.ones((1, 1), dtype=np.int64)
+
+    segments_info = [
+        {"id": i, "isthing": is_thing_map[cat], "category_id": int(cat), "area": a}
+        for i, (cat, a) in enumerate(zip(cur_classes, area))
+    ]
+    del cur_classes
+
+    with io.BytesIO() as out:
+        PIL.Image.fromarray(seg_img).save(out, format="PNG")
+        predictions = {"png_string": out.getvalue(), "segments_info": segments_info}
+
+    return predictions
+
+
+# Copied from transformers.models.detr.image_processing_detr.resize_annotation
+def resize_annotation(
+    annotation: Dict[str, Any],
+    orig_size: Tuple[int, int],
+    target_size: Tuple[int, int],
+    threshold: float = 0.5,
+    resample: PILImageResampling = PILImageResampling.NEAREST,
+):
+    """
+    Resizes an annotation to a target size.
+
+    Args:
+        annotation (`Dict[str, Any]`):
+            The annotation dictionary.
+        orig_size (`Tuple[int, int]`):
+            The original size of the input image.
+        target_size (`Tuple[int, int]`):
+            The target size of the image, as returned by the preprocessing `resize` step.
+        threshold (`float`, *optional*, defaults to 0.5):
+            The threshold used to binarize the segmentation masks.
+        resample (`PILImageResampling`, defaults to `PILImageResampling.NEAREST`):
+            The resampling filter to use when resizing the masks.
+    """
+    ratios = tuple(float(s) / float(s_orig) for s, s_orig in zip(target_size, orig_size))
+    ratio_height, ratio_width = ratios
+
+    new_annotation = {}
+    new_annotation["size"] = target_size
+
+    for key, value in annotation.items():
+        if key == "boxes":
+            boxes = value
+            scaled_boxes = boxes * np.asarray([ratio_width, ratio_height, ratio_width, ratio_height], dtype=np.float32)
+            new_annotation["boxes"] = scaled_boxes
+        elif key == "area":
+            area = value
+            scaled_area = area * (ratio_width * ratio_height)
+            new_annotation["area"] = scaled_area
+        elif key == "masks":
+            masks = value[:, None]
+            masks = np.array([resize(mask, target_size, resample=resample) for mask in masks])
+            masks = masks.astype(np.float32)
+            masks = masks[:, 0] > threshold
+            new_annotation["masks"] = masks
+        elif key == "size":
+            new_annotation["size"] = target_size
+        else:
+            new_annotation[key] = value
+
+    return new_annotation
+
+
+# Copied from transformers.models.detr.image_processing_detr.binary_mask_to_rle
+def binary_mask_to_rle(mask):
+    """
+    Converts given binary mask of shape `(height, width)` to the run-length encoding (RLE) format.
+
+    Args:
+        mask (`torch.Tensor` or `numpy.array`):
+            A binary mask tensor of shape `(height, width)` where 0 denotes background and 1 denotes the target
+            segment_id or class_id.
+    Returns:
+        `List`: Run-length encoded list of the binary mask. Refer to COCO API for more information about the RLE
+        format.
+    """
+    if is_torch_tensor(mask):
+        mask = mask.numpy()
+
+    pixels = mask.flatten()
+    pixels = np.concatenate([[0], pixels, [0]])
+    runs = np.where(pixels[1:] != pixels[:-1])[0] + 1
+    runs[1::2] -= runs[::2]
+    return list(runs)
+
+
+# Copied from transformers.models.detr.image_processing_detr.convert_segmentation_to_rle
+def convert_segmentation_to_rle(segmentation):
+    """
+    Converts given segmentation map of shape `(height, width)` to the run-length encoding (RLE) format.
+
+    Args:
+        segmentation (`torch.Tensor` or `numpy.array`):
+            A segmentation map of shape `(height, width)` where each value denotes a segment or class id.
+    Returns:
+        `List[List]`: A list of lists, where each list is the run-length encoding of a segment / class id.
+    """
+    segment_ids = torch.unique(segmentation)
+
+    run_length_encodings = []
+    for idx in segment_ids:
+        mask = torch.where(segmentation == idx, 1, 0)
+        rle = binary_mask_to_rle(mask)
+        run_length_encodings.append(rle)
+
+    return run_length_encodings
+
+
+# Copied from transformers.models.detr.image_processing_detr.remove_low_and_no_objects
+def remove_low_and_no_objects(masks, scores, labels, object_mask_threshold, num_labels):
+    """
+    Binarize the given masks using `object_mask_threshold`, it returns the associated values of `masks`, `scores` and
+    `labels`.
+
+    Args:
+        masks (`torch.Tensor`):
+            A tensor of shape `(num_queries, height, width)`.
+        scores (`torch.Tensor`):
+            A tensor of shape `(num_queries)`.
+        labels (`torch.Tensor`):
+            A tensor of shape `(num_queries)`.
+        object_mask_threshold (`float`):
+            A number between 0 and 1 used to binarize the masks.
+    Raises:
+        `ValueError`: Raised when the first dimension doesn't match in all input tensors.
+    Returns:
+        `Tuple[`torch.Tensor`, `torch.Tensor`, `torch.Tensor`]`: The `masks`, `scores` and `labels` without the region
+        < `object_mask_threshold`.
+    """
+    if not (masks.shape[0] == scores.shape[0] == labels.shape[0]):
+        raise ValueError("mask, scores and labels must have the same shape!")
+
+    to_keep = labels.ne(num_labels) & (scores > object_mask_threshold)
+
+    return masks[to_keep], scores[to_keep], labels[to_keep]
+
+
+# Copied from transformers.models.detr.image_processing_detr.check_segment_validity
+def check_segment_validity(mask_labels, mask_probs, k, mask_threshold=0.5, overlap_mask_area_threshold=0.8):
+    # Get the mask associated with the k class
+    mask_k = mask_labels == k
+    mask_k_area = mask_k.sum()
+
+    # Compute the area of all the stuff in query k
+    original_area = (mask_probs[k] >= mask_threshold).sum()
+    mask_exists = mask_k_area > 0 and original_area > 0
+
+    # Eliminate disconnected tiny segments
+    if mask_exists:
+        area_ratio = mask_k_area / original_area
+        if not area_ratio.item() > overlap_mask_area_threshold:
+            mask_exists = False
+
+    return mask_exists, mask_k
+
+
+# Copied from transformers.models.detr.image_processing_detr.compute_segments
+def compute_segments(
+    mask_probs,
+    pred_scores,
+    pred_labels,
+    mask_threshold: float = 0.5,
+    overlap_mask_area_threshold: float = 0.8,
+    label_ids_to_fuse: Optional[Set[int]] = None,
+    target_size: Tuple[int, int] = None,
+):
+    height = mask_probs.shape[1] if target_size is None else target_size[0]
+    width = mask_probs.shape[2] if target_size is None else target_size[1]
+
+    segmentation = torch.zeros((height, width), dtype=torch.int32, device=mask_probs.device)
+    segments: List[Dict] = []
+
+    if target_size is not None:
+        mask_probs = nn.functional.interpolate(
+            mask_probs.unsqueeze(0), size=target_size, mode="bilinear", align_corners=False
+        )[0]
+
+    current_segment_id = 0
+
+    # Weigh each mask by its prediction score
+    mask_probs *= pred_scores.view(-1, 1, 1)
+    mask_labels = mask_probs.argmax(0)  # [height, width]
+
+    # Keep track of instances of each class
+    stuff_memory_list: Dict[str, int] = {}
+    for k in range(pred_labels.shape[0]):
+        pred_class = pred_labels[k].item()
+        should_fuse = pred_class in label_ids_to_fuse
+
+        # Check if mask exists and large enough to be a segment
+        mask_exists, mask_k = check_segment_validity(
+            mask_labels, mask_probs, k, mask_threshold, overlap_mask_area_threshold
+        )
+
+        if mask_exists:
+            if pred_class in stuff_memory_list:
+                current_segment_id = stuff_memory_list[pred_class]
+            else:
+                current_segment_id += 1
+
+            # Add current object segment to final segmentation map
+            segmentation[mask_k] = current_segment_id
+            segment_score = round(pred_scores[k].item(), 6)
+            segments.append(
+                {
+                    "id": current_segment_id,
+                    "label_id": pred_class,
+                    "was_fused": should_fuse,
+                    "score": segment_score,
+                }
+            )
+            if should_fuse:
+                stuff_memory_list[pred_class] = current_segment_id
+
+    return segmentation, segments
+
+
+class ConditionalDetrImageProcessor(BaseImageProcessor):
+    r"""
+    Constructs a Conditional Detr image processor.
