Add files using upload-large-folder tool

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

@@ -0,0 +1,325 @@
+# Copyright 2020 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 . import (
+    albert,
+    align,
+    altclip,
+    aria,
+    audio_spectrogram_transformer,
+    auto,
+    autoformer,
+    aya_vision,
+    bamba,
+    bark,
+    bart,
+    barthez,
+    bartpho,
+    beit,
+    bert,
+    bert_generation,
+    bert_japanese,
+    bertweet,
+    big_bird,
+    bigbird_pegasus,
+    biogpt,
+    bit,
+    blenderbot,
+    blenderbot_small,
+    blip,
+    blip_2,
+    bloom,
+    bridgetower,
+    bros,
+    byt5,
+    camembert,
+    canine,
+    chameleon,
+    chinese_clip,
+    clap,
+    clip,
+    clipseg,
+    clvp,
+    code_llama,
+    codegen,
+    cohere,
+    cohere2,
+    colpali,
+    conditional_detr,
+    convbert,
+    convnext,
+    convnextv2,
+    cpm,
+    cpmant,
+    ctrl,
+    cvt,
+    dab_detr,
+    dac,
+    data2vec,
+    dbrx,
+    deberta,
+    deberta_v2,
+    decision_transformer,
+    deepseek_v3,
+    deformable_detr,
+    deit,
+    deprecated,
+    depth_anything,
+    depth_pro,
+    detr,
+    dialogpt,
+    diffllama,
+    dinat,
+    dinov2,
+    dinov2_with_registers,
+    distilbert,
+    dit,
+    donut,
+    dpr,
+    dpt,
+    efficientnet,
+    electra,
+    emu3,
+    encodec,
+    encoder_decoder,
+    ernie,
+    esm,
+    falcon,
+    falcon_mamba,
+    fastspeech2_conformer,
+    flaubert,
+    flava,
+    fnet,
+    focalnet,
+    fsmt,
+    funnel,
+    fuyu,
+    gemma,
+    gemma2,
+    gemma3,
+    git,
+    glm,
+    glpn,
+    got_ocr2,
+    gpt2,
+    gpt_bigcode,
+    gpt_neo,
+    gpt_neox,
+    gpt_neox_japanese,
+    gpt_sw3,
+    gptj,
+    granite,
+    granitemoe,
+    granitemoeshared,
+    grounding_dino,
+    groupvit,
+    helium,
+    herbert,
+    hiera,
+    hubert,
+    ibert,
+    idefics,
+    idefics2,
+    idefics3,
+    ijepa,
+    imagegpt,
+    informer,
+    instructblip,
+    instructblipvideo,
+    jamba,
+    jetmoe,
+    kosmos2,
+    layoutlm,
+    layoutlmv2,
+    layoutlmv3,
+    layoutxlm,
+    led,
+    levit,
+    lilt,
+    llama,
+    llama4,
+    llava,
+    llava_next,
+    llava_next_video,
+    llava_onevision,
+    longformer,
+    longt5,
+    luke,
+    lxmert,
+    m2m_100,
+    mamba,
+    mamba2,
+    marian,
+    markuplm,
+    mask2former,
+    maskformer,
+    mbart,
+    mbart50,
+    megatron_bert,
+    megatron_gpt2,
+    mgp_str,
+    mimi,
+    mistral,
+    mistral3,
+    mixtral,
+    mllama,
+    mluke,
+    mobilebert,
+    mobilenet_v1,
+    mobilenet_v2,
+    mobilevit,
+    mobilevitv2,
+    modernbert,
+    moonshine,
+    moshi,
+    mpnet,
+    mpt,
+    mra,
+    mt5,
+    musicgen,
+    musicgen_melody,
+    mvp,
+    myt5,
+    nemotron,
+    nllb,
+    nllb_moe,
+    nougat,
+    nystromformer,
+    olmo,
+    olmo2,
+    olmoe,
+    omdet_turbo,
+    oneformer,
+    openai,
+    opt,
+    owlv2,
+    owlvit,
+    paligemma,
+    patchtsmixer,
+    patchtst,
+    pegasus,
+    pegasus_x,
+    perceiver,
+    persimmon,
+    phi,
+    phi3,
+    phi4_multimodal,
+    phimoe,
+    phobert,
+    pix2struct,
+    pixtral,
+    plbart,
+    poolformer,
+    pop2piano,
+    prompt_depth_anything,
+    prophetnet,
+    pvt,
+    pvt_v2,
+    qwen2,
+    qwen2_5_vl,
+    qwen2_audio,
+    qwen2_moe,
+    qwen2_vl,
+    qwen3,
+    qwen3_moe,
+    rag,
+    recurrent_gemma,
+    reformer,
+    regnet,
+    rembert,
+    resnet,
+    roberta,
+    roberta_prelayernorm,
+    roc_bert,
+    roformer,
+    rt_detr,
+    rt_detr_v2,
+    rwkv,
+    sam,
+    seamless_m4t,
+    seamless_m4t_v2,
+    segformer,
+    seggpt,
+    sew,
+    sew_d,
+    shieldgemma2,
+    siglip,
+    siglip2,
+    smolvlm,
+    speech_encoder_decoder,
+    speech_to_text,
+    speecht5,
+    splinter,
+    squeezebert,
+    stablelm,
+    starcoder2,
+    superglue,
+    superpoint,
+    swiftformer,
+    swin,
+    swin2sr,
+    swinv2,
+    switch_transformers,
+    t5,
+    table_transformer,
+    tapas,
+    textnet,
+    time_series_transformer,
+    timesformer,
+    timm_backbone,
+    timm_wrapper,
+    trocr,
+    tvp,
+    udop,
+    umt5,
+    unispeech,
+    unispeech_sat,
+    univnet,
+    upernet,
+    video_llava,
+    videomae,
+    vilt,
+    vipllava,
+    vision_encoder_decoder,
+    vision_text_dual_encoder,
+    visual_bert,
+    vit,
+    vit_mae,
+    vit_msn,
+    vitdet,
+    vitmatte,
+    vitpose,
+    vitpose_backbone,
+    vits,
+    vivit,
+    wav2vec2,
+    wav2vec2_bert,
+    wav2vec2_conformer,
+    wav2vec2_phoneme,
+    wav2vec2_with_lm,
+    wavlm,
+    whisper,
+    x_clip,
+    xglm,
+    xlm,
+    xlm_roberta,
+    xlm_roberta_xl,
+    xlnet,
+    xmod,
+    yolos,
+    yoso,
+    zamba,
+    zamba2,
+    zoedepth,
+)

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

@@ -0,0 +1,31 @@
+# Copyright 2020 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_albert import *
+    from .modeling_albert import *
+    from .modeling_flax_albert import *
+    from .modeling_tf_albert import *
+    from .tokenization_albert import *
+    from .tokenization_albert_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/albert/configuration_albert.py

@@ -0,0 +1,170 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  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.
+"""ALBERT model configuration"""
+
+from collections import OrderedDict
+from typing import Mapping
+
+from ...configuration_utils import PretrainedConfig
+from ...onnx import OnnxConfig
+
+
+class AlbertConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`AlbertModel`] or a [`TFAlbertModel`]. It is used
+    to instantiate an ALBERT 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 ALBERT
+    [albert/albert-xxlarge-v2](https://huggingface.co/albert/albert-xxlarge-v2) 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 30000):
+            Vocabulary size of the ALBERT model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`AlbertModel`] or [`TFAlbertModel`].
+        embedding_size (`int`, *optional*, defaults to 128):
+            Dimensionality of vocabulary embeddings.
+        hidden_size (`int`, *optional*, defaults to 4096):
+            Dimensionality of the encoder layers and the pooler layer.
+        num_hidden_layers (`int`, *optional*, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        num_hidden_groups (`int`, *optional*, defaults to 1):
+            Number of groups for the hidden layers, parameters in the same group are shared.
+        num_attention_heads (`int`, *optional*, defaults to 64):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        intermediate_size (`int`, *optional*, defaults to 16384):
+            The dimensionality of the "intermediate" (often named feed-forward) layer in the Transformer encoder.
+        inner_group_num (`int`, *optional*, defaults to 1):
+            The number of inner repetition of attention and ffn.
+        hidden_act (`str` or `Callable`, *optional*, defaults to `"gelu_new"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"silu"` and `"gelu_new"` are supported.
+        hidden_dropout_prob (`float`, *optional*, defaults to 0):
+            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_probs_dropout_prob (`float`, *optional*, defaults to 0):
+            The dropout ratio for the attention probabilities.
+        max_position_embeddings (`int`, *optional*, defaults to 512):
+            The maximum sequence length that this model might ever be used with. Typically set this to something large
+            (e.g., 512 or 1024 or 2048).
+        type_vocab_size (`int`, *optional*, defaults to 2):
+            The vocabulary size of the `token_type_ids` passed when calling [`AlbertModel`] or [`TFAlbertModel`].
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        layer_norm_eps (`float`, *optional*, defaults to 1e-12):
+            The epsilon used by the layer normalization layers.
+        classifier_dropout_prob (`float`, *optional*, defaults to 0.1):
+            The dropout ratio for attached classifiers.
+        position_embedding_type (`str`, *optional*, defaults to `"absolute"`):
+            Type of position embedding. Choose one of `"absolute"`, `"relative_key"`, `"relative_key_query"`. For
+            positional embeddings use `"absolute"`. For more information on `"relative_key"`, please refer to
+            [Self-Attention with Relative Position Representations (Shaw et al.)](https://arxiv.org/abs/1803.02155).
+            For more information on `"relative_key_query"`, please refer to *Method 4* in [Improve Transformer Models
+            with Better Relative Position Embeddings (Huang et al.)](https://arxiv.org/abs/2009.13658).
+        pad_token_id (`int`, *optional*, defaults to 0):
+            Padding token id.
+        bos_token_id (`int`, *optional*, defaults to 2):
+            Beginning of stream token id.
+        eos_token_id (`int`, *optional*, defaults to 3):
+            End of stream token id.
+
+    Examples:
+
+    ```python
+    >>> from transformers import AlbertConfig, AlbertModel
+
+    >>> # Initializing an ALBERT-xxlarge style configuration
+    >>> albert_xxlarge_configuration = AlbertConfig()
+
+    >>> # Initializing an ALBERT-base style configuration
+    >>> albert_base_configuration = AlbertConfig(
+    ...     hidden_size=768,
+    ...     num_attention_heads=12,
+    ...     intermediate_size=3072,
+    ... )
+
+    >>> # Initializing a model (with random weights) from the ALBERT-base style configuration
+    >>> model = AlbertModel(albert_xxlarge_configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "albert"
+
+    def __init__(
+        self,
+        vocab_size=30000,
+        embedding_size=128,
+        hidden_size=4096,
+        num_hidden_layers=12,
+        num_hidden_groups=1,
+        num_attention_heads=64,
+        intermediate_size=16384,
+        inner_group_num=1,
+        hidden_act="gelu_new",
+        hidden_dropout_prob=0,
+        attention_probs_dropout_prob=0,
+        max_position_embeddings=512,
+        type_vocab_size=2,
+        initializer_range=0.02,
+        layer_norm_eps=1e-12,
+        classifier_dropout_prob=0.1,
+        position_embedding_type="absolute",
+        pad_token_id=0,
+        bos_token_id=2,
+        eos_token_id=3,
+        **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.embedding_size = embedding_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_hidden_groups = num_hidden_groups
+        self.num_attention_heads = num_attention_heads
+        self.inner_group_num = inner_group_num
+        self.hidden_act = hidden_act
+        self.intermediate_size = intermediate_size
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.max_position_embeddings = max_position_embeddings
+        self.type_vocab_size = type_vocab_size
+        self.initializer_range = initializer_range
+        self.layer_norm_eps = layer_norm_eps
+        self.classifier_dropout_prob = classifier_dropout_prob
+        self.position_embedding_type = position_embedding_type
+
+
+# Copied from transformers.models.bert.configuration_bert.BertOnnxConfig with Roberta->Albert
+class AlbertOnnxConfig(OnnxConfig):
+    @property
+    def inputs(self) -> Mapping[str, Mapping[int, str]]:
+        if self.task == "multiple-choice":
+            dynamic_axis = {0: "batch", 1: "choice", 2: "sequence"}
+        else:
+            dynamic_axis = {0: "batch", 1: "sequence"}
+        return OrderedDict(
+            [
+                ("input_ids", dynamic_axis),
+                ("attention_mask", dynamic_axis),
+                ("token_type_ids", dynamic_axis),
+            ]
+        )
+
+
+__all__ = ["AlbertConfig", "AlbertOnnxConfig"]

lib/python3.12/site-packages/transformers/models/albert/modeling_albert.py

@@ -0,0 +1,1483 @@
+# coding=utf-8
+# Copyright 2018 Google AI, Google Brain and the HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""PyTorch ALBERT model."""
+
+import math
+import os
+from dataclasses import dataclass
+from typing import Dict, List, Optional, Tuple, Union
+
+import torch
+from torch import nn
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+
+from ...activations import ACT2FN
+from ...modeling_attn_mask_utils import _prepare_4d_attention_mask_for_sdpa
+from ...modeling_outputs import (
+    BaseModelOutput,
+    BaseModelOutputWithPooling,
+    MaskedLMOutput,
+    MultipleChoiceModelOutput,
+    QuestionAnsweringModelOutput,
+    SequenceClassifierOutput,
+    TokenClassifierOutput,
+)
+from ...modeling_utils import PreTrainedModel
+from ...pytorch_utils import (
+    apply_chunking_to_forward,
+    find_pruneable_heads_and_indices,
+    is_torch_greater_or_equal_than_2_2,
+    prune_linear_layer,
+)
+from ...utils import (
+    ModelOutput,
+    add_code_sample_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    logging,
+    replace_return_docstrings,
+)
+from .configuration_albert import AlbertConfig
+
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "albert/albert-base-v2"
+_CONFIG_FOR_DOC = "AlbertConfig"
+
+
+def load_tf_weights_in_albert(model, config, tf_checkpoint_path):
+    """Load tf checkpoints in a pytorch model."""
+    try:
+        import re
+
+        import numpy as np
+        import tensorflow as tf
+    except ImportError:
+        logger.error(
+            "Loading a TensorFlow model in PyTorch, requires TensorFlow to be installed. Please see "
+            "https://www.tensorflow.org/install/ for installation instructions."
+        )
+        raise
+    tf_path = os.path.abspath(tf_checkpoint_path)
+    logger.info(f"Converting TensorFlow checkpoint from {tf_path}")
+    # Load weights from TF model
+    init_vars = tf.train.list_variables(tf_path)
+    names = []
+    arrays = []
+    for name, shape in init_vars:
+        logger.info(f"Loading TF weight {name} with shape {shape}")
+        array = tf.train.load_variable(tf_path, name)
+        names.append(name)
+        arrays.append(array)
+
+    for name, array in zip(names, arrays):
+        print(name)
+
+    for name, array in zip(names, arrays):
+        original_name = name
+
+        # If saved from the TF HUB module
+        name = name.replace("module/", "")
+
+        # Renaming and simplifying
+        name = name.replace("ffn_1", "ffn")
+        name = name.replace("bert/", "albert/")
+        name = name.replace("attention_1", "attention")
+        name = name.replace("transform/", "")
+        name = name.replace("LayerNorm_1", "full_layer_layer_norm")
+        name = name.replace("LayerNorm", "attention/LayerNorm")
+        name = name.replace("transformer/", "")
+
+        # The feed forward layer had an 'intermediate' step which has been abstracted away
+        name = name.replace("intermediate/dense/", "")
+        name = name.replace("ffn/intermediate/output/dense/", "ffn_output/")
+
+        # ALBERT attention was split between self and output which have been abstracted away
+        name = name.replace("/output/", "/")
+        name = name.replace("/self/", "/")
+
+        # The pooler is a linear layer
+        name = name.replace("pooler/dense", "pooler")
+
+        # The classifier was simplified to predictions from cls/predictions
+        name = name.replace("cls/predictions", "predictions")
+        name = name.replace("predictions/attention", "predictions")
+
+        # Naming was changed to be more explicit
+        name = name.replace("embeddings/attention", "embeddings")
+        name = name.replace("inner_group_", "albert_layers/")
+        name = name.replace("group_", "albert_layer_groups/")
+
+        # Classifier
+        if len(name.split("/")) == 1 and ("output_bias" in name or "output_weights" in name):
+            name = "classifier/" + name
+
+        # No ALBERT model currently handles the next sentence prediction task
+        if "seq_relationship" in name:
+            name = name.replace("seq_relationship/output_", "sop_classifier/classifier/")
+            name = name.replace("weights", "weight")
+
+        name = name.split("/")
+
+        # Ignore the gradients applied by the LAMB/ADAM optimizers.
+        if (
+            "adam_m" in name
+            or "adam_v" in name
+            or "AdamWeightDecayOptimizer" in name
+            or "AdamWeightDecayOptimizer_1" in name
+            or "global_step" in name
+        ):
+            logger.info(f"Skipping {'/'.join(name)}")
+            continue
+
+        pointer = model
+        for m_name in name:
+            if re.fullmatch(r"[A-Za-z]+_\d+", m_name):
+                scope_names = re.split(r"_(\d+)", m_name)
+            else:
+                scope_names = [m_name]
+
+            if scope_names[0] == "kernel" or scope_names[0] == "gamma":
+                pointer = getattr(pointer, "weight")
+            elif scope_names[0] == "output_bias" or scope_names[0] == "beta":
+                pointer = getattr(pointer, "bias")
+            elif scope_names[0] == "output_weights":
+                pointer = getattr(pointer, "weight")
+            elif scope_names[0] == "squad":
+                pointer = getattr(pointer, "classifier")
+            else:
+                try:
+                    pointer = getattr(pointer, scope_names[0])
+                except AttributeError:
+                    logger.info(f"Skipping {'/'.join(name)}")
+                    continue
+            if len(scope_names) >= 2:
+                num = int(scope_names[1])
+                pointer = pointer[num]
+
+        if m_name[-11:] == "_embeddings":
+            pointer = getattr(pointer, "weight")
+        elif m_name == "kernel":
+            array = np.transpose(array)
+        try:
+            if pointer.shape != array.shape:
+                raise ValueError(f"Pointer shape {pointer.shape} and array shape {array.shape} mismatched")
+        except ValueError as e:
+            e.args += (pointer.shape, array.shape)
+            raise
+        print(f"Initialize PyTorch weight {name} from {original_name}")
+        pointer.data = torch.from_numpy(array)
+
+    return model
+
+
+class AlbertEmbeddings(nn.Module):
+    """
+    Construct the embeddings from word, position and token_type embeddings.
+    """
+
+    def __init__(self, config: AlbertConfig):
+        super().__init__()
+        self.word_embeddings = nn.Embedding(config.vocab_size, config.embedding_size, padding_idx=config.pad_token_id)
+        self.position_embeddings = nn.Embedding(config.max_position_embeddings, config.embedding_size)
+        self.token_type_embeddings = nn.Embedding(config.type_vocab_size, config.embedding_size)
+
+        # self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load
+        # any TensorFlow checkpoint file
+        self.LayerNorm = nn.LayerNorm(config.embedding_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+        # position_ids (1, len position emb) is contiguous in memory and exported when serialized
+        self.register_buffer(
+            "position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)), persistent=False
+        )
+        self.position_embedding_type = getattr(config, "position_embedding_type", "absolute")
+        self.register_buffer(
+            "token_type_ids", torch.zeros(self.position_ids.size(), dtype=torch.long), persistent=False
+        )
+
+    # Copied from transformers.models.bert.modeling_bert.BertEmbeddings.forward
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        past_key_values_length: int = 0,
+    ) -> torch.Tensor:
+        if input_ids is not None:
+            input_shape = input_ids.size()
+        else:
+            input_shape = inputs_embeds.size()[:-1]
+
+        seq_length = input_shape[1]
+
+        if position_ids is None:
+            position_ids = self.position_ids[:, past_key_values_length : seq_length + past_key_values_length]
+
+        # Setting the token_type_ids to the registered buffer in constructor where it is all zeros, which usually occurs
+        # when its auto-generated, registered buffer helps users when tracing the model without passing token_type_ids, solves
+        # issue #5664
+        if token_type_ids is None:
+            if hasattr(self, "token_type_ids"):
+                buffered_token_type_ids = self.token_type_ids[:, :seq_length]
+                buffered_token_type_ids_expanded = buffered_token_type_ids.expand(input_shape[0], seq_length)
+                token_type_ids = buffered_token_type_ids_expanded
+            else:
+                token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=self.position_ids.device)
+
+        if inputs_embeds is None:
+            inputs_embeds = self.word_embeddings(input_ids)
+        token_type_embeddings = self.token_type_embeddings(token_type_ids)
+
+        embeddings = inputs_embeds + token_type_embeddings
+        if self.position_embedding_type == "absolute":
+            position_embeddings = self.position_embeddings(position_ids)
+            embeddings += position_embeddings
+        embeddings = self.LayerNorm(embeddings)
+        embeddings = self.dropout(embeddings)
+        return embeddings
+
+
+class AlbertAttention(nn.Module):
+    def __init__(self, config: AlbertConfig):
+        super().__init__()
+        if config.hidden_size % config.num_attention_heads != 0 and not hasattr(config, "embedding_size"):
+            raise ValueError(
+                f"The hidden size ({config.hidden_size}) is not a multiple of the number of attention "
+                f"heads ({config.num_attention_heads}"
+            )
+
+        self.num_attention_heads = config.num_attention_heads
+        self.hidden_size = config.hidden_size
+        self.attention_head_size = config.hidden_size // config.num_attention_heads
+        self.all_head_size = self.num_attention_heads * self.attention_head_size
+
+        self.query = nn.Linear(config.hidden_size, self.all_head_size)
+        self.key = nn.Linear(config.hidden_size, self.all_head_size)
+        self.value = nn.Linear(config.hidden_size, self.all_head_size)
+
+        self.attention_dropout = nn.Dropout(config.attention_probs_dropout_prob)
+        self.output_dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.pruned_heads = set()
+
+        self.position_embedding_type = getattr(config, "position_embedding_type", "absolute")
+        if self.position_embedding_type == "relative_key" or self.position_embedding_type == "relative_key_query":
+            self.max_position_embeddings = config.max_position_embeddings
+            self.distance_embedding = nn.Embedding(2 * config.max_position_embeddings - 1, self.attention_head_size)
+
+    # Copied from transformers.models.bert.modeling_bert.BertSelfAttention.transpose_for_scores
+    def transpose_for_scores(self, x: torch.Tensor) -> torch.Tensor:
+        new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
+        x = x.view(new_x_shape)
+        return x.permute(0, 2, 1, 3)
+
+    def prune_heads(self, heads: List[int]) -> None:
+        if len(heads) == 0:
+            return
+        heads, index = find_pruneable_heads_and_indices(
+            heads, self.num_attention_heads, self.attention_head_size, self.pruned_heads
+        )
+
+        # Prune linear layers
+        self.query = prune_linear_layer(self.query, index)
+        self.key = prune_linear_layer(self.key, index)
+        self.value = prune_linear_layer(self.value, index)
+        self.dense = prune_linear_layer(self.dense, index, dim=1)
+
+        # Update hyper params and store pruned heads
+        self.num_attention_heads = self.num_attention_heads - len(heads)
+        self.all_head_size = self.attention_head_size * self.num_attention_heads
+        self.pruned_heads = self.pruned_heads.union(heads)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        output_attentions: bool = False,
+    ) -> Union[Tuple[torch.Tensor], Tuple[torch.Tensor, torch.Tensor]]:
+        mixed_query_layer = self.query(hidden_states)
+        mixed_key_layer = self.key(hidden_states)
+        mixed_value_layer = self.value(hidden_states)
+
+        query_layer = self.transpose_for_scores(mixed_query_layer)
+        key_layer = self.transpose_for_scores(mixed_key_layer)
+        value_layer = self.transpose_for_scores(mixed_value_layer)
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
+        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
+
+        if attention_mask is not None:
+            # Apply the attention mask is (precomputed for all layers in BertModel forward() function)
+            attention_scores = attention_scores + attention_mask
+
+        if self.position_embedding_type == "relative_key" or self.position_embedding_type == "relative_key_query":
+            seq_length = hidden_states.size()[1]
+            position_ids_l = torch.arange(seq_length, dtype=torch.long, device=hidden_states.device).view(-1, 1)
+            position_ids_r = torch.arange(seq_length, dtype=torch.long, device=hidden_states.device).view(1, -1)
+            distance = position_ids_l - position_ids_r
+            positional_embedding = self.distance_embedding(distance + self.max_position_embeddings - 1)
+            positional_embedding = positional_embedding.to(dtype=query_layer.dtype)  # fp16 compatibility
+
+            if self.position_embedding_type == "relative_key":
+                relative_position_scores = torch.einsum("bhld,lrd->bhlr", query_layer, positional_embedding)
+                attention_scores = attention_scores + relative_position_scores
+            elif self.position_embedding_type == "relative_key_query":
+                relative_position_scores_query = torch.einsum("bhld,lrd->bhlr", query_layer, positional_embedding)
+                relative_position_scores_key = torch.einsum("bhrd,lrd->bhlr", key_layer, positional_embedding)
+                attention_scores = attention_scores + relative_position_scores_query + relative_position_scores_key
+
+        # Normalize the attention scores to probabilities.
+        attention_probs = nn.functional.softmax(attention_scores, dim=-1)
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        attention_probs = self.attention_dropout(attention_probs)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attention_probs = attention_probs * head_mask
+
+        context_layer = torch.matmul(attention_probs, value_layer)
+        context_layer = context_layer.transpose(2, 1).flatten(2)
+
+        projected_context_layer = self.dense(context_layer)
+        projected_context_layer_dropout = self.output_dropout(projected_context_layer)
+        layernormed_context_layer = self.LayerNorm(hidden_states + projected_context_layer_dropout)
+        return (layernormed_context_layer, attention_probs) if output_attentions else (layernormed_context_layer,)
+
+
+class AlbertSdpaAttention(AlbertAttention):
+    def __init__(self, config):
+        super().__init__(config)
+        self.dropout_prob = config.attention_probs_dropout_prob
+        self.require_contiguous_qkv = not is_torch_greater_or_equal_than_2_2
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        output_attentions: bool = False,
+    ) -> Union[Tuple[torch.Tensor], Tuple[torch.Tensor, torch.Tensor]]:
+        if self.position_embedding_type != "absolute" or output_attentions or head_mask is not None:
+            logger.warning(
+                "AlbertSdpaAttention is used but `torch.nn.functional.scaled_dot_product_attention` does not support "
+                "non-absolute `position_embedding_type` or `output_attentions=True` or `head_mask`. Falling back to "
+                "the eager attention implementation, but specifying the eager implementation will be required from "
+                "Transformers version v5.0.0 onwards. This warning can be removed using the argument "
+                '`attn_implementation="eager"` when loading the model.'
+            )
+            return super().forward(hidden_states, attention_mask, head_mask, output_attentions)
+
+        batch_size, seq_len, _ = hidden_states.size()
+        query_layer = self.transpose_for_scores(self.query(hidden_states))
+        key_layer = self.transpose_for_scores(self.key(hidden_states))
+        value_layer = self.transpose_for_scores(self.value(hidden_states))
+
+        # SDPA with memory-efficient backend is broken in torch==2.1.2 when using non-contiguous inputs and a custom
+        # attn_mask, so we need to call `.contiguous()` here. This was fixed in torch==2.2.0.
+        # Reference: https://github.com/pytorch/pytorch/issues/112577
+        if self.require_contiguous_qkv and query_layer.device.type == "cuda" and attention_mask is not None:
+            query_layer = query_layer.contiguous()
+            key_layer = key_layer.contiguous()
+            value_layer = value_layer.contiguous()
+
+        attention_output = torch.nn.functional.scaled_dot_product_attention(
+            query=query_layer,
+            key=key_layer,
+            value=value_layer,
+            attn_mask=attention_mask,
+            dropout_p=self.dropout_prob if self.training else 0.0,
+            is_causal=False,
+        )
+
+        attention_output = attention_output.transpose(1, 2)
+        attention_output = attention_output.reshape(batch_size, seq_len, self.all_head_size)
+
+        projected_context_layer = self.dense(attention_output)
+        projected_context_layer_dropout = self.output_dropout(projected_context_layer)
+        layernormed_context_layer = self.LayerNorm(hidden_states + projected_context_layer_dropout)
+        return (layernormed_context_layer,)
+
+
+ALBERT_ATTENTION_CLASSES = {
+    "eager": AlbertAttention,
+    "sdpa": AlbertSdpaAttention,
+}
+
+
+class AlbertLayer(nn.Module):
+    def __init__(self, config: AlbertConfig):
+        super().__init__()
+
+        self.config = config
+        self.chunk_size_feed_forward = config.chunk_size_feed_forward
+        self.seq_len_dim = 1
+        self.full_layer_layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.attention = ALBERT_ATTENTION_CLASSES[config._attn_implementation](config)
+        self.ffn = nn.Linear(config.hidden_size, config.intermediate_size)
+        self.ffn_output = nn.Linear(config.intermediate_size, config.hidden_size)
+        self.activation = ACT2FN[config.hidden_act]
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        attention_output = self.attention(hidden_states, attention_mask, head_mask, output_attentions)
+
+        ffn_output = apply_chunking_to_forward(
+            self.ff_chunk,
+            self.chunk_size_feed_forward,
+            self.seq_len_dim,
+            attention_output[0],
+        )
+        hidden_states = self.full_layer_layer_norm(ffn_output + attention_output[0])
+
+        return (hidden_states,) + attention_output[1:]  # add attentions if we output them
+
+    def ff_chunk(self, attention_output: torch.Tensor) -> torch.Tensor:
+        ffn_output = self.ffn(attention_output)
+        ffn_output = self.activation(ffn_output)
+        ffn_output = self.ffn_output(ffn_output)
+        return ffn_output
+
+
+class AlbertLayerGroup(nn.Module):
+    def __init__(self, config: AlbertConfig):
+        super().__init__()
+
+        self.albert_layers = nn.ModuleList([AlbertLayer(config) for _ in range(config.inner_group_num)])
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+    ) -> Tuple[Union[torch.Tensor, Tuple[torch.Tensor]], ...]:
+        layer_hidden_states = ()
+        layer_attentions = ()
+
+        for layer_index, albert_layer in enumerate(self.albert_layers):
+            layer_output = albert_layer(hidden_states, attention_mask, head_mask[layer_index], output_attentions)
+            hidden_states = layer_output[0]
+
+            if output_attentions:
+                layer_attentions = layer_attentions + (layer_output[1],)
+
+            if output_hidden_states:
+                layer_hidden_states = layer_hidden_states + (hidden_states,)
+
+        outputs = (hidden_states,)
+        if output_hidden_states:
+            outputs = outputs + (layer_hidden_states,)
+        if output_attentions:
+            outputs = outputs + (layer_attentions,)
+        return outputs  # last-layer hidden state, (layer hidden states), (layer attentions)
+
+
+class AlbertTransformer(nn.Module):
+    def __init__(self, config: AlbertConfig):
+        super().__init__()
+
+        self.config = config
+        self.embedding_hidden_mapping_in = nn.Linear(config.embedding_size, config.hidden_size)
+        self.albert_layer_groups = nn.ModuleList([AlbertLayerGroup(config) for _ in range(config.num_hidden_groups)])
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ) -> Union[BaseModelOutput, Tuple]:
+        hidden_states = self.embedding_hidden_mapping_in(hidden_states)
+
+        all_hidden_states = (hidden_states,) if output_hidden_states else None
+        all_attentions = () if output_attentions else None
+
+        head_mask = [None] * self.config.num_hidden_layers if head_mask is None else head_mask
+
+        for i in range(self.config.num_hidden_layers):
+            # Number of layers in a hidden group
+            layers_per_group = int(self.config.num_hidden_layers / self.config.num_hidden_groups)
+
+            # Index of the hidden group
+            group_idx = int(i / (self.config.num_hidden_layers / self.config.num_hidden_groups))
+
+            layer_group_output = self.albert_layer_groups[group_idx](
+                hidden_states,
+                attention_mask,
+                head_mask[group_idx * layers_per_group : (group_idx + 1) * layers_per_group],
+                output_attentions,
+                output_hidden_states,
+            )
+            hidden_states = layer_group_output[0]
+
+            if output_attentions:
+                all_attentions = all_attentions + layer_group_output[-1]
+
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, all_hidden_states, all_attentions] if v is not None)
+        return BaseModelOutput(
+            last_hidden_state=hidden_states, hidden_states=all_hidden_states, attentions=all_attentions
+        )
+
+
+class AlbertPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = AlbertConfig
+    load_tf_weights = load_tf_weights_in_albert
+    base_model_prefix = "albert"
+    _supports_sdpa = True
+
+    def _init_weights(self, module):
+        """Initialize the weights."""
+        if isinstance(module, nn.Linear):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+        elif isinstance(module, AlbertMLMHead):
+            module.bias.data.zero_()
+
+
+@dataclass
+class AlbertForPreTrainingOutput(ModelOutput):
+    """
+    Output type of [`AlbertForPreTraining`].
+
+    Args:
+        loss (*optional*, returned when `labels` is provided, `torch.FloatTensor` of shape `(1,)`):
+            Total loss as the sum of the masked language modeling loss and the next sequence prediction
+            (classification) loss.
+        prediction_logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        sop_logits (`torch.FloatTensor` of shape `(batch_size, 2)`):
+            Prediction scores of the next sequence prediction (classification) head (scores of True/False continuation
+            before SoftMax).
+        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer) of
+            shape `(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
+            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
+            sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    prediction_logits: Optional[torch.FloatTensor] = None
+    sop_logits: Optional[torch.FloatTensor] = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+ALBERT_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.
+
+    Args:
+        config ([`AlbertConfig`]): 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.
+"""
+
+ALBERT_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `({0})`):
+            Indices of input sequence tokens in the vocabulary.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.__call__`] and
+            [`PreTrainedTokenizer.encode`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`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.max_position_embeddings - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
+            Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        inputs_embeds (`torch.FloatTensor` of shape `({0}, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
+            model's internal embedding lookup matrix.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+@add_start_docstrings(
+    "The bare ALBERT Model transformer outputting raw hidden-states without any specific head on top.",
+    ALBERT_START_DOCSTRING,
+)
+class AlbertModel(AlbertPreTrainedModel):
+    config_class = AlbertConfig
+    base_model_prefix = "albert"
+
+    def __init__(self, config: AlbertConfig, add_pooling_layer: bool = True):
+        super().__init__(config)
+
+        self.config = config
+        self.embeddings = AlbertEmbeddings(config)
+        self.encoder = AlbertTransformer(config)
+        if add_pooling_layer:
+            self.pooler = nn.Linear(config.hidden_size, config.hidden_size)
+            self.pooler_activation = nn.Tanh()
+        else:
+            self.pooler = None
+            self.pooler_activation = None
+
+        self.attn_implementation = config._attn_implementation
+        self.position_embedding_type = config.position_embedding_type
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self) -> nn.Embedding:
+        return self.embeddings.word_embeddings
+
+    def set_input_embeddings(self, value: nn.Embedding) -> None:
+        self.embeddings.word_embeddings = value
+
+    def _prune_heads(self, heads_to_prune: Dict[int, List[int]]) -> None:
+        """
+        Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} ALBERT has
+        a different architecture in that its layers are shared across groups, which then has inner groups. If an ALBERT
+        model has 12 hidden layers and 2 hidden groups, with two inner groups, there is a total of 4 different layers.
+
+        These layers are flattened: the indices [0,1] correspond to the two inner groups of the first hidden layer,
+        while [2,3] correspond to the two inner groups of the second hidden layer.
+
+        Any layer with in index other than [0,1,2,3] will result in an error. See base class PreTrainedModel for more
+        information about head pruning
+        """
+        for layer, heads in heads_to_prune.items():
+            group_idx = int(layer / self.config.inner_group_num)
+            inner_group_idx = int(layer - group_idx * self.config.inner_group_num)
+            self.encoder.albert_layer_groups[group_idx].albert_layers[inner_group_idx].attention.prune_heads(heads)
+
+    @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=BaseModelOutputWithPooling,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = 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,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[BaseModelOutputWithPooling, 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
+
+        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()
+        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")
+
+        batch_size, seq_length = input_shape
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+
+        if attention_mask is None:
+            attention_mask = torch.ones(input_shape, device=device)
+        if token_type_ids is None:
+            if hasattr(self.embeddings, "token_type_ids"):
+                buffered_token_type_ids = self.embeddings.token_type_ids[:, :seq_length]
+                buffered_token_type_ids_expanded = buffered_token_type_ids.expand(batch_size, seq_length)
+                token_type_ids = buffered_token_type_ids_expanded
+            else:
+                token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
+
+        embedding_output = self.embeddings(
+            input_ids, position_ids=position_ids, token_type_ids=token_type_ids, inputs_embeds=inputs_embeds
+        )
+
+        use_sdpa_attention_mask = (
+            self.attn_implementation == "sdpa"
+            and self.position_embedding_type == "absolute"
+            and head_mask is None
+            and not output_attentions
+        )
+
+        if use_sdpa_attention_mask:
+            extended_attention_mask = _prepare_4d_attention_mask_for_sdpa(
+                attention_mask, embedding_output.dtype, tgt_len=seq_length
+            )
+        else:
+            extended_attention_mask = attention_mask.unsqueeze(1).unsqueeze(2)
+            extended_attention_mask = extended_attention_mask.to(dtype=self.dtype)  # fp16 compatibility
+            extended_attention_mask = (1.0 - extended_attention_mask) * torch.finfo(self.dtype).min
+
+        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
+
+        encoder_outputs = self.encoder(
+            embedding_output,
+            extended_attention_mask,
+            head_mask=head_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        sequence_output = encoder_outputs[0]
+
+        pooled_output = self.pooler_activation(self.pooler(sequence_output[:, 0])) if self.pooler is not None else None
+
+        if not return_dict:
+            return (sequence_output, pooled_output) + encoder_outputs[1:]
+
+        return BaseModelOutputWithPooling(
+            last_hidden_state=sequence_output,
+            pooler_output=pooled_output,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    Albert Model with two heads on top as done during the pretraining: a `masked language modeling` head and a
+    `sentence order prediction (classification)` head.
+    """,
+    ALBERT_START_DOCSTRING,
+)
+class AlbertForPreTraining(AlbertPreTrainedModel):
+    _tied_weights_keys = ["predictions.decoder.bias", "predictions.decoder.weight"]
+
+    def __init__(self, config: AlbertConfig):
+        super().__init__(config)
+
+        self.albert = AlbertModel(config)
+        self.predictions = AlbertMLMHead(config)
+        self.sop_classifier = AlbertSOPHead(config)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_output_embeddings(self) -> nn.Linear:
+        return self.predictions.decoder
+
+    def set_output_embeddings(self, new_embeddings: nn.Linear) -> None:
+        self.predictions.decoder = new_embeddings
+
+    def get_input_embeddings(self) -> nn.Embedding:
+        return self.albert.embeddings.word_embeddings
+
+    @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @replace_return_docstrings(output_type=AlbertForPreTrainingOutput, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = 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,
+        sentence_order_label: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[AlbertForPreTrainingOutput, Tuple]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should be in `[-100, 0, ...,
+            config.vocab_size]` (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]`
+        sentence_order_label (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the next sequence prediction (classification) loss. Input should be a sequence pair
+            (see `input_ids` docstring) Indices should be in `[0, 1]`. `0` indicates original order (sequence A, then
+            sequence B), `1` indicates switched order (sequence B, then sequence A).
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer, AlbertForPreTraining
+        >>> import torch
+
+        >>> tokenizer = AutoTokenizer.from_pretrained("albert/albert-base-v2")
+        >>> model = AlbertForPreTraining.from_pretrained("albert/albert-base-v2")
+
+        >>> input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)
+        >>> # Batch size 1
+        >>> outputs = model(input_ids)
+
+        >>> prediction_logits = outputs.prediction_logits
+        >>> sop_logits = outputs.sop_logits
+        ```"""
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.albert(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        sequence_output, pooled_output = outputs[:2]
+
+        prediction_scores = self.predictions(sequence_output)
+        sop_scores = self.sop_classifier(pooled_output)
+
+        total_loss = None
+        if labels is not None and sentence_order_label is not None:
+            loss_fct = CrossEntropyLoss()
+            masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), labels.view(-1))
+            sentence_order_loss = loss_fct(sop_scores.view(-1, 2), sentence_order_label.view(-1))
+            total_loss = masked_lm_loss + sentence_order_loss
+
+        if not return_dict:
+            output = (prediction_scores, sop_scores) + outputs[2:]
+            return ((total_loss,) + output) if total_loss is not None else output
+
+        return AlbertForPreTrainingOutput(
+            loss=total_loss,
+            prediction_logits=prediction_scores,
+            sop_logits=sop_scores,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+class AlbertMLMHead(nn.Module):
+    def __init__(self, config: AlbertConfig):
+        super().__init__()
+
+        self.LayerNorm = nn.LayerNorm(config.embedding_size, eps=config.layer_norm_eps)
+        self.bias = nn.Parameter(torch.zeros(config.vocab_size))
+        self.dense = nn.Linear(config.hidden_size, config.embedding_size)
+        self.decoder = nn.Linear(config.embedding_size, config.vocab_size)
+        self.activation = ACT2FN[config.hidden_act]
+        self.decoder.bias = self.bias
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.activation(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states)
+        hidden_states = self.decoder(hidden_states)
+
+        prediction_scores = hidden_states
+
+        return prediction_scores
+
+    def _tie_weights(self) -> None:
+        # For accelerate compatibility and to not break backward compatibility
+        if self.decoder.bias.device.type == "meta":
+            self.decoder.bias = self.bias
+        else:
+            # To tie those two weights if they get disconnected (on TPU or when the bias is resized)
+            self.bias = self.decoder.bias
+
+
+class AlbertSOPHead(nn.Module):
+    def __init__(self, config: AlbertConfig):
+        super().__init__()
+
+        self.dropout = nn.Dropout(config.classifier_dropout_prob)
+        self.classifier = nn.Linear(config.hidden_size, config.num_labels)
+
+    def forward(self, pooled_output: torch.Tensor) -> torch.Tensor:
+        dropout_pooled_output = self.dropout(pooled_output)
+        logits = self.classifier(dropout_pooled_output)
+        return logits
+
+
+@add_start_docstrings(
+    "Albert Model with a `language modeling` head on top.",
+    ALBERT_START_DOCSTRING,
+)
+class AlbertForMaskedLM(AlbertPreTrainedModel):
+    _tied_weights_keys = ["predictions.decoder.bias", "predictions.decoder.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.albert = AlbertModel(config, add_pooling_layer=False)
+        self.predictions = AlbertMLMHead(config)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_output_embeddings(self) -> nn.Linear:
+        return self.predictions.decoder
+
+    def set_output_embeddings(self, new_embeddings: nn.Linear) -> None:
+        self.predictions.decoder = new_embeddings
+        self.predictions.bias = new_embeddings.bias
+
+    def get_input_embeddings(self) -> nn.Embedding:
+        return self.albert.embeddings.word_embeddings
+
+    @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @replace_return_docstrings(output_type=MaskedLMOutput, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = 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,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[MaskedLMOutput, Tuple]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should be in `[-100, 0, ...,
+            config.vocab_size]` (see `input_ids` docstring) Tokens with indices set to `-100` are ignored (masked), the
+            loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> import torch
+        >>> from transformers import AutoTokenizer, AlbertForMaskedLM
+
+        >>> tokenizer = AutoTokenizer.from_pretrained("albert/albert-base-v2")
+        >>> model = AlbertForMaskedLM.from_pretrained("albert/albert-base-v2")
+
+        >>> # add mask_token
+        >>> inputs = tokenizer("The capital of [MASK] is Paris.", return_tensors="pt")
+        >>> with torch.no_grad():
+        ...     logits = model(**inputs).logits
+
+        >>> # retrieve index of [MASK]
+        >>> mask_token_index = (inputs.input_ids == tokenizer.mask_token_id)[0].nonzero(as_tuple=True)[0]
+        >>> predicted_token_id = logits[0, mask_token_index].argmax(axis=-1)
+        >>> tokenizer.decode(predicted_token_id)
+        'france'
+        ```
+
+        ```python
+        >>> labels = tokenizer("The capital of France is Paris.", return_tensors="pt")["input_ids"]
+        >>> labels = torch.where(inputs.input_ids == tokenizer.mask_token_id, labels, -100)
+        >>> outputs = model(**inputs, labels=labels)
+        >>> round(outputs.loss.item(), 2)
+        0.81
+        ```
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.albert(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        sequence_outputs = outputs[0]
+
+        prediction_scores = self.predictions(sequence_outputs)
+
+        masked_lm_loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), labels.view(-1))
+
+        if not return_dict:
+            output = (prediction_scores,) + outputs[2:]
+            return ((masked_lm_loss,) + output) if masked_lm_loss is not None else output
+
+        return MaskedLMOutput(
+            loss=masked_lm_loss,
+            logits=prediction_scores,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    Albert Model transformer with a sequence classification/regression head on top (a linear layer on top of the pooled
+    output) e.g. for GLUE tasks.
+    """,
+    ALBERT_START_DOCSTRING,
+)
+class AlbertForSequenceClassification(AlbertPreTrainedModel):
+    def __init__(self, config: AlbertConfig):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.config = config
+
+        self.albert = AlbertModel(config)
+        self.dropout = nn.Dropout(config.classifier_dropout_prob)
+        self.classifier = nn.Linear(config.hidden_size, self.config.num_labels)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        checkpoint="textattack/albert-base-v2-imdb",
+        output_type=SequenceClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+        expected_output="'LABEL_1'",
+        expected_loss=0.12,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = 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,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[SequenceClassifierOutput, Tuple]:
+        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).
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.albert(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        pooled_output = outputs[1]
+
+        pooled_output = self.dropout(pooled_output)
+        logits = self.classifier(pooled_output)
+
+        loss = None
+        if labels is not None:
+            if self.config.problem_type is None:
+                if self.num_labels == 1:
+                    self.config.problem_type = "regression"
+                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
+                    self.config.problem_type = "single_label_classification"
+                else:
+                    self.config.problem_type = "multi_label_classification"
+
+            if self.config.problem_type == "regression":
+                loss_fct = MSELoss()
+                if self.num_labels == 1:
+                    loss = loss_fct(logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(logits, labels)
+            elif self.config.problem_type == "single_label_classification":
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            elif self.config.problem_type == "multi_label_classification":
+                loss_fct = BCEWithLogitsLoss()
+                loss = loss_fct(logits, labels)
+
+        if not return_dict:
+            output = (logits,) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return SequenceClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    Albert Model with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for
+    Named-Entity-Recognition (NER) tasks.
+    """,
+    ALBERT_START_DOCSTRING,
+)
+class AlbertForTokenClassification(AlbertPreTrainedModel):
+    def __init__(self, config: AlbertConfig):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.albert = AlbertModel(config, add_pooling_layer=False)
+        classifier_dropout_prob = (
+            config.classifier_dropout_prob
+            if config.classifier_dropout_prob is not None
+            else config.hidden_dropout_prob
+        )
+        self.dropout = nn.Dropout(classifier_dropout_prob)
+        self.classifier = nn.Linear(config.hidden_size, self.config.num_labels)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=TokenClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = 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,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[TokenClassifierOutput, Tuple]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the token classification loss. Indices should be in `[0, ..., config.num_labels - 1]`.
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.albert(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        sequence_output = outputs[0]
+
+        sequence_output = self.dropout(sequence_output)
+        logits = self.classifier(sequence_output)
+
+        loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+
+        if not return_dict:
+            output = (logits,) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TokenClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    Albert Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear
+    layers on top of the hidden-states output to compute `span start logits` and `span end logits`).
+    """,
+    ALBERT_START_DOCSTRING,
+)
+class AlbertForQuestionAnswering(AlbertPreTrainedModel):
+    def __init__(self, config: AlbertConfig):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.albert = AlbertModel(config, add_pooling_layer=False)
+        self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        checkpoint="twmkn9/albert-base-v2-squad2",
+        output_type=QuestionAnsweringModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+        qa_target_start_index=12,
+        qa_target_end_index=13,
+        expected_output="'a nice puppet'",
+        expected_loss=7.36,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = 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,
+        start_positions: Optional[torch.LongTensor] = None,
+        end_positions: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[AlbertForPreTrainingOutput, Tuple]:
+        r"""
+        start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for position (index) of the start of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
+            are not taken into account for computing the loss.
+        end_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for position (index) of the end of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
+            are not taken into account for computing the loss.
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.albert(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        sequence_output = outputs[0]
+
+        logits: torch.Tensor = self.qa_outputs(sequence_output)
+        start_logits, end_logits = logits.split(1, dim=-1)
+        start_logits = start_logits.squeeze(-1).contiguous()
+        end_logits = end_logits.squeeze(-1).contiguous()
+
+        total_loss = None
+        if start_positions is not None and end_positions is not None:
+            # If we are on multi-GPU, split add a dimension
+            if len(start_positions.size()) > 1:
+                start_positions = start_positions.squeeze(-1)
+            if len(end_positions.size()) > 1:
+                end_positions = end_positions.squeeze(-1)
+            # sometimes the start/end positions are outside our model inputs, we ignore these terms
+            ignored_index = start_logits.size(1)
+            start_positions = start_positions.clamp(0, ignored_index)
+            end_positions = end_positions.clamp(0, ignored_index)
+
+            loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
+            start_loss = loss_fct(start_logits, start_positions)
+            end_loss = loss_fct(end_logits, end_positions)
+            total_loss = (start_loss + end_loss) / 2
+
+        if not return_dict:
+            output = (start_logits, end_logits) + outputs[2:]
+            return ((total_loss,) + output) if total_loss is not None else output
+
+        return QuestionAnsweringModelOutput(
+            loss=total_loss,
+            start_logits=start_logits,
+            end_logits=end_logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    Albert Model with a multiple choice classification head on top (a linear layer on top of the pooled output and a
+    softmax) e.g. for RocStories/SWAG tasks.
+    """,
+    ALBERT_START_DOCSTRING,
+)
+class AlbertForMultipleChoice(AlbertPreTrainedModel):
+    def __init__(self, config: AlbertConfig):
+        super().__init__(config)
+
+        self.albert = AlbertModel(config)
+        self.dropout = nn.Dropout(config.classifier_dropout_prob)
+        self.classifier = nn.Linear(config.hidden_size, 1)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length"))
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=MultipleChoiceModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = 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,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[AlbertForPreTrainingOutput, Tuple]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the multiple choice classification loss. Indices should be in `[0, ...,
+            num_choices-1]` where *num_choices* is the size of the second dimension of the input tensors. (see
+            *input_ids* above)
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        num_choices = input_ids.shape[1] if input_ids is not None else inputs_embeds.shape[1]
+
+        input_ids = input_ids.view(-1, input_ids.size(-1)) if input_ids is not None else None
+        attention_mask = attention_mask.view(-1, attention_mask.size(-1)) if attention_mask is not None else None
+        token_type_ids = token_type_ids.view(-1, token_type_ids.size(-1)) if token_type_ids is not None else None
+        position_ids = position_ids.view(-1, position_ids.size(-1)) if position_ids is not None else None
+        inputs_embeds = (
+            inputs_embeds.view(-1, inputs_embeds.size(-2), inputs_embeds.size(-1))
+            if inputs_embeds is not None
+            else None
+        )
+        outputs = self.albert(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        pooled_output = outputs[1]
+
+        pooled_output = self.dropout(pooled_output)
+        logits: torch.Tensor = self.classifier(pooled_output)
+        reshaped_logits = logits.view(-1, num_choices)
+
+        loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(reshaped_logits, labels)
+
+        if not return_dict:
+            output = (reshaped_logits,) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return MultipleChoiceModelOutput(
+            loss=loss,
+            logits=reshaped_logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+__all__ = [
+    "load_tf_weights_in_albert",
+    "AlbertPreTrainedModel",
+    "AlbertModel",
+    "AlbertForPreTraining",
+    "AlbertForMaskedLM",
+    "AlbertForSequenceClassification",
+    "AlbertForTokenClassification",
+    "AlbertForQuestionAnswering",
+    "AlbertForMultipleChoice",
+]

lib/python3.12/site-packages/transformers/models/albert/modeling_flax_albert.py

@@ -0,0 +1,1132 @@
+# coding=utf-8
+# Copyright 2021 Google AI, Google Brain and the HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Callable, Optional, Tuple
+
+import flax
+import flax.linen as nn
+import jax
+import jax.numpy as jnp
+import numpy as np
+from flax.core.frozen_dict import FrozenDict, freeze, unfreeze
+from flax.linen.attention import dot_product_attention_weights
+from flax.traverse_util import flatten_dict, unflatten_dict
+from jax import lax
+
+from ...modeling_flax_outputs import (
+    FlaxBaseModelOutput,
+    FlaxBaseModelOutputWithPooling,
+    FlaxMaskedLMOutput,
+    FlaxMultipleChoiceModelOutput,
+    FlaxQuestionAnsweringModelOutput,
+    FlaxSequenceClassifierOutput,
+    FlaxTokenClassifierOutput,
+)
+from ...modeling_flax_utils import (
+    ACT2FN,
+    FlaxPreTrainedModel,
+    append_call_sample_docstring,
+    append_replace_return_docstrings,
+    overwrite_call_docstring,
+)
+from ...utils import ModelOutput, add_start_docstrings, add_start_docstrings_to_model_forward, logging
+from .configuration_albert import AlbertConfig
+
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "albert/albert-base-v2"
+_CONFIG_FOR_DOC = "AlbertConfig"
+
+
+@flax.struct.dataclass
+class FlaxAlbertForPreTrainingOutput(ModelOutput):
+    """
+    Output type of [`FlaxAlbertForPreTraining`].
+
+    Args:
+        prediction_logits (`jnp.ndarray` of shape `(batch_size, sequence_length, config.vocab_size)`):
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        sop_logits (`jnp.ndarray` of shape `(batch_size, 2)`):
+            Prediction scores of the next sequence prediction (classification) head (scores of True/False continuation
+            before SoftMax).
+        hidden_states (`tuple(jnp.ndarray)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `jnp.ndarray` (one for the output of the embeddings + one for the output of each layer) of shape
+            `(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (`tuple(jnp.ndarray)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
+            Tuple of `jnp.ndarray` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
+            sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    prediction_logits: jnp.ndarray = None
+    sop_logits: jnp.ndarray = None
+    hidden_states: Optional[Tuple[jnp.ndarray]] = None
+    attentions: Optional[Tuple[jnp.ndarray]] = None
+
+
+ALBERT_START_DOCSTRING = r"""
+
+    This model inherits from [`FlaxPreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading, saving and converting weights from PyTorch models)
+
+    This model is also a
+    [flax.linen.Module](https://flax.readthedocs.io/en/latest/api_reference/flax.linen/module.html) subclass. Use it as
+    a regular Flax linen Module and refer to the Flax documentation for all matter related to general usage and
+    behavior.
+
+    Finally, this model supports inherent JAX features such as:
+
+    - [Just-In-Time (JIT) compilation](https://jax.readthedocs.io/en/latest/jax.html#just-in-time-compilation-jit)
+    - [Automatic Differentiation](https://jax.readthedocs.io/en/latest/jax.html#automatic-differentiation)
+    - [Vectorization](https://jax.readthedocs.io/en/latest/jax.html#vectorization-vmap)
+    - [Parallelization](https://jax.readthedocs.io/en/latest/jax.html#parallelization-pmap)
+
+    Parameters:
+        config ([`AlbertConfig`]): 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`].
+"""
+
+ALBERT_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`numpy.ndarray` 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 (`numpy.ndarray` 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 (`numpy.ndarray` 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 (`numpy.ndarray` of shape `({0})`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.max_position_embeddings - 1]`.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+
+"""
+
+
+class FlaxAlbertEmbeddings(nn.Module):
+    """Construct the embeddings from word, position and token_type embeddings."""
+
+    config: AlbertConfig
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
+
+    def setup(self):
+        self.word_embeddings = nn.Embed(
+            self.config.vocab_size,
+            self.config.embedding_size,
+            embedding_init=jax.nn.initializers.normal(stddev=self.config.initializer_range),
+        )
+        self.position_embeddings = nn.Embed(
+            self.config.max_position_embeddings,
+            self.config.embedding_size,
+            embedding_init=jax.nn.initializers.normal(stddev=self.config.initializer_range),
+        )
+        self.token_type_embeddings = nn.Embed(
+            self.config.type_vocab_size,
+            self.config.embedding_size,
+            embedding_init=jax.nn.initializers.normal(stddev=self.config.initializer_range),
+        )
+        self.LayerNorm = nn.LayerNorm(epsilon=self.config.layer_norm_eps, dtype=self.dtype)
+        self.dropout = nn.Dropout(rate=self.config.hidden_dropout_prob)
+
+    def __call__(self, input_ids, token_type_ids, position_ids, deterministic: bool = True):
+        # Embed
+        inputs_embeds = self.word_embeddings(input_ids.astype("i4"))
+        position_embeds = self.position_embeddings(position_ids.astype("i4"))
+        token_type_embeddings = self.token_type_embeddings(token_type_ids.astype("i4"))
+
+        # Sum all embeddings
+        hidden_states = inputs_embeds + token_type_embeddings + position_embeds
+
+        # Layer Norm
+        hidden_states = self.LayerNorm(hidden_states)
+        hidden_states = self.dropout(hidden_states, deterministic=deterministic)
+        return hidden_states
+
+
+class FlaxAlbertSelfAttention(nn.Module):
+    config: AlbertConfig
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
+
+    def setup(self):
+        if self.config.hidden_size % self.config.num_attention_heads != 0:
+            raise ValueError(
+                "`config.hidden_size`: {self.config.hidden_size} has to be a multiple of `config.num_attention_heads` "
+                "                   : {self.config.num_attention_heads}"
+            )
+
+        self.query = nn.Dense(
+            self.config.hidden_size,
+            dtype=self.dtype,
+            kernel_init=jax.nn.initializers.normal(self.config.initializer_range),
+        )
+        self.key = nn.Dense(
+            self.config.hidden_size,
+            dtype=self.dtype,
+            kernel_init=jax.nn.initializers.normal(self.config.initializer_range),
+        )
+        self.value = nn.Dense(
+            self.config.hidden_size,
+            dtype=self.dtype,
+            kernel_init=jax.nn.initializers.normal(self.config.initializer_range),
+        )
+        self.dense = nn.Dense(
+            self.config.hidden_size,
+            kernel_init=jax.nn.initializers.normal(self.config.initializer_range),
+            dtype=self.dtype,
+        )
+        self.LayerNorm = nn.LayerNorm(epsilon=self.config.layer_norm_eps, dtype=self.dtype)
+        self.dropout = nn.Dropout(rate=self.config.hidden_dropout_prob)
+
+    def __call__(self, hidden_states, attention_mask, deterministic=True, output_attentions: bool = False):
+        head_dim = self.config.hidden_size // self.config.num_attention_heads
+
+        query_states = self.query(hidden_states).reshape(
+            hidden_states.shape[:2] + (self.config.num_attention_heads, head_dim)
+        )
+        value_states = self.value(hidden_states).reshape(
+            hidden_states.shape[:2] + (self.config.num_attention_heads, head_dim)
+        )
+        key_states = self.key(hidden_states).reshape(
+            hidden_states.shape[:2] + (self.config.num_attention_heads, head_dim)
+        )
+
+        # Convert the boolean attention mask to an attention bias.
+        if attention_mask is not None:
+            # attention mask in the form of attention bias
+            attention_mask = jnp.expand_dims(attention_mask, axis=(-3, -2))
+            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.config.attention_probs_dropout_prob > 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.config.attention_probs_dropout_prob,
+            broadcast_dropout=True,
+            deterministic=deterministic,
+            dtype=self.dtype,
+            precision=None,
+        )
+
+        attn_output = jnp.einsum("...hqk,...khd->...qhd", attn_weights, value_states)
+        attn_output = attn_output.reshape(attn_output.shape[:2] + (-1,))
+
+        projected_attn_output = self.dense(attn_output)
+        projected_attn_output = self.dropout(projected_attn_output, deterministic=deterministic)
+        layernormed_attn_output = self.LayerNorm(projected_attn_output + hidden_states)
+        outputs = (layernormed_attn_output, attn_weights) if output_attentions else (layernormed_attn_output,)
+        return outputs
+
+
+class FlaxAlbertLayer(nn.Module):
+    config: AlbertConfig
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
+
+    def setup(self):
+        self.attention = FlaxAlbertSelfAttention(self.config, dtype=self.dtype)
+        self.ffn = nn.Dense(
+            self.config.intermediate_size,
+            kernel_init=jax.nn.initializers.normal(self.config.initializer_range),
+            dtype=self.dtype,
+        )
+        self.activation = ACT2FN[self.config.hidden_act]
+        self.ffn_output = nn.Dense(
+            self.config.hidden_size,
+            kernel_init=jax.nn.initializers.normal(self.config.initializer_range),
+            dtype=self.dtype,
+        )
+        self.full_layer_layer_norm = nn.LayerNorm(epsilon=self.config.layer_norm_eps, dtype=self.dtype)
+        self.dropout = nn.Dropout(rate=self.config.hidden_dropout_prob)
+
+    def __call__(
+        self,
+        hidden_states,
+        attention_mask,
+        deterministic: bool = True,
+        output_attentions: bool = False,
+    ):
+        attention_outputs = self.attention(
+            hidden_states, attention_mask, deterministic=deterministic, output_attentions=output_attentions
+        )
+        attention_output = attention_outputs[0]
+        ffn_output = self.ffn(attention_output)
+        ffn_output = self.activation(ffn_output)
+        ffn_output = self.ffn_output(ffn_output)
+        ffn_output = self.dropout(ffn_output, deterministic=deterministic)
+        hidden_states = self.full_layer_layer_norm(ffn_output + attention_output)
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (attention_outputs[1],)
+        return outputs
+
+
+class FlaxAlbertLayerCollection(nn.Module):
+    config: AlbertConfig
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
+
+    def setup(self):
+        self.layers = [
+            FlaxAlbertLayer(self.config, name=str(i), dtype=self.dtype) for i in range(self.config.inner_group_num)
+        ]
+
+    def __call__(
+        self,
+        hidden_states,
+        attention_mask,
+        deterministic: bool = True,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+    ):
+        layer_hidden_states = ()
+        layer_attentions = ()
+
+        for layer_index, albert_layer in enumerate(self.layers):
+            layer_output = albert_layer(
+                hidden_states,
+                attention_mask,
+                deterministic=deterministic,
+                output_attentions=output_attentions,
+            )
+            hidden_states = layer_output[0]
+
+            if output_attentions:
+                layer_attentions = layer_attentions + (layer_output[1],)
+
+            if output_hidden_states:
+                layer_hidden_states = layer_hidden_states + (hidden_states,)
+
+        outputs = (hidden_states,)
+        if output_hidden_states:
+            outputs = outputs + (layer_hidden_states,)
+        if output_attentions:
+            outputs = outputs + (layer_attentions,)
+        return outputs  # last-layer hidden state, (layer hidden states), (layer attentions)
+
+
+class FlaxAlbertLayerCollections(nn.Module):
+    config: AlbertConfig
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
+    layer_index: Optional[str] = None
+
+    def setup(self):
+        self.albert_layers = FlaxAlbertLayerCollection(self.config, dtype=self.dtype)
+
+    def __call__(
+        self,
+        hidden_states,
+        attention_mask,
+        deterministic: bool = True,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+    ):
+        outputs = self.albert_layers(
+            hidden_states,
+            attention_mask,
+            deterministic=deterministic,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+        )
+        return outputs
+
+
+class FlaxAlbertLayerGroups(nn.Module):
+    config: AlbertConfig
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
+
+    def setup(self):
+        self.layers = [
+            FlaxAlbertLayerCollections(self.config, name=str(i), layer_index=str(i), dtype=self.dtype)
+            for i in range(self.config.num_hidden_groups)
+        ]
+
+    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 = (hidden_states,) if output_hidden_states else None
+
+        for i in range(self.config.num_hidden_layers):
+            # Index of the hidden group
+            group_idx = int(i / (self.config.num_hidden_layers / self.config.num_hidden_groups))
+            layer_group_output = self.layers[group_idx](
+                hidden_states,
+                attention_mask,
+                deterministic=deterministic,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+            )
+            hidden_states = layer_group_output[0]
+
+            if output_attentions:
+                all_attentions = all_attentions + layer_group_output[-1]
+
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, all_hidden_states, all_attentions] if v is not None)
+        return FlaxBaseModelOutput(
+            last_hidden_state=hidden_states, hidden_states=all_hidden_states, attentions=all_attentions
+        )
+
+
+class FlaxAlbertEncoder(nn.Module):
+    config: AlbertConfig
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
+
+    def setup(self):
+        self.embedding_hidden_mapping_in = nn.Dense(
+            self.config.hidden_size,
+            kernel_init=jax.nn.initializers.normal(self.config.initializer_range),
+            dtype=self.dtype,
+        )
+        self.albert_layer_groups = FlaxAlbertLayerGroups(self.config, dtype=self.dtype)
+
+    def __call__(
+        self,
+        hidden_states,
+        attention_mask,
+        deterministic: bool = True,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ):
+        hidden_states = self.embedding_hidden_mapping_in(hidden_states)
+        return self.albert_layer_groups(
+            hidden_states,
+            attention_mask,
+            deterministic=deterministic,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+        )
+
+
+class FlaxAlbertOnlyMLMHead(nn.Module):
+    config: AlbertConfig
+    dtype: jnp.dtype = jnp.float32
+    bias_init: Callable[..., np.ndarray] = jax.nn.initializers.zeros
+
+    def setup(self):
+        self.dense = nn.Dense(self.config.embedding_size, dtype=self.dtype)
+        self.activation = ACT2FN[self.config.hidden_act]
+        self.LayerNorm = nn.LayerNorm(epsilon=self.config.layer_norm_eps, dtype=self.dtype)
+        self.decoder = nn.Dense(self.config.vocab_size, dtype=self.dtype, use_bias=False)
+        self.bias = self.param("bias", self.bias_init, (self.config.vocab_size,))
+
+    def __call__(self, hidden_states, shared_embedding=None):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.activation(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states)
+
+        if shared_embedding is not None:
+            hidden_states = self.decoder.apply({"params": {"kernel": shared_embedding.T}}, hidden_states)
+        else:
+            hidden_states = self.decoder(hidden_states)
+
+        hidden_states += self.bias
+        return hidden_states
+
+
+class FlaxAlbertSOPHead(nn.Module):
+    config: AlbertConfig
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        self.dropout = nn.Dropout(self.config.classifier_dropout_prob)
+        self.classifier = nn.Dense(2, dtype=self.dtype)
+
+    def __call__(self, pooled_output, deterministic=True):
+        pooled_output = self.dropout(pooled_output, deterministic=deterministic)
+        logits = self.classifier(pooled_output)
+        return logits
+
+
+class FlaxAlbertPreTrainedModel(FlaxPreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = AlbertConfig
+    base_model_prefix = "albert"
+    module_class: nn.Module = None
+
+    def __init__(
+        self,
+        config: AlbertConfig,
+        input_shape: Tuple = (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")
+        token_type_ids = jnp.zeros_like(input_ids)
+        position_ids = jnp.broadcast_to(jnp.arange(jnp.atleast_2d(input_ids).shape[-1]), input_shape)
+        attention_mask = jnp.ones_like(input_ids)
+
+        params_rng, dropout_rng = jax.random.split(rng)
+        rngs = {"params": params_rng, "dropout": dropout_rng}
+
+        random_params = self.module.init(
+            rngs, input_ids, attention_mask, token_type_ids, position_ids, return_dict=False
+        )["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
+
+    @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    def __call__(
+        self,
+        input_ids,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        params: dict = None,
+        dropout_rng: jax.random.PRNGKey = None,
+        train: bool = False,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ):
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.return_dict
+
+        # init input tensors if not passed
+        if token_type_ids is None:
+            token_type_ids = jnp.zeros_like(input_ids)
+
+        if position_ids is None:
+            position_ids = jnp.broadcast_to(jnp.arange(jnp.atleast_2d(input_ids).shape[-1]), input_ids.shape)
+
+        if attention_mask is None:
+            attention_mask = jnp.ones_like(input_ids)
+
+        # Handle any PRNG if needed
+        rngs = {}
+        if dropout_rng is not None:
+            rngs["dropout"] = dropout_rng
+
+        return self.module.apply(
+            {"params": params or self.params},
+            jnp.array(input_ids, dtype="i4"),
+            jnp.array(attention_mask, dtype="i4"),
+            jnp.array(token_type_ids, dtype="i4"),
+            jnp.array(position_ids, dtype="i4"),
+            not train,
+            output_attentions,
+            output_hidden_states,
+            return_dict,
+            rngs=rngs,
+        )
+
+
+class FlaxAlbertModule(nn.Module):
+    config: AlbertConfig
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
+    add_pooling_layer: bool = True
+
+    def setup(self):
+        self.embeddings = FlaxAlbertEmbeddings(self.config, dtype=self.dtype)
+        self.encoder = FlaxAlbertEncoder(self.config, dtype=self.dtype)
+        if self.add_pooling_layer:
+            self.pooler = nn.Dense(
+                self.config.hidden_size,
+                kernel_init=jax.nn.initializers.normal(self.config.initializer_range),
+                dtype=self.dtype,
+                name="pooler",
+            )
+            self.pooler_activation = nn.tanh
+        else:
+            self.pooler = None
+            self.pooler_activation = None
+
+    def __call__(
+        self,
+        input_ids,
+        attention_mask,
+        token_type_ids: Optional[np.ndarray] = None,
+        position_ids: Optional[np.ndarray] = None,
+        deterministic: bool = True,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ):
+        # make sure `token_type_ids` is correctly initialized when not passed
+        if token_type_ids is None:
+            token_type_ids = jnp.zeros_like(input_ids)
+
+        # make sure `position_ids` is correctly initialized when not passed
+        if position_ids is None:
+            position_ids = jnp.broadcast_to(jnp.arange(jnp.atleast_2d(input_ids).shape[-1]), input_ids.shape)
+
+        hidden_states = self.embeddings(input_ids, token_type_ids, position_ids, deterministic=deterministic)
+
+        outputs = self.encoder(
+            hidden_states,
+            attention_mask,
+            deterministic=deterministic,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        hidden_states = outputs[0]
+        if self.add_pooling_layer:
+            pooled = self.pooler(hidden_states[:, 0])
+            pooled = self.pooler_activation(pooled)
+        else:
+            pooled = None
+
+        if not return_dict:
+            # if pooled is None, don't return it
+            if pooled is None:
+                return (hidden_states,) + outputs[1:]
+            return (hidden_states, pooled) + outputs[1:]
+
+        return FlaxBaseModelOutputWithPooling(
+            last_hidden_state=hidden_states,
+            pooler_output=pooled,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    "The bare Albert Model transformer outputting raw hidden-states without any specific head on top.",
+    ALBERT_START_DOCSTRING,
+)
+class FlaxAlbertModel(FlaxAlbertPreTrainedModel):
+    module_class = FlaxAlbertModule
+
+
+append_call_sample_docstring(FlaxAlbertModel, _CHECKPOINT_FOR_DOC, FlaxBaseModelOutputWithPooling, _CONFIG_FOR_DOC)
+
+
+class FlaxAlbertForPreTrainingModule(nn.Module):
+    config: AlbertConfig
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        self.albert = FlaxAlbertModule(config=self.config, dtype=self.dtype)
+        self.predictions = FlaxAlbertOnlyMLMHead(config=self.config, dtype=self.dtype)
+        self.sop_classifier = FlaxAlbertSOPHead(config=self.config, dtype=self.dtype)
+
+    def __call__(
+        self,
+        input_ids,
+        attention_mask,
+        token_type_ids,
+        position_ids,
+        deterministic: bool = True,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ):
+        # Model
+        outputs = self.albert(
+            input_ids,
+            attention_mask,
+            token_type_ids,
+            position_ids,
+            deterministic=deterministic,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        if self.config.tie_word_embeddings:
+            shared_embedding = self.albert.variables["params"]["embeddings"]["word_embeddings"]["embedding"]
+        else:
+            shared_embedding = None
+
+        hidden_states = outputs[0]
+        pooled_output = outputs[1]
+
+        prediction_scores = self.predictions(hidden_states, shared_embedding=shared_embedding)
+        sop_scores = self.sop_classifier(pooled_output, deterministic=deterministic)
+
+        if not return_dict:
+            return (prediction_scores, sop_scores) + outputs[2:]
+
+        return FlaxAlbertForPreTrainingOutput(
+            prediction_logits=prediction_scores,
+            sop_logits=sop_scores,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    Albert Model with two heads on top as done during the pretraining: a `masked language modeling` head and a
+    `sentence order prediction (classification)` head.
+    """,
+    ALBERT_START_DOCSTRING,
+)
+class FlaxAlbertForPreTraining(FlaxAlbertPreTrainedModel):
+    module_class = FlaxAlbertForPreTrainingModule
+
+
+FLAX_ALBERT_FOR_PRETRAINING_DOCSTRING = """
+    Returns:
+
+    Example:
+
+    ```python
+    >>> from transformers import AutoTokenizer, FlaxAlbertForPreTraining
+
+    >>> tokenizer = AutoTokenizer.from_pretrained("albert/albert-base-v2")
+    >>> model = FlaxAlbertForPreTraining.from_pretrained("albert/albert-base-v2")
+
+    >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="np")
+    >>> outputs = model(**inputs)
+
+    >>> prediction_logits = outputs.prediction_logits
+    >>> seq_relationship_logits = outputs.sop_logits
+    ```
+"""
+
+overwrite_call_docstring(
+    FlaxAlbertForPreTraining,
+    ALBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length") + FLAX_ALBERT_FOR_PRETRAINING_DOCSTRING,
+)
+append_replace_return_docstrings(
+    FlaxAlbertForPreTraining, output_type=FlaxAlbertForPreTrainingOutput, config_class=_CONFIG_FOR_DOC
+)
+
+
+class FlaxAlbertForMaskedLMModule(nn.Module):
+    config: AlbertConfig
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        self.albert = FlaxAlbertModule(config=self.config, add_pooling_layer=False, dtype=self.dtype)
+        self.predictions = FlaxAlbertOnlyMLMHead(config=self.config, dtype=self.dtype)
+
+    def __call__(
+        self,
+        input_ids,
+        attention_mask,
+        token_type_ids,
+        position_ids,
+        deterministic: bool = True,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ):
+        # Model
+        outputs = self.albert(
+            input_ids,
+            attention_mask,
+            token_type_ids,
+            position_ids,
+            deterministic=deterministic,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = outputs[0]
+        if self.config.tie_word_embeddings:
+            shared_embedding = self.albert.variables["params"]["embeddings"]["word_embeddings"]["embedding"]
+        else:
+            shared_embedding = None
+
+        # Compute the prediction scores
+        logits = self.predictions(hidden_states, shared_embedding=shared_embedding)
+
+        if not return_dict:
+            return (logits,) + outputs[1:]
+
+        return FlaxMaskedLMOutput(
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings("""Albert Model with a `language modeling` head on top.""", ALBERT_START_DOCSTRING)
+class FlaxAlbertForMaskedLM(FlaxAlbertPreTrainedModel):
+    module_class = FlaxAlbertForMaskedLMModule
+
+
+append_call_sample_docstring(
+    FlaxAlbertForMaskedLM, _CHECKPOINT_FOR_DOC, FlaxMaskedLMOutput, _CONFIG_FOR_DOC, revision="refs/pr/11"
+)
+
+
+class FlaxAlbertForSequenceClassificationModule(nn.Module):
+    config: AlbertConfig
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        self.albert = FlaxAlbertModule(config=self.config, dtype=self.dtype)
+        classifier_dropout = (
+            self.config.classifier_dropout_prob
+            if self.config.classifier_dropout_prob is not None
+            else self.config.hidden_dropout_prob
+        )
+        self.dropout = nn.Dropout(rate=classifier_dropout)
+        self.classifier = nn.Dense(
+            self.config.num_labels,
+            dtype=self.dtype,
+        )
+
+    def __call__(
+        self,
+        input_ids,
+        attention_mask,
+        token_type_ids,
+        position_ids,
+        deterministic: bool = True,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ):
+        # Model
+        outputs = self.albert(
+            input_ids,
+            attention_mask,
+            token_type_ids,
+            position_ids,
+            deterministic=deterministic,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        pooled_output = outputs[1]
+        pooled_output = self.dropout(pooled_output, deterministic=deterministic)
+        logits = self.classifier(pooled_output)
+
+        if not return_dict:
+            return (logits,) + outputs[2:]
+
+        return FlaxSequenceClassifierOutput(
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    Albert Model transformer with a sequence classification/regression head on top (a linear layer on top of the pooled
+    output) e.g. for GLUE tasks.
+    """,
+    ALBERT_START_DOCSTRING,
+)
+class FlaxAlbertForSequenceClassification(FlaxAlbertPreTrainedModel):
+    module_class = FlaxAlbertForSequenceClassificationModule
+
+
+append_call_sample_docstring(
+    FlaxAlbertForSequenceClassification,
+    _CHECKPOINT_FOR_DOC,
+    FlaxSequenceClassifierOutput,
+    _CONFIG_FOR_DOC,
+)
+
+
+class FlaxAlbertForMultipleChoiceModule(nn.Module):
+    config: AlbertConfig
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        self.albert = FlaxAlbertModule(config=self.config, dtype=self.dtype)
+        self.dropout = nn.Dropout(rate=self.config.hidden_dropout_prob)
+        self.classifier = nn.Dense(1, dtype=self.dtype)
+
+    def __call__(
+        self,
+        input_ids,
+        attention_mask,
+        token_type_ids,
+        position_ids,
+        deterministic: bool = True,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ):
+        num_choices = input_ids.shape[1]
+        input_ids = input_ids.reshape(-1, input_ids.shape[-1]) if input_ids is not None else None
+        attention_mask = attention_mask.reshape(-1, attention_mask.shape[-1]) if attention_mask is not None else None
+        token_type_ids = token_type_ids.reshape(-1, token_type_ids.shape[-1]) if token_type_ids is not None else None
+        position_ids = position_ids.reshape(-1, position_ids.shape[-1]) if position_ids is not None else None
+
+        # Model
+        outputs = self.albert(
+            input_ids,
+            attention_mask,
+            token_type_ids,
+            position_ids,
+            deterministic=deterministic,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        pooled_output = outputs[1]
+        pooled_output = self.dropout(pooled_output, deterministic=deterministic)
+        logits = self.classifier(pooled_output)
+
+        reshaped_logits = logits.reshape(-1, num_choices)
+
+        if not return_dict:
+            return (reshaped_logits,) + outputs[2:]
+
+        return FlaxMultipleChoiceModelOutput(
+            logits=reshaped_logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    Albert Model with a multiple choice classification head on top (a linear layer on top of the pooled output and a
+    softmax) e.g. for RocStories/SWAG tasks.
+    """,
+    ALBERT_START_DOCSTRING,
+)
+class FlaxAlbertForMultipleChoice(FlaxAlbertPreTrainedModel):
+    module_class = FlaxAlbertForMultipleChoiceModule
+
+
+overwrite_call_docstring(
+    FlaxAlbertForMultipleChoice, ALBERT_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length")
+)
+append_call_sample_docstring(
+    FlaxAlbertForMultipleChoice,
+    _CHECKPOINT_FOR_DOC,
+    FlaxMultipleChoiceModelOutput,
+    _CONFIG_FOR_DOC,
+)
+
+
+class FlaxAlbertForTokenClassificationModule(nn.Module):
+    config: AlbertConfig
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        self.albert = FlaxAlbertModule(config=self.config, dtype=self.dtype, add_pooling_layer=False)
+        classifier_dropout = (
+            self.config.classifier_dropout_prob
+            if self.config.classifier_dropout_prob is not None
+            else self.config.hidden_dropout_prob
+        )
+        self.dropout = nn.Dropout(rate=classifier_dropout)
+        self.classifier = nn.Dense(self.config.num_labels, dtype=self.dtype)
+
+    def __call__(
+        self,
+        input_ids,
+        attention_mask,
+        token_type_ids,
+        position_ids,
+        deterministic: bool = True,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ):
+        # Model
+        outputs = self.albert(
+            input_ids,
+            attention_mask,
+            token_type_ids,
+            position_ids,
+            deterministic=deterministic,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = outputs[0]
+        hidden_states = self.dropout(hidden_states, deterministic=deterministic)
+        logits = self.classifier(hidden_states)
+
+        if not return_dict:
+            return (logits,) + outputs[1:]
+
+        return FlaxTokenClassifierOutput(
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    Albert Model with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for
+    Named-Entity-Recognition (NER) tasks.
+    """,
+    ALBERT_START_DOCSTRING,
+)
+class FlaxAlbertForTokenClassification(FlaxAlbertPreTrainedModel):
+    module_class = FlaxAlbertForTokenClassificationModule
+
+
+append_call_sample_docstring(
+    FlaxAlbertForTokenClassification,
+    _CHECKPOINT_FOR_DOC,
+    FlaxTokenClassifierOutput,
+    _CONFIG_FOR_DOC,
+)
+
+
+class FlaxAlbertForQuestionAnsweringModule(nn.Module):
+    config: AlbertConfig
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        self.albert = FlaxAlbertModule(config=self.config, dtype=self.dtype, add_pooling_layer=False)
+        self.qa_outputs = nn.Dense(self.config.num_labels, dtype=self.dtype)
+
+    def __call__(
+        self,
+        input_ids,
+        attention_mask,
+        token_type_ids,
+        position_ids,
+        deterministic: bool = True,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ):
+        # Model
+        outputs = self.albert(
+            input_ids,
+            attention_mask,
+            token_type_ids,
+            position_ids,
+            deterministic=deterministic,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = outputs[0]
+
+        logits = self.qa_outputs(hidden_states)
+        start_logits, end_logits = jnp.split(logits, self.config.num_labels, axis=-1)
+        start_logits = start_logits.squeeze(-1)
+        end_logits = end_logits.squeeze(-1)
+
+        if not return_dict:
+            return (start_logits, end_logits) + outputs[1:]
+
+        return FlaxQuestionAnsweringModelOutput(
+            start_logits=start_logits,
+            end_logits=end_logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    Albert Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear
+    layers on top of the hidden-states output to compute `span start logits` and `span end logits`).
+    """,
+    ALBERT_START_DOCSTRING,
+)
+class FlaxAlbertForQuestionAnswering(FlaxAlbertPreTrainedModel):
+    module_class = FlaxAlbertForQuestionAnsweringModule
+
+
+append_call_sample_docstring(
+    FlaxAlbertForQuestionAnswering,
+    _CHECKPOINT_FOR_DOC,
+    FlaxQuestionAnsweringModelOutput,
+    _CONFIG_FOR_DOC,
+)
+
+__all__ = [
+    "FlaxAlbertPreTrainedModel",
+    "FlaxAlbertModel",
+    "FlaxAlbertForPreTraining",
+    "FlaxAlbertForMaskedLM",
+    "FlaxAlbertForSequenceClassification",
+    "FlaxAlbertForMultipleChoice",
+    "FlaxAlbertForTokenClassification",
+    "FlaxAlbertForQuestionAnswering",
+]

lib/python3.12/site-packages/transformers/models/albert/modeling_tf_albert.py

@@ -0,0 +1,1573 @@
+# coding=utf-8
+# Copyright 2018 The OpenAI Team Authors and HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  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 ALBERT model."""
+
+from __future__ import annotations
+
+import math
+from dataclasses import dataclass
+from typing import Dict, Optional, Tuple, Union
+
+import numpy as np
+import tensorflow as tf
+
+from ...activations_tf import get_tf_activation
+from ...modeling_tf_outputs import (
+    TFBaseModelOutput,
+    TFBaseModelOutputWithPooling,
+    TFMaskedLMOutput,
+    TFMultipleChoiceModelOutput,
+    TFQuestionAnsweringModelOutput,
+    TFSequenceClassifierOutput,
+    TFTokenClassifierOutput,
+)
+from ...modeling_tf_utils import (
+    TFMaskedLanguageModelingLoss,
+    TFModelInputType,
+    TFMultipleChoiceLoss,
+    TFPreTrainedModel,
+    TFQuestionAnsweringLoss,
+    TFSequenceClassificationLoss,
+    TFTokenClassificationLoss,
+    get_initializer,
+    keras,
+    keras_serializable,
+    unpack_inputs,
+)
+from ...tf_utils import check_embeddings_within_bounds, shape_list, stable_softmax
+from ...utils import (
+    ModelOutput,
+    add_code_sample_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    logging,
+    replace_return_docstrings,
+)
+from .configuration_albert import AlbertConfig
+
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "albert/albert-base-v2"
+_CONFIG_FOR_DOC = "AlbertConfig"
+
+
+class TFAlbertPreTrainingLoss:
+    """
+    Loss function suitable for ALBERT pretraining, that is, the task of pretraining a language model by combining SOP +
+    MLM. .. note:: Any label of -100 will be ignored (along with the corresponding logits) in the loss computation.
+    """
+
+    def hf_compute_loss(self, labels: tf.Tensor, logits: tf.Tensor) -> tf.Tensor:
+        loss_fn = keras.losses.SparseCategoricalCrossentropy(from_logits=True, reduction=keras.losses.Reduction.NONE)
+        if self.config.tf_legacy_loss:
+            # make sure only labels that are not equal to -100
+            # are taken into account as loss
+            masked_lm_active_loss = tf.not_equal(tf.reshape(tensor=labels["labels"], shape=(-1,)), -100)
+            masked_lm_reduced_logits = tf.boolean_mask(
+                tensor=tf.reshape(tensor=logits[0], shape=(-1, shape_list(logits[0])[2])),
+                mask=masked_lm_active_loss,
+            )
+            masked_lm_labels = tf.boolean_mask(
+                tensor=tf.reshape(tensor=labels["labels"], shape=(-1,)), mask=masked_lm_active_loss
+            )
+            sentence_order_active_loss = tf.not_equal(
+                tf.reshape(tensor=labels["sentence_order_label"], shape=(-1,)), -100
+            )
+            sentence_order_reduced_logits = tf.boolean_mask(
+                tensor=tf.reshape(tensor=logits[1], shape=(-1, 2)), mask=sentence_order_active_loss
+            )
+            sentence_order_label = tf.boolean_mask(
+                tensor=tf.reshape(tensor=labels["sentence_order_label"], shape=(-1,)), mask=sentence_order_active_loss
+            )
+            masked_lm_loss = loss_fn(y_true=masked_lm_labels, y_pred=masked_lm_reduced_logits)
+            sentence_order_loss = loss_fn(y_true=sentence_order_label, y_pred=sentence_order_reduced_logits)
+            masked_lm_loss = tf.reshape(tensor=masked_lm_loss, shape=(-1, shape_list(sentence_order_loss)[0]))
+            masked_lm_loss = tf.reduce_mean(input_tensor=masked_lm_loss, axis=0)
+
+            return masked_lm_loss + sentence_order_loss
+
+        # Clip negative labels to zero here to avoid NaNs and errors - those positions will get masked later anyway
+        unmasked_lm_losses = loss_fn(y_true=tf.nn.relu(labels["labels"]), y_pred=logits[0])
+        # make sure only labels that are not equal to -100
+        # are taken into account for the loss computation
+        lm_loss_mask = tf.cast(labels["labels"] != -100, dtype=unmasked_lm_losses.dtype)
+        masked_lm_losses = unmasked_lm_losses * lm_loss_mask
+        reduced_masked_lm_loss = tf.reduce_sum(masked_lm_losses) / tf.reduce_sum(lm_loss_mask)
+
+        sop_logits = tf.reshape(logits[1], (-1, 2))
+        # Clip negative labels to zero here to avoid NaNs and errors - those positions will get masked later anyway
+        unmasked_sop_loss = loss_fn(y_true=tf.nn.relu(labels["sentence_order_label"]), y_pred=sop_logits)
+        sop_loss_mask = tf.cast(labels["sentence_order_label"] != -100, dtype=unmasked_sop_loss.dtype)
+
+        masked_sop_loss = unmasked_sop_loss * sop_loss_mask
+        reduced_masked_sop_loss = tf.reduce_sum(masked_sop_loss) / tf.reduce_sum(sop_loss_mask)
+
+        return tf.reshape(reduced_masked_lm_loss + reduced_masked_sop_loss, (1,))
+
+
+class TFAlbertEmbeddings(keras.layers.Layer):
+    """Construct the embeddings from word, position and token_type embeddings."""
+
+    def __init__(self, config: AlbertConfig, **kwargs):
+        super().__init__(**kwargs)
+
+        self.config = config
+        self.embedding_size = config.embedding_size
+        self.max_position_embeddings = config.max_position_embeddings
+        self.initializer_range = config.initializer_range
+        self.LayerNorm = keras.layers.LayerNormalization(epsilon=config.layer_norm_eps, name="LayerNorm")
+        self.dropout = keras.layers.Dropout(rate=config.hidden_dropout_prob)
+
+    def build(self, input_shape=None):
+        with tf.name_scope("word_embeddings"):
+            self.weight = self.add_weight(
+                name="weight",
+                shape=[self.config.vocab_size, self.embedding_size],
+                initializer=get_initializer(self.initializer_range),
+            )
+
+        with tf.name_scope("token_type_embeddings"):
+            self.token_type_embeddings = self.add_weight(
+                name="embeddings",
+                shape=[self.config.type_vocab_size, self.embedding_size],
+                initializer=get_initializer(self.initializer_range),
+            )
+
+        with tf.name_scope("position_embeddings"):
+            self.position_embeddings = self.add_weight(
+                name="embeddings",
+                shape=[self.max_position_embeddings, self.embedding_size],
+                initializer=get_initializer(self.initializer_range),
+            )
+
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "LayerNorm", None) is not None:
+            with tf.name_scope(self.LayerNorm.name):
+                self.LayerNorm.build([None, None, self.config.embedding_size])
+
+    # Copied from transformers.models.bert.modeling_tf_bert.TFBertEmbeddings.call
+    def call(
+        self,
+        input_ids: Optional[tf.Tensor] = None,
+        position_ids: Optional[tf.Tensor] = None,
+        token_type_ids: Optional[tf.Tensor] = None,
+        inputs_embeds: Optional[tf.Tensor] = None,
+        past_key_values_length=0,
+        training: bool = False,
+    ) -> tf.Tensor:
+        """
+        Applies embedding based on inputs tensor.
+
+        Returns:
+            final_embeddings (`tf.Tensor`): output embedding tensor.
+        """
+        if input_ids is None and inputs_embeds is None:
+            raise ValueError("Need to provide either `input_ids` or `input_embeds`.")
+
+        if input_ids is not None:
+            check_embeddings_within_bounds(input_ids, self.config.vocab_size)
+            inputs_embeds = tf.gather(params=self.weight, indices=input_ids)
+
+        input_shape = shape_list(inputs_embeds)[:-1]
+
+        if token_type_ids is None:
+            token_type_ids = tf.fill(dims=input_shape, value=0)
+
+        if position_ids is None:
+            position_ids = tf.expand_dims(
+                tf.range(start=past_key_values_length, limit=input_shape[1] + past_key_values_length), axis=0
+            )
+
+        position_embeds = tf.gather(params=self.position_embeddings, indices=position_ids)
+        token_type_embeds = tf.gather(params=self.token_type_embeddings, indices=token_type_ids)
+        final_embeddings = inputs_embeds + position_embeds + token_type_embeds
+        final_embeddings = self.LayerNorm(inputs=final_embeddings)
+        final_embeddings = self.dropout(inputs=final_embeddings, training=training)
+
+        return final_embeddings
+
+
+class TFAlbertAttention(keras.layers.Layer):
+    """Contains the complete attention sublayer, including both dropouts and layer norm."""
+