+
+    Args:
+        format (`str`, *optional*, defaults to `"coco_detection"`):
+            Data format of the annotations. One of "coco_detection" or "coco_panoptic".
+        do_resize (`bool`, *optional*, defaults to `True`):
+            Controls whether to resize the image's (height, width) dimensions to the specified `size`. Can be
+            overridden by the `do_resize` parameter in the `preprocess` method.
+        size (`Dict[str, int]` *optional*, defaults to `{"shortest_edge": 800, "longest_edge": 1333}`):
+            Size of the image's `(height, width)` dimensions after resizing. Can be overridden by the `size` parameter
+            in the `preprocess` method. Available options are:
+                - `{"height": int, "width": int}`: The image will be resized to the exact size `(height, width)`.
+                    Do NOT keep the aspect ratio.
+                - `{"shortest_edge": int, "longest_edge": int}`: The image will be resized to a maximum size respecting
+                    the aspect ratio and keeping the shortest edge less or equal to `shortest_edge` and the longest edge
+                    less or equal to `longest_edge`.
+                - `{"max_height": int, "max_width": int}`: The image will be resized to the maximum size respecting the
+                    aspect ratio and keeping the height less or equal to `max_height` and the width less or equal to
+                    `max_width`.
+        resample (`PILImageResampling`, *optional*, defaults to `PILImageResampling.BILINEAR`):
+            Resampling filter to use if resizing the image.
+        do_rescale (`bool`, *optional*, defaults to `True`):
+            Controls whether to rescale the image by the specified scale `rescale_factor`. Can be overridden by the
+            `do_rescale` parameter in the `preprocess` method.
+        rescale_factor (`int` or `float`, *optional*, defaults to `1/255`):
+            Scale factor to use if rescaling the image. Can be overridden by the `rescale_factor` parameter in the
+            `preprocess` method.
+        do_normalize:
+            Controls whether to normalize the image. Can be overridden by the `do_normalize` parameter in the
+            `preprocess` method.
+        image_mean (`float` or `List[float]`, *optional*, defaults to `IMAGENET_DEFAULT_MEAN`):
+            Mean values to use when normalizing the image. Can be a single value or a list of values, one for each
+            channel. Can be overridden by the `image_mean` parameter in the `preprocess` method.
+        image_std (`float` or `List[float]`, *optional*, defaults to `IMAGENET_DEFAULT_STD`):
+            Standard deviation values to use when normalizing the image. Can be a single value or a list of values, one
+            for each channel. Can be overridden by the `image_std` parameter in the `preprocess` method.
+        do_convert_annotations (`bool`, *optional*, defaults to `True`):
+            Controls whether to convert the annotations to the format expected by the DETR model. Converts the
+            bounding boxes to the format `(center_x, center_y, width, height)` and in the range `[0, 1]`.
+            Can be overridden by the `do_convert_annotations` parameter in the `preprocess` method.
+        do_pad (`bool`, *optional*, defaults to `True`):
+            Controls whether to pad the image. Can be overridden by the `do_pad` parameter in the `preprocess`
+            method. If `True`, padding will be applied to the bottom and right of the image with zeros.
+            If `pad_size` is provided, the image will be padded to the specified dimensions.
+            Otherwise, the image will be padded to the maximum height and width of the batch.
+        pad_size (`Dict[str, int]`, *optional*):
+            The size `{"height": int, "width" int}` to pad the images to. Must be larger than any image size
+            provided for preprocessing. If `pad_size` is not provided, images will be padded to the largest
+            height and width in the batch.