+    def __init__(self, config: AlbertConfig, **kwargs):
+        super().__init__(**kwargs)
+
+        if config.hidden_size % config.num_attention_heads != 0:
+            raise ValueError(
+                f"The hidden size ({config.hidden_size}) is not a multiple of the number "
+                f"of attention heads ({config.num_attention_heads})"
+            )
+
+        self.num_attention_heads = config.num_attention_heads
+        self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
+        self.all_head_size = self.num_attention_heads * self.attention_head_size
+        self.sqrt_att_head_size = math.sqrt(self.attention_head_size)
+        self.output_attentions = config.output_attentions
+
+        self.query = keras.layers.Dense(
+            units=self.all_head_size, kernel_initializer=get_initializer(config.initializer_range), name="query"
+        )
+        self.key = keras.layers.Dense(
+            units=self.all_head_size, kernel_initializer=get_initializer(config.initializer_range), name="key"
+        )
+        self.value = keras.layers.Dense(
+            units=self.all_head_size, kernel_initializer=get_initializer(config.initializer_range), name="value"
+        )
+        self.dense = keras.layers.Dense(
+            units=config.hidden_size, kernel_initializer=get_initializer(config.initializer_range), name="dense"
+        )
+        self.LayerNorm = keras.layers.LayerNormalization(epsilon=config.layer_norm_eps, name="LayerNorm")
+        # Two different dropout probabilities; see https://github.com/google-research/albert/blob/master/modeling.py#L971-L993
+        self.attention_dropout = keras.layers.Dropout(rate=config.attention_probs_dropout_prob)
+        self.output_dropout = keras.layers.Dropout(rate=config.hidden_dropout_prob)
+        self.config = config
+
+    def transpose_for_scores(self, tensor: tf.Tensor, batch_size: int) -> tf.Tensor:
+        # Reshape from [batch_size, seq_length, all_head_size] to [batch_size, seq_length, num_attention_heads, attention_head_size]
+        tensor = tf.reshape(tensor=tensor, shape=(batch_size, -1, self.num_attention_heads, self.attention_head_size))
+
+        # Transpose the tensor from [batch_size, seq_length, num_attention_heads, attention_head_size] to [batch_size, num_attention_heads, seq_length, attention_head_size]
+        return tf.transpose(tensor, perm=[0, 2, 1, 3])
+
+    def call(
+        self,
+        input_tensor: tf.Tensor,
+        attention_mask: tf.Tensor,
+        head_mask: tf.Tensor,
+        output_attentions: bool,
+        training: bool = False,
+    ) -> Tuple[tf.Tensor]:
+        batch_size = shape_list(input_tensor)[0]
+        mixed_query_layer = self.query(inputs=input_tensor)
+        mixed_key_layer = self.key(inputs=input_tensor)
+        mixed_value_layer = self.value(inputs=input_tensor)
+        query_layer = self.transpose_for_scores(mixed_query_layer, batch_size)
+        key_layer = self.transpose_for_scores(mixed_key_layer, batch_size)
+        value_layer = self.transpose_for_scores(mixed_value_layer, batch_size)
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        # (batch size, num_heads, seq_len_q, seq_len_k)
+        attention_scores = tf.matmul(query_layer, key_layer, transpose_b=True)
+        dk = tf.cast(self.sqrt_att_head_size, dtype=attention_scores.dtype)
+        attention_scores = tf.divide(attention_scores, dk)
+
+        if attention_mask is not None:
+            # Apply the attention mask is (precomputed for all layers in TFAlbertModel call() function)
+            attention_scores = tf.add(attention_scores, attention_mask)
+
+        # Normalize the attention scores to probabilities.
+        attention_probs = stable_softmax(logits=attention_scores, axis=-1)
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        attention_probs = self.attention_dropout(inputs=attention_probs, training=training)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attention_probs = tf.multiply(attention_probs, head_mask)
+
+        context_layer = tf.matmul(attention_probs, value_layer)
+        context_layer = tf.transpose(context_layer, perm=[0, 2, 1, 3])
+
+        # (batch_size, seq_len_q, all_head_size)
+        context_layer = tf.reshape(tensor=context_layer, shape=(batch_size, -1, self.all_head_size))
+        self_outputs = (context_layer, attention_probs) if output_attentions else (context_layer,)
+        hidden_states = self_outputs[0]
+        hidden_states = self.dense(inputs=hidden_states)
+        hidden_states = self.output_dropout(inputs=hidden_states, training=training)
+        attention_output = self.LayerNorm(inputs=hidden_states + input_tensor)
+
+        # add attentions if we output them
+        outputs = (attention_output,) + self_outputs[1:]
+
+        return outputs
+
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "query", None) is not None:
+            with tf.name_scope(self.query.name):
+                self.query.build([None, None, self.config.hidden_size])
+        if getattr(self, "key", None) is not None:
+            with tf.name_scope(self.key.name):
+                self.key.build([None, None, self.config.hidden_size])
+        if getattr(self, "value", None) is not None:
+            with tf.name_scope(self.value.name):
+                self.value.build([None, None, self.config.hidden_size])
+        if getattr(self, "dense", None) is not None:
+            with tf.name_scope(self.dense.name):
+                self.dense.build([None, None, self.config.hidden_size])
+        if getattr(self, "LayerNorm", None) is not None:
+            with tf.name_scope(self.LayerNorm.name):
+                self.LayerNorm.build([None, None, self.config.hidden_size])
+
+
+class TFAlbertLayer(keras.layers.Layer):
+    def __init__(self, config: AlbertConfig, **kwargs):
+        super().__init__(**kwargs)
+
+        self.attention = TFAlbertAttention(config, name="attention")
+        self.ffn = keras.layers.Dense(
+            units=config.intermediate_size, kernel_initializer=get_initializer(config.initializer_range), name="ffn"
+        )
+
+        if isinstance(config.hidden_act, str):
+            self.activation = get_tf_activation(config.hidden_act)
+        else:
+            self.activation = config.hidden_act
+
+        self.ffn_output = keras.layers.Dense(
+            units=config.hidden_size, kernel_initializer=get_initializer(config.initializer_range), name="ffn_output"
+        )
+        self.full_layer_layer_norm = keras.layers.LayerNormalization(
+            epsilon=config.layer_norm_eps, name="full_layer_layer_norm"
+        )
+        self.dropout = keras.layers.Dropout(rate=config.hidden_dropout_prob)
+        self.config = config
+
+    def call(
+        self,
+        hidden_states: tf.Tensor,
+        attention_mask: tf.Tensor,
+        head_mask: tf.Tensor,
+        output_attentions: bool,
+        training: bool = False,
+    ) -> Tuple[tf.Tensor]:
+        attention_outputs = self.attention(
+            input_tensor=hidden_states,
+            attention_mask=attention_mask,
+            head_mask=head_mask,
+            output_attentions=output_attentions,
+            training=training,
+        )
+        ffn_output = self.ffn(inputs=attention_outputs[0])
+        ffn_output = self.activation(ffn_output)
+        ffn_output = self.ffn_output(inputs=ffn_output)
+        ffn_output = self.dropout(inputs=ffn_output, training=training)
+        hidden_states = self.full_layer_layer_norm(inputs=ffn_output + attention_outputs[0])
+
+        # add attentions if we output them
+        outputs = (hidden_states,) + attention_outputs[1:]
+
+        return outputs
+
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "attention", None) is not None:
+            with tf.name_scope(self.attention.name):
+                self.attention.build(None)
+        if getattr(self, "ffn", None) is not None:
+            with tf.name_scope(self.ffn.name):
+                self.ffn.build([None, None, self.config.hidden_size])
+        if getattr(self, "ffn_output", None) is not None:
+            with tf.name_scope(self.ffn_output.name):
+                self.ffn_output.build([None, None, self.config.intermediate_size])
+        if getattr(self, "full_layer_layer_norm", None) is not None:
+            with tf.name_scope(self.full_layer_layer_norm.name):
+                self.full_layer_layer_norm.build([None, None, self.config.hidden_size])
+
+
+class TFAlbertLayerGroup(keras.layers.Layer):
+    def __init__(self, config: AlbertConfig, **kwargs):
+        super().__init__(**kwargs)
+
+        self.albert_layers = [
+            TFAlbertLayer(config, name=f"albert_layers_._{i}") for i in range(config.inner_group_num)
+        ]
+
+    def call(
+        self,
+        hidden_states: tf.Tensor,
+        attention_mask: tf.Tensor,
+        head_mask: tf.Tensor,
+        output_attentions: bool,
+        output_hidden_states: bool,
+        training: bool = False,
+    ) -> Union[TFBaseModelOutput, Tuple[tf.Tensor]]:
+        layer_hidden_states = () if output_hidden_states else None
+        layer_attentions = () if output_attentions else None
+
+        for layer_index, albert_layer in enumerate(self.albert_layers):
+            if output_hidden_states:
+                layer_hidden_states = layer_hidden_states + (hidden_states,)
+
+            layer_output = albert_layer(
+                hidden_states=hidden_states,
+                attention_mask=attention_mask,
+                head_mask=head_mask[layer_index],
+                output_attentions=output_attentions,
+                training=training,
+            )
+            hidden_states = layer_output[0]
+
+            if output_attentions:
+                layer_attentions = layer_attentions + (layer_output[1],)
+
+        # Add last layer
+        if output_hidden_states:
+            layer_hidden_states = layer_hidden_states + (hidden_states,)
+
+        return tuple(v for v in [hidden_states, layer_hidden_states, layer_attentions] if v is not None)
+
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "albert_layers", None) is not None:
+            for layer in self.albert_layers:
+                with tf.name_scope(layer.name):
+                    layer.build(None)
+
+
+class TFAlbertTransformer(keras.layers.Layer):
+    def __init__(self, config: AlbertConfig, **kwargs):
+        super().__init__(**kwargs)
+
+        self.num_hidden_layers = config.num_hidden_layers
+        self.num_hidden_groups = config.num_hidden_groups
+        # Number of layers in a hidden group
+        self.layers_per_group = int(config.num_hidden_layers / config.num_hidden_groups)
+        self.embedding_hidden_mapping_in = keras.layers.Dense(
+            units=config.hidden_size,
+            kernel_initializer=get_initializer(config.initializer_range),
+            name="embedding_hidden_mapping_in",
+        )
+        self.albert_layer_groups = [
+            TFAlbertLayerGroup(config, name=f"albert_layer_groups_._{i}") for i in range(config.num_hidden_groups)
+        ]
+        self.config = config
+
+    def call(
+        self,
+        hidden_states: tf.Tensor,
+        attention_mask: tf.Tensor,
+        head_mask: tf.Tensor,
+        output_attentions: bool,
+        output_hidden_states: bool,
+        return_dict: bool,
+        training: bool = False,
+    ) -> Union[TFBaseModelOutput, Tuple[tf.Tensor]]:
+        hidden_states = self.embedding_hidden_mapping_in(inputs=hidden_states)
+        all_attentions = () if output_attentions else None
+        all_hidden_states = (hidden_states,) if output_hidden_states else None
+
+        for i in range(self.num_hidden_layers):
+            # Index of the hidden group
+            group_idx = int(i / (self.num_hidden_layers / self.num_hidden_groups))
+            layer_group_output = self.albert_layer_groups[group_idx](
+                hidden_states=hidden_states,
+                attention_mask=attention_mask,
+                head_mask=head_mask[group_idx * self.layers_per_group : (group_idx + 1) * self.layers_per_group],
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                training=training,
+            )
+            hidden_states = layer_group_output[0]
+
+            if output_attentions:
+                all_attentions = all_attentions + layer_group_output[-1]
+
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, all_hidden_states, all_attentions] if v is not None)
+
+        return TFBaseModelOutput(
+            last_hidden_state=hidden_states, hidden_states=all_hidden_states, attentions=all_attentions
+        )
+
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "embedding_hidden_mapping_in", None) is not None:
+            with tf.name_scope(self.embedding_hidden_mapping_in.name):
+                self.embedding_hidden_mapping_in.build([None, None, self.config.embedding_size])
+        if getattr(self, "albert_layer_groups", None) is not None:
+            for layer in self.albert_layer_groups:
+                with tf.name_scope(layer.name):
+                    layer.build(None)
+
+
+class TFAlbertPreTrainedModel(TFPreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = AlbertConfig
+    base_model_prefix = "albert"
+
+
+class TFAlbertMLMHead(keras.layers.Layer):
+    def __init__(self, config: AlbertConfig, input_embeddings: keras.layers.Layer, **kwargs):
+        super().__init__(**kwargs)
+
+        self.config = config
+        self.embedding_size = config.embedding_size
+        self.dense = keras.layers.Dense(
+            config.embedding_size, kernel_initializer=get_initializer(config.initializer_range), name="dense"
+        )
+        if isinstance(config.hidden_act, str):
+            self.activation = get_tf_activation(config.hidden_act)
+        else:
+            self.activation = config.hidden_act
+
+        self.LayerNorm = keras.layers.LayerNormalization(epsilon=config.layer_norm_eps, name="LayerNorm")
+
+        # The output weights are the same as the input embeddings, but there is
+        # an output-only bias for each token.
+        self.decoder = input_embeddings
+
+    def build(self, input_shape=None):
+        self.bias = self.add_weight(shape=(self.config.vocab_size,), initializer="zeros", trainable=True, name="bias")
+        self.decoder_bias = self.add_weight(
+            shape=(self.config.vocab_size,), initializer="zeros", trainable=True, name="decoder/bias"
+        )
+
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "dense", None) is not None:
+            with tf.name_scope(self.dense.name):
+                self.dense.build([None, None, self.config.hidden_size])
+        if getattr(self, "LayerNorm", None) is not None:
+            with tf.name_scope(self.LayerNorm.name):
+                self.LayerNorm.build([None, None, self.config.embedding_size])
+
+    def get_output_embeddings(self) -> keras.layers.Layer:
+        return self.decoder
+
+    def set_output_embeddings(self, value: tf.Variable):
+        self.decoder.weight = value
+        self.decoder.vocab_size = shape_list(value)[0]
+
+    def get_bias(self) -> Dict[str, tf.Variable]:
+        return {"bias": self.bias, "decoder_bias": self.decoder_bias}
+
+    def set_bias(self, value: tf.Variable):
+        self.bias = value["bias"]
+        self.decoder_bias = value["decoder_bias"]
+        self.config.vocab_size = shape_list(value["bias"])[0]
+
+    def call(self, hidden_states: tf.Tensor) -> tf.Tensor:
+        hidden_states = self.dense(inputs=hidden_states)
+        hidden_states = self.activation(hidden_states)
+        hidden_states = self.LayerNorm(inputs=hidden_states)
+        seq_length = shape_list(tensor=hidden_states)[1]
+        hidden_states = tf.reshape(tensor=hidden_states, shape=[-1, self.embedding_size])
+        hidden_states = tf.matmul(a=hidden_states, b=self.decoder.weight, transpose_b=True)
+        hidden_states = tf.reshape(tensor=hidden_states, shape=[-1, seq_length, self.config.vocab_size])
+        hidden_states = tf.nn.bias_add(value=hidden_states, bias=self.decoder_bias)
+
+        return hidden_states
+
+
+@keras_serializable
+class TFAlbertMainLayer(keras.layers.Layer):
+    config_class = AlbertConfig
+
+    def __init__(self, config: AlbertConfig, add_pooling_layer: bool = True, **kwargs):
+        super().__init__(**kwargs)
+
+        self.config = config
+
+        self.embeddings = TFAlbertEmbeddings(config, name="embeddings")
+        self.encoder = TFAlbertTransformer(config, name="encoder")
+        self.pooler = (
+            keras.layers.Dense(
+                units=config.hidden_size,
+                kernel_initializer=get_initializer(config.initializer_range),
+                activation="tanh",
+                name="pooler",
+            )
+            if add_pooling_layer
+            else None
+        )
+
+    def get_input_embeddings(self) -> keras.layers.Layer:
+        return self.embeddings
+
+    def set_input_embeddings(self, value: tf.Variable):
+        self.embeddings.weight = value
+        self.embeddings.vocab_size = shape_list(value)[0]
+
+    def _prune_heads(self, heads_to_prune):
+        """
+        Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
+        class PreTrainedModel
+        """
+        raise NotImplementedError
+
+    @unpack_inputs
+    def call(
+        self,
+        input_ids: TFModelInputType | None = None,
+        attention_mask: np.ndarray | tf.Tensor | None = None,
+        token_type_ids: np.ndarray | tf.Tensor | None = None,
+        position_ids: np.ndarray | tf.Tensor | None = None,
+        head_mask: np.ndarray | tf.Tensor | None = None,
+        inputs_embeds: np.ndarray | tf.Tensor | None = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        training: bool = False,
+    ) -> Union[TFBaseModelOutputWithPooling, Tuple[tf.Tensor]]:
+        if input_ids is 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 attention_mask is None:
+            attention_mask = tf.fill(dims=input_shape, value=1)
+
+        if token_type_ids is None:
+            token_type_ids = tf.fill(dims=input_shape, value=0)
+
+        embedding_output = self.embeddings(
+            input_ids=input_ids,
+            position_ids=position_ids,
+            token_type_ids=token_type_ids,
+            inputs_embeds=inputs_embeds,
+            training=training,
+        )
+
+        # We create a 3D attention mask from a 2D tensor mask.
+        # Sizes are [batch_size, 1, 1, to_seq_length]
+        # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
+        # this attention mask is more simple than the triangular masking of causal attention
+        # used in OpenAI GPT, we just need to prepare the broadcast dimension here.
+        extended_attention_mask = tf.reshape(attention_mask, (input_shape[0], 1, 1, input_shape[1]))
+
+        # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
+        # masked positions, this operation will create a tensor which is 0.0 for
+        # positions we want to attend and -10000.0 for masked positions.
+        # Since we are adding it to the raw scores before the softmax, this is
+        # effectively the same as removing these entirely.
+        extended_attention_mask = tf.cast(extended_attention_mask, dtype=embedding_output.dtype)
+        one_cst = tf.constant(1.0, dtype=embedding_output.dtype)
+        ten_thousand_cst = tf.constant(-10000.0, dtype=embedding_output.dtype)
+        extended_attention_mask = tf.multiply(tf.subtract(one_cst, extended_attention_mask), ten_thousand_cst)
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+        if head_mask is not None:
+            raise NotImplementedError
+        else:
+            head_mask = [None] * self.config.num_hidden_layers
+
+        encoder_outputs = self.encoder(
+            hidden_states=embedding_output,
+            attention_mask=extended_attention_mask,
+            head_mask=head_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            training=training,
+        )
+
+        sequence_output = encoder_outputs[0]
+        pooled_output = self.pooler(inputs=sequence_output[:, 0]) if self.pooler is not None else None
+
+        if not return_dict:
+            return (
+                sequence_output,
+                pooled_output,
+            ) + encoder_outputs[1:]
+
+        return TFBaseModelOutputWithPooling(
+            last_hidden_state=sequence_output,
+            pooler_output=pooled_output,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )
+
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "embeddings", None) is not None:
+            with tf.name_scope(self.embeddings.name):
+                self.embeddings.build(None)
+        if getattr(self, "encoder", None) is not None:
+            with tf.name_scope(self.encoder.name):
+                self.encoder.build(None)
+        if getattr(self, "pooler", None) is not None:
+            with tf.name_scope(self.pooler.name):
+                self.pooler.build([None, None, self.config.hidden_size])
+
+
+@dataclass
+class TFAlbertForPreTrainingOutput(ModelOutput):
+    """
+    Output type of [`TFAlbertForPreTraining`].
+
+    Args:
+        prediction_logits (`tf.Tensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        sop_logits (`tf.Tensor` of shape `(batch_size, 2)`):
+            Prediction scores of the next sequence prediction (classification) head (scores of True/False continuation
+            before SoftMax).
+        hidden_states (`tuple(tf.Tensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `tf.Tensor` (one for the output of the embeddings + one for the output of each layer) of shape
+            `(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (`tuple(tf.Tensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
+            Tuple of `tf.Tensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
+            sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    loss: Optional[tf.Tensor] = None
+    prediction_logits: Optional[tf.Tensor] = None
+    sop_logits: Optional[tf.Tensor] = None
+    hidden_states: Tuple[tf.Tensor] | None = None
+    attentions: Tuple[tf.Tensor] | None = None
+
+
+ALBERT_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 ([`AlbertConfig`]): 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.
+"""
+
+ALBERT_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`Numpy array` or `tf.Tensor` of shape `({0})`):
+            Indices of input sequence tokens in the vocabulary.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.__call__`] and
+            [`PreTrainedTokenizer.encode`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`Numpy array` or `tf.Tensor` of shape `({0})`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+        token_type_ids (`Numpy array` or `tf.Tensor` 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 (`Numpy array` or `tf.Tensor` of shape `({0})`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.max_position_embeddings - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        head_mask (`Numpy array` or `tf.Tensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
+            Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        inputs_embeds (`tf.Tensor` of shape `({0}, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
+            model's internal embedding lookup matrix.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail. 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).
+"""
+
+
+@add_start_docstrings(
+    "The bare Albert Model transformer outputting raw hidden-states without any specific head on top.",
+    ALBERT_START_DOCSTRING,
+)
+class TFAlbertModel(TFAlbertPreTrainedModel):
+    def __init__(self, config: AlbertConfig, *inputs, **kwargs):
+        super().__init__(config, *inputs, **kwargs)
+
+        self.albert = TFAlbertMainLayer(config, name="albert")
+
+    @unpack_inputs
+    @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=TFBaseModelOutputWithPooling,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def call(
+        self,
+        input_ids: TFModelInputType | None = None,
+        attention_mask: np.ndarray | tf.Tensor | None = None,
+        token_type_ids: np.ndarray | tf.Tensor | None = None,
+        position_ids: np.ndarray | tf.Tensor | None = None,
+        head_mask: np.ndarray | tf.Tensor | None = None,
+        inputs_embeds: np.ndarray | tf.Tensor | None = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        training: Optional[bool] = False,
+    ) -> Union[TFBaseModelOutputWithPooling, Tuple[tf.Tensor]]:
+        outputs = self.albert(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            training=training,
+        )
+
+        return outputs
+
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "albert", None) is not None:
+            with tf.name_scope(self.albert.name):
+                self.albert.build(None)
+
+
+@add_start_docstrings(
+    """
+    Albert Model with two heads on top for pretraining: a `masked language modeling` head and a `sentence order
+    prediction` (classification) head.
+    """,
+    ALBERT_START_DOCSTRING,
+)
+class TFAlbertForPreTraining(TFAlbertPreTrainedModel, TFAlbertPreTrainingLoss):
+    # names with a '.' represents the authorized unexpected/missing layers when a TF model is loaded from a PT model
+    _keys_to_ignore_on_load_unexpected = [r"predictions.decoder.weight"]
+
+    def __init__(self, config: AlbertConfig, *inputs, **kwargs):
+        super().__init__(config, *inputs, **kwargs)
+
+        self.num_labels = config.num_labels
+
+        self.albert = TFAlbertMainLayer(config, name="albert")
+        self.predictions = TFAlbertMLMHead(config, input_embeddings=self.albert.embeddings, name="predictions")
+        self.sop_classifier = TFAlbertSOPHead(config, name="sop_classifier")
+
+    def get_lm_head(self) -> keras.layers.Layer:
+        return self.predictions
+
+    @unpack_inputs
+    @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @replace_return_docstrings(output_type=TFAlbertForPreTrainingOutput, config_class=_CONFIG_FOR_DOC)
+    def call(
+        self,
+        input_ids: TFModelInputType | None = None,
+        attention_mask: np.ndarray | tf.Tensor | None = None,
+        token_type_ids: np.ndarray | tf.Tensor | None = None,
+        position_ids: np.ndarray | tf.Tensor | None = None,
+        head_mask: np.ndarray | tf.Tensor | None = None,
+        inputs_embeds: np.ndarray | tf.Tensor | None = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        labels: np.ndarray | tf.Tensor | None = None,
+        sentence_order_label: np.ndarray | tf.Tensor | None = None,
+        training: Optional[bool] = False,
+    ) -> Union[TFAlbertForPreTrainingOutput, Tuple[tf.Tensor]]:
+        r"""
+        Return:
+
+        Example:
+
+        ```python
+        >>> import tensorflow as tf
+        >>> from transformers import AutoTokenizer, TFAlbertForPreTraining
+
+        >>> tokenizer = AutoTokenizer.from_pretrained("albert/albert-base-v2")
+        >>> model = TFAlbertForPreTraining.from_pretrained("albert/albert-base-v2")
+
+        >>> input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]
+        >>> # Batch size 1
+        >>> outputs = model(input_ids)
+
+        >>> prediction_logits = outputs.prediction_logits
+        >>> sop_logits = outputs.sop_logits
+        ```"""
+
+        outputs = self.albert(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            training=training,
+        )
+        sequence_output, pooled_output = outputs[:2]
+        prediction_scores = self.predictions(hidden_states=sequence_output)
+        sop_scores = self.sop_classifier(pooled_output=pooled_output, training=training)
+        total_loss = None
+
+        if labels is not None and sentence_order_label is not None:
+            d_labels = {"labels": labels}
+            d_labels["sentence_order_label"] = sentence_order_label
+            total_loss = self.hf_compute_loss(labels=d_labels, logits=(prediction_scores, sop_scores))
+
+        if not return_dict:
+            output = (prediction_scores, sop_scores) + outputs[2:]
+            return ((total_loss,) + output) if total_loss is not None else output
+
+        return TFAlbertForPreTrainingOutput(
+            loss=total_loss,
+            prediction_logits=prediction_scores,
+            sop_logits=sop_scores,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "albert", None) is not None:
+            with tf.name_scope(self.albert.name):
+                self.albert.build(None)
+        if getattr(self, "predictions", None) is not None:
+            with tf.name_scope(self.predictions.name):
+                self.predictions.build(None)
+        if getattr(self, "sop_classifier", None) is not None:
+            with tf.name_scope(self.sop_classifier.name):
+                self.sop_classifier.build(None)
+
+
+class TFAlbertSOPHead(keras.layers.Layer):
+    def __init__(self, config: AlbertConfig, **kwargs):
+        super().__init__(**kwargs)
+
+        self.dropout = keras.layers.Dropout(rate=config.classifier_dropout_prob)
+        self.classifier = keras.layers.Dense(
+            units=config.num_labels,
+            kernel_initializer=get_initializer(config.initializer_range),
+            name="classifier",
+        )
+        self.config = config
+
+    def call(self, pooled_output: tf.Tensor, training: bool) -> tf.Tensor:
+        dropout_pooled_output = self.dropout(inputs=pooled_output, training=training)
+        logits = self.classifier(inputs=dropout_pooled_output)
+
+        return logits
+
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "classifier", None) is not None:
+            with tf.name_scope(self.classifier.name):
+                self.classifier.build([None, None, self.config.hidden_size])
+
+
+@add_start_docstrings("""Albert Model with a `language modeling` head on top.""", ALBERT_START_DOCSTRING)
+class TFAlbertForMaskedLM(TFAlbertPreTrainedModel, TFMaskedLanguageModelingLoss):
+    # names with a '.' represents the authorized unexpected/missing layers when a TF model is loaded from a PT model
+    _keys_to_ignore_on_load_unexpected = [r"pooler", r"predictions.decoder.weight"]
+
+    def __init__(self, config: AlbertConfig, *inputs, **kwargs):
+        super().__init__(config, *inputs, **kwargs)
+
+        self.albert = TFAlbertMainLayer(config, add_pooling_layer=False, name="albert")
+        self.predictions = TFAlbertMLMHead(config, input_embeddings=self.albert.embeddings, name="predictions")
+
+    def get_lm_head(self) -> keras.layers.Layer:
+        return self.predictions
+
+    @unpack_inputs
+    @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @replace_return_docstrings(output_type=TFMaskedLMOutput, config_class=_CONFIG_FOR_DOC)
+    def call(
+        self,
+        input_ids: TFModelInputType | None = None,
+        attention_mask: np.ndarray | tf.Tensor | None = None,
+        token_type_ids: np.ndarray | tf.Tensor | None = None,
+        position_ids: np.ndarray | tf.Tensor | None = None,
+        head_mask: np.ndarray | tf.Tensor | None = None,
+        inputs_embeds: np.ndarray | tf.Tensor | None = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        labels: np.ndarray | tf.Tensor | None = None,
+        training: Optional[bool] = False,
+    ) -> Union[TFMaskedLMOutput, Tuple[tf.Tensor]]:
+        r"""
+        labels (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should be in `[-100, 0, ...,
+            config.vocab_size]` (see `input_ids` docstring) Tokens with indices set to `-100` are ignored (masked), the
+            loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> import tensorflow as tf
+        >>> from transformers import AutoTokenizer, TFAlbertForMaskedLM
+
+        >>> tokenizer = AutoTokenizer.from_pretrained("albert/albert-base-v2")
+        >>> model = TFAlbertForMaskedLM.from_pretrained("albert/albert-base-v2")
+
+        >>> # add mask_token
+        >>> inputs = tokenizer(f"The capital of [MASK] is Paris.", return_tensors="tf")
+        >>> logits = model(**inputs).logits
+
+        >>> # retrieve index of [MASK]
+        >>> mask_token_index = tf.where(inputs.input_ids == tokenizer.mask_token_id)[0][1]
+        >>> predicted_token_id = tf.math.argmax(logits[0, mask_token_index], axis=-1)
+        >>> tokenizer.decode(predicted_token_id)
+        'france'
+        ```
+
+        ```python
+        >>> labels = tokenizer("The capital of France is Paris.", return_tensors="tf")["input_ids"]
+        >>> labels = tf.where(inputs.input_ids == tokenizer.mask_token_id, labels, -100)
+        >>> outputs = model(**inputs, labels=labels)
+        >>> round(float(outputs.loss), 2)
+        0.81
+        ```
+        """
+        outputs = self.albert(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            training=training,
+        )
+        sequence_output = outputs[0]
+        prediction_scores = self.predictions(hidden_states=sequence_output, training=training)
+        loss = None if labels is None else self.hf_compute_loss(labels=labels, logits=prediction_scores)
+
+        if not return_dict:
+            output = (prediction_scores,) + outputs[2:]
+
+            return ((loss,) + output) if loss is not None else output
+
+        return TFMaskedLMOutput(
+            loss=loss,
+            logits=prediction_scores,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "albert", None) is not None:
+            with tf.name_scope(self.albert.name):
+                self.albert.build(None)
+        if getattr(self, "predictions", None) is not None:
+            with tf.name_scope(self.predictions.name):
+                self.predictions.build(None)
+
+
+@add_start_docstrings(
+    """
+    Albert Model transformer with a sequence classification/regression head on top (a linear layer on top of the pooled
+    output) e.g. for GLUE tasks.
+    """,
+    ALBERT_START_DOCSTRING,
+)
+class TFAlbertForSequenceClassification(TFAlbertPreTrainedModel, TFSequenceClassificationLoss):
+    # names with a '.' represents the authorized unexpected/missing layers when a TF model is loaded from a PT model
+    _keys_to_ignore_on_load_unexpected = [r"predictions"]
+    _keys_to_ignore_on_load_missing = [r"dropout"]
+
+    def __init__(self, config: AlbertConfig, *inputs, **kwargs):
+        super().__init__(config, *inputs, **kwargs)
+
+        self.num_labels = config.num_labels
+
+        self.albert = TFAlbertMainLayer(config, name="albert")
+        self.dropout = keras.layers.Dropout(rate=config.classifier_dropout_prob)
+        self.classifier = keras.layers.Dense(
+            units=config.num_labels, kernel_initializer=get_initializer(config.initializer_range), name="classifier"
+        )
+        self.config = config
+
+    @unpack_inputs
+    @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        checkpoint="vumichien/albert-base-v2-imdb",
+        output_type=TFSequenceClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+        expected_output="'LABEL_1'",
+        expected_loss=0.12,
+    )
+    def call(
+        self,
+        input_ids: TFModelInputType | None = None,
+        attention_mask: np.ndarray | tf.Tensor | None = None,
+        token_type_ids: np.ndarray | tf.Tensor | None = None,
+        position_ids: np.ndarray | tf.Tensor | None = None,
+        head_mask: np.ndarray | tf.Tensor | None = None,
+        inputs_embeds: np.ndarray | tf.Tensor | None = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        labels: np.ndarray | tf.Tensor | None = None,
+        training: Optional[bool] = False,
+    ) -> Union[TFSequenceClassifierOutput, Tuple[tf.Tensor]]:
+        r"""
+        labels (`tf.Tensor` 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).
+        """
+        outputs = self.albert(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            training=training,
+        )
+        pooled_output = outputs[1]
+        pooled_output = self.dropout(inputs=pooled_output, training=training)
+        logits = self.classifier(inputs=pooled_output)
+        loss = None if labels is None else self.hf_compute_loss(labels=labels, logits=logits)
+
+        if not return_dict:
+            output = (logits,) + outputs[2:]
+
+            return ((loss,) + output) if loss is not None else output
+
+        return TFSequenceClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "albert", None) is not None:
+            with tf.name_scope(self.albert.name):
+                self.albert.build(None)
+        if getattr(self, "classifier", None) is not None:
+            with tf.name_scope(self.classifier.name):
+                self.classifier.build([None, None, self.config.hidden_size])
+
+
+@add_start_docstrings(
+    """
+    Albert Model with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for
+    Named-Entity-Recognition (NER) tasks.
+    """,
+    ALBERT_START_DOCSTRING,
+)
+class TFAlbertForTokenClassification(TFAlbertPreTrainedModel, TFTokenClassificationLoss):
+    # names with a '.' represents the authorized unexpected/missing layers when a TF model is loaded from a PT model
+    _keys_to_ignore_on_load_unexpected = [r"pooler", r"predictions"]
+    _keys_to_ignore_on_load_missing = [r"dropout"]
+
+    def __init__(self, config: AlbertConfig, *inputs, **kwargs):
+        super().__init__(config, *inputs, **kwargs)
+
+        self.num_labels = config.num_labels
+
+        self.albert = TFAlbertMainLayer(config, add_pooling_layer=False, name="albert")
+        classifier_dropout_prob = (
+            config.classifier_dropout_prob
+            if config.classifier_dropout_prob is not None
+            else config.hidden_dropout_prob
+        )
+        self.dropout = keras.layers.Dropout(rate=classifier_dropout_prob)
+        self.classifier = keras.layers.Dense(
+            units=config.num_labels, kernel_initializer=get_initializer(config.initializer_range), name="classifier"
+        )
+        self.config = config
+
+    @unpack_inputs
+    @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=TFTokenClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def call(
+        self,
+        input_ids: TFModelInputType | None = None,
+        attention_mask: np.ndarray | tf.Tensor | None = None,
+        token_type_ids: np.ndarray | tf.Tensor | None = None,
+        position_ids: np.ndarray | tf.Tensor | None = None,
+        head_mask: np.ndarray | tf.Tensor | None = None,
+        inputs_embeds: np.ndarray | tf.Tensor | None = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        labels: np.ndarray | tf.Tensor | None = None,
+        training: Optional[bool] = False,
+    ) -> Union[TFTokenClassifierOutput, Tuple[tf.Tensor]]:
+        r"""
+        labels (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the token classification loss. Indices should be in `[0, ..., config.num_labels - 1]`.
+        """
+        outputs = self.albert(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            training=training,
+        )
+        sequence_output = outputs[0]
+        sequence_output = self.dropout(inputs=sequence_output, training=training)
+        logits = self.classifier(inputs=sequence_output)
+        loss = None if labels is None else self.hf_compute_loss(labels=labels, logits=logits)
+
+        if not return_dict:
+            output = (logits,) + outputs[2:]
+
+            return ((loss,) + output) if loss is not None else output
+
+        return TFTokenClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "albert", None) is not None:
+            with tf.name_scope(self.albert.name):
+                self.albert.build(None)
+        if getattr(self, "classifier", None) is not None:
+            with tf.name_scope(self.classifier.name):
+                self.classifier.build([None, None, self.config.hidden_size])
+
+
+@add_start_docstrings(
+    """
+    Albert Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear
+    layer on top of the hidden-states output to compute `span start logits` and `span end logits`).
+    """,
+    ALBERT_START_DOCSTRING,
+)
+class TFAlbertForQuestionAnswering(TFAlbertPreTrainedModel, TFQuestionAnsweringLoss):
+    # names with a '.' represents the authorized unexpected/missing layers when a TF model is loaded from a PT model
+    _keys_to_ignore_on_load_unexpected = [r"pooler", r"predictions"]
+
+    def __init__(self, config: AlbertConfig, *inputs, **kwargs):
+        super().__init__(config, *inputs, **kwargs)
+
+        self.num_labels = config.num_labels
+
+        self.albert = TFAlbertMainLayer(config, add_pooling_layer=False, name="albert")
+        self.qa_outputs = keras.layers.Dense(
+            units=config.num_labels, kernel_initializer=get_initializer(config.initializer_range), name="qa_outputs"
+        )
+        self.config = config
+
+    @unpack_inputs
+    @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        checkpoint="vumichien/albert-base-v2-squad2",
+        output_type=TFQuestionAnsweringModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+        qa_target_start_index=12,
+        qa_target_end_index=13,
+        expected_output="'a nice puppet'",
+        expected_loss=7.36,
+    )
+    def call(
+        self,
+        input_ids: TFModelInputType | None = None,
+        attention_mask: np.ndarray | tf.Tensor | None = None,
+        token_type_ids: np.ndarray | tf.Tensor | None = None,
+        position_ids: np.ndarray | tf.Tensor | None = None,
+        head_mask: np.ndarray | tf.Tensor | None = None,
+        inputs_embeds: np.ndarray | tf.Tensor | None = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        start_positions: np.ndarray | tf.Tensor | None = None,
+        end_positions: np.ndarray | tf.Tensor | None = None,
+        training: Optional[bool] = False,
+    ) -> Union[TFQuestionAnsweringModelOutput, Tuple[tf.Tensor]]:
+        r"""
+        start_positions (`tf.Tensor` of shape `(batch_size,)`, *optional*):
+            Labels for position (index) of the start of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
+            are not taken into account for computing the loss.
+        end_positions (`tf.Tensor` of shape `(batch_size,)`, *optional*):
+            Labels for position (index) of the end of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
+            are not taken into account for computing the loss.
+        """
+        outputs = self.albert(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            training=training,
+        )
+        sequence_output = outputs[0]
+        logits = self.qa_outputs(inputs=sequence_output)
+        start_logits, end_logits = tf.split(value=logits, num_or_size_splits=2, axis=-1)
+        start_logits = tf.squeeze(input=start_logits, axis=-1)
+        end_logits = tf.squeeze(input=end_logits, axis=-1)
+        loss = None
+
+        if start_positions is not None and end_positions is not None:
+            labels = {"start_position": start_positions}
+            labels["end_position"] = end_positions
+            loss = self.hf_compute_loss(labels=labels, logits=(start_logits, end_logits))
+
+        if not return_dict:
+            output = (start_logits, end_logits) + outputs[2:]
+
+            return ((loss,) + output) if loss is not None else output
+
+        return TFQuestionAnsweringModelOutput(
+            loss=loss,
+            start_logits=start_logits,
+            end_logits=end_logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "albert", None) is not None:
+            with tf.name_scope(self.albert.name):
+                self.albert.build(None)
+        if getattr(self, "qa_outputs", None) is not None:
+            with tf.name_scope(self.qa_outputs.name):
+                self.qa_outputs.build([None, None, self.config.hidden_size])
+
+
+@add_start_docstrings(
+    """
+    Albert Model with a multiple choice classification head on top (a linear layer on top of the pooled output and a
+    softmax) e.g. for RocStories/SWAG tasks.
+    """,
+    ALBERT_START_DOCSTRING,
+)
+class TFAlbertForMultipleChoice(TFAlbertPreTrainedModel, TFMultipleChoiceLoss):
+    # names with a '.' represents the authorized unexpected/missing layers when a TF model is loaded from a PT model
+    _keys_to_ignore_on_load_unexpected = [r"pooler", r"predictions"]
+    _keys_to_ignore_on_load_missing = [r"dropout"]
+
+    def __init__(self, config: AlbertConfig, *inputs, **kwargs):
+        super().__init__(config, *inputs, **kwargs)
+
+        self.albert = TFAlbertMainLayer(config, name="albert")
+        self.dropout = keras.layers.Dropout(rate=config.hidden_dropout_prob)
+        self.classifier = keras.layers.Dense(
+            units=1, kernel_initializer=get_initializer(config.initializer_range), name="classifier"
+        )
+        self.config = config
+
+    @unpack_inputs
+    @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length"))
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=TFMultipleChoiceModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def call(
+        self,
+        input_ids: TFModelInputType | None = None,
+        attention_mask: np.ndarray | tf.Tensor | None = None,
+        token_type_ids: np.ndarray | tf.Tensor | None = None,
+        position_ids: np.ndarray | tf.Tensor | None = None,
+        head_mask: np.ndarray | tf.Tensor | None = None,
+        inputs_embeds: np.ndarray | tf.Tensor | None = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        labels: np.ndarray | tf.Tensor | None = None,
+        training: Optional[bool] = False,
+    ) -> Union[TFMultipleChoiceModelOutput, Tuple[tf.Tensor]]:
+        r"""
+        labels (`tf.Tensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the multiple choice classification loss. Indices should be in `[0, ..., num_choices]`
+            where `num_choices` is the size of the second dimension of the input tensors. (See `input_ids` above)
+        """
+
+        if input_ids is not None:
+            num_choices = shape_list(input_ids)[1]
+            seq_length = shape_list(input_ids)[2]
+        else:
+            num_choices = shape_list(inputs_embeds)[1]
+            seq_length = shape_list(inputs_embeds)[2]
+
+        flat_input_ids = tf.reshape(input_ids, (-1, seq_length)) if input_ids is not None else None
+        flat_attention_mask = (
+            tf.reshape(tensor=attention_mask, shape=(-1, seq_length)) if attention_mask is not None else None
+        )
+        flat_token_type_ids = (
+            tf.reshape(tensor=token_type_ids, shape=(-1, seq_length)) if token_type_ids is not None else None
+        )
+        flat_position_ids = (
+            tf.reshape(tensor=position_ids, shape=(-1, seq_length)) if position_ids is not None else None
+        )
+        flat_inputs_embeds = (
+            tf.reshape(tensor=inputs_embeds, shape=(-1, seq_length, shape_list(inputs_embeds)[3]))
+            if inputs_embeds is not None
+            else None
+        )
+        outputs = self.albert(
+            input_ids=flat_input_ids,
+            attention_mask=flat_attention_mask,
+            token_type_ids=flat_token_type_ids,
+            position_ids=flat_position_ids,
+            head_mask=head_mask,
+            inputs_embeds=flat_inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            training=training,
+        )
+        pooled_output = outputs[1]
+        pooled_output = self.dropout(inputs=pooled_output, training=training)
+        logits = self.classifier(inputs=pooled_output)
+        reshaped_logits = tf.reshape(tensor=logits, shape=(-1, num_choices))
+        loss = None if labels is None else self.hf_compute_loss(labels=labels, logits=reshaped_logits)
+
+        if not return_dict:
+            output = (reshaped_logits,) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TFMultipleChoiceModelOutput(
+            loss=loss,
+            logits=reshaped_logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "albert", None) is not None:
+            with tf.name_scope(self.albert.name):
+                self.albert.build(None)
+        if getattr(self, "classifier", None) is not None:
+            with tf.name_scope(self.classifier.name):
+                self.classifier.build([None, None, self.config.hidden_size])
+
+
+__all__ = [
+    "TFAlbertPreTrainedModel",
+    "TFAlbertModel",
+    "TFAlbertForPreTraining",
+    "TFAlbertForMaskedLM",
+    "TFAlbertForSequenceClassification",
+    "TFAlbertForTokenClassification",
+    "TFAlbertForQuestionAnswering",
+    "TFAlbertForMultipleChoice",
+    "TFAlbertMainLayer",
+]

lib/python3.12/site-packages/transformers/models/albert/tokenization_albert.py

@@ -0,0 +1,350 @@
+# coding=utf-8
+# Copyright 2018 Google AI, Google Brain 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 ALBERT model."""
+
+import os
+import unicodedata
+from shutil import copyfile
+from typing import Any, Dict, List, Optional, Tuple
+
+import sentencepiece as spm
+
+from ...tokenization_utils import AddedToken, PreTrainedTokenizer
+from ...utils import logging
+from ...utils.import_utils import export
+
+
+logger = logging.get_logger(__name__)
+VOCAB_FILES_NAMES = {"vocab_file": "spiece.model"}
+
+
+SPIECE_UNDERLINE = "▁"
+
+
+@export(backends=("sentencepiece",))
+class AlbertTokenizer(PreTrainedTokenizer):
+    """
+    Construct an ALBERT tokenizer. Based on [SentencePiece](https://github.com/google/sentencepiece).
+
+    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`):
+            [SentencePiece](https://github.com/google/sentencepiece) file (generally has a *.spm* extension) that
+            contains the vocabulary necessary to instantiate a tokenizer.
+        do_lower_case (`bool`, *optional*, defaults to `True`):
+            Whether or not to lowercase the input when tokenizing.
+        remove_space (`bool`, *optional*, defaults to `True`):
+            Whether or not to strip the text when tokenizing (removing excess spaces before and after the string).
+        keep_accents (`bool`, *optional*, defaults to `False`):
+            Whether or not to keep accents when tokenizing.
+        bos_token (`str`, *optional*, defaults to `"[CLS]"`):
+            The beginning of sequence token that was used during pretraining. Can be used a sequence classifier token.
+
+            <Tip>
+
+            When building a sequence using special tokens, this is not the token that is used for the beginning of
+            sequence. The token used is the `cls_token`.
+
+            </Tip>
+
+        eos_token (`str`, *optional*, defaults to `"[SEP]"`):
+            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.
+        sep_token (`str`, *optional*, defaults to `"[SEP]"`):
+            The separator token, which is used when building a sequence from multiple sequences, e.g. two sequences for
+            sequence classification or for a text and a question for question answering. It is also used as the last
+            token of a sequence built with special tokens.
+        pad_token (`str`, *optional*, defaults to `"<pad>"`):
+            The token used for padding, for example when batching sequences of different lengths.
+        cls_token (`str`, *optional*, defaults to `"[CLS]"`):
+            The classifier token which is used when doing sequence classification (classification of the whole sequence
+            instead of per-token classification). It is the first token of the sequence when built with special tokens.
+        mask_token (`str`, *optional*, defaults to `"[MASK]"`):
+            The token used for masking values. This is the token used when training this model with masked language
+            modeling. This is the token which the model will try to predict.
+        sp_model_kwargs (`dict`, *optional*):
+            Will be passed to the `SentencePieceProcessor.__init__()` method. The [Python wrapper for
+            SentencePiece](https://github.com/google/sentencepiece/tree/master/python) can be used, among other things,
+            to set:
+
+            - `enable_sampling`: Enable subword regularization.
+            - `nbest_size`: Sampling parameters for unigram. Invalid for BPE-Dropout.
+
+              - `nbest_size = {0,1}`: No sampling is performed.
+              - `nbest_size > 1`: samples from the nbest_size results.
+              - `nbest_size < 0`: assuming that nbest_size is infinite and samples from the all hypothesis (lattice)
+                using forward-filtering-and-backward-sampling algorithm.
+
+            - `alpha`: Smoothing parameter for unigram sampling, and dropout probability of merge operations for
+              BPE-dropout.
+
+    Attributes:
+        sp_model (`SentencePieceProcessor`):
+            The *SentencePiece* processor that is used for every conversion (string, tokens and IDs).
+    """
+
+    vocab_files_names = VOCAB_FILES_NAMES
+
+    def __init__(
+        self,
+        vocab_file,
+        do_lower_case=True,
+        remove_space=True,
+        keep_accents=False,
+        bos_token="[CLS]",
+        eos_token="[SEP]",
+        unk_token="<unk>",
+        sep_token="[SEP]",
+        pad_token="<pad>",
+        cls_token="[CLS]",
+        mask_token="[MASK]",
+        sp_model_kwargs: Optional[Dict[str, Any]] = None,
+        **kwargs,
+    ) -> None:
+        # Mask token behave like a normal word, i.e. include the space before it and
+        # is included in the raw text, there should be a match in a non-normalized sentence.
+        mask_token = (
+            AddedToken(mask_token, lstrip=True, rstrip=False, normalized=False)
+            if isinstance(mask_token, str)
+            else mask_token
+        )
+
+        self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs
+
+        self.do_lower_case = do_lower_case
+        self.remove_space = remove_space
+        self.keep_accents = keep_accents
+        self.vocab_file = vocab_file
+
+        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
+        self.sp_model.Load(vocab_file)
+
+        super().__init__(
+            do_lower_case=do_lower_case,
+            remove_space=remove_space,
+            keep_accents=keep_accents,
+            bos_token=bos_token,
+            eos_token=eos_token,
+            unk_token=unk_token,
+            sep_token=sep_token,
+            pad_token=pad_token,
+            cls_token=cls_token,
+            mask_token=mask_token,
+            sp_model_kwargs=self.sp_model_kwargs,
+            **kwargs,
+        )
+
+    @property
+    def vocab_size(self) -> int:
+        return len(self.sp_model)
+
+    def get_vocab(self) -> Dict[str, int]:
+        vocab = {self.convert_ids_to_tokens(i): i for i in range(self.vocab_size)}
+        vocab.update(self.added_tokens_encoder)
+        return vocab
+
+    def __getstate__(self):
+        state = self.__dict__.copy()
+        state["sp_model"] = None
+        return state
+
+    def __setstate__(self, d):
+        self.__dict__ = d
+
+        # for backward compatibility
+        if not hasattr(self, "sp_model_kwargs"):
+            self.sp_model_kwargs = {}
+
+        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
+        self.sp_model.Load(self.vocab_file)
+
+    def preprocess_text(self, inputs):
+        if self.remove_space:
+            outputs = " ".join(inputs.strip().split())
+        else:
+            outputs = inputs
+        outputs = outputs.replace("``", '"').replace("''", '"')
+
+        if not self.keep_accents:
+            outputs = unicodedata.normalize("NFKD", outputs)
+            outputs = "".join([c for c in outputs if not unicodedata.combining(c)])
+        if self.do_lower_case:
+            outputs = outputs.lower()
+
+        return outputs
+
+    def _tokenize(self, text: str) -> List[str]:
+        """Tokenize a string."""
+        text = self.preprocess_text(text)
+        pieces = self.sp_model.encode(text, out_type=str)
+        new_pieces = []
+        for piece in pieces:
+            if len(piece) > 1 and piece[-1] == str(",") and piece[-2].isdigit():
+                # Logic to handle special cases see https://github.com/google-research/bert/blob/master/README.md#tokenization
+                # `9,9` -> ['▁9', ',', '9'] instead of [`_9,`, '9']
+                cur_pieces = self.sp_model.EncodeAsPieces(piece[:-1].replace(SPIECE_UNDERLINE, ""))
+                if piece[0] != SPIECE_UNDERLINE and cur_pieces[0][0] == SPIECE_UNDERLINE:
+                    if len(cur_pieces[0]) == 1:
+                        cur_pieces = cur_pieces[1:]
+                    else:
+                        cur_pieces[0] = cur_pieces[0][1:]
+                cur_pieces.append(piece[-1])
+                new_pieces.extend(cur_pieces)
+            else:
+                new_pieces.append(piece)
+
+        return new_pieces
+
+    def _convert_token_to_id(self, token):
+        """Converts a token (str) in an id using the vocab."""
+        return self.sp_model.PieceToId(token)
+
+    def _convert_id_to_token(self, index):
+        """Converts an index (integer) in a token (str) using the vocab."""
+        return self.sp_model.IdToPiece(index)
+
+    def convert_tokens_to_string(self, tokens):
+        """Converts a sequence of tokens (string) in a single string."""
+        current_sub_tokens = []
+        out_string = ""
+        prev_is_special = False
+        for token in tokens:
+            # make sure that special tokens are not decoded using sentencepiece model
+            if token in self.all_special_tokens:
+                if not prev_is_special:
+                    out_string += " "
+                out_string += self.sp_model.decode(current_sub_tokens) + token
+                prev_is_special = True
+                current_sub_tokens = []
+            else:
+                current_sub_tokens.append(token)
+                prev_is_special = False
+        out_string += self.sp_model.decode(current_sub_tokens)
+        return out_string.strip()
+
+    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. An ALBERT sequence has the following format:
+
+        - single sequence: `[CLS] X [SEP]`
+        - pair of sequences: `[CLS] A [SEP] B [SEP]`
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs to which the special tokens will be added.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            `List[int]`: List of [input IDs](../glossary#input-ids) with the appropriate special tokens.
+        """
+        sep = [self.sep_token_id]
+        cls = [self.cls_token_id]
+        if token_ids_1 is None:
+            return cls + token_ids_0 + sep
+        return cls + token_ids_0 + sep + token_ids_1 + sep
+
+    def get_special_tokens_mask(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None, already_has_special_tokens: bool = False
+    ) -> List[int]:
+        """
+        Retrieve sequence ids from a token list that has no special tokens added. This method is called when adding
+        special tokens using the tokenizer `prepare_for_model` method.
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+            already_has_special_tokens (`bool`, *optional*, defaults to `False`):
+                Whether or not the token list is already formatted with special tokens for the model.
+
+        Returns:
+            `List[int]`: A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
+        """
+
+        if already_has_special_tokens:
+            return super().get_special_tokens_mask(
+                token_ids_0=token_ids_0, token_ids_1=token_ids_1, already_has_special_tokens=True
+            )
+
+        if token_ids_1 is not None:
+            return [1] + ([0] * len(token_ids_0)) + [1] + ([0] * len(token_ids_1)) + [1]
+        return [1] + ([0] * len(token_ids_0)) + [1]
+
+    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. An ALBERT
+        sequence pair mask has the following format:
+
+        ```
+        0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1
+        | first sequence    | second sequence |
+        ```
+
+        If `token_ids_1` is `None`, this method only returns the first portion of the mask (0s).
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            `List[int]`: List of [token type IDs](../glossary#token-type-ids) according to the given sequence(s).
+        """
+        sep = [self.sep_token_id]
+        cls = [self.cls_token_id]
+
+        if token_ids_1 is None:
+            return len(cls + token_ids_0 + sep) * [0]
+        return len(cls + token_ids_0 + sep) * [0] + len(token_ids_1 + sep) * [1]
+
+    def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
+        if not os.path.isdir(save_directory):
+            logger.error(f"Vocabulary path ({save_directory}) should be a directory")
+            return
+        out_vocab_file = os.path.join(
+            save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]
+        )
+
+        if os.path.abspath(self.vocab_file) != os.path.abspath(out_vocab_file) and os.path.isfile(self.vocab_file):
+            copyfile(self.vocab_file, out_vocab_file)
+        elif not os.path.isfile(self.vocab_file):
+            with open(out_vocab_file, "wb") as fi:
+                content_spiece_model = self.sp_model.serialized_model_proto()
+                fi.write(content_spiece_model)
+
+        return (out_vocab_file,)
+
+
+__all__ = ["AlbertTokenizer"]

lib/python3.12/site-packages/transformers/models/albert/tokenization_albert_fast.py

@@ -0,0 +1,212 @@
+# coding=utf-8
+# Copyright 2018 Google AI, Google Brain 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 ALBERT model."""
+
+import os
+from shutil import copyfile
+from typing import List, Optional, Tuple
+
+from ...tokenization_utils import AddedToken
+from ...tokenization_utils_fast import PreTrainedTokenizerFast
+from ...utils import is_sentencepiece_available, logging
+
+
+if is_sentencepiece_available():
+    from .tokenization_albert import AlbertTokenizer
+else:
+    AlbertTokenizer = None
+
+logger = logging.get_logger(__name__)
+VOCAB_FILES_NAMES = {"vocab_file": "spiece.model", "tokenizer_file": "tokenizer.json"}
+
+
+SPIECE_UNDERLINE = "▁"
+
+
+class AlbertTokenizerFast(PreTrainedTokenizerFast):
+    """
+    Construct a "fast" ALBERT tokenizer (backed by HuggingFace's *tokenizers* library). Based on
+    [Unigram](https://huggingface.co/docs/tokenizers/python/latest/components.html?highlight=unigram#models). 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`):
+            [SentencePiece](https://github.com/google/sentencepiece) file (generally has a *.spm* extension) that
+            contains the vocabulary necessary to instantiate a tokenizer.
+        do_lower_case (`bool`, *optional*, defaults to `True`):
+            Whether or not to lowercase the input when tokenizing.
+        remove_space (`bool`, *optional*, defaults to `True`):
+            Whether or not to strip the text when tokenizing (removing excess spaces before and after the string).
+        keep_accents (`bool`, *optional*, defaults to `False`):
+            Whether or not to keep accents when tokenizing.
+        bos_token (`str`, *optional*, defaults to `"[CLS]"`):
+            The beginning of sequence token that was used during pretraining. Can be used a sequence classifier token.
+
+            <Tip>
+
+            When building a sequence using special tokens, this is not the token that is used for the beginning of
+            sequence. The token used is the `cls_token`.
+
+            </Tip>
+
+        eos_token (`str`, *optional*, defaults to `"[SEP]"`):
+            The end of sequence token. .. note:: 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`.
+        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.
+        sep_token (`str`, *optional*, defaults to `"[SEP]"`):
+            The separator token, which is used when building a sequence from multiple sequences, e.g. two sequences for
+            sequence classification or for a text and a question for question answering. It is also used as the last
+            token of a sequence built with special tokens.
+        pad_token (`str`, *optional*, defaults to `"<pad>"`):
+            The token used for padding, for example when batching sequences of different lengths.
+        cls_token (`str`, *optional*, defaults to `"[CLS]"`):
+            The classifier token which is used when doing sequence classification (classification of the whole sequence
+            instead of per-token classification). It is the first token of the sequence when built with special tokens.
+        mask_token (`str`, *optional*, defaults to `"[MASK]"`):
+            The token used for masking values. This is the token used when training this model with masked language
+            modeling. This is the token which the model will try to predict.
+    """
+
+    vocab_files_names = VOCAB_FILES_NAMES
+    slow_tokenizer_class = AlbertTokenizer
+
+    def __init__(
+        self,
+        vocab_file=None,
+        tokenizer_file=None,
+        do_lower_case=True,
+        remove_space=True,
+        keep_accents=False,
+        bos_token="[CLS]",
+        eos_token="[SEP]",
+        unk_token="<unk>",
+        sep_token="[SEP]",
+        pad_token="<pad>",
+        cls_token="[CLS]",
+        mask_token="[MASK]",
+        **kwargs,
+    ):
+        # Mask token behave like a normal word, i.e. include the space before it and
+        # is included in the raw text, there should be a match in a non-normalized sentence.
+        mask_token = (
+            AddedToken(mask_token, lstrip=True, rstrip=False, normalized=False)
+            if isinstance(mask_token, str)
+            else mask_token
+        )
+
+        super().__init__(
+            vocab_file,
+            tokenizer_file=tokenizer_file,
+            do_lower_case=do_lower_case,
+            remove_space=remove_space,
+            keep_accents=keep_accents,
+            bos_token=bos_token,
+            eos_token=eos_token,
+            unk_token=unk_token,
+            sep_token=sep_token,
+            pad_token=pad_token,
+            cls_token=cls_token,
+            mask_token=mask_token,
+            **kwargs,
+        )
+
+        self.do_lower_case = do_lower_case
+        self.remove_space = remove_space
+        self.keep_accents = keep_accents
+        self.vocab_file = vocab_file
+
+    @property
+    def can_save_slow_tokenizer(self) -> bool:
+        return os.path.isfile(self.vocab_file) if self.vocab_file else False
+
+    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. An ALBERT sequence has the following format:
+
+        - single sequence: `[CLS] X [SEP]`
+        - pair of sequences: `[CLS] A [SEP] B [SEP]`
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs to which the special tokens will be added
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            `List[int]`: list of [input IDs](../glossary#input-ids) with the appropriate special tokens.
+        """
+        sep = [self.sep_token_id]
+        cls = [self.cls_token_id]
+        if token_ids_1 is None:
+            return cls + token_ids_0 + sep
+        return cls + token_ids_0 + sep + token_ids_1 + sep
+
+    def create_token_type_ids_from_sequences(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        Creates a mask from the two sequences passed to be used in a sequence-pair classification task. An ALBERT
+        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, only returns the first portion of the mask (0s).
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of ids.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            `List[int]`: List of [token type IDs](../glossary#token-type-ids) according to the given sequence(s).
+        """
+        sep = [self.sep_token_id]
+        cls = [self.cls_token_id]
+
+        if token_ids_1 is None:
+            return len(cls + token_ids_0 + sep) * [0]
+        return len(cls + token_ids_0 + sep) * [0] + len(token_ids_1 + sep) * [1]
+
+    def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
+        if not self.can_save_slow_tokenizer:
+            raise ValueError(
+                "Your fast tokenizer does not have the necessary information to save the vocabulary for a slow "
+                "tokenizer."
+            )
+
+        if not os.path.isdir(save_directory):
+            logger.error(f"Vocabulary path ({save_directory}) should be a directory")
+            return
+        out_vocab_file = os.path.join(
+            save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]
+        )
+
+        if os.path.abspath(self.vocab_file) != os.path.abspath(out_vocab_file):
+            copyfile(self.vocab_file, out_vocab_file)
+
+        return (out_vocab_file,)
+
+
+__all__ = ["AlbertTokenizerFast"]

lib/python3.12/site-packages/transformers/models/bertweet/__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_bertweet 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/bertweet/tokenization_bertweet.py