+    """
+
+    model_input_names = ["pixel_values", "pixel_mask"]
+
+    # Copied from transformers.models.detr.image_processing_detr.DetrImageProcessor.__init__
+    def __init__(
+        self,
+        format: Union[str, AnnotationFormat] = AnnotationFormat.COCO_DETECTION,
+        do_resize: bool = True,
+        size: Dict[str, int] = None,
+        resample: PILImageResampling = PILImageResampling.BILINEAR,
+        do_rescale: bool = True,
+        rescale_factor: Union[int, float] = 1 / 255,
+        do_normalize: bool = True,
+        image_mean: Union[float, List[float]] = None,
+        image_std: Union[float, List[float]] = None,
+        do_convert_annotations: Optional[bool] = None,
+        do_pad: bool = True,
+        pad_size: Optional[Dict[str, int]] = None,
+        **kwargs,
+    ) -> None:
+        if "pad_and_return_pixel_mask" in kwargs:
+            do_pad = kwargs.pop("pad_and_return_pixel_mask")
+
+        if "max_size" in kwargs:
+            logger.warning_once(
+                "The `max_size` parameter is deprecated and will be removed in v4.26. "
+                "Please specify in `size['longest_edge'] instead`.",
+            )
+            max_size = kwargs.pop("max_size")
+        else:
+            max_size = None if size is None else 1333
+
+        size = size if size is not None else {"shortest_edge": 800, "longest_edge": 1333}
+        size = get_size_dict(size, max_size=max_size, default_to_square=False)
+
+        # Backwards compatibility
+        if do_convert_annotations is None:
+            do_convert_annotations = do_normalize
+
+        super().__init__(**kwargs)
+        self.format = format
+        self.do_resize = do_resize
+        self.size = size
+        self.resample = resample
+        self.do_rescale = do_rescale
+        self.rescale_factor = rescale_factor
+        self.do_normalize = do_normalize
+        self.do_convert_annotations = do_convert_annotations
+        self.image_mean = image_mean if image_mean is not None else IMAGENET_DEFAULT_MEAN
+        self.image_std = image_std if image_std is not None else IMAGENET_DEFAULT_STD
+        self.do_pad = do_pad
+        self.pad_size = pad_size
+        self._valid_processor_keys = [
+            "images",
+            "annotations",
+            "return_segmentation_masks",
+            "masks_path",
+            "do_resize",
+            "size",
+            "resample",
+            "do_rescale",
+            "rescale_factor",
+            "do_normalize",
+            "do_convert_annotations",
+            "image_mean",
+            "image_std",
+            "do_pad",
+            "pad_size",
+            "format",
+            "return_tensors",
+            "data_format",
+            "input_data_format",
+        ]
+
+    @classmethod
+    # Copied from transformers.models.detr.image_processing_detr.DetrImageProcessor.from_dict with Detr->ConditionalDetr
+    def from_dict(cls, image_processor_dict: Dict[str, Any], **kwargs):
+        """
+        Overrides the `from_dict` method from the base class to make sure parameters are updated if image processor is
+        created using from_dict and kwargs e.g. `ConditionalDetrImageProcessor.from_pretrained(checkpoint, size=600,
+        max_size=800)`
+        """
+        image_processor_dict = image_processor_dict.copy()
+        if "max_size" in kwargs:
+            image_processor_dict["max_size"] = kwargs.pop("max_size")
+        if "pad_and_return_pixel_mask" in kwargs:
+            image_processor_dict["pad_and_return_pixel_mask"] = kwargs.pop("pad_and_return_pixel_mask")
+        return super().from_dict(image_processor_dict, **kwargs)
+
+    # Copied from transformers.models.detr.image_processing_detr.DetrImageProcessor.prepare_annotation with DETR->ConditionalDetr
+    def prepare_annotation(
+        self,
+        image: np.ndarray,
+        target: Dict,
+        format: Optional[AnnotationFormat] = None,
+        return_segmentation_masks: Optional[bool] = None,
+        masks_path: Optional[Union[str, pathlib.Path]] = None,
+        input_data_format: Optional[Union[str, ChannelDimension]] = None,
+    ) -> Dict:
+        """
+        Prepare an annotation for feeding into ConditionalDetr model.
+        """
+        format = format if format is not None else self.format
+
+        if format == AnnotationFormat.COCO_DETECTION:
+            return_segmentation_masks = False if return_segmentation_masks is None else return_segmentation_masks
+            target = prepare_coco_detection_annotation(
+                image, target, return_segmentation_masks, input_data_format=input_data_format
+            )
+        elif format == AnnotationFormat.COCO_PANOPTIC:
+            return_segmentation_masks = True if return_segmentation_masks is None else return_segmentation_masks
+            target = prepare_coco_panoptic_annotation(
+                image,
+                target,
+                masks_path=masks_path,
+                return_masks=return_segmentation_masks,
+                input_data_format=input_data_format,
+            )
+        else:
+            raise ValueError(f"Format {format} is not supported.")
+        return target
+
+    # Copied from transformers.models.detr.image_processing_detr.DetrImageProcessor.resize
+    def resize(
+        self,
+        image: np.ndarray,
+        size: Dict[str, int],
+        resample: PILImageResampling = PILImageResampling.BILINEAR,
+        data_format: Optional[ChannelDimension] = None,
+        input_data_format: Optional[Union[str, ChannelDimension]] = None,
+        **kwargs,
+    ) -> np.ndarray:
+        """
+        Resize the image to the given size. Size can be `min_size` (scalar) or `(height, width)` tuple. If size is an
+        int, smaller edge of the image will be matched to this number.
+
+        Args:
+            image (`np.ndarray`):
+                Image to resize.
+            size (`Dict[str, int]`):
+                Size of the image's `(height, width)` dimensions after resizing. Available options are:
+                    - `{"height": int, "width": int}`: The image will be resized to the exact size `(height, width)`.
+                        Do NOT keep the aspect ratio.
+                    - `{"shortest_edge": int, "longest_edge": int}`: The image will be resized to a maximum size respecting
+                        the aspect ratio and keeping the shortest edge less or equal to `shortest_edge` and the longest edge
+                        less or equal to `longest_edge`.
+                    - `{"max_height": int, "max_width": int}`: The image will be resized to the maximum size respecting the
+                        aspect ratio and keeping the height less or equal to `max_height` and the width less or equal to
+                        `max_width`.
+            resample (`PILImageResampling`, *optional*, defaults to `PILImageResampling.BILINEAR`):
+                Resampling filter to use if resizing the image.
+            data_format (`str` or `ChannelDimension`, *optional*):
+                The channel dimension format for the output image. If unset, the channel dimension format of the input
+                image is used.
+            input_data_format (`ChannelDimension` or `str`, *optional*):
+                The channel dimension format of the input image. If not provided, it will be inferred.
+        """
+        if "max_size" in kwargs:
+            logger.warning_once(
+                "The `max_size` parameter is deprecated and will be removed in v4.26. "
+                "Please specify in `size['longest_edge'] instead`.",
+            )
+            max_size = kwargs.pop("max_size")
+        else:
+            max_size = None
+        size = get_size_dict(size, max_size=max_size, default_to_square=False)
+        if "shortest_edge" in size and "longest_edge" in size:
+            new_size = get_resize_output_image_size(
+                image, size["shortest_edge"], size["longest_edge"], input_data_format=input_data_format
+            )
+        elif "max_height" in size and "max_width" in size:
+            new_size = get_image_size_for_max_height_width(
+                image, size["max_height"], size["max_width"], input_data_format=input_data_format
+            )
+        elif "height" in size and "width" in size:
+            new_size = (size["height"], size["width"])
+        else:
+            raise ValueError(
+                "Size must contain 'height' and 'width' keys or 'shortest_edge' and 'longest_edge' keys. Got"
+                f" {size.keys()}."
+            )
+        image = resize(
+            image,
+            size=new_size,
+            resample=resample,
+            data_format=data_format,
+            input_data_format=input_data_format,
+            **kwargs,
+        )
+        return image
+
+    # Copied from transformers.models.detr.image_processing_detr.DetrImageProcessor.resize_annotation
+    def resize_annotation(
+        self,
+        annotation,
+        orig_size,
+        size,
+        resample: PILImageResampling = PILImageResampling.NEAREST,
+    ) -> Dict:
+        """
+        Resize the annotation to match the resized image. If size is an int, smaller edge of the mask will be matched
+        to this number.
+        """
+        return resize_annotation(annotation, orig_size=orig_size, target_size=size, resample=resample)
+
+    # Copied from transformers.models.detr.image_processing_detr.DetrImageProcessor.rescale
+    def rescale(
+        self,
+        image: np.ndarray,
+        rescale_factor: float,
+        data_format: Optional[Union[str, ChannelDimension]] = None,
+        input_data_format: Optional[Union[str, ChannelDimension]] = None,
+    ) -> np.ndarray:
+        """
+        Rescale the image by the given factor. image = image * rescale_factor.
+
+        Args:
+            image (`np.ndarray`):
+                Image to rescale.
+            rescale_factor (`float`):
+                The value to use for rescaling.
+            data_format (`str` or `ChannelDimension`, *optional*):
+                The channel dimension format for the output image. If unset, the channel dimension format of the input
+                image is used. Can be one of:
+                - `"channels_first"` or `ChannelDimension.FIRST`: image in (num_channels, height, width) format.
+                - `"channels_last"` or `ChannelDimension.LAST`: image in (height, width, num_channels) format.
+            input_data_format (`str` or `ChannelDimension`, *optional*):
+                The channel dimension format for the input image. If unset, 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.
+        """
+        return rescale(image, rescale_factor, data_format=data_format, input_data_format=input_data_format)
+
+    # Copied from transformers.models.detr.image_processing_detr.DetrImageProcessor.normalize_annotation
+    def normalize_annotation(self, annotation: Dict, image_size: Tuple[int, int]) -> Dict:
+        """
+        Normalize the boxes in the annotation from `[top_left_x, top_left_y, bottom_right_x, bottom_right_y]` to
+        `[center_x, center_y, width, height]` format and from absolute to relative pixel values.
+        """
+        return normalize_annotation(annotation, image_size=image_size)
+
+    # Copied from transformers.models.detr.image_processing_detr.DetrImageProcessor._update_annotation_for_padded_image
+    def _update_annotation_for_padded_image(
+        self,
+        annotation: Dict,
+        input_image_size: Tuple[int, int],
+        output_image_size: Tuple[int, int],
+        padding,
+        update_bboxes,
+    ) -> Dict:
+        """
+        Update the annotation for a padded image.
+        """
+        new_annotation = {}
+        new_annotation["size"] = output_image_size
+
+        for key, value in annotation.items():
+            if key == "masks":
+                masks = value
+                masks = pad(
+                    masks,
+                    padding,
+                    mode=PaddingMode.CONSTANT,
+                    constant_values=0,
+                    input_data_format=ChannelDimension.FIRST,
+                )
+                masks = safe_squeeze(masks, 1)
+                new_annotation["masks"] = masks
+            elif key == "boxes" and update_bboxes:
+                boxes = value
+                boxes *= np.asarray(
+                    [
+                        input_image_size[1] / output_image_size[1],
+                        input_image_size[0] / output_image_size[0],
+                        input_image_size[1] / output_image_size[1],
+                        input_image_size[0] / output_image_size[0],
+                    ]
+                )
+                new_annotation["boxes"] = boxes
+            elif key == "size":
+                new_annotation["size"] = output_image_size
+            else:
+                new_annotation[key] = value
+        return new_annotation
+
+    # Copied from transformers.models.detr.image_processing_detr.DetrImageProcessor._pad_image
+    def _pad_image(
+        self,
+        image: np.ndarray,
+        output_size: Tuple[int, int],
+        annotation: Optional[Dict[str, Any]] = None,
+        constant_values: Union[float, Iterable[float]] = 0,
+        data_format: Optional[ChannelDimension] = None,
+        input_data_format: Optional[Union[str, ChannelDimension]] = None,
+        update_bboxes: bool = True,
+    ) -> np.ndarray:
+        """
+        Pad an image with zeros to the given size.
+        """
+        input_height, input_width = get_image_size(image, channel_dim=input_data_format)
+        output_height, output_width = output_size
+
+        pad_bottom = output_height - input_height
+        pad_right = output_width - input_width
+        padding = ((0, pad_bottom), (0, pad_right))
+        padded_image = pad(
+            image,
+            padding,
+            mode=PaddingMode.CONSTANT,
+            constant_values=constant_values,
+            data_format=data_format,
+            input_data_format=input_data_format,
+        )
+        if annotation is not None:
+            annotation = self._update_annotation_for_padded_image(
+                annotation, (input_height, input_width), (output_height, output_width), padding, update_bboxes
+            )
+        return padded_image, annotation
+
+    # Copied from transformers.models.detr.image_processing_detr.DetrImageProcessor.pad
+    def pad(
+        self,
+        images: List[np.ndarray],
+        annotations: Optional[Union[AnnotationType, List[AnnotationType]]] = None,
+        constant_values: Union[float, Iterable[float]] = 0,
+        return_pixel_mask: bool = True,
+        return_tensors: Optional[Union[str, TensorType]] = None,
+        data_format: Optional[ChannelDimension] = None,
+        input_data_format: Optional[Union[str, ChannelDimension]] = None,
+        update_bboxes: bool = True,
+        pad_size: Optional[Dict[str, int]] = None,
+    ) -> BatchFeature:
+        """
+        Pads a batch of images to the bottom and right of the image with zeros to the size of largest height and width
+        in the batch and optionally returns their corresponding pixel mask.
+
+        Args:
+            images (List[`np.ndarray`]):
+                Images to pad.
+            annotations (`AnnotationType` or `List[AnnotationType]`, *optional*):
+                Annotations to transform according to the padding that is applied to the images.
+            constant_values (`float` or `Iterable[float]`, *optional*):
+                The value to use for the padding if `mode` is `"constant"`.
+            return_pixel_mask (`bool`, *optional*, defaults to `True`):
+                Whether to return a pixel mask.
+            return_tensors (`str` or `TensorType`, *optional*):
+                The type of tensors to return. Can be one of:
+                    - Unset: Return a list of `np.ndarray`.
+                    - `TensorType.TENSORFLOW` or `'tf'`: Return a batch of type `tf.Tensor`.
+                    - `TensorType.PYTORCH` or `'pt'`: Return a batch of type `torch.Tensor`.
+                    - `TensorType.NUMPY` or `'np'`: Return a batch of type `np.ndarray`.
+                    - `TensorType.JAX` or `'jax'`: Return a batch of type `jax.numpy.ndarray`.
+            data_format (`str` or `ChannelDimension`, *optional*):
+                The channel dimension format of the image. If not provided, it will be the same as the input image.
+            input_data_format (`ChannelDimension` or `str`, *optional*):
+                The channel dimension format of the input image. If not provided, it will be inferred.
+            update_bboxes (`bool`, *optional*, defaults to `True`):
+                Whether to update the bounding boxes in the annotations to match the padded images. If the
+                bounding boxes have not been converted to relative coordinates and `(centre_x, centre_y, width, height)`
+                format, the bounding boxes will not be updated.
+            pad_size (`Dict[str, int]`, *optional*):
+                The size `{"height": int, "width" int}` to pad the images to. Must be larger than any image size
+                provided for preprocessing. If `pad_size` is not provided, images will be padded to the largest
+                height and width in the batch.