@@ -0,0 +1,769 @@
+# coding=utf-8
+# Copyright (c) 2020, VinAI Research and the HuggingFace Inc. team.
+# Copyright 2018 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 BERTweet"""
+
+import html
+import os
+import re
+from shutil import copyfile
+from typing import List, Optional, Tuple
+
+import regex
+
+from ...tokenization_utils import PreTrainedTokenizer
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+VOCAB_FILES_NAMES = {
+    "vocab_file": "vocab.txt",
+    "merges_file": "bpe.codes",
+}
+
+
+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 BertweetTokenizer(PreTrainedTokenizer):
+    """
+    Constructs a BERTweet tokenizer, using Byte-Pair-Encoding.
+
+    This tokenizer inherits from [`PreTrainedTokenizer`] which contains most of the main methods. Users should refer to
+    this superclass for more information regarding those methods.
+
+    Args:
+        vocab_file (`str`):
+            Path to the vocabulary file.
+        merges_file (`str`):
+            Path to the merges file.
+        normalization (`bool`, *optional*, defaults to `False`):
+            Whether or not to apply a normalization preprocess.
+        bos_token (`str`, *optional*, defaults to `"<s>"`):
+            The beginning of sequence token that was used during pretraining. Can be used a sequence classifier token.
+
+            <Tip>
+
+            When building a sequence using special tokens, this is not the token that is used for the beginning of
+            sequence. The token used is the `cls_token`.
+
+            </Tip>
+
+        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>
+
+        sep_token (`str`, *optional*, defaults to `"</s>"`):
+            The separator token, which is used when building a sequence from multiple sequences, e.g. two sequences for
+            sequence classification or for a text and a question for question answering. It is also used as the last
+            token of a sequence built with special tokens.
+        cls_token (`str`, *optional*, defaults to `"<s>"`):
+            The classifier token which is used when doing sequence classification (classification of the whole sequence
+            instead of per-token classification). It is the first token of the sequence when built with special tokens.
+        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.
+        mask_token (`str`, *optional*, defaults to `"<mask>"`):
+            The token used for masking values. This is the token used when training this model with masked language
+            modeling. This is the token which the model will try to predict.
+    """
+
+    vocab_files_names = VOCAB_FILES_NAMES
+
+    def __init__(
+        self,
+        vocab_file,
+        merges_file,
+        normalization=False,
+        bos_token="<s>",
+        eos_token="</s>",
+        sep_token="</s>",
+        cls_token="<s>",
+        unk_token="<unk>",
+        pad_token="<pad>",
+        mask_token="<mask>",
+        **kwargs,
+    ):
+        try:
+            from emoji import demojize
+
+            self.demojizer = demojize
+        except ImportError:
+            logger.warning(
+                "emoji is not installed, thus not converting emoticons or emojis into text. Install emoji: pip3"
+                " install emoji==0.6.0"
+            )
+            self.demojizer = None
+
+        self.vocab_file = vocab_file
+        self.merges_file = merges_file
+
+        self.encoder = {}
+        self.encoder[str(bos_token)] = 0
+        self.encoder[str(pad_token)] = 1
+        self.encoder[str(eos_token)] = 2
+        self.encoder[str(unk_token)] = 3
+
+        self.add_from_file(vocab_file)
+
+        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]
+        merges = [tuple(merge.split()[:-1]) for merge in merges]
+        self.bpe_ranks = dict(zip(merges, range(len(merges))))
+        self.cache = {}
+
+        self.normalization = normalization
+        self.tweetPreprocessor = TweetTokenizer()
+        self.special_puncts = {"’": "'", "…": "..."}
+
+        super().__init__(
+            normalization=normalization,
+            bos_token=bos_token,
+            eos_token=eos_token,
+            sep_token=sep_token,
+            cls_token=cls_token,
+            unk_token=unk_token,
+            pad_token=pad_token,
+            mask_token=mask_token,
+            **kwargs,
+        )
+
+    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 BERTweet sequence has the following format:
+
+        - single sequence: `<s> X </s>`
+        - pair of sequences: `<s> A </s></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.
+        """
+
+        if token_ids_1 is None:
+            return [self.cls_token_id] + token_ids_0 + [self.sep_token_id]
+        cls = [self.cls_token_id]
+        sep = [self.sep_token_id]
+        return cls + token_ids_0 + sep + sep + token_ids_1 + sep
+
+    def get_special_tokens_mask(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None, already_has_special_tokens: bool = False
+    ) -> List[int]:
+        """
+        Retrieve sequence ids from a token list that has no special tokens added. This method is called when adding
+        special tokens using the tokenizer `prepare_for_model` method.
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+            already_has_special_tokens (`bool`, *optional*, defaults to `False`):
+                Whether or not the token list is already formatted with special tokens for the model.
+
+        Returns:
+            `List[int]`: A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
+        """
+
+        if already_has_special_tokens:
+            return super().get_special_tokens_mask(
+                token_ids_0=token_ids_0, token_ids_1=token_ids_1, already_has_special_tokens=True
+            )
+
+        if token_ids_1 is None:
+            return [1] + ([0] * len(token_ids_0)) + [1]
+        return [1] + ([0] * len(token_ids_0)) + [1, 1] + ([0] * len(token_ids_1)) + [1]
+
+    def create_token_type_ids_from_sequences(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        Create a mask from the two sequences passed to be used in a sequence-pair classification task. BERTweet 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]
+
+    @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)
+        word = tuple(list(word[:-1]) + [word[-1] + "</w>"])
+        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)
+        word = word[:-4]
+        self.cache[token] = word
+        return word
+
+    def _tokenize(self, text):
+        """Tokenize a string."""
+        if self.normalization:  # Perform Tweet normalization before performing BPE
+            text = self.normalizeTweet(text)
+
+        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 normalizeTweet(self, tweet):
+        """
+        Normalize a raw Tweet
+        """
+        for punct in self.special_puncts:
+            tweet = tweet.replace(punct, self.special_puncts[punct])
+
+        tokens = self.tweetPreprocessor.tokenize(tweet)
+        normTweet = " ".join([self.normalizeToken(token) for token in tokens])
+
+        normTweet = (
+            normTweet.replace("cannot ", "can not ")
+            .replace("n't ", " n't ")
+            .replace("n 't ", " n't ")
+            .replace("ca n't", "can't")
+            .replace("ai n't", "ain't")
+        )
+        normTweet = (
+            normTweet.replace("'m ", " 'm ")
+            .replace("'re ", " 're ")
+            .replace("'s ", " 's ")
+            .replace("'ll ", " 'll ")
+            .replace("'d ", " 'd ")
+            .replace("'ve ", " 've ")
+        )
+        normTweet = (
+            normTweet.replace(" p . m .", "  p.m.")
+            .replace(" p . m ", " p.m ")
+            .replace(" a . m .", " a.m.")
+            .replace(" a . m ", " a.m ")
+        )
+
+        return " ".join(normTweet.split())
+
+    def normalizeToken(self, token):
+        """
+        Normalize tokens in a Tweet
+        """
+        lowercased_token = token.lower()
+        if token.startswith("@"):
+            return "@USER"
+        elif lowercased_token.startswith("http") or lowercased_token.startswith("www"):
+            return "HTTPURL"
+        elif len(token) == 1:
+            if token in self.special_puncts:
+                return self.special_puncts[token]
+            if self.demojizer is not None:
+                return self.demojizer(token)
+            else:
+                return token
+        else:
+            return token
+
+    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, self.unk_token)
+
+    def convert_tokens_to_string(self, tokens):
+        """Converts a sequence of tokens (string) 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
+        out_vocab_file = os.path.join(
+            save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]
+        )
+        out_merge_file = os.path.join(
+            save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["merges_file"]
+        )
+
+        if os.path.abspath(self.vocab_file) != os.path.abspath(out_vocab_file) and os.path.isfile(self.vocab_file):
+            copyfile(self.vocab_file, out_vocab_file)
+        elif not os.path.isfile(self.vocab_file):
+            with open(out_vocab_file, "wb") as fi:
+                content_spiece_model = self.sp_model.serialized_model_proto()
+                fi.write(content_spiece_model)
+
+        if os.path.abspath(self.merges_file) != os.path.abspath(out_merge_file):
+            copyfile(self.merges_file, out_merge_file)
+
+        return out_vocab_file, out_merge_file
+
+    # def decode(self, token_ids, skip_special_tokens=False, clean_up_tokenization_spaces=True):
+    #     filtered_tokens = ' '.join(self.convert_ids_to_tokens(token_ids, skip_special_tokens=skip_special_tokens))
+    #     tokens_generated_so_far = re.sub('(@@ )', '', string=filtered_tokens)
+    #     tokens_generated_so_far = re.sub('(@@ ?$)', '', string=tokens_generated_so_far)
+    #     return ''.join(tokens_generated_so_far)
+
+    def add_from_file(self, f):
+        """
+        Loads a pre-existing dictionary from a text file and adds its symbols to this instance.
+        """
+        if isinstance(f, str):
+            try:
+                with open(f, "r", encoding="utf-8") as fd:
+                    self.add_from_file(fd)
+            except FileNotFoundError as fnfe:
+                raise fnfe
+            except UnicodeError:
+                raise Exception(f"Incorrect encoding detected in {f}, please rebuild the dataset")
+            return
+
+        lines = f.readlines()
+        for lineTmp in lines:
+            line = lineTmp.strip()
+            idx = line.rfind(" ")
+            if idx == -1:
+                raise ValueError("Incorrect dictionary format, expected '<token> <cnt>'")
+            word = line[:idx]
+            self.encoder[word] = len(self.encoder)
+
+
+# Natural Language Toolkit: Twitter Tokenizer
+#
+# Copyright (C) 2001-2020 NLTK Project
+# Author: Christopher Potts <cgpotts@stanford.edu>
+#         Ewan Klein <ewan@inf.ed.ac.uk> (modifications)
+#         Pierpaolo Pantone <> (modifications)
+# URL: http://nltk.org/
+# For license information, see LICENSE.TXT
+#
+
+
+"""
+Twitter-aware tokenizer, designed to be flexible and easy to adapt to new domains and tasks. The basic logic is this:
+
+1. The tuple regex_strings defines a list of regular expression strings.
+
+2. The regex_strings strings are put, in order, into a compiled regular expression object called word_re.
+
+3. The tokenization is done by word_re.findall(s), where s is the user-supplied string, inside the tokenize() method of
+   the class Tokenizer.
+
+4. When instantiating Tokenizer objects, there is a single option: preserve_case. By default, it is set to True. If it
+   is set to False, then the tokenizer will lowercase everything except for emoticons.
+
+"""
+
+
+######################################################################
+#
+# import regex  # https://github.com/nltk/nltk/issues/2409
+# import html
+#
+######################################################################
+# The following strings are components in the regular expression
+# that is used for tokenizing. It's important that phone_number
+# appears first in the final regex (since it can contain whitespace).
+# It also could matter that tags comes after emoticons, due to the
+# possibility of having text like
+#
+#     <:| and some text >:)
+#
+# Most importantly, the final element should always be last, since it
+# does a last ditch whitespace-based tokenization of whatever is left.
+
+# ToDo: Update with http://en.wikipedia.org/wiki/List_of_emoticons ?
+
+# This particular element is used in a couple ways, so we define it
+# with a name:
+# docstyle-ignore
+EMOTICONS = r"""
+    (?:
+      [<>]?
+      [:;=8]                     # eyes
+      [\-o\*\']?                 # optional nose
+      [\)\]\(\[dDpP/\:\}\{@\|\\] # mouth
+      |
+      [\)\]\(\[dDpP/\:\}\{@\|\\] # mouth
+      [\-o\*\']?                 # optional nose
+      [:;=8]                     # eyes
+      [<>]?
+      |
+      <3                         # heart
+    )"""
+
+# URL pattern due to John Gruber, modified by Tom Winzig. See
+# https://gist.github.com/winzig/8894715
+# docstyle-ignore
+URLS = r"""			# Capture 1: entire matched URL
+  (?:
+  https?:				# URL protocol and colon
+    (?:
+      /{1,3}				# 1-3 slashes
+      |					#   or
+      [a-z0-9%]				# Single letter or digit or '%'
+                                       # (Trying not to match e.g. "URI::Escape")
+    )
+    |					#   or
+                                       # looks like domain name followed by a slash:
+    [a-z0-9.\-]+[.]
+    (?:[a-z]{2,13})
+    /
+  )
+  (?:					# One or more:
+    [^\s()<>{}\[\]]+			# Run of non-space, non-()<>{}[]
+    |					#   or
+    \([^\s()]*?\([^\s()]+\)[^\s()]*?\) # balanced parens, one level deep: (...(...)...)
+    |
+    \([^\s]+?\)				# balanced parens, non-recursive: (...)
+  )+
+  (?:					# End with:
+    \([^\s()]*?\([^\s()]+\)[^\s()]*?\) # balanced parens, one level deep: (...(...)...)
+    |
+    \([^\s]+?\)				# balanced parens, non-recursive: (...)
+    |					#   or
+    [^\s`!()\[\]{};:'".,<>?«»“”‘’]	# not a space or one of these punct chars
+  )
+  |					# OR, the following to match naked domains:
+  (?:
+    (?<!@)			        # not preceded by a @, avoid matching foo@_gmail.com_
+    [a-z0-9]+
+    (?:[.\-][a-z0-9]+)*
+    [.]
+    (?:[a-z]{2,13})
+    \b
+    /?
+    (?!@)			        # not succeeded by a @,
+                            # avoid matching "foo.na" in "foo.na@example.com"
+  )
+"""
+
+# docstyle-ignore
+# The components of the tokenizer:
+REGEXPS = (
+    URLS,
+    # Phone numbers:
+    r"""
+    (?:
+      (?:            # (international)
+        \+?[01]
+        [ *\-.\)]*
+      )?
+      (?:            # (area code)
+        [\(]?
+        \d{3}
+        [ *\-.\)]*
+      )?
+      \d{3}          # exchange
+      [ *\-.\)]*
+      \d{4}          # base
+    )""",
+    # ASCII Emoticons
+    EMOTICONS,
+    # HTML tags:
+    r"""<[^>\s]+>""",
+    # ASCII Arrows
+    r"""[\-]+>|<[\-]+""",
+    # Twitter username:
+    r"""(?:@[\w_]+)""",
+    # Twitter hashtags:
+    r"""(?:\#+[\w_]+[\w\'_\-]*[\w_]+)""",
+    # email addresses
+    r"""[\w.+-]+@[\w-]+\.(?:[\w-]\.?)+[\w-]""",
+    # docstyle-ignore
+    # Remaining word types:
+    r"""
+    (?:[^\W\d_](?:[^\W\d_]|['\-_])+[^\W\d_]) # Words with apostrophes or dashes.
+    |
+    (?:[+\-]?\d+[,/.:-]\d+[+\-]?)  # Numbers, including fractions, decimals.
+    |
+    (?:[\w_]+)                     # Words without apostrophes or dashes.
+    |
+    (?:\.(?:\s*\.){1,})            # Ellipsis dots.
+    |
+    (?:\S)                         # Everything else that isn't whitespace.
+    """,
+)
+
+######################################################################
+# This is the core tokenizing regex:
+
+WORD_RE = regex.compile(r"""(%s)""" % "|".join(REGEXPS), regex.VERBOSE | regex.I | regex.UNICODE)
+
+# WORD_RE performs poorly on these patterns:
+HANG_RE = regex.compile(r"([^a-zA-Z0-9])\1{3,}")
+
+# The emoticon string gets its own regex so that we can preserve case for
+# them as needed:
+EMOTICON_RE = regex.compile(EMOTICONS, regex.VERBOSE | regex.I | regex.UNICODE)
+
+# These are for regularizing HTML entities to Unicode:
+ENT_RE = regex.compile(r"&(#?(x?))([^&;\s]+);")
+
+
+######################################################################
+# Functions for converting html entities
+######################################################################
+
+
+def _str_to_unicode(text, encoding=None, errors="strict"):
+    if encoding is None:
+        encoding = "utf-8"
+    if isinstance(text, bytes):
+        return text.decode(encoding, errors)
+    return text
+
+
+def _replace_html_entities(text, keep=(), remove_illegal=True, encoding="utf-8"):
+    """
+    Remove entities from text by converting them to their corresponding unicode character.
+
+    Args:
+        text:
+            A unicode string or a byte string encoded in the given *encoding* (which defaults to 'utf-8').
+        keep (list):
+            List of entity names which should not be replaced. This supports both numeric entities (`&#nnnn;` and
+            `&#hhhh;`) and named entities (such as `&nbsp;` or `&gt;`).
+        remove_illegal (bool):
+            If `True`, entities that can't be converted are removed. Otherwise, entities that can't be converted are
+            kept "as is".
+
+    Returns: A unicode string with the entities removed.
+
+    See https://github.com/scrapy/w3lib/blob/master/w3lib/html.py
+
+    Examples:
+
+    ```python
+    >>> from nltk.tokenize.casual import _replace_html_entities
+
+    >>> _replace_html_entities(b"Price: &pound;100")
+    'Price: \\xa3100'
+
+    >>> print(_replace_html_entities(b"Price: &pound;100"))
+    Price: £100
+    ```"""
+
+    def _convert_entity(match):
+        entity_body = match.group(3)
+        if match.group(1):
+            try:
+                if match.group(2):
+                    number = int(entity_body, 16)
+                else:
+                    number = int(entity_body, 10)
+                # Numeric character references in the 80-9F range are typically
+                # interpreted by browsers as representing the characters mapped
+                # to bytes 80-9F in the Windows-1252 encoding. For more info
+                # see: https://en.wikipedia.org/wiki/ISO/IEC_8859-1#Similar_character_sets
+                if 0x80 <= number <= 0x9F:
+                    return bytes((number,)).decode("cp1252")
+            except ValueError:
+                number = None
+        else:
+            if entity_body in keep:
+                return match.group(0)
+            else:
+                number = html.entities.name2codepoint.get(entity_body)
+        if number is not None:
+            try:
+                return chr(number)
+            except (ValueError, OverflowError):
+                pass
+
+        return "" if remove_illegal else match.group(0)
+
+    return ENT_RE.sub(_convert_entity, _str_to_unicode(text, encoding))
+
+
+######################################################################
+
+
+class TweetTokenizer:
+    r"""
+    Examples:
+
+    ```python
+    >>> # Tokenizer for tweets.
+    >>> from nltk.tokenize import TweetTokenizer
+
+    >>> tknzr = TweetTokenizer()
+    >>> s0 = "This is a cooool #dummysmiley: :-) :-P <3 and some arrows < > -> <--"
+    >>> tknzr.tokenize(s0)
+    ['This', 'is', 'a', 'cooool', '#dummysmiley', ':', ':-)', ':-P', '<3', 'and', 'some', 'arrows', '<', '>', '->', '<--']
+
+    >>> # Examples using *strip_handles* and *reduce_len parameters*:
+    >>> tknzr = TweetTokenizer(strip_handles=True, reduce_len=True)
+    >>> s1 = "@remy: This is waaaaayyyy too much for you!!!!!!"
+    >>> tknzr.tokenize(s1)
+    [':', 'This', 'is', 'waaayyy', 'too', 'much', 'for', 'you', '!', '!', '!']
+    ```"""
+
+    def __init__(self, preserve_case=True, reduce_len=False, strip_handles=False):
+        self.preserve_case = preserve_case
+        self.reduce_len = reduce_len
+        self.strip_handles = strip_handles
+
+    def tokenize(self, text):
+        """
+        Args:
+            text: str
+
+        Returns: list(str) A tokenized list of strings; concatenating this list returns the original string if
+        `preserve_case=False`
+        """
+        # Fix HTML character entities:
+        text = _replace_html_entities(text)
+        # Remove username handles
+        if self.strip_handles:
+            text = remove_handles(text)
+        # Normalize word lengthening
+        if self.reduce_len:
+            text = reduce_lengthening(text)
+        # Shorten problematic sequences of characters
+        safe_text = HANG_RE.sub(r"\1\1\1", text)
+        # Tokenize:
+        words = WORD_RE.findall(safe_text)
+        # Possibly alter the case, but avoid changing emoticons like :D into :d:
+        if not self.preserve_case:
+            words = [x if EMOTICON_RE.search(x) else x.lower() for x in words]
+        return words
+
+
+######################################################################
+# Normalization Functions
+######################################################################
+
+
+def reduce_lengthening(text):
+    """
+    Replace repeated character sequences of length 3 or greater with sequences of length 3.
+    """
+    pattern = regex.compile(r"(.)\1{2,}")
+    return pattern.sub(r"\1\1\1", text)
+
+
+def remove_handles(text):
+    """
+    Remove Twitter username handles from text.
+    """
+    pattern = regex.compile(
+        r"(?<![A-Za-z0-9_!@#\$%&*])@(([A-Za-z0-9_]){20}(?!@))|(?<![A-Za-z0-9_!@#\$%&*])@(([A-Za-z0-9_]){1,19})(?![A-Za-z0-9_]*@)"
+    )
+    # Substitute handles with ' ' to ensure that text on either side of removed handles are tokenized correctly
+    return pattern.sub(" ", text)
+
+
+######################################################################
+# Tokenization Function
+######################################################################
+
+
+def casual_tokenize(text, preserve_case=True, reduce_len=False, strip_handles=False):
+    """
+    Convenience function for wrapping the tokenizer.
+    """
+    return TweetTokenizer(preserve_case=preserve_case, reduce_len=reduce_len, strip_handles=strip_handles).tokenize(
+        text
+    )
+
+
+###############################################################################
+
+
+__all__ = ["BertweetTokenizer"]

lib/python3.12/site-packages/transformers/models/biogpt/__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_biogpt import *
+    from .modeling_biogpt import *
+    from .tokenization_biogpt 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/biogpt/configuration_biogpt.py

@@ -0,0 +1,134 @@
+# coding=utf-8
+# Copyright 2022 The HuggingFace Team and Microsoft Research AI4Science 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.
+"""BioGPT model configuration"""
+
+from ...configuration_utils import PretrainedConfig
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class BioGptConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`BioGptModel`]. It is used to instantiate an
+    BioGPT 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 BioGPT
+    [microsoft/biogpt](https://huggingface.co/microsoft/biogpt) 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 42384):
+            Vocabulary size of the BioGPT model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`BioGptModel`].
+        hidden_size (`int`, *optional*, defaults to 1024):
+            Dimension of the encoder layers and the pooler layer.
+        num_hidden_layers (`int`, *optional*, defaults to 24):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 16):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        intermediate_size (`int`, *optional*, defaults to 4096):
+            Dimension of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
+        hidden_act (`str` or `function`, *optional*, defaults to `"gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"selu"` and `"gelu_new"` are supported.
+        hidden_dropout_prob (`float`, *optional*, defaults to 0.1):
+            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_probs_dropout_prob (`float`, *optional*, defaults to 0.1):
+            The dropout ratio for the attention probabilities.
+        max_position_embeddings (`int`, *optional*, defaults to 1024):
+            The maximum sequence length that this model might ever be used with. Typically set this to something large
+            just in case (e.g., 512 or 1024 or 2048).
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        layer_norm_eps (`float`, *optional*, defaults to 1e-12):
+            The epsilon used by the layer normalization layers.
+        scale_embedding (`bool`, *optional*, defaults to `True`):
+            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). Only
+            relevant if `config.is_decoder=True`.
+        layerdrop (`float`, *optional*, defaults to 0.0):
+            Please refer to the paper about LayerDrop: https://arxiv.org/abs/1909.11556 for further details
+        activation_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for activations inside the fully connected layer.
+        pad_token_id (`int`, *optional*, defaults to 1):
+            Padding token id.
+        bos_token_id (`int`, *optional*, defaults to 0):
+            Beginning of stream token id.
+        eos_token_id (`int`, *optional*, defaults to 2):
+            End of stream token id.
+
+    Example:
+
+    ```python
+    >>> from transformers import BioGptModel, BioGptConfig
+
+    >>> # Initializing a BioGPT microsoft/biogpt style configuration
+    >>> configuration = BioGptConfig()
+
+    >>> # Initializing a model from the microsoft/biogpt style configuration
+    >>> model = BioGptModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "biogpt"
+
+    def __init__(
+        self,
+        vocab_size=42384,
+        hidden_size=1024,
+        num_hidden_layers=24,
+        num_attention_heads=16,
+        intermediate_size=4096,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        max_position_embeddings=1024,
+        initializer_range=0.02,
+        layer_norm_eps=1e-12,
+        scale_embedding=True,
+        use_cache=True,
+        layerdrop=0.0,
+        activation_dropout=0.0,
+        pad_token_id=1,
+        bos_token_id=0,
+        eos_token_id=2,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.intermediate_size = intermediate_size
+        self.hidden_act = hidden_act
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.initializer_range = initializer_range
+        self.layer_norm_eps = layer_norm_eps
+        self.scale_embedding = scale_embedding
+        self.use_cache = use_cache
+        self.layerdrop = layerdrop
+        self.activation_dropout = activation_dropout
+        super().__init__(pad_token_id=pad_token_id, bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)
+
+
+__all__ = ["BioGptConfig"]

lib/python3.12/site-packages/transformers/models/biogpt/modeling_biogpt.py

@@ -0,0 +1,1042 @@
+# coding=utf-8
+# Copyright 2022 The HuggingFace Team and Microsoft Research AI4Science 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 BioGPT model."""
+
+import math
+from typing import Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+
+from ...activations import ACT2FN
+from ...generation import GenerationMixin
+from ...modeling_attn_mask_utils import _prepare_4d_causal_attention_mask, _prepare_4d_causal_attention_mask_for_sdpa
+from ...modeling_outputs import (
+    BaseModelOutputWithPastAndCrossAttentions,
+    CausalLMOutputWithCrossAttentions,
+    SequenceClassifierOutputWithPast,
+    TokenClassifierOutput,
+)
+from ...modeling_utils import PreTrainedModel
+from ...utils import (
+    add_code_sample_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    logging,
+)
+from .configuration_biogpt import BioGptConfig
+
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "microsoft/biogpt"
+_CONFIG_FOR_DOC = "BioGptConfig"
+
+
+# copied from transformers.models.opt.modeling_opt.OPTLearnedPositionalEmbedding with OPT->BioGpt
+# TODO @ArthurZucker bring copied from back
+class BioGptLearnedPositionalEmbedding(nn.Embedding):
+    """
+    This module learns positional embeddings up to a fixed maximum size.
+    """
+
+    def __init__(self, num_embeddings: int, embedding_dim: int):
+        # BioGpt is set up so that if padding_idx is specified then offset the embedding ids by 2
+        # and adjust num_embeddings appropriately. Other models don't have this hack
+        self.offset = 2
+        super().__init__(num_embeddings + self.offset, embedding_dim)
+
+    def forward(self, attention_mask: torch.LongTensor, past_key_values_length: int = 0):
+        """`input_ids_shape` is expected to be [bsz x seqlen]."""
+        attention_mask = attention_mask.long()
+
+        # create positions depending on attention_mask
+        positions = (torch.cumsum(attention_mask, dim=1).type_as(attention_mask) * attention_mask).long() - 1
+
+        # cut positions if `past_key_values_length` is > 0
+        positions = positions[:, past_key_values_length:]
+
+        return super().forward(positions + self.offset)
+
+
+# Copied from transformers.models.bart.modeling_bart.BartScaledWordEmbedding with Bart->BioGpt
+class BioGptScaledWordEmbedding(nn.Embedding):
+    """
+    This module overrides nn.Embeddings' forward by multiplying with embeddings scale.
+    """
+
+    def __init__(self, num_embeddings: int, embedding_dim: int, padding_idx: int, embed_scale: Optional[float] = 1.0):
+        super().__init__(num_embeddings, embedding_dim, padding_idx)
+        self.embed_scale = embed_scale
+
+    def forward(self, input_ids: torch.Tensor):
+        return super().forward(input_ids) * self.embed_scale
+
+
+# Copied from transformers.models.bart.modeling_bart.BartAttention with Bart->BioGpt
+class BioGptAttention(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[BioGptConfig] = 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.BartSdpaAttention with Bart->BioGpt
+class BioGptSdpaAttention(BioGptAttention):
+    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 output_attentions or layer_head_mask is not None:
+            # TODO: Improve this warning with e.g. `model.config._attn_implementation = "manual"` once this is implemented.
+            logger.warning_once(
+                "BioGptModel is using BioGptSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True` or `layer_head_mask` not None. Falling back to the manual attention"
+                ' implementation, but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
+            )
+            return super().forward(
+                hidden_states,
+                key_value_states=key_value_states,
+                past_key_value=past_key_value,
+                attention_mask=attention_mask,
+                layer_head_mask=layer_head_mask,
+                output_attentions=output_attentions,
+            )
+
+        # 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)
+        # 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)
+
+        query_states = self._shape(query_states, tgt_len, bsz)
+
+        # We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment
+        # in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling.
+        # The tgt_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case tgt_len == 1.
+        is_causal = True if self.is_causal and attention_mask is None and tgt_len > 1 else False
+
+        # NOTE: SDPA with memory-efficient backend is currently (torch==2.1.2) bugged when using non-contiguous inputs and a custom attn_mask,
+        # but we are fine here as `_shape` do call `.contiguous()`. Reference: https://github.com/pytorch/pytorch/issues/112577
+        attn_output = torch.nn.functional.scaled_dot_product_attention(
+            query_states,
+            key_states,
+            value_states,
+            attn_mask=attention_mask,
+            dropout_p=self.dropout if self.training else 0.0,
+            is_causal=is_causal,
+        )
+
+        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.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, None, past_key_value
+
+
+BIOGPT_ATTENTION_CLASSES = {
+    "eager": BioGptAttention,
+    "sdpa": BioGptSdpaAttention,
+}
+
+
+class BioGptDecoderLayer(nn.Module):
+    def __init__(self, config: BioGptConfig):
+        super().__init__()
+        self.embed_dim = config.hidden_size
+
+        self.self_attn = BIOGPT_ATTENTION_CLASSES[config._attn_implementation](
+            embed_dim=self.embed_dim,
+            num_heads=config.num_attention_heads,
+            dropout=config.attention_probs_dropout_prob,
+            is_decoder=True,
+            is_causal=True,
+        )
+        self.dropout = config.hidden_dropout_prob
+        self.activation_fn = ACT2FN[config.hidden_act]
+        self.activation_dropout = config.activation_dropout
+
+        self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
+
+        self.fc1 = nn.Linear(self.embed_dim, config.intermediate_size)
+        self.fc2 = nn.Linear(config.intermediate_size, 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,
+        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.
+            layer_head_mask (`torch.FloatTensor`): mask for attention heads in a given layer of size
+                `(encoder_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.
+            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`).
+        """
+        residual = hidden_states
+
+        hidden_states = self.self_attn_layer_norm(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
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.final_layer_norm(hidden_states)
+        hidden_states = self.fc1(hidden_states)
+        hidden_states = self.activation_fn(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
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        if use_cache:
+            outputs += (present_key_value,)
+
+        return outputs
+
+
+class BioGptPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = BioGptConfig
+    base_model_prefix = "biogpt"
+    supports_gradient_checkpointing = True
+    _supports_sdpa = True
+
+    def _init_weights(self, module):
+        """Initialize the weights"""
+        if isinstance(module, nn.Linear):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+
+
+BIOGPT_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 ([`~BioGptConfig`]): 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.
+"""
+
+BIOGPT_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` 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 (`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)
+        head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
+            Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        inputs_embeds (`torch.FloatTensor` of shape `({0}, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+            is useful if you want more control over how to convert *input_ids* indices into associated vectors than the
+            model's internal embedding lookup matrix.
+        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.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+            `past_key_values`).
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+@add_start_docstrings(
+    "The bare BioGPT Model transformer outputting raw hidden-states without any specific head on top.",
+    BIOGPT_START_DOCSTRING,
+)
+class BioGptModel(BioGptPreTrainedModel):
+    def __init__(self, config: BioGptConfig):
+        super().__init__(config)
+        self.config = config
+        self.layerdrop = config.layerdrop
+        self.dropout = config.hidden_dropout_prob
+        self.embed_dim = config.hidden_size
+        self.padding_idx = config.pad_token_id
+        embed_scale = math.sqrt(config.hidden_size) if config.scale_embedding else 1.0
+
+        self.embed_tokens = BioGptScaledWordEmbedding(
+            config.vocab_size, self.embed_dim, self.padding_idx, embed_scale=embed_scale
+        )
+        self.embed_positions = BioGptLearnedPositionalEmbedding(config.max_position_embeddings, self.embed_dim)
+
+        self.layers = nn.ModuleList([BioGptDecoderLayer(config) for _ in range(config.num_hidden_layers)])
+        self.layer_norm = nn.LayerNorm(self.embed_dim)
+
+        self.gradient_checkpointing = False
+        self._use_sdpa = config._attn_implementation == "sdpa"
+        # 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
+
+    @add_start_docstrings_to_model_forward(BIOGPT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=BaseModelOutputWithPastAndCrossAttentions,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        **kwargs,  # NOOP kwargs, for now
+    ) -> Union[Tuple, BaseModelOutputWithPastAndCrossAttentions]:
+        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 input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input = input_ids
+            input_shape = input.size()
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+            input = inputs_embeds[:, :, -1]
+        else:
+            raise ValueError("You have to specify either input_ids or 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)
+
+        if attention_mask is None:
+            attention_mask = torch.ones(
+                (inputs_embeds.shape[0], inputs_embeds.shape[1] + past_key_values_length),
+                dtype=torch.bool,
+                device=inputs_embeds.device,
+            )
+        elif attention_mask.shape[1] != past_key_values_length + input_shape[1]:
+            raise ValueError(
+                f"The provided attention mask has length {attention_mask.shape[1]}, but its length should be "
+                f"{past_key_values_length + input_shape[1]} (sum of the lengths of current and past inputs)"
+            )
+
+        # embed positions
+        positions = self.embed_positions(attention_mask, past_key_values_length)
+
+        if self._use_sdpa and not output_attentions and head_mask is None:
+            # output_attentions=True & head_mask can not be supported when using SDPA, fall back to
+            # the manual implementation that requires a 4D causal mask in all cases.
+            attention_mask = _prepare_4d_causal_attention_mask_for_sdpa(
+                attention_mask, input_shape, inputs_embeds, past_key_values_length
+            )
+        else:
+            attention_mask = _prepare_4d_causal_attention_mask(
+                attention_mask, input_shape, inputs_embeds, past_key_values_length
+            )
+
+        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
+
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        all_cross_attentions = None
+        next_decoder_cache = () if use_cache else None
+
+        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,
+                    head_mask[idx] if head_mask is not None else None,
+                    None,
+                    output_attentions,
+                    use_cache,
+                )
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    layer_head_mask=(head_mask[idx] if 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[2 if output_attentions else 1],)
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        hidden_states = self.layer_norm(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(
+    """BioGPT Model with a `language modeling` head on top for CLM fine-tuning.""", BIOGPT_START_DOCSTRING
+)
+class BioGptForCausalLM(BioGptPreTrainedModel, GenerationMixin):
+    _tied_weights_keys = ["output_projection.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.biogpt = BioGptModel(config)
+        self.output_projection = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_output_embeddings(self):
+        return self.output_projection
+
+    def set_output_embeddings(self, new_embeddings):
+        self.output_projection = new_embeddings
+
+    @add_start_docstrings_to_model_forward(BIOGPT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=CausalLMOutputWithCrossAttentions,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = 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,
+        **kwargs,
+    ) -> Union[Tuple, CausalLMOutputWithCrossAttentions]:
+        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
+
+        outputs = self.biogpt(
+            input_ids,
+            attention_mask=attention_mask,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        sequence_output = outputs[0]
+        prediction_scores = self.output_projection(sequence_output)
+
+        lm_loss = None
+        if labels is not None:
+            lm_loss = self.loss_function(
+                prediction_scores,
+                labels,
+                vocab_size=self.config.vocab_size,
+                **kwargs,
+            )
+
+        if not return_dict:
+            output = (prediction_scores,) + outputs[1:]
+            return ((lm_loss,) + output) if lm_loss is not None else output
+
+        return CausalLMOutputWithCrossAttentions(
+            loss=lm_loss,
+            logits=prediction_scores,
+            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
+
+
+@add_start_docstrings(
+    """
+    BioGPT Model with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for
+    Named-Entity-Recognition (NER) tasks.
+    """,
+    BIOGPT_START_DOCSTRING,
+)
+class BioGptForTokenClassification(BioGptPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.biogpt = BioGptModel(config)
+        if hasattr(config, "classifier_dropout") and config.classifier_dropout is not None:
+            classifier_dropout = config.classifier_dropout
+        else:
+            classifier_dropout = config.hidden_dropout_prob
+        self.dropout = nn.Dropout(classifier_dropout)
+        self.classifier = nn.Linear(config.hidden_size, config.num_labels)
+
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(BIOGPT_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=TokenClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = 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, TokenClassifierOutput]:
+        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).
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        transformer_outputs = self.biogpt(
+            input_ids,
+            past_key_values=past_key_values,
+            attention_mask=attention_mask,
+            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,
+        )
+
+        hidden_states = transformer_outputs[0]
+        hidden_states = self.dropout(hidden_states)
+        logits = self.classifier(hidden_states)
+
+        loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            # Only keep active parts of the loss
+            if attention_mask is not None:
+                active_loss = attention_mask.view(-1) == 1
+                active_logits = logits.view(-1, self.num_labels)
+                active_labels = torch.where(
+                    active_loss, labels.view(-1), torch.tensor(loss_fct.ignore_index).type_as(labels)
+                )
+                loss = loss_fct(active_logits, active_labels)
+            else:
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+
+        if not return_dict:
+            output = (logits,) + transformer_outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TokenClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    The BioGpt Model transformer with a sequence classification head on top (linear layer).
+
+    [`BioGptForSequenceClassification`] uses the last token in order to do the classification, as other causal models
+    (e.g. GPT-2) do.
+
+    Since it does classification on the last token, it is required to know the position of the last token. If a
+    `pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
+    no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
+    padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in
+    each row of the batch).
+    """,
+    BIOGPT_START_DOCSTRING,
+)
+class BioGptForSequenceClassification(BioGptPreTrainedModel):
+    def __init__(self, config: BioGptConfig):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.biogpt = BioGptModel(config)
+        self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(BIOGPT_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=SequenceClassifierOutputWithPast,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = 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, SequenceClassifierOutputWithPast]:
+        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).
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        transformer_outputs = self.biogpt(
+            input_ids,
+            past_key_values=past_key_values,
+            attention_mask=attention_mask,
+            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,
+        )
+        hidden_states = transformer_outputs[0]
+        logits = self.score(hidden_states)
+
+        if input_ids is not None:
+            batch_size, sequence_length = input_ids.shape[:2]
+        else:
+            batch_size, sequence_length = inputs_embeds.shape[:2]
+
+        if self.config.pad_token_id is None:
+            sequence_length = -1
+        else:
+            if input_ids is not None:
+                sequence_length = (torch.ne(input_ids, self.config.pad_token_id).sum(-1) - 1).to(logits.device)
+            else:
+                sequence_length = -1
+                logger.warning_once(
+                    f"{self.__class__.__name__} will not detect padding tokens in `inputs_embeds`. Results may be "
+                    "unexpected if using padding tokens in conjunction with `inputs_embeds.`"
+                )
+
+        pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_length]
+
+        loss = None
+        if labels is not None:
+            if self.config.problem_type is None:
+                if self.num_labels == 1:
+                    self.config.problem_type = "regression"
+                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
+                    self.config.problem_type = "single_label_classification"
+                else:
+                    self.config.problem_type = "multi_label_classification"
+
+            if self.config.problem_type == "regression":
+                loss_fct = MSELoss()
+                if self.num_labels == 1:
+                    loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(pooled_logits, labels)
+            elif self.config.problem_type == "single_label_classification":
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(pooled_logits.view(-1, self.num_labels), labels.view(-1))
+            elif self.config.problem_type == "multi_label_classification":
+                loss_fct = BCEWithLogitsLoss()
+                loss = loss_fct(pooled_logits, labels)
+        if not return_dict:
+            output = (pooled_logits,) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return SequenceClassifierOutputWithPast(
+            loss=loss,
+            logits=pooled_logits,
+            past_key_values=transformer_outputs.past_key_values,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
+
+    def get_input_embeddings(self):
+        return self.biogpt.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.biogpt.embed_tokens = value
+
+
+__all__ = [
+    "BioGptForCausalLM",
+    "BioGptForTokenClassification",
+    "BioGptForSequenceClassification",
+    "BioGptModel",
+    "BioGptPreTrainedModel",
+]

lib/python3.12/site-packages/transformers/models/biogpt/tokenization_biogpt.py

@@ -0,0 +1,361 @@
+# coding=utf-8
+# Copyright 2022 The HuggingFace Team and Microsoft Research AI4Science. 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 classes for BioGPT."""
+
+import json
+import os
+from typing import List, Optional, Tuple
+
+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",
+}
+
+
+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 BioGptTokenizer(PreTrainedTokenizer):
+    """
+    Construct an FAIRSEQ Transformer tokenizer. Moses tokenization followed by Byte-Pair Encoding.
+
+    This tokenizer inherits from [`PreTrainedTokenizer`] which contains most of the main methods. Users should refer to
+    this superclass for more information regarding those methods.
+
+    Args:
+        vocab_file (`str`):
+            Path to the vocabulary file.
+        merges_file (`str`):
+            Merges file.
+        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.
+        bos_token (`str`, *optional*, defaults to `"<s>"`):
+            The beginning of sequence token that was used during pretraining. Can be used a sequence classifier token.
+
+            <Tip>
+
+            When building a sequence using special tokens, this is not the token that is used for the beginning of
+            sequence. The token used is the `cls_token`.
+
+            </Tip>
+
+        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>
+
+        sep_token (`str`, *optional*, defaults to `"</s>"`):
+            The separator token, which is used when building a sequence from multiple sequences, e.g. two sequences for
+            sequence classification or for a text and a question for question answering. It is also used as the last
+            token of a sequence built with special tokens.
+        pad_token (`str`, *optional*, defaults to `"<pad>"`):
+            The token used for padding, for example when batching sequences of different lengths.
+    """
+
+    vocab_files_names = VOCAB_FILES_NAMES
+    model_input_names = ["input_ids", "attention_mask"]
+
+    def __init__(
+        self,
+        vocab_file,
+        merges_file,
+        unk_token="<unk>",
+        bos_token="<s>",
+        eos_token="</s>",
+        sep_token="</s>",
+        pad_token="<pad>",
+        **kwargs,
+    ):
+        try:
+            import sacremoses
+        except ImportError:
+            raise ImportError(
+                "You need to install sacremoses to use BioGptTokenizer. "
+                "See https://pypi.org/project/sacremoses/ for installation."
+            )
+
+        self.lang = "en"
+        self.sm = sacremoses
+        # cache of sm.MosesTokenizer instance
+        self.cache_moses_tokenizer = {}
+        self.cache_moses_detokenizer = {}
+
+        """ Initialisation"""
+        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]
+        merges = [tuple(merge.split()[:2]) for merge in merges]
+        self.bpe_ranks = dict(zip(merges, range(len(merges))))
+        self.cache = {}
+
+        super().__init__(
+            bos_token=bos_token,
+            eos_token=eos_token,
+            sep_token=sep_token,
+            unk_token=unk_token,
+            pad_token=pad_token,
+            **kwargs,
+        )
+
+    @property
+    def vocab_size(self):
+        """Returns vocab size"""
+        return len(self.encoder)
+
+    def get_vocab(self):
+        return dict(self.encoder, **self.added_tokens_encoder)
+
+    def moses_tokenize(self, text, lang):
+        if lang not in self.cache_moses_tokenizer:
+            moses_tokenizer = self.sm.MosesTokenizer(lang=lang)
+            self.cache_moses_tokenizer[lang] = moses_tokenizer
+        return self.cache_moses_tokenizer[lang].tokenize(
+            text, aggressive_dash_splits=True, return_str=False, escape=True
+        )
+
+    def moses_detokenize(self, tokens, lang):
+        if lang not in self.cache_moses_detokenizer:
+            moses_detokenizer = self.sm.MosesDetokenizer(lang=lang)
+            self.cache_moses_detokenizer[lang] = moses_detokenizer
+        return self.cache_moses_detokenizer[lang].detokenize(tokens)
+
+    def bpe(self, token):
+        word = tuple(token[:-1]) + (token[-1] + "</w>",)
+        if token in self.cache:
+            return self.cache[token]
+        pairs = get_pairs(word)
+
+        if not pairs:
+            return token + "</w>"
+
+        while True:
+            bigram = min(pairs, key=lambda pair: self.bpe_ranks.get(pair, float("inf")))
+            if bigram not in self.bpe_ranks:
+                break
+            first, second = bigram
+            new_word = []
+            i = 0
+            while i < len(word):
+                try:
+                    j = word.index(first, i)
+                except ValueError:
+                    new_word.extend(word[i:])
+                    break
+                else:
+                    new_word.extend(word[i:j])
+                    i = j
+
+                if word[i] == first and i < len(word) - 1 and word[i + 1] == second:
+                    new_word.append(first + second)
+                    i += 2
+                else:
+                    new_word.append(word[i])
+                    i += 1
+            new_word = tuple(new_word)
+            word = new_word
+            if len(word) == 1:
+                break
+            else:
+                pairs = get_pairs(word)
+        word = " ".join(word)
+        if word == "\n  </w>":
+            word = "\n</w>"
+        self.cache[token] = word
+        return word
+
+    def _tokenize(self, text, bypass_tokenizer=False):
+        """Returns a tokenized string."""
+        if bypass_tokenizer:
+            text = text.split()
+        else:
+            text = self.moses_tokenize(text, self.lang)
+
+        split_tokens = []
+        for token in text:
+            if token:
+                split_tokens.extend(list(self.bpe(token).split(" ")))
+
+        return split_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, self.unk_token)
+
+    def convert_tokens_to_string(self, tokens):
+        """Converts a sequence of tokens (string) in a single string."""
+        # remove BPE
+        tokens = [t.replace(" ", "").replace("</w>", " ") for t in tokens]
+        tokens = "".join(tokens).split()
+        # detokenize
+        text = self.moses_detokenize(tokens, self.lang)
+        return text
+
+    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 BioGPT sequence has the following format:
+
+        - single sequence: `</s> X `
+        - pair of sequences: `</s> A </s> B `
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs to which the special tokens will be added.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            `List[int]`: List of [input IDs](../glossary#input-ids) with the appropriate special tokens.
+        """
+        if token_ids_1 is None:
+            return [self.sep_token_id] + token_ids_0
+        sep = [self.sep_token_id]
+        return sep + token_ids_0 + sep + token_ids_1
+
+    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
+            )
+        # no bos used in fairseq
+        if token_ids_1 is not None:
+            return [1] + ([0] * len(token_ids_0)) + [1] + ([0] * len(token_ids_1))
+        return [1] + ([0] * len(token_ids_0))
+
+    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 FAIRSEQ
+        Transformer 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]
+
+        # no bos used in fairseq
+        if token_ids_1 is None:
+            return len(token_ids_0 + sep) * [0]
+        return len(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:
+            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 __getstate__(self):
+        state = self.__dict__.copy()
+        state["sm"] = None
+        return state
+
+    def __setstate__(self, d):
+        self.__dict__ = d
+
+        try:
+            import sacremoses
+        except ImportError:
+            raise ImportError(
+                "You need to install sacremoses to use XLMTokenizer. "
+                "See https://pypi.org/project/sacremoses/ for installation."
+            )
+
+        self.sm = sacremoses
+
+
+__all__ = ["BioGptTokenizer"]

lib/python3.12/site-packages/transformers/models/blenderbot/__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 import *
+    from .modeling_blenderbot import *
+    from .modeling_flax_blenderbot import *
+    from .modeling_tf_blenderbot import *
+    from .tokenization_blenderbot import *
+    from .tokenization_blenderbot_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/configuration_blenderbot.py

@@ -0,0 +1,395 @@
+# 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.
+"""Blenderbot 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 BlenderbotConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`BlenderbotModel`]. It is used to instantiate an
+    Blenderbot 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 Blenderbot
+    [facebook/blenderbot-3B](https://huggingface.co/facebook/blenderbot-3B) 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 Blenderbot model. Defines the number of different tokens that can be represented by
+            the `inputs_ids` passed when calling [`BlenderbotModel`] or [`TFBlenderbotModel`].
+        d_model (`int`, *optional*, defaults to 1024):
+            Dimensionality of the layers and the pooler layer.
+        encoder_layers (`int`, *optional*, defaults to 12):
+            Number of encoder layers.
+        decoder_layers (`int`, *optional*, defaults to 12):
+            Number of decoder layers.
+        encoder_attention_heads (`int`, *optional*, defaults to 16):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        decoder_attention_heads (`int`, *optional*, defaults to 16):
+            Number of attention heads for each attention layer in the Transformer decoder.
+        decoder_ffn_dim (`int`, *optional*, defaults to 4096):
+            Dimensionality of the "intermediate" (often named feed-forward) layer in decoder.
+        encoder_ffn_dim (`int`, *optional*, defaults to 4096):
+            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 128):
+            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 BlenderbotConfig, BlenderbotModel
+
+    >>> # Initializing a Blenderbot facebook/blenderbot-3B style configuration
+    >>> configuration = BlenderbotConfig()
+
+    >>> # Initializing a model (with random weights) from the facebook/blenderbot-3B style configuration
+    >>> model = BlenderbotModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "blenderbot"
+    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=8008,
+        max_position_embeddings=128,
+        encoder_layers=2,
+        encoder_ffn_dim=10240,
+        encoder_attention_heads=32,
+        decoder_layers=24,
+        decoder_ffn_dim=10240,
+        decoder_attention_heads=32,
+        encoder_layerdrop=0.0,
+        decoder_layerdrop=0.0,
+        use_cache=True,
+        is_encoder_decoder=True,
+        activation_function="gelu",
+        d_model=2560,
+        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,
+        encoder_no_repeat_ngram_size=3,
+        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,
+            encoder_no_repeat_ngram_size=encoder_no_repeat_ngram_size,
+            forced_eos_token_id=forced_eos_token_id,
+            **kwargs,
+        )
+
+
+class BlenderbotOnnxConfig(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":
+            common_inputs = OrderedDict(
+                [
+                    ("input_ids", {0: "batch", 1: "encoder_sequence"}),
+                    ("attention_mask", {0: "batch", 1: "encoder_sequence"}),
+                ]
+            )
+            if self.use_past:
+                _, num_decoder_layers = self.num_layers
+                for i in range(num_decoder_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
+    # Copied from transformers.models.bart.configuration_bart.BartOnnxConfig.outputs
+    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
+            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"] = []
+            _, num_decoder_layers = self.num_layers
+
+            for _ in range(num_decoder_layers):
+                common_inputs["past_key_values"].append(
+                    (
+                        torch.zeros(decoder_shape),
+                        torch.zeros(decoder_shape),
+                        torch.zeros(encoder_shape),
+                        torch.zeros(encoder_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
+            past_key_values_length = seqlen
+            _, num_decoder_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_decoder_layers)
+            ]
+        return common_inputs
+
+    # Copied from transformers.models.bart.configuration_bart.BartOnnxConfig._generate_dummy_inputs_for_sequence_classification_and_question_answering
+    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
+
+    # Copied from transformers.models.bart.configuration_bart.BartOnnxConfig.generate_dummy_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
+
+    # Copied from transformers.models.bart.configuration_bart.BartOnnxConfig._flatten_past_key_values_
+    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
+            )
+
+    def fill_with_past_key_values_(self, inputs_or_outputs: Mapping[str, Mapping[int, str]], direction: str):
+        if direction not in ["inputs", "outputs"]:
+            raise ValueError(f'direction must either be "inputs" or "outputs", but {direction} was given')
+
+        name = "past_key_values" if direction == "inputs" else "present"
+        _, num_decoder_layers = self.num_layers
+
+        encoder_sequence = "past_encoder_sequence"
+        decoder_sequence = "past_decoder_sequence" if direction == "inputs" else "past_decoder_sequence + sequence"
+
+        for i in range(num_decoder_layers):
+            inputs_or_outputs[f"{name}.{i}.decoder.key"] = {0: "batch", 2: decoder_sequence}
+            inputs_or_outputs[f"{name}.{i}.decoder.value"] = {0: "batch", 2: decoder_sequence}
+            inputs_or_outputs[f"{name}.{i}.encoder.key"] = {0: "batch", 2: encoder_sequence}
+            inputs_or_outputs[f"{name}.{i}.encoder.value"] = {0: "batch", 2: encoder_sequence}
+
+
+__all__ = ["BlenderbotConfig", "BlenderbotOnnxConfig"]

lib/python3.12/site-packages/transformers/models/blenderbot/modeling_blenderbot.py

@@ -0,0 +1,1557 @@
+# 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 Blenderbot model."""
+
+import copy
+import math
+import os
+import warnings
+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 ..blenderbot_small import BlenderbotSmallForConditionalGeneration, BlenderbotSmallModel
+from .configuration_blenderbot import BlenderbotConfig
+
+
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = "BlenderbotConfig"
+_CHECKPOINT_FOR_DOC = "facebook/blenderbot-400M-distill"
+
+
+# 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
+
+
+class BlenderbotLearnedPositionalEmbedding(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.BartScaledWordEmbedding with Bart->Blenderbot
+class BlenderbotScaledWordEmbedding(nn.Embedding):
+    """
+    This module overrides nn.Embeddings' forward by multiplying with embeddings scale.
+    """
+
+    def __init__(self, num_embeddings: int, embedding_dim: int, padding_idx: int, embed_scale: Optional[float] = 1.0):
+        super().__init__(num_embeddings, embedding_dim, padding_idx)
+        self.embed_scale = embed_scale
+
+    def forward(self, input_ids: torch.Tensor):
+        return super().forward(input_ids) * self.embed_scale
+
+
+# Copied from transformers.models.bart.modeling_bart.BartAttention with Bart->Blenderbot
+class BlenderbotAttention(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[BlenderbotConfig] = 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
+
+
+BLENDERBOT_ATTENTION_CLASSES = {"eager": BlenderbotAttention}
+
+
+# Copied from transformers.models.mbart.modeling_mbart.MBartEncoderLayer with MBart->Blenderbot, MBART->BLENDERBOT
+class BlenderbotEncoderLayer(nn.Module):
+    def __init__(self, config: BlenderbotConfig):
+        super().__init__()
+        self.embed_dim = config.d_model
+
+        self.self_attn = BLENDERBOT_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.Tensor,
+        attention_mask: torch.Tensor,
+        layer_head_mask: torch.Tensor,
+        output_attentions: bool = False,
+    ) -> torch.Tensor:
+        """
+        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 = self.self_attn_layer_norm(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
+
+        residual = hidden_states
+        hidden_states = self.final_layer_norm(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
+
+        if hidden_states.dtype == torch.float16:
+            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
+
+
+# Copied from transformers.models.mbart.modeling_mbart.MBartDecoderLayer with MBart->Blenderbot, MBART->BLENDERBOT
+class BlenderbotDecoderLayer(nn.Module):
+    def __init__(self, config: BlenderbotConfig):
+        super().__init__()
+        self.embed_dim = config.d_model
+
+        self.self_attn = BLENDERBOT_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_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,
+    ) -> torch.Tensor:
+        """
+        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
+        hidden_states = self.self_attn_layer_norm(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
+
+        # Cross-Attention Block
+        cross_attn_present_key_value = None
+        cross_attn_weights = None
+        if encoder_hidden_states is not None:
+            residual = hidden_states
+            hidden_states = self.encoder_attn_layer_norm(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
+
+            # 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.final_layer_norm(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
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights, cross_attn_weights)
+
+        if use_cache:
+            outputs += (present_key_value,)
+
+        return outputs
+
+
+class BlenderbotPreTrainedModel(PreTrainedModel):
+    config_class = BlenderbotConfig
+    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_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 ([`BlenderbotConfig`]):
+            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_GENERATION_EXAMPLE = r"""
+    Conversation example:
+
+    ```python
+    >>> from transformers import AutoTokenizer, BlenderbotForConditionalGeneration
+
+    >>> mname = "facebook/blenderbot-400M-distill"
+    >>> model = BlenderbotForConditionalGeneration.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: That's unfortunate. Are they trying to lose weight or are they just trying to be healthier?
+
+    >>> 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.</s> <s>That's unfortunate. "
+    ...     "Are they trying to lose weight or are they just trying to be healthier?</s> "
+    ...     "<s> 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:   I see. Well, it's good that they're trying to change their eating habits.
+    ```
+"""
+
+BLENDERBOT_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)
+
+            Blenderbot 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 BlenderbotEncoder(BlenderbotPreTrainedModel):
+    """
+    Transformer encoder consisting of *config.encoder_layers* self attention layers. Each layer is a
+    [`BlenderbotEncoderLayer`].
+
+    Args:
+        config: BlenderbotConfig
+        embed_tokens (nn.Embedding): output embedding
+    """
+
+    def __init__(self, config: BlenderbotConfig, 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
+        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 = BlenderbotScaledWordEmbedding(
+                config.vocab_size, embed_dim, self.padding_idx, embed_scale=embed_scale
+            )
+
+        self.embed_positions = BlenderbotLearnedPositionalEmbedding(
+            config.max_position_embeddings,
+            embed_dim,
+        )
+        self.layers = nn.ModuleList([BlenderbotEncoderLayer(config) for _ in range(config.encoder_layers)])
+        self.layer_norm = nn.LayerNorm(config.d_model)
+
+        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)
+
+        embed_pos = self.embed_positions(input_shape)
+
+        hidden_states = inputs_embeds + embed_pos
+        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],)
+
+        # add final layer norm
+        hidden_states = self.layer_norm(hidden_states)
+
+        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 BlenderbotDecoder(BlenderbotPreTrainedModel):
+    """
+    Transformer decoder consisting of *config.decoder_layers* layers. Each layer is a [`BlenderbotDecoderLayer`]
+
+    Args:
+        config: BlenderbotConfig
+        embed_tokens (nn.Embedding): output embedding
+    """
+
+    def __init__(self, config: BlenderbotConfig, 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
+        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 = BlenderbotScaledWordEmbedding(
+                config.vocab_size, config.d_model, self.padding_idx, embed_scale=embed_scale
+            )
+
+        self.embed_positions = BlenderbotLearnedPositionalEmbedding(
+            config.max_position_embeddings,
+            config.d_model,
+        )
+        self.layers = nn.ModuleList([BlenderbotDecoderLayer(config) for _ in range(config.decoder_layers)])
+        self.layer_norm = 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 `(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**.
+
+            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)
+
+        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)
+
+        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(
+                    "`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 final layer norm
+        hidden_states = self.layer_norm(hidden_states)
+
+        # 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 Blenderbot Model outputting raw hidden-states without any specific head on top.",
+    BLENDERBOT_START_DOCSTRING,
+)
+class BlenderbotModel(BlenderbotPreTrainedModel):
+    _tied_weights_keys = ["decoder.embed_tokens.weight", "encoder.embed_tokens.weight"]
+
+    def __init__(self, config: BlenderbotConfig):
+        super().__init__(config)
+
+        padding_idx, vocab_size = config.pad_token_id, config.vocab_size
+        embed_scale = math.sqrt(config.d_model) if config.scale_embedding else 1.0
+        self.shared = BlenderbotScaledWordEmbedding(vocab_size, config.d_model, padding_idx, embed_scale=embed_scale)
+        self.encoder = BlenderbotEncoder(config, self.shared)
+        self.decoder = BlenderbotDecoder(config, self.shared)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], *model_args, **kwargs):
+        if pretrained_model_name_or_path == "facebook/blenderbot-90M":
+            warnings.warn(
+                "The checkpoint `facebook/blenderbot-90M` is deprecated. In the future, please use the identical"
+                " checkpoint `facebook/small_blenderbot-90M` with"
+                " `BlenderbotSmallModel.from_pretrained('facebook/small_blenderbot-90M')` instead.",
+                FutureWarning,
+            )
+            return BlenderbotSmallModel.from_pretrained(pretrained_model_name_or_path)
+
+        return super(BlenderbotModel, cls).from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+
+    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_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, BlenderbotModel
+
+        >>> model = BlenderbotModel.from_pretrained("facebook/blenderbot-400M-distill")
+        >>> tokenizer = AutoTokenizer.from_pretrained("facebook/blenderbot-400M-distill")
+
+        >>> inputs = tokenizer("Studies have been shown that owning a dog is good for you", return_tensors="pt")
+        >>> decoder_input_ids = tokenizer("Studies show that", return_tensors="pt").input_ids  # Batch size 1
+        >>> outputs = model(input_ids=inputs.input_ids, decoder_input_ids=decoder_input_ids)
+
+        >>> last_hidden_states = outputs.last_hidden_state
+        >>> list(last_hidden_states.shape)
+        [1, 6, 1280]
+        ```"""
+        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 Blenderbot Model with a language modeling head. Can be used for summarization.", BLENDERBOT_START_DOCSTRING
+)
+class BlenderbotForConditionalGeneration(BlenderbotPreTrainedModel, 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: BlenderbotConfig):
+        super().__init__(config)
+        self.model = BlenderbotModel(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()
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], *model_args, **kwargs):
+        if pretrained_model_name_or_path == "facebook/blenderbot-90M":
+            warnings.warn(
+                "The checkpoint `facebook/blenderbot-90M` is deprecated. In the future, please use the identical"
+                " checkpoint `facebook/small_blenderbot-90M` with"
+                " `BlenderbotSmallForConditionalGeneration.from_pretrained('facebook/small_blenderbot-90M')` instead.",
+                FutureWarning,
+            )
+            return BlenderbotSmallForConditionalGeneration.from_pretrained(pretrained_model_name_or_path)
+
+        return super(BlenderbotForConditionalGeneration, cls).from_pretrained(
+            pretrained_model_name_or_path, *model_args, **kwargs
+        )
+
+    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_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=Seq2SeqLMOutput, config_class=_CONFIG_FOR_DOC)
+    @add_end_docstrings(BLENDERBOT_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->Blenderbot
+class BlenderbotDecoderWrapper(BlenderbotPreTrainedModel):
+    """
+    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 = BlenderbotDecoder(config)
+
+    def forward(self, *args, **kwargs):
+        return self.decoder(*args, **kwargs)
+
+
+# Copied from transformers.models.bart.modeling_bart.BartForCausalLM with Bart->Blenderbot, facebook/bart-base->facebook/blenderbot-400M-distill
+class BlenderbotForCausalLM(BlenderbotPreTrainedModel, 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 = BlenderbotDecoderWrapper(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, BlenderbotForCausalLM
+
+        >>> tokenizer = AutoTokenizer.from_pretrained("facebook/blenderbot-400M-distill")
+        >>> model = BlenderbotForCausalLM.from_pretrained("facebook/blenderbot-400M-distill", 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__ = [
+    "BlenderbotForCausalLM",
+    "BlenderbotForConditionalGeneration",
+    "BlenderbotModel",
+    "BlenderbotPreTrainedModel",
+]