+        """
+        pad_size = pad_size if pad_size is not None else self.pad_size
+        if pad_size is not None:
+            padded_size = (pad_size["height"], pad_size["width"])
+        else:
+            padded_size = get_max_height_width(images, input_data_format=input_data_format)
+
+        annotation_list = annotations if annotations is not None else [None] * len(images)
+        padded_images = []
+        padded_annotations = []
+        for image, annotation in zip(images, annotation_list):
+            padded_image, padded_annotation = self._pad_image(
+                image,
+                padded_size,
+                annotation,
+                constant_values=constant_values,
+                data_format=data_format,
+                input_data_format=input_data_format,
+                update_bboxes=update_bboxes,
+            )
+            padded_images.append(padded_image)
+            padded_annotations.append(padded_annotation)
+
+        data = {"pixel_values": padded_images}
+
+        if return_pixel_mask:
+            masks = [
+                make_pixel_mask(image=image, output_size=padded_size, input_data_format=input_data_format)
+                for image in images
+            ]
+            data["pixel_mask"] = masks
+
+        encoded_inputs = BatchFeature(data=data, tensor_type=return_tensors)
+
+        if annotations is not None:
+            encoded_inputs["labels"] = [
+                BatchFeature(annotation, tensor_type=return_tensors) for annotation in padded_annotations
+            ]
+
+        return encoded_inputs
+
+    # Copied from transformers.models.detr.image_processing_detr.DetrImageProcessor.preprocess
+    def preprocess(
+        self,
+        images: ImageInput,
+        annotations: Optional[Union[AnnotationType, List[AnnotationType]]] = None,
+        return_segmentation_masks: Optional[bool] = None,
+        masks_path: Optional[Union[str, pathlib.Path]] = None,
+        do_resize: Optional[bool] = None,
+        size: Optional[Dict[str, int]] = None,
+        resample=None,  # PILImageResampling
+        do_rescale: Optional[bool] = None,
+        rescale_factor: Optional[Union[int, float]] = None,
+        do_normalize: Optional[bool] = None,
+        do_convert_annotations: Optional[bool] = None,
+        image_mean: Optional[Union[float, List[float]]] = None,
+        image_std: Optional[Union[float, List[float]]] = None,
+        do_pad: Optional[bool] = None,
+        format: Optional[Union[str, AnnotationFormat]] = None,
+        return_tensors: Optional[Union[TensorType, str]] = None,
+        data_format: Union[str, ChannelDimension] = ChannelDimension.FIRST,
+        input_data_format: Optional[Union[str, ChannelDimension]] = None,
+        pad_size: Optional[Dict[str, int]] = None,
+        **kwargs,
+    ) -> BatchFeature:
+        """
+        Preprocess an image or a batch of images so that it can be used by the model.
+
+        Args:
+            images (`ImageInput`):
+                Image or batch of images 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`.
+            annotations (`AnnotationType` or `List[AnnotationType]`, *optional*):
+                List of annotations associated with the image or batch of images. If annotation is for object
+                detection, the annotations should be a dictionary with the following keys:
+                - "image_id" (`int`): The image id.
+                - "annotations" (`List[Dict]`): List of annotations for an image. Each annotation should be a
+                  dictionary. An image can have no annotations, in which case the list should be empty.
+                If annotation is for segmentation, the annotations should be a dictionary with the following keys:
+                - "image_id" (`int`): The image id.
+                - "segments_info" (`List[Dict]`): List of segments for an image. Each segment should be a dictionary.
+                  An image can have no segments, in which case the list should be empty.
+                - "file_name" (`str`): The file name of the image.
+            return_segmentation_masks (`bool`, *optional*, defaults to self.return_segmentation_masks):
+                Whether to return segmentation masks.
+            masks_path (`str` or `pathlib.Path`, *optional*):
+                Path to the directory containing the segmentation masks.
+            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's `(height, width)` dimensions after resizing. Available options are:
+                    - `{"height": int, "width": int}`: The image will be resized to the exact size `(height, width)`.
+                        Do NOT keep the aspect ratio.
+                    - `{"shortest_edge": int, "longest_edge": int}`: The image will be resized to a maximum size respecting
+                        the aspect ratio and keeping the shortest edge less or equal to `shortest_edge` and the longest edge
+                        less or equal to `longest_edge`.
+                    - `{"max_height": int, "max_width": int}`: The image will be resized to the maximum size respecting the
+                        aspect ratio and keeping the height less or equal to `max_height` and the width less or equal to
+                        `max_width`.
+            resample (`PILImageResampling`, *optional*, defaults to self.resample):
+                Resampling filter to use when resizing the image.
+            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 use when rescaling the image.
+            do_normalize (`bool`, *optional*, defaults to self.do_normalize):
+                Whether to normalize the image.
+            do_convert_annotations (`bool`, *optional*, defaults to self.do_convert_annotations):
+                Whether to convert the annotations to the format expected by the model. Converts the bounding
+                boxes from the format `(top_left_x, top_left_y, width, height)` to `(center_x, center_y, width, height)`
+                and in relative coordinates.
+            image_mean (`float` or `List[float]`, *optional*, defaults to self.image_mean):
+                Mean to use when normalizing the image.
+            image_std (`float` or `List[float]`, *optional*, defaults to self.image_std):
+                Standard deviation to use when normalizing the image.
+            do_pad (`bool`, *optional*, defaults to self.do_pad):
+                Whether to pad the image. If `True`, padding will be applied to the bottom and right of
+                the image with zeros. If `pad_size` is provided, the image will be padded to the specified
+                dimensions. Otherwise, the image will be padded to the maximum height and width of the batch.
+            format (`str` or `AnnotationFormat`, *optional*, defaults to self.format):
+                Format of the annotations.
+            return_tensors (`str` or `TensorType`, *optional*, defaults to self.return_tensors):
+                Type of tensors to return. If `None`, will return the list of images.
+            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.
+            pad_size (`Dict[str, int]`, *optional*):
+                The size `{"height": int, "width" int}` to pad the images to. Must be larger than any image size
+                provided for preprocessing. If `pad_size` is not provided, images will be padded to the largest
+                height and width in the batch.
+        """
+        if "pad_and_return_pixel_mask" in kwargs:
+            logger.warning_once(
+                "The `pad_and_return_pixel_mask` argument is deprecated and will be removed in a future version, "
+                "use `do_pad` instead."
+            )
+            do_pad = kwargs.pop("pad_and_return_pixel_mask")
+
+        if "max_size" in kwargs:
+            logger.warning_once(
+                "The `max_size` argument is deprecated and will be removed in a future version, use"
+                " `size['longest_edge']` instead."
+            )
+            size = kwargs.pop("max_size")
+
+        do_resize = self.do_resize if do_resize is None else do_resize
+        size = self.size if size is None else size
+        size = get_size_dict(size=size, default_to_square=False)
+        resample = self.resample if resample is None else resample
+        do_rescale = self.do_rescale if do_rescale is None else do_rescale
+        rescale_factor = self.rescale_factor if rescale_factor is None else rescale_factor
+        do_normalize = self.do_normalize if do_normalize is None else do_normalize
+        image_mean = self.image_mean if image_mean is None else image_mean
+        image_std = self.image_std if image_std is None else image_std
+        do_convert_annotations = (
+            self.do_convert_annotations if do_convert_annotations is None else do_convert_annotations
+        )
+        do_pad = self.do_pad if do_pad is None else do_pad
+        pad_size = self.pad_size if pad_size is None else pad_size
+        format = self.format if format is None else format
+
+        images = make_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_kwargs(captured_kwargs=kwargs.keys(), valid_processor_keys=self._valid_processor_keys)
+
+        # Here, the pad() method pads to the maximum of (width, height). It does not need to be validated.
+        validate_preprocess_arguments(
+            do_rescale=do_rescale,
+            rescale_factor=rescale_factor,
+            do_normalize=do_normalize,
+            image_mean=image_mean,
+            image_std=image_std,
+            do_resize=do_resize,
+            size=size,
+            resample=resample,
+        )
+
+        if annotations is not None and isinstance(annotations, dict):
+            annotations = [annotations]
+
+        if annotations is not None and len(images) != len(annotations):
+            raise ValueError(
+                f"The number of images ({len(images)}) and annotations ({len(annotations)}) do not match."
+            )
+
+        format = AnnotationFormat(format)
+        if annotations is not None:
+            validate_annotations(format, SUPPORTED_ANNOTATION_FORMATS, annotations)
+
+        if (
+            masks_path is not None
+            and format == AnnotationFormat.COCO_PANOPTIC
+            and not isinstance(masks_path, (pathlib.Path, str))
+        ):
+            raise ValueError(
+                "The path to the directory containing the mask PNG files should be provided as a"
+                f" `pathlib.Path` or string object, but is {type(masks_path)} instead."
+            )
+
+        # 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])
+
+        # prepare (COCO annotations as a list of Dict -> DETR target as a single Dict per image)
+        if annotations is not None:
+            prepared_images = []
+            prepared_annotations = []
+            for image, target in zip(images, annotations):
+                target = self.prepare_annotation(
+                    image,
+                    target,
+                    format,
+                    return_segmentation_masks=return_segmentation_masks,
+                    masks_path=masks_path,
+                    input_data_format=input_data_format,
+                )
+                prepared_images.append(image)
+                prepared_annotations.append(target)
+            images = prepared_images
+            annotations = prepared_annotations
+            del prepared_images, prepared_annotations
+
+        # transformations
+        if do_resize:
+            if annotations is not None:
+                resized_images, resized_annotations = [], []
+                for image, target in zip(images, annotations):
+                    orig_size = get_image_size(image, input_data_format)
+                    resized_image = self.resize(
+                        image, size=size, resample=resample, input_data_format=input_data_format
+                    )
+                    resized_annotation = self.resize_annotation(
+                        target, orig_size, get_image_size(resized_image, input_data_format)
+                    )
+                    resized_images.append(resized_image)
+                    resized_annotations.append(resized_annotation)
+                images = resized_images
+                annotations = resized_annotations
+                del resized_images, resized_annotations
+            else:
+                images = [
+                    self.resize(image, size=size, resample=resample, input_data_format=input_data_format)
+                    for image in images
+                ]
+
+        if do_rescale:
+            images = [self.rescale(image, rescale_factor, input_data_format=input_data_format) for image in images]
+
+        if do_normalize:
+            images = [
+                self.normalize(image, image_mean, image_std, input_data_format=input_data_format) for image in images
+            ]
+
+        if do_convert_annotations and annotations is not None:
+            annotations = [
+                self.normalize_annotation(annotation, get_image_size(image, input_data_format))
+                for annotation, image in zip(annotations, images)
+            ]
+
+        if do_pad:
+            # Pads images and returns their mask: {'pixel_values': ..., 'pixel_mask': ...}
+            encoded_inputs = self.pad(
+                images,
+                annotations=annotations,
+                return_pixel_mask=True,
+                data_format=data_format,
+                input_data_format=input_data_format,
+                update_bboxes=do_convert_annotations,
+                return_tensors=return_tensors,
+                pad_size=pad_size,
+            )
+        else:
+            images = [
+                to_channel_dimension_format(image, data_format, input_channel_dim=input_data_format)
+                for image in images
+            ]
+            encoded_inputs = BatchFeature(data={"pixel_values": images}, tensor_type=return_tensors)
+            if annotations is not None:
+                encoded_inputs["labels"] = [
+                    BatchFeature(annotation, tensor_type=return_tensors) for annotation in annotations
+                ]
+
+        return encoded_inputs
+
+    # POSTPROCESSING METHODS - TODO: add support for other frameworks
+    def post_process(self, outputs, target_sizes):
+        """
+        Converts the output of [`ConditionalDetrForObjectDetection`] into the format expected by the Pascal VOC format (xmin, ymin, xmax, ymax).
+        Only supports PyTorch.
+
+        Args:
+            outputs ([`ConditionalDetrObjectDetectionOutput`]):
+                Raw outputs of the model.
+            target_sizes (`torch.Tensor` of shape `(batch_size, 2)`):
+                Tensor containing the size (h, w) of each image of the batch. For evaluation, this must be the original
+                image size (before any data augmentation). For visualization, this should be the image size after data
+                augment, but before padding.
+        Returns:
+            `List[Dict]`: A list of dictionaries, each dictionary containing the scores, labels and boxes for an image
+            in the batch as predicted by the model.
+        """
+        logging.warning_once(
+            "`post_process` is deprecated and will be removed in v5 of Transformers, please use"
+            " `post_process_object_detection` instead, with `threshold=0.` for equivalent results.",
+        )
+
+        out_logits, out_bbox = outputs.logits, outputs.pred_boxes
+
+        if len(out_logits) != len(target_sizes):
+            raise ValueError("Make sure that you pass in as many target sizes as the batch dimension of the logits")
+        if target_sizes.shape[1] != 2:
+            raise ValueError("Each element of target_sizes must contain the size (h, w) of each image of the batch")
+
+        prob = out_logits.sigmoid()
+        topk_values, topk_indexes = torch.topk(prob.view(out_logits.shape[0], -1), 300, dim=1)
+        scores = topk_values
+        topk_boxes = torch.div(topk_indexes, out_logits.shape[2], rounding_mode="floor")
+        labels = topk_indexes % out_logits.shape[2]
+        boxes = center_to_corners_format(out_bbox)
+        boxes = torch.gather(boxes, 1, topk_boxes.unsqueeze(-1).repeat(1, 1, 4))
+
+        # and from relative [0, 1] to absolute [0, height] coordinates
+        img_h, img_w = target_sizes.unbind(1)
+        scale_fct = torch.stack([img_w, img_h, img_w, img_h], dim=1)
+        boxes = boxes * scale_fct[:, None, :]
+
+        results = [{"scores": s, "labels": l, "boxes": b} for s, l, b in zip(scores, labels, boxes)]
+
+        return results
+
+    # Copied from transformers.models.deformable_detr.image_processing_deformable_detr.DeformableDetrImageProcessor.post_process_object_detection with DeformableDetr->ConditionalDetr
+    def post_process_object_detection(
+        self, outputs, threshold: float = 0.5, target_sizes: Union[TensorType, List[Tuple]] = None, top_k: int = 100
+    ):
+        """
+        Converts the raw output of [`ConditionalDetrForObjectDetection`] into final bounding boxes in (top_left_x,
+        top_left_y, bottom_right_x, bottom_right_y) format. Only supports PyTorch.