lib/python3.12/site-packages/transformers/models/blenderbot/modeling_flax_blenderbot.py

@@ -0,0 +1,1508 @@
+# coding=utf-8
+# Copyright 2021 The Fairseq Authors and The Google Flax Team Authors 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 Blenderbot 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, add_start_docstrings_to_model_forward, logging, replace_return_docstrings
+from .configuration_blenderbot import BlenderbotConfig
+
+
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = "BlenderbotConfig"
+_CHECKPOINT_FOR_DOC = "facebook/blenderbot-400M-distill"
+
+
+BLENDERBOT_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 ([`BlenderbotConfig`]): 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.
+"""
+
+BLENDERBOT_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_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_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->Blenderbot
+class FlaxBlenderbotAttention(nn.Module):
+    config: BlenderbotConfig
+    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.mbart.modeling_flax_mbart.FlaxMBartEncoderLayer with MBart->Blenderbot
+class FlaxBlenderbotEncoderLayer(nn.Module):
+    config: BlenderbotConfig
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self) -> None:
+        self.embed_dim = self.config.d_model
+        self.self_attn = FlaxBlenderbotAttention(
+            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 = self.self_attn_layer_norm(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
+
+        residual = hidden_states
+        hidden_states = self.final_layer_norm(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
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (attn_weights,)
+
+        return outputs
+
+
+# Copied from transformers.models.bart.modeling_flax_bart.FlaxBartEncoderLayerCollection with Bart->Blenderbot
+class FlaxBlenderbotEncoderLayerCollection(nn.Module):
+    config: BlenderbotConfig
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
+
+    def setup(self):
+        self.layers = [
+            FlaxBlenderbotEncoderLayer(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.mbart.modeling_flax_mbart.FlaxMBartDecoderLayer with MBart->Blenderbot
+class FlaxBlenderbotDecoderLayer(nn.Module):
+    config: BlenderbotConfig
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self) -> None:
+        self.embed_dim = self.config.d_model
+        self.self_attn = FlaxBlenderbotAttention(
+            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 = FlaxBlenderbotAttention(
+            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
+        hidden_states = self.self_attn_layer_norm(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
+
+        # Cross-Attention Block
+        cross_attn_weights = None
+        if encoder_hidden_states is not None:
+            residual = hidden_states
+
+            hidden_states = self.encoder_attn_layer_norm(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
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.final_layer_norm(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
+
+        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->Blenderbot
+class FlaxBlenderbotDecoderLayerCollection(nn.Module):
+    config: BlenderbotConfig
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
+
+    def setup(self):
+        self.layers = [
+            FlaxBlenderbotDecoderLayer(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 FlaxBlenderbotEncoder(nn.Module):
+    config: BlenderbotConfig
+    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 = FlaxBlenderbotEncoderLayerCollection(self.config, self.dtype)
+        self.layer_norm = 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.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,
+        )
+        last_hidden_states = outputs[0]
+        last_hidden_states = self.layer_norm(last_hidden_states)
+
+        # update the last element in `hidden_states` after applying `layernorm` above
+        hidden_states = None
+        if output_hidden_states:
+            hidden_states = outputs[1]
+            hidden_states = hidden_states[:-1] + (last_hidden_states,)
+
+        if not return_dict:
+            outputs = (last_hidden_states, hidden_states) + (outputs[2:] if output_hidden_states else outputs[1:])
+            return tuple(v for v in outputs if v is not None)
+
+        return FlaxBaseModelOutput(
+            last_hidden_state=last_hidden_states,
+            hidden_states=hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+class FlaxBlenderbotDecoder(nn.Module):
+    config: BlenderbotConfig
+    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 = FlaxBlenderbotDecoderLayerCollection(self.config, self.dtype)
+        self.layer_norm = 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)
+
+        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,
+        )
+
+        last_hidden_states = outputs[0]
+        last_hidden_states = self.layer_norm(last_hidden_states)
+
+        # update the last element in `hidden_states` after applying `layernorm` above
+        hidden_states = None
+        if output_hidden_states:
+            hidden_states = outputs[1]
+            hidden_states = hidden_states[:-1] + (last_hidden_states,)
+
+        if not return_dict:
+            outputs = (last_hidden_states, hidden_states) + (outputs[2:] if output_hidden_states else outputs[1:])
+            return tuple(v for v in outputs if v is not None)
+
+        return FlaxBaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=last_hidden_states,
+            hidden_states=hidden_states,
+            attentions=outputs.attentions,
+            cross_attentions=outputs.cross_attentions,
+        )
+
+
+# Copied from transformers.models.bart.modeling_flax_bart.FlaxBartModule with Bart->Blenderbot
+class FlaxBlenderbotModule(nn.Module):
+    config: BlenderbotConfig
+    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 = FlaxBlenderbotEncoder(self.config, dtype=self.dtype, embed_tokens=self.shared)
+        self.decoder = FlaxBlenderbotDecoder(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 FlaxBlenderbotPreTrainedModel(FlaxPreTrainedModel):
+    config_class = BlenderbotConfig
+    base_model_prefix: str = "model"
+    module_class: nn.Module = None
+
+    def __init__(
+        self,
+        config: BlenderbotConfig,
+        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 FlaxBlenderbotForSequenceClassificationModule
+        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_ENCODE_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=FlaxBaseModelOutput, config_class=BlenderbotConfig)
+    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, FlaxBlenderbotForConditionalGeneration
+
+        >>> model = FlaxBlenderbotForConditionalGeneration.from_pretrained("facebook/blenderbot-400M-distill")
+        >>> tokenizer = AutoTokenizer.from_pretrained("facebook/blenderbot-400M-distill")
+
+        >>> text = "My friends are cool but they eat too many carbs."
+        >>> inputs = tokenizer(text, max_length=1024, return_tensors="jax")
+        >>> 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_DECODE_INPUTS_DOCSTRING)
+    @replace_return_docstrings(
+        output_type=FlaxBaseModelOutputWithPastAndCrossAttentions, config_class=BlenderbotConfig
+    )
+    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, FlaxBlenderbotForConditionalGeneration
+
+        >>> model = FlaxBlenderbotForConditionalGeneration.from_pretrained("facebook/blenderbot-400M-distill")
+        >>> tokenizer = AutoTokenizer.from_pretrained("facebook/blenderbot-400M-distill")
+
+        >>> text = "My friends are cool but they eat too many carbs."
+        >>> inputs = tokenizer(text, max_length=1024, return_tensors="jax")
+        >>> 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 FlaxBlenderbotAttention 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
+
+    @add_start_docstrings_to_model_forward(BLENDERBOT_INPUTS_DOCSTRING)
+    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 MBart Model transformer outputting raw hidden-states without any specific head on top.",
+    BLENDERBOT_START_DOCSTRING,
+)
+class FlaxBlenderbotModel(FlaxBlenderbotPreTrainedModel):
+    config: BlenderbotConfig
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
+    module_class = FlaxBlenderbotModule
+
+
+append_call_sample_docstring(FlaxBlenderbotModel, _CHECKPOINT_FOR_DOC, FlaxSeq2SeqModelOutput, _CONFIG_FOR_DOC)
+
+
+# Copied from transformers.models.bart.modeling_flax_bart.FlaxBartForConditionalGenerationModule with Bart->Blenderbot
+class FlaxBlenderbotForConditionalGenerationModule(nn.Module):
+    config: BlenderbotConfig
+    dtype: jnp.dtype = jnp.float32
+    bias_init: Callable[..., jnp.ndarray] = jax.nn.initializers.zeros
+
+    def setup(self):
+        self.model = FlaxBlenderbotModule(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 Model with a language modeling head. Can be used for summarization.", BLENDERBOT_START_DOCSTRING
+)
+class FlaxBlenderbotForConditionalGeneration(FlaxBlenderbotPreTrainedModel):
+    module_class = FlaxBlenderbotForConditionalGenerationModule
+    dtype: jnp.dtype = jnp.float32
+
+    @add_start_docstrings(BLENDERBOT_DECODE_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=FlaxCausalLMOutputWithCrossAttentions, config_class=BlenderbotConfig)
+    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, FlaxBlenderbotForConditionalGeneration
+
+        >>> model = FlaxBlenderbotForConditionalGeneration.from_pretrained("facebook/blenderbot-400M-distill")
+        >>> tokenizer = AutoTokenizer.from_pretrained("facebook/blenderbot-400M-distill")
+
+        >>> text = "My friends are cool but they eat too many carbs."
+        >>> inputs = tokenizer(text, max_length=1024, return_tensors="jax")
+        >>> 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 FlaxBlenderbotAttention 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
+            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=not train,
+            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_CONDITIONAL_GENERATION_DOCSTRING = r"""
+    Returns:
+
+    Conversation example::
+
+    ```py
+    >>> from transformers import AutoTokenizer, FlaxBlenderbotForConditionalGeneration
+
+    >>> model = FlaxBlenderbotForConditionalGeneration.from_pretrained("facebook/blenderbot-400M-distill")
+    >>> tokenizer = AutoTokenizer.from_pretrained("facebook/blenderbot-400M-distill")
+
+    >>> UTTERANCE = "My friends are cool but they eat too many carbs."
+    >>> inputs = tokenizer([UTTERANCE], max_length=1024, return_tensors="np")
+
+    >>> # Generate Reply
+    >>> reply_ids = model.generate(inputs["input_ids"], num_beams=4, max_length=5, early_stopping=True).sequences
+    >>> print([tokenizer.decode(g, skip_special_tokens=True, clean_up_tokenization_spaces=False) for g in reply_ids])
+    ```
+"""
+
+overwrite_call_docstring(
+    FlaxBlenderbotForConditionalGeneration,
+    BLENDERBOT_INPUTS_DOCSTRING + FLAX_BLENDERBOT_CONDITIONAL_GENERATION_DOCSTRING,
+)
+append_replace_return_docstrings(
+    FlaxBlenderbotForConditionalGeneration, output_type=FlaxSeq2SeqLMOutput, config_class=_CONFIG_FOR_DOC
+)
+
+
+__all__ = ["FlaxBlenderbotForConditionalGeneration", "FlaxBlenderbotModel", "FlaxBlenderbotPreTrainedModel"]

lib/python3.12/site-packages/transformers/models/blenderbot/modeling_tf_blenderbot.py