+
+        Args:
+            outputs ([`DetrObjectDetectionOutput`]):
+                Raw outputs of the model.
+            threshold (`float`, *optional*):
+                Score threshold to keep object detection predictions.
+            target_sizes (`torch.Tensor` or `List[Tuple[int, int]]`, *optional*):
+                Tensor of shape `(batch_size, 2)` or list of tuples (`Tuple[int, int]`) containing the target size
+                (height, width) of each image in the batch. If left to None, predictions will not be resized.
+            top_k (`int`, *optional*, defaults to 100):
+                Keep only top k bounding boxes before filtering by thresholding.
+
+        Returns:
+            `List[Dict]`: A list of dictionaries, each dictionary containing the scores, labels and boxes for an image
+            in the batch as predicted by the model.
+        """
+        out_logits, out_bbox = outputs.logits, outputs.pred_boxes
+
+        if target_sizes is not None:
+            if len(out_logits) != len(target_sizes):
+                raise ValueError(
+                    "Make sure that you pass in as many target sizes as the batch dimension of the logits"
+                )
+
+        prob = out_logits.sigmoid()
+        prob = prob.view(out_logits.shape[0], -1)
+        k_value = min(top_k, prob.size(1))
+        topk_values, topk_indexes = torch.topk(prob, k_value, dim=1)
+        scores = topk_values
+        topk_boxes = torch.div(topk_indexes, out_logits.shape[2], rounding_mode="floor")
+        labels = topk_indexes % out_logits.shape[2]
+        boxes = center_to_corners_format(out_bbox)
+        boxes = torch.gather(boxes, 1, topk_boxes.unsqueeze(-1).repeat(1, 1, 4))
+
+        # and from relative [0, 1] to absolute [0, height] coordinates
+        if target_sizes is not None:
+            if isinstance(target_sizes, List):
+                img_h = torch.Tensor([i[0] for i in target_sizes])
+                img_w = torch.Tensor([i[1] for i in target_sizes])
+            else:
+                img_h, img_w = target_sizes.unbind(1)
+            scale_fct = torch.stack([img_w, img_h, img_w, img_h], dim=1).to(boxes.device)
+            boxes = boxes * scale_fct[:, None, :]
+
+        results = []
+        for s, l, b in zip(scores, labels, boxes):
+            score = s[s > threshold]
+            label = l[s > threshold]
+            box = b[s > threshold]
+            results.append({"scores": score, "labels": label, "boxes": box})
+
+        return results
+
+    # Copied from transformers.models.detr.image_processing_detr.DetrImageProcessor.post_process_semantic_segmentation with Detr->ConditionalDetr
+    def post_process_semantic_segmentation(self, outputs, target_sizes: List[Tuple[int, int]] = None):
+        """
+        Converts the output of [`ConditionalDetrForSegmentation`] into semantic segmentation maps. Only supports PyTorch.
+
+        Args:
+            outputs ([`ConditionalDetrForSegmentation`]):
+                Raw outputs of the model.
+            target_sizes (`List[Tuple[int, int]]`, *optional*):
+                A list of tuples (`Tuple[int, int]`) containing the target size (height, width) of each image in the
+                batch. If unset, predictions will not be resized.
+        Returns:
+            `List[torch.Tensor]`:
+                A list of length `batch_size`, where each item is a semantic segmentation map of shape (height, width)
+                corresponding to the target_sizes entry (if `target_sizes` is specified). Each entry of each
+                `torch.Tensor` correspond to a semantic class id.
+        """
+        class_queries_logits = outputs.logits  # [batch_size, num_queries, num_classes+1]
+        masks_queries_logits = outputs.pred_masks  # [batch_size, num_queries, height, width]
+
+        # Remove the null class `[..., :-1]`
+        masks_classes = class_queries_logits.softmax(dim=-1)[..., :-1]
+        masks_probs = masks_queries_logits.sigmoid()  # [batch_size, num_queries, height, width]
+
+        # Semantic segmentation logits of shape (batch_size, num_classes, height, width)
+        segmentation = torch.einsum("bqc, bqhw -> bchw", masks_classes, masks_probs)
+        batch_size = class_queries_logits.shape[0]
+
+        # Resize logits and compute semantic segmentation maps
+        if target_sizes is not None:
+            if batch_size != len(target_sizes):
+                raise ValueError(
+                    "Make sure that you pass in as many target sizes as the batch dimension of the logits"
+                )
+
+            semantic_segmentation = []
+            for idx in range(batch_size):
+                resized_logits = nn.functional.interpolate(
+                    segmentation[idx].unsqueeze(dim=0), size=target_sizes[idx], mode="bilinear", align_corners=False
+                )
+                semantic_map = resized_logits[0].argmax(dim=0)
+                semantic_segmentation.append(semantic_map)
+        else:
+            semantic_segmentation = segmentation.argmax(dim=1)
+            semantic_segmentation = [semantic_segmentation[i] for i in range(semantic_segmentation.shape[0])]
+
+        return semantic_segmentation
+
+    # Copied from transformers.models.detr.image_processing_detr.DetrImageProcessor.post_process_instance_segmentation with Detr->ConditionalDetr
+    def post_process_instance_segmentation(
+        self,
+        outputs,
+        threshold: float = 0.5,
+        mask_threshold: float = 0.5,
+        overlap_mask_area_threshold: float = 0.8,
+        target_sizes: Optional[List[Tuple[int, int]]] = None,
+        return_coco_annotation: Optional[bool] = False,
+    ) -> List[Dict]:
+        """
+        Converts the output of [`ConditionalDetrForSegmentation`] into instance segmentation predictions. Only supports PyTorch.
+
+        Args:
+            outputs ([`ConditionalDetrForSegmentation`]):
+                Raw outputs of the model.
+            threshold (`float`, *optional*, defaults to 0.5):
+                The probability score threshold to keep predicted instance masks.
+            mask_threshold (`float`, *optional*, defaults to 0.5):
+                Threshold to use when turning the predicted masks into binary values.
+            overlap_mask_area_threshold (`float`, *optional*, defaults to 0.8):
+                The overlap mask area threshold to merge or discard small disconnected parts within each binary
+                instance mask.
+            target_sizes (`List[Tuple]`, *optional*):
+                List of length (batch_size), where each list item (`Tuple[int, int]]`) corresponds to the requested
+                final size (height, width) of each prediction. If unset, predictions will not be resized.
+            return_coco_annotation (`bool`, *optional*):
+                Defaults to `False`. If set to `True`, segmentation maps are returned in COCO run-length encoding (RLE)
+                format.
+        Returns:
+            `List[Dict]`: A list of dictionaries, one per image, each dictionary containing two keys:
+            - **segmentation** -- A tensor of shape `(height, width)` where each pixel represents a `segment_id` or
+              `List[List]` run-length encoding (RLE) of the segmentation map if return_coco_annotation is set to
+              `True`. Set to `None` if no mask if found above `threshold`.
+            - **segments_info** -- A dictionary that contains additional information on each segment.
+                - **id** -- An integer representing the `segment_id`.
+                - **label_id** -- An integer representing the label / semantic class id corresponding to `segment_id`.
+                - **score** -- Prediction score of segment with `segment_id`.
+        """
+        class_queries_logits = outputs.logits  # [batch_size, num_queries, num_classes+1]
+        masks_queries_logits = outputs.pred_masks  # [batch_size, num_queries, height, width]
+
+        batch_size = class_queries_logits.shape[0]
+        num_labels = class_queries_logits.shape[-1] - 1
+
+        mask_probs = masks_queries_logits.sigmoid()  # [batch_size, num_queries, height, width]
+
+        # Predicted label and score of each query (batch_size, num_queries)
+        pred_scores, pred_labels = nn.functional.softmax(class_queries_logits, dim=-1).max(-1)
+
+        # Loop over items in batch size
+        results: List[Dict[str, TensorType]] = []
+
+        for i in range(batch_size):
+            mask_probs_item, pred_scores_item, pred_labels_item = remove_low_and_no_objects(
+                mask_probs[i], pred_scores[i], pred_labels[i], threshold, num_labels
+            )
+
+            # No mask found
+            if mask_probs_item.shape[0] <= 0:
+                height, width = target_sizes[i] if target_sizes is not None else mask_probs_item.shape[1:]
+                segmentation = torch.zeros((height, width)) - 1
+                results.append({"segmentation": segmentation, "segments_info": []})
+                continue
+
+            # Get segmentation map and segment information of batch item
+            target_size = target_sizes[i] if target_sizes is not None else None
+            segmentation, segments = compute_segments(
+                mask_probs=mask_probs_item,
+                pred_scores=pred_scores_item,
+                pred_labels=pred_labels_item,
+                mask_threshold=mask_threshold,
+                overlap_mask_area_threshold=overlap_mask_area_threshold,
+                label_ids_to_fuse=[],
+                target_size=target_size,
+            )
+
+            # Return segmentation map in run-length encoding (RLE) format
+            if return_coco_annotation:
+                segmentation = convert_segmentation_to_rle(segmentation)
+
+            results.append({"segmentation": segmentation, "segments_info": segments})
+        return results
+
+    # Copied from transformers.models.detr.image_processing_detr.DetrImageProcessor.post_process_panoptic_segmentation with Detr->ConditionalDetr
+    def post_process_panoptic_segmentation(
+        self,
+        outputs,
+        threshold: float = 0.5,
+        mask_threshold: float = 0.5,
+        overlap_mask_area_threshold: float = 0.8,
+        label_ids_to_fuse: Optional[Set[int]] = None,
+        target_sizes: Optional[List[Tuple[int, int]]] = None,
+    ) -> List[Dict]:
+        """
+        Converts the output of [`ConditionalDetrForSegmentation`] into image panoptic segmentation predictions. Only supports
+        PyTorch.
+
+        Args:
+            outputs ([`ConditionalDetrForSegmentation`]):
+                The outputs from [`ConditionalDetrForSegmentation`].
+            threshold (`float`, *optional*, defaults to 0.5):
+                The probability score threshold to keep predicted instance masks.
+            mask_threshold (`float`, *optional*, defaults to 0.5):
+                Threshold to use when turning the predicted masks into binary values.
+            overlap_mask_area_threshold (`float`, *optional*, defaults to 0.8):
+                The overlap mask area threshold to merge or discard small disconnected parts within each binary
+                instance mask.
+            label_ids_to_fuse (`Set[int]`, *optional*):
+                The labels in this state will have all their instances be fused together. For instance we could say
+                there can only be one sky in an image, but several persons, so the label ID for sky would be in that
+                set, but not the one for person.
+            target_sizes (`List[Tuple]`, *optional*):
+                List of length (batch_size), where each list item (`Tuple[int, int]]`) corresponds to the requested
+                final size (height, width) of each prediction in batch. If unset, predictions will not be resized.
+        Returns:
+            `List[Dict]`: A list of dictionaries, one per image, each dictionary containing two keys:
+            - **segmentation** -- a tensor of shape `(height, width)` where each pixel represents a `segment_id` or
+              `None` if no mask if found above `threshold`. If `target_sizes` is specified, segmentation is resized to
+              the corresponding `target_sizes` entry.
+            - **segments_info** -- A dictionary that contains additional information on each segment.
+                - **id** -- an integer representing the `segment_id`.
+                - **label_id** -- An integer representing the label / semantic class id corresponding to `segment_id`.
+                - **was_fused** -- a boolean, `True` if `label_id` was in `label_ids_to_fuse`, `False` otherwise.
+                  Multiple instances of the same class / label were fused and assigned a single `segment_id`.
+                - **score** -- Prediction score of segment with `segment_id`.
+        """
+
+        if label_ids_to_fuse is None:
+            logger.warning_once("`label_ids_to_fuse` unset. No instance will be fused.")
+            label_ids_to_fuse = set()
+
+        class_queries_logits = outputs.logits  # [batch_size, num_queries, num_classes+1]
+        masks_queries_logits = outputs.pred_masks  # [batch_size, num_queries, height, width]
+
+        batch_size = class_queries_logits.shape[0]
+        num_labels = class_queries_logits.shape[-1] - 1
+
+        mask_probs = masks_queries_logits.sigmoid()  # [batch_size, num_queries, height, width]
+
+        # Predicted label and score of each query (batch_size, num_queries)
+        pred_scores, pred_labels = nn.functional.softmax(class_queries_logits, dim=-1).max(-1)
+
+        # Loop over items in batch size
+        results: List[Dict[str, TensorType]] = []
+
+        for i in range(batch_size):
+            mask_probs_item, pred_scores_item, pred_labels_item = remove_low_and_no_objects(
+                mask_probs[i], pred_scores[i], pred_labels[i], threshold, num_labels
+            )
+
+            # No mask found
+            if mask_probs_item.shape[0] <= 0:
+                height, width = target_sizes[i] if target_sizes is not None else mask_probs_item.shape[1:]
+                segmentation = torch.zeros((height, width)) - 1
+                results.append({"segmentation": segmentation, "segments_info": []})
+                continue
+
+            # Get segmentation map and segment information of batch item
+            target_size = target_sizes[i] if target_sizes is not None else None
+            segmentation, segments = compute_segments(
+                mask_probs=mask_probs_item,
+                pred_scores=pred_scores_item,
+                pred_labels=pred_labels_item,
+                mask_threshold=mask_threshold,
+                overlap_mask_area_threshold=overlap_mask_area_threshold,
+                label_ids_to_fuse=label_ids_to_fuse,
+                target_size=target_size,
+            )
+
+            results.append({"segmentation": segmentation, "segments_info": segments})
+        return results
+
+
+__all__ = ["ConditionalDetrImageProcessor"]

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

@@ -0,0 +1,31 @@
+# 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_convnext import *
+    from .feature_extraction_convnext import *
+    from .image_processing_convnext import *
+    from .image_processing_convnext_fast import *
+    from .modeling_convnext import *
+    from .modeling_tf_convnext import *
+else:
+    import sys
+
+    _file = globals()["__file__"]
+    sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)