@@ -0,0 +1,1558 @@
+# 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 Blenderbot model."""
+
+from __future__ import annotations
+
+import os
+import random
+import warnings
+from typing import List, Optional, Tuple, Union
+
+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 import BlenderbotConfig
+
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "facebook/blenderbot-400M-distill"
+_CONFIG_FOR_DOC = "BlenderbotConfig"
+
+
+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
+
+
+class TFBlenderbotLearnedPositionalEmbedding(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->Blenderbot
+class TFBlenderbotAttention(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.mbart.modeling_tf_mbart.TFMBartEncoderLayer with MBart->Blenderbot
+class TFBlenderbotEncoderLayer(keras.layers.Layer):
+    def __init__(self, config: BlenderbotConfig, **kwargs):
+        super().__init__(**kwargs)
+        self.embed_dim = config.d_model
+        self.self_attn = TFBlenderbotAttention(
+            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: tf.Tensor,
+        layer_head_mask: tf.Tensor,
+        training: Optional[bool] = False,
+    ):
+        """
+        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.self_attn_layer_norm(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
+
+        residual = hidden_states
+        hidden_states = self.final_layer_norm(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
+
+        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.mbart.modeling_tf_mbart.TFMBartDecoderLayer with MBart->Blenderbot
+class TFBlenderbotDecoderLayer(keras.layers.Layer):
+    def __init__(self, config: BlenderbotConfig, **kwargs):
+        super().__init__(**kwargs)
+        self.embed_dim = config.d_model
+        self.self_attn = TFBlenderbotAttention(
+            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 = TFBlenderbotAttention(
+            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: tf.Tensor | None = None,
+        encoder_hidden_states: tf.Tensor | None = None,
+        encoder_attention_mask: tf.Tensor | None = None,
+        layer_head_mask: tf.Tensor | None = None,
+        cross_attn_layer_head_mask: tf.Tensor | None = None,
+        past_key_value: Tuple[tf.Tensor] | None = 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
+        hidden_states = self.self_attn_layer_norm(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
+
+        # Cross-Attention Block
+        cross_attn_present_key_value = None
+        cross_attn_weights = None
+        if encoder_hidden_states is not None:
+            residual = hidden_states
+            hidden_states = self.encoder_attn_layer_norm(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
+
+            # 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.final_layer_norm(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
+
+        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 TFBlenderbotPreTrainedModel(TFPreTrainedModel):
+    config_class = BlenderbotConfig
+    base_model_prefix = "model"
+
+
+BLENDERBOT_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 ([`BlenderbotConfig`]): 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_GENERATION_EXAMPLE = r"""
+    Conversation example::
+
+    ```py
+    >>> from transformers import AutoTokenizer, TFBlenderbotForConditionalGeneration
+
+    >>> mname = "facebook/blenderbot-400M-distill"
+    >>> model = TFBlenderbotForConditionalGeneration.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])
+
+    >>> REPLY = "I'm not sure"
+    >>> print("Human: ", REPLY)
+    >>> NEXT_UTTERANCE = (
+    ...     "My friends are cool but they eat too many carbs.</s> <s>That's unfortunate. "
+    ...     "Are they trying to lose weight or are they just trying to be healthier?</s> "
+    ...     "<s> I'm not sure."
+    ... )
+    >>> inputs = tokenizer([NEXT_UTTERANCE], return_tensors="tf")
+    >>> next_reply_ids = model.generate(**inputs)
+    >>> print("Bot: ", tokenizer.batch_decode(next_reply_ids, skip_special_tokens=True)[0])
+    ```
+"""
+
+BLENDERBOT_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)
+
+            Blenderbot 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 TFBlenderbotEncoder(keras.layers.Layer):
+    config_class = BlenderbotConfig
+    """
+    Transformer encoder consisting of *config.encoder_layers* self attention layers. Each layer is a
+    [`TFBlenderbotEncoderLayer`].
+
+    Args:
+        config: BlenderbotConfig
+    """
+
+    def __init__(self, config: BlenderbotConfig, 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 = TFBlenderbotLearnedPositionalEmbedding(
+            config.max_position_embeddings,
+            config.d_model,
+            name="embed_positions",
+        )
+        self.layers = [TFBlenderbotEncoderLayer(config, name=f"layers.{i}") for i in range(config.encoder_layers)]
+        self.layer_norm = keras.layers.LayerNormalization(epsilon=1e-5, name="layer_norm")
+
+    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.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,)
+
+        hidden_states = self.layer_norm(hidden_states)
+
+        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, "layer_norm", None) is not None:
+            with tf.name_scope(self.layer_norm.name):
+                self.layer_norm.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 TFBlenderbotDecoder(keras.layers.Layer):
+    config_class = BlenderbotConfig
+    """
+    Transformer decoder consisting of *config.decoder_layers* layers. Each layer is a [`TFBlenderbotDecoderLayer`]
+
+    Args:
+        config: BlenderbotConfig
+        embed_tokens: output embedding
+    """
+
+    def __init__(self, config: BlenderbotConfig, 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 = TFBlenderbotLearnedPositionalEmbedding(
+            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 = [TFBlenderbotDecoderLayer(config, name=f"layers.{i}") for i in range(config.decoder_layers)]
+        self.layer_norm = keras.layers.LayerNormalization(epsilon=1e-5, name="layer_norm")
+
+        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
+
+        # 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)
+
+        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
+
+        hidden_states = inputs_embeds
+
+        # [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])
+
+        hidden_states = hidden_states + 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,)
+
+        hidden_states = self.layer_norm(hidden_states)
+
+        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, "layer_norm", None) is not None:
+            with tf.name_scope(self.layer_norm.name):
+                self.layer_norm.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 TFBlenderbotMainLayer(keras.layers.Layer):
+    config_class = BlenderbotConfig
+
+    def __init__(self, config: BlenderbotConfig, **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 = TFBlenderbotEncoder(config, self.shared, name="encoder")
+        self.decoder = TFBlenderbotDecoder(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 Model outputting raw hidden-states without any specific head on top.",
+    BLENDERBOT_START_DOCSTRING,
+)
+class TFBlenderbotModel(TFBlenderbotPreTrainedModel):
+    def __init__(self, config: BlenderbotConfig, *inputs, **kwargs):
+        super().__init__(config, *inputs, **kwargs)
+
+        self.model = TFBlenderbotMainLayer(config, name="model")
+
+    def get_encoder(self):
+        return self.model.encoder
+
+    def get_decoder(self):
+        return self.model.decoder
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], *model_args, **kwargs):
+        if pretrained_model_name_or_path == "facebook/blenderbot-90M":
+            from ..blenderbot_small import TFBlenderbotSmallModel
+
+            warnings.warn(
+                "The checkpoint `facebook/blenderbot-90M` is deprecated. In the future, please use the identical"
+                " checkpoint `facebook/small_blenderbot-90M` with"
+                " `TFBlenderbotSmallForConditionalGeneration.from_pretrained('facebook/small_blenderbot-90M')`"
+                " instead.",
+                FutureWarning,
+            )
+            return TFBlenderbotSmallModel.from_pretrained(pretrained_model_name_or_path)
+
+        return super().from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+
+    @unpack_inputs
+    @add_start_docstrings_to_model_forward(BLENDERBOT_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 Model with a language modeling head. Can be used for summarization.",
+    BLENDERBOT_START_DOCSTRING,
+)
+class TFBlenderbotForConditionalGeneration(TFBlenderbotPreTrainedModel, 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 = TFBlenderbotMainLayer(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"])
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], *model_args, **kwargs):
+        if pretrained_model_name_or_path == "facebook/blenderbot-90M":
+            from ..blenderbot_small import TFBlenderbotSmallForConditionalGeneration
+
+            warnings.warn(
+                "The checkpoint `facebook/blenderbot-90M` is deprecated. In the future, please use the identical"
+                " checkpoint `facebook/small_blenderbot-90M` with"
+                " `TFBlenderbotSmallForConditionalGeneration.from_pretrained('facebook/small_blenderbot-90M')`"
+                " instead.",
+                FutureWarning,
+            )
+            return TFBlenderbotSmallForConditionalGeneration.from_pretrained(pretrained_model_name_or_path)
+
+        return super().from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+
+    @unpack_inputs
+    @add_start_docstrings_to_model_forward(BLENDERBOT_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=TFSeq2SeqLMOutput, config_class=_CONFIG_FOR_DOC)
+    @add_end_docstrings(BLENDERBOT_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[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,
+        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,
+            encoder_outputs=encoder_outputs,
+            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,
+            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__ = ["TFBlenderbotForConditionalGeneration", "TFBlenderbotModel", "TFBlenderbotPreTrainedModel"]

lib/python3.12/site-packages/transformers/models/blenderbot/tokenization_blenderbot.py

@@ -0,0 +1,410 @@
+# 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 Blenderbot."""
+
+import json
+import os
+from functools import lru_cache
+from typing import List, Optional, Tuple
+
+import regex as re
+
+from ...tokenization_utils import AddedToken, 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",
+}
+
+
+@lru_cache()
+# Copied from transformers.models.roberta.tokenization_roberta.bytes_to_unicode
+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))
+
+
+# Copied from transformers.models.roberta.tokenization_roberta.get_pairs
+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 BlenderbotTokenizer(PreTrainedTokenizer):
+    """
+    Constructs a Blenderbot tokenizer, derived from the GPT-2 tokenizer, using 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 BlenderbotTokenizer
+
+    >>> tokenizer = BlenderbotTokenizer.from_pretrained("facebook/blenderbot-3B")
+    >>> tokenizer.add_prefix_space = False
+    >>> tokenizer("Hello world")["input_ids"]
+    [47, 921, 86, 1085, 2]
+
+    >>> tokenizer(" Hello world")["input_ids"]
+    [6950, 1085, 2]
+    ```
+
+    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.
+        bos_token (`str`, *optional*, defaults to `"<s>"`):
+            The beginning of sequence token that was used during pretraining. Can be used a sequence classifier token.
+
+            <Tip>
+
+            When building a sequence using special tokens, this is not the token that is used for the beginning of
+            sequence. The token used is the `cls_token`.
+
+            </Tip>
+
+        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>
+
+        sep_token (`str`, *optional*, defaults to `"</s>"`):
+            The separator token, which is used when building a sequence from multiple sequences, e.g. two sequences for
+            sequence classification or for a text and a question for question answering. It is also used as the last
+            token of a sequence built with special tokens.
+        cls_token (`str`, *optional*, defaults to `"<s>"`):
+            The classifier token which is used when doing sequence classification (classification of the whole sequence
+            instead of per-token classification). It is the first token of the sequence when built with special tokens.
+        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.
+        mask_token (`str`, *optional*, defaults to `"<mask>"`):
+            The token used for masking values. This is the token used when training this model with masked language
+            modeling. This is the token which the model will try to predict.
+        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. (Blenderbot tokenizer detect beginning of words by the preceding space).
+    """
+
+    vocab_files_names = VOCAB_FILES_NAMES
+    model_input_names = ["input_ids", "attention_mask"]
+
+    # Copied from transformers.models.roberta.tokenization_roberta.RobertaTokenizer.__init__ with Roberta->Blenderbot, RoBERTa->Blenderbot
+    def __init__(
+        self,
+        vocab_file,
+        merges_file,
+        errors="replace",
+        bos_token="<s>",
+        eos_token="</s>",
+        sep_token="</s>",
+        cls_token="<s>",
+        unk_token="<unk>",
+        pad_token="<pad>",
+        mask_token="<mask>",
+        add_prefix_space=False,
+        **kwargs,
+    ):
+        bos_token = AddedToken(bos_token, lstrip=False, rstrip=False) if isinstance(bos_token, str) else bos_token
+        pad_token = AddedToken(pad_token, lstrip=False, rstrip=False) if isinstance(pad_token, str) else pad_token
+        eos_token = AddedToken(eos_token, lstrip=False, rstrip=False) if isinstance(eos_token, str) else eos_token
+        unk_token = AddedToken(unk_token, lstrip=False, rstrip=False) if isinstance(unk_token, str) else unk_token
+        sep_token = AddedToken(sep_token, lstrip=False, rstrip=False) if isinstance(sep_token, str) else sep_token
+        cls_token = AddedToken(cls_token, lstrip=False, rstrip=False) if isinstance(cls_token, str) else cls_token
+
+        # Mask token behave like a normal word, i.e. include the space before it
+        mask_token = (
+            AddedToken(mask_token, lstrip=True, rstrip=False, normalized=False)
+            if isinstance(mask_token, str)
+            else mask_token
+        )
+
+        # these special tokens are not part of the vocab.json, let's add them in the correct order
+
+        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,
+            bos_token=bos_token,
+            eos_token=eos_token,
+            unk_token=unk_token,
+            sep_token=sep_token,
+            cls_token=cls_token,
+            pad_token=pad_token,
+            mask_token=mask_token,
+            add_prefix_space=add_prefix_space,
+            **kwargs,
+        )
+
+    @property
+    # Copied from transformers.models.roberta.tokenization_roberta.RobertaTokenizer.vocab_size with Roberta->Blenderbot, RoBERTa->Blenderbot
+    def vocab_size(self):
+        return len(self.encoder)
+
+    # Copied from transformers.models.roberta.tokenization_roberta.RobertaTokenizer.get_vocab with Roberta->Blenderbot, RoBERTa->Blenderbot
+    def get_vocab(self):
+        vocab = dict(self.encoder).copy()
+        vocab.update(self.added_tokens_encoder)
+        return vocab
+
+    # Copied from transformers.models.roberta.tokenization_roberta.RobertaTokenizer.bpe with Roberta->Blenderbot, RoBERTa->Blenderbot
+    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
+
+    # Copied from transformers.models.roberta.tokenization_roberta.RobertaTokenizer._tokenize with Roberta->Blenderbot, RoBERTa->Blenderbot
+    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
+
+    # Copied from transformers.models.roberta.tokenization_roberta.RobertaTokenizer._convert_token_to_id with Roberta->Blenderbot, RoBERTa->Blenderbot
+    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))
+
+    # Copied from transformers.models.roberta.tokenization_roberta.RobertaTokenizer._convert_id_to_token with Roberta->Blenderbot, RoBERTa->Blenderbot
+    def _convert_id_to_token(self, index):
+        """Converts an index (integer) in a token (str) using the vocab."""
+        return self.decoder.get(index)
+
+    # Copied from transformers.models.roberta.tokenization_roberta.RobertaTokenizer.convert_tokens_to_string with Roberta->Blenderbot, RoBERTa->Blenderbot
+    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
+
+    # Copied from transformers.models.roberta.tokenization_roberta.RobertaTokenizer.save_vocabulary with Roberta->Blenderbot, RoBERTa->Blenderbot
+    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
+
+    # Copied from transformers.models.roberta.tokenization_roberta.RobertaTokenizer.get_special_tokens_mask with Roberta->Blenderbot, RoBERTa->Blenderbot
+    def get_special_tokens_mask(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None, already_has_special_tokens: bool = False
+    ) -> List[int]:
+        """
+        Retrieve sequence ids from a token list that has no special tokens added. This method is called when adding
+        special tokens using the tokenizer `prepare_for_model` method.
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+            already_has_special_tokens (`bool`, *optional*, defaults to `False`):
+                Whether or not the token list is already formatted with special tokens for the model.
+
+        Returns:
+            `List[int]`: A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
+        """
+        if already_has_special_tokens:
+            return super().get_special_tokens_mask(
+                token_ids_0=token_ids_0, token_ids_1=token_ids_1, already_has_special_tokens=True
+            )
+
+        if token_ids_1 is None:
+            return [1] + ([0] * len(token_ids_0)) + [1]
+        return [1] + ([0] * len(token_ids_0)) + [1, 1] + ([0] * len(token_ids_1)) + [1]
+
+    # Copied from transformers.models.roberta.tokenization_roberta.RobertaTokenizer.create_token_type_ids_from_sequences with Roberta->Blenderbot, RoBERTa->Blenderbot
+    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. Blenderbot 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]
+
+    # Copied from transformers.models.roberta.tokenization_roberta.RobertaTokenizer.prepare_for_tokenization with Roberta->Blenderbot, RoBERTa->Blenderbot
+    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) and (len(text) > 0 and not text[0].isspace()):
+            text = " " + text
+        return (text, kwargs)
+
+    def build_inputs_with_special_tokens(self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None):
+        """
+        Build model inputs from a sequence or a pair of sequence for sequence classification tasks by concatenating and
+        adding special tokens. A Blenderbot sequence has the following format:
+        - single sequence: ` X </s>`
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs to which the special tokens will be added
+            token_ids_1 (`List[int]`, *optional*):
+                Will be ignored
+        Returns:
+            `List[int]`: list of [input IDs](../glossary#input-ids) with the appropriate special tokens.
+        """
+        return token_ids_0 + [self.eos_token_id]
+
+
+__all__ = ["BlenderbotTokenizer"]

lib/python3.12/site-packages/transformers/models/blenderbot/tokenization_blenderbot_fast.py

@@ -0,0 +1,284 @@
+# 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 Blenderbot."""
+
+import json
+from typing import List, Optional, Tuple
+
+from tokenizers import processors
+
+from ...tokenization_utils_base import AddedToken, BatchEncoding
+from ...tokenization_utils_fast import PreTrainedTokenizerFast
+from ...utils import logging
+from .tokenization_blenderbot import BlenderbotTokenizer
+
+
+logger = logging.get_logger(__name__)
+
+
+VOCAB_FILES_NAMES = {
+    "vocab_file": "vocab.json",
+    "merges_file": "merges.txt",
+    "tokenizer_config_file": "tokenizer_config.json",
+}
+
+
+class BlenderbotTokenizerFast(PreTrainedTokenizerFast):
+    """
+    Construct a "fast" Blenderbot tokenizer (backed by HuggingFace's *tokenizers* library), derived from the GPT-2
+    tokenizer, using 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 BlenderbotTokenizerFast
+
+    >>> tokenizer = BlenderbotTokenizerFast.from_pretrained("facebook/blenderbot-3B")
+    >>> tokenizer("Hello world")["input_ids"]
+    [6950, 1085, 2]
+
+    >>> tokenizer(" Hello world")["input_ids"]
+    [6950, 1085, 2]
+    ```
+
+    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 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`):
+            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.
+        bos_token (`str`, *optional*, defaults to `"<s>"`):
+            The beginning of sequence token that was used during pretraining. Can be used a sequence classifier token.
+
+            <Tip>
+
+            When building a sequence using special tokens, this is not the token that is used for the beginning of
+            sequence. The token used is the `cls_token`.
+
+            </Tip>
+
+        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>
+
+        sep_token (`str`, *optional*, defaults to `"</s>"`):
+            The separator token, which is used when building a sequence from multiple sequences, e.g. two sequences for
+            sequence classification or for a text and a question for question answering. It is also used as the last
+            token of a sequence built with special tokens.
+        cls_token (`str`, *optional*, defaults to `"<s>"`):
+            The classifier token which is used when doing sequence classification (classification of the whole sequence
+            instead of per-token classification). It is the first token of the sequence when built with special tokens.
+        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.
+        mask_token (`str`, *optional*, defaults to `"<mask>"`):
+            The token used for masking values. This is the token used when training this model with masked language
+            modeling. This is the token which the model will try to predict.
+        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. (Blenderbot tokenizer detect beginning of words by the preceding space).
+        trim_offsets (`bool`, *optional*, defaults to `True`):
+            Whether the post processing step should trim offsets to avoid including whitespaces.
+    """
+
+    vocab_files_names = VOCAB_FILES_NAMES
+    model_input_names = ["input_ids", "attention_mask"]
+    slow_tokenizer_class = BlenderbotTokenizer
+
+    # Copied from transformers.models.roberta.tokenization_roberta_fast.RobertaTokenizerFast.__init__ with Roberta->Blenderbot, RoBERTa->Blenderbot
+    def __init__(
+        self,
+        vocab_file=None,
+        merges_file=None,
+        tokenizer_file=None,
+        errors="replace",
+        bos_token="<s>",
+        eos_token="</s>",
+        sep_token="</s>",
+        cls_token="<s>",
+        unk_token="<unk>",
+        pad_token="<pad>",
+        mask_token="<mask>",
+        add_prefix_space=False,
+        trim_offsets=True,
+        **kwargs,
+    ):
+        mask_token = (
+            AddedToken(mask_token, lstrip=True, rstrip=False, normalized=False)
+            if isinstance(mask_token, str)
+            else mask_token
+        )
+        super().__init__(
+            vocab_file,
+            merges_file,
+            tokenizer_file=tokenizer_file,
+            errors=errors,
+            bos_token=bos_token,
+            eos_token=eos_token,
+            sep_token=sep_token,
+            cls_token=cls_token,
+            unk_token=unk_token,
+            pad_token=pad_token,
+            mask_token=mask_token,
+            add_prefix_space=add_prefix_space,
+            trim_offsets=trim_offsets,
+            **kwargs,
+        )
+
+        tokenizer_component = "post_processor"
+        tokenizer_component_instance = getattr(self.backend_tokenizer, tokenizer_component, None)
+        if tokenizer_component_instance:
+            state = json.loads(tokenizer_component_instance.__getstate__())
+
+            # The lists 'sep' and 'cls' must be cased in tuples for the object `post_processor_class`
+            if "sep" in state:
+                state["sep"] = tuple(state["sep"])
+            if "cls" in state:
+                state["cls"] = tuple(state["cls"])
+
+            changes_to_apply = False
+
+            if state.get("add_prefix_space", add_prefix_space) != add_prefix_space:
+                state["add_prefix_space"] = add_prefix_space
+                changes_to_apply = True
+
+            if state.get("trim_offsets", trim_offsets) != trim_offsets:
+                state["trim_offsets"] = trim_offsets
+                changes_to_apply = True
+
+            if changes_to_apply:
+                component_class = getattr(processors, state.pop("type"))
+                new_value = component_class(**state)
+                setattr(self.backend_tokenizer, tokenizer_component, new_value)
+
+    @property
+    # Copied from transformers.models.roberta.tokenization_roberta_fast.RobertaTokenizerFast.mask_token with Roberta->Blenderbot, RoBERTa->Blenderbot
+    def mask_token(self) -> str:
+        """
+        `str`: Mask token, to use when training a model with masked-language modeling. Log an error if used while not
+        having been set.
+
+        Blenderbot tokenizer has a special mask token to be usable in the fill-mask pipeline. The mask token will greedily
+        comprise the space before the *<mask>*.
+        """
+        if self._mask_token is None:
+            if self.verbose:
+                logger.error("Using mask_token, but it is not set yet.")
+            return None
+        return str(self._mask_token)
+
+    @mask_token.setter
+    def mask_token(self, value):
+        """
+        Overriding the default behavior of the mask token to have it eat the space before it.
+
+        This is needed to preserve backward compatibility with all the previously used models based on Roberta.
+        """
+        # Mask token behave like a normal word, i.e. include the space before it
+        # So we set lstrip to True
+        value = AddedToken(value, lstrip=True, rstrip=False) if isinstance(value, str) else value
+        self._mask_token = value
+
+    # Copied from transformers.models.roberta.tokenization_roberta_fast.RobertaTokenizerFast._batch_encode_plus with Roberta->Blenderbot, RoBERTa->Blenderbot
+    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)
+
+    # Copied from transformers.models.roberta.tokenization_roberta_fast.RobertaTokenizerFast._encode_plus with Roberta->Blenderbot, RoBERTa->Blenderbot
+    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.roberta.tokenization_roberta_fast.RobertaTokenizerFast.save_vocabulary with Roberta->Blenderbot, RoBERTa->Blenderbot
+    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)
+
+    # Copied from transformers.models.roberta.tokenization_roberta_fast.RobertaTokenizerFast.create_token_type_ids_from_sequences with Roberta->Blenderbot, RoBERTa->Blenderbot
+    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. Blenderbot 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]
+
+    def build_inputs_with_special_tokens(self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None):
+        """
+        Build model inputs from a sequence or a pair of sequence for sequence classification tasks by concatenating and
+        adding special tokens. A Blenderbot sequence has the following format:
+        - single sequence: ` X </s>`
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs to which the special tokens will be added
+            token_ids_1 (`List[int]`, *optional*):
+                Will be ignored
+        Returns:
+            `List[int]`: list of [input IDs](../glossary#input-ids) with the appropriate special tokens.
+        """
+        return token_ids_0 + [self.eos_token_id]
+
+
+__all__ = ["BlenderbotTokenizerFast"]

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

@@ -0,0 +1,27 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import TYPE_CHECKING
+
+from ...utils import _LazyModule
+from ...utils.import_utils import define_import_structure
+
+
+if TYPE_CHECKING:
+    from .tokenization_code_llama import *
+    from .tokenization_code_llama_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/code_llama/tokenization_code_llama.py

@@ -0,0 +1,452 @@
+# coding=utf-8
+# Copyright 2023 MetaAI 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 classes for Code LLaMA."""
+
+import os
+from shutil import copyfile
+from typing import Any, Dict, List, Optional, Tuple
+
+import sentencepiece as spm
+
+from ...convert_slow_tokenizer import import_protobuf
+from ...tokenization_utils import AddedToken, PreTrainedTokenizer
+from ...utils import logging, requires_backends
+
+
+logger = logging.get_logger(__name__)
+
+VOCAB_FILES_NAMES = {"vocab_file": "tokenizer.model"}
+
+SPIECE_UNDERLINE = "▁"
+
+B_INST, E_INST = "[INST]", "[/INST]"
+B_SYS, E_SYS = "<<SYS>>\n", "\n<</SYS>>\n\n"
+
+# fmt: off
+DEFAULT_SYSTEM_PROMPT = """You are a helpful, respectful and honest assistant. Always answer as helpfully as possible, while being safe. Your \
+answers should not include any harmful, unethical, racist, sexist, toxic, dangerous, or illegal content. Please ensure\
+ that your responses are socially unbiased and positive in nature.
+
+If a question does not make any sense, or is not factually coherent, explain why instead of answering something not \
+correct. If you don't know the answer to a question, please don't share false information."""
+# fmt: on
+
+
+class CodeLlamaTokenizer(PreTrainedTokenizer):
+    """
+    Construct a CodeLlama tokenizer. Based on byte-level Byte-Pair-Encoding. The default padding token is unset as
+    there is no padding token in the original model.
+
+    The default configuration match that of
+    [codellama/CodeLlama-7b-Instruct-hf](https://huggingface.co/meta-llama/CodeLlama-7b-Instruct-hf/blob/main/tokenizer_config.json)
+    which supports prompt infilling.
+
+    Args:
+        vocab_file (`str`):
+            Path to the vocabulary file.
+        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.
+        bos_token (`str`, *optional*, defaults to `"<s>"`):
+            The beginning of sequence token that was used during pretraining. Can be used a sequence classifier token.
+        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>
+
+        prefix_token (`str`, *optional*, defaults to `"▁<PRE>"`):
+            Prefix token used for infilling.
+        middle_token (`str`, *optional*, defaults to `"▁<MID>"`):
+            Middle token used for infilling.
+        suffix_token (`str`, *optional*, defaults to `"▁<SUF>"`):
+            Suffix token used for infilling.
+        eot_token (`str`, *optional*, defaults to `"▁<EOT>"`):
+            End of text token used for infilling.
+        fill_token (`str`, *optional*, defaults to `"<FILL_ME>"`):
+            The token used to split the input between the prefix and suffix.
+        suffix_first (`bool`, *optional*, defaults to `False`):
+            Whether the input prompt and suffix should be formatted with the suffix first.
+        sp_model_kwargs (`dict`, *optional*):
+            Will be passed to the `SentencePieceProcessor.__init__()` method. The [Python wrapper for
+            SentencePiece](https://github.com/google/sentencepiece/tree/master/python) can be used, among other things,
+            to set:
+
+            - `enable_sampling`: Enable subword regularization.
+            - `nbest_size`: Sampling parameters for unigram. Invalid for BPE-Dropout.
+
+              - `nbest_size = {0,1}`: No sampling is performed.
+              - `nbest_size > 1`: samples from the nbest_size results.
+              - `nbest_size < 0`: assuming that nbest_size is infinite and samples from the all hypothesis (lattice)
+                using forward-filtering-and-backward-sampling algorithm.
+
+            - `alpha`: Smoothing parameter for unigram sampling, and dropout probability of merge operations for
+              BPE-dropout.
+        add_bos_token (`bool`, *optional*, defaults to `True`):
+            Whether to add a beginning of sequence token at the start of sequences.
+        add_eos_token (`bool`, *optional*, defaults to `False`):
+            Whether to add an end of sequence token at the end of sequences.
+        clean_up_tokenization_spaces (`bool`, *optional*, defaults to `False`):
+            Whether or not to clean up the tokenization spaces.
+        additional_special_tokens (`List[str]`, *optional*):
+            Additional special tokens used by the tokenizer.
+        use_default_system_prompt (`bool`, *optional*, defaults to `False`):
+            Whether or not the default system prompt for Llama should be used.
+    """
+
+    vocab_files_names = VOCAB_FILES_NAMES
+    model_input_names = ["input_ids", "attention_mask"]
+
+    def __init__(
+        self,
+        vocab_file,
+        unk_token="<unk>",
+        bos_token="<s>",
+        eos_token="</s>",
+        prefix_token="▁<PRE>",
+        middle_token="▁<MID>",
+        suffix_token="▁<SUF>",
+        eot_token="▁<EOT>",
+        fill_token="<FILL_ME>",
+        suffix_first=False,
+        sp_model_kwargs: Optional[Dict[str, Any]] = None,
+        add_bos_token=True,
+        add_eos_token=False,
+        clean_up_tokenization_spaces=False,
+        additional_special_tokens=None,
+        use_default_system_prompt=False,
+        **kwargs,
+    ):
+        requires_backends(self, "protobuf")
+        self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs
+        bos_token = AddedToken(bos_token, normalized=False, special=True) if isinstance(bos_token, str) else bos_token
+        eos_token = AddedToken(eos_token, normalized=False, special=True) if isinstance(eos_token, str) else eos_token
+        unk_token = AddedToken(unk_token, normalized=False, special=True) if isinstance(unk_token, str) else unk_token
+
+        self.use_default_system_prompt = use_default_system_prompt
+        # mark tokens special to skip them
+        additional_special_tokens = additional_special_tokens or []
+        for token in [prefix_token, middle_token, suffix_token, eot_token]:
+            additional_special_tokens += [token] if token is not None else []
+
+        self.vocab_file = vocab_file
+        self.add_bos_token = add_bos_token
+        self.add_eos_token = add_eos_token
+        self._prefix_token = prefix_token
+        self._middle_token = middle_token
+        self._suffix_token = suffix_token
+        self._eot_token = eot_token
+        self.fill_token = fill_token
+        self.suffix_first = suffix_first
+        self.sp_model = self.get_spm_processor()
+
+        super().__init__(
+            bos_token=bos_token,
+            eos_token=eos_token,
+            unk_token=unk_token,
+            add_bos_token=add_bos_token,
+            add_eos_token=add_eos_token,
+            prefix_token=prefix_token,
+            middle_token=middle_token,
+            suffix_token=suffix_token,
+            eot_token=eot_token,
+            fill_token=fill_token,
+            sp_model_kwargs=self.sp_model_kwargs,
+            suffix_first=suffix_first,
+            clean_up_tokenization_spaces=clean_up_tokenization_spaces,
+            additional_special_tokens=additional_special_tokens,
+            use_default_system_prompt=use_default_system_prompt,
+            **kwargs,
+        )
+
+    @property
+    def unk_token_length(self):
+        return len(self.sp_model.encode(str(self.unk_token)))
+
+    def get_spm_processor(self):
+        tokenizer = spm.SentencePieceProcessor(**self.sp_model_kwargs)
+        with open(self.vocab_file, "rb") as f:
+            sp_model = f.read()
+            model_pb2 = import_protobuf()
+            model = model_pb2.ModelProto.FromString(sp_model)
+            normalizer_spec = model_pb2.NormalizerSpec()
+            normalizer_spec.add_dummy_prefix = False
+            model.normalizer_spec.MergeFrom(normalizer_spec)
+            sp_model = model.SerializeToString()
+            tokenizer.LoadFromSerializedProto(sp_model)
+        return tokenizer
+
+    @property
+    def prefix_token(self):
+        return self._prefix_token
+
+    @property
+    def prefix_id(self):
+        if self._prefix_token is None:
+            return None
+        return self.convert_tokens_to_ids(self.prefix_token)
+
+    @property
+    def middle_token(self):
+        return self._middle_token
+
+    @property
+    def middle_id(self):
+        if self._middle_token is None:
+            return None
+        return self.convert_tokens_to_ids(self.middle_token)
+
+    @property
+    def suffix_token(self):
+        return self._suffix_token
+
+    @property
+    def suffix_id(self):
+        if self._suffix_token is None:
+            return None
+        return self.convert_tokens_to_ids(self.suffix_token)
+
+    @property
+    def eot_token(self):
+        return self._eot_token
+
+    @property
+    def eot_id(self):
+        if self._eot_token is None:
+            return None
+        return self.convert_tokens_to_ids(self.eot_token)
+
+    @property
+    def vocab_size(self):
+        """Returns vocab size"""
+        return self.sp_model.get_piece_size()
+
+    # Copied from transformers.models.llama.tokenization_llama.LlamaTokenizer.get_vocab
+    def get_vocab(self):
+        """Returns vocab as a dict"""
+        vocab = {self.convert_ids_to_tokens(i): i for i in range(self.vocab_size)}
+        vocab.update(self.added_tokens_encoder)
+        return vocab
+
+    def tokenize(self, prefix, suffix=None, suffix_first=False, **kwargs) -> List[int]:
+        # add a prefix space to `prefix`
+        if self.fill_token is not None and self.fill_token in prefix and suffix is None:
+            prefix, suffix = prefix.split(self.fill_token)
+
+        if len(prefix) > 0:
+            prefix = SPIECE_UNDERLINE + prefix.replace(SPIECE_UNDERLINE, " ")
+
+        if suffix is None or len(suffix) < 1:
+            tokens = super().tokenize(prefix, **kwargs)
+            if len(tokens) > 1 and tokens[0] == SPIECE_UNDERLINE and tokens[1] in self.all_special_tokens:
+                tokens = tokens[1:]
+            return tokens
+
+        prefix_tokens = self._tokenize(prefix)  # prefix has an extra `SPIECE_UNDERLINE`
+
+        if None in (self.prefix_id, self.middle_id, self.suffix_id):
+            raise ValueError(
+                "The input either includes a `prefix` and a `suffix` used for the infilling task,"
+                f"  or can be split on the {self.fill_token} token, creating a suffix and prefix,"
+                " but the model does not support `infilling`."
+            )
+        suffix_tokens = self._tokenize(suffix)  # make sure CodeLlama sp model does not mess up
+
+        suffix_first = suffix_first if suffix_first is not None else self.suffix_first
+        if suffix_first:
+            # format as " <PRE> <SUF>{suf} <MID> {pre}"
+            return [self.prefix_token, self.suffix_token] + suffix_tokens + [self.middle_token] + prefix_tokens
+        else:
+            # format as " <PRE> {pre} <SUF>{suf} <MID>"
+            return [self.prefix_token] + prefix_tokens + [self.suffix_token] + suffix_tokens + [self.middle_token]
+
+    def _tokenize(self, text, **kwargs):
+        """
+        Returns a tokenized string.
+
+        We de-activated the `add_dummy_prefix` option, thus the sentencepiece internals will always strip any
+        SPIECE_UNDERLINE. For example: `self.sp_model.encode(f"{SPIECE_UNDERLINE}Hey", out_type = str)` will give
+        `['H', 'e', 'y']` instead of `['▁He', 'y']`. Thus we always encode `f"{unk_token}text"` and strip the
+        `unk_token`. Here is an example with `unk_token = "<unk>"` and `unk_token_length = 4`.
+        `self.tokenizer.sp_model.encode("<unk> Hey", out_type = str)[4:]`.
+        """
+        tokens = self.sp_model.encode(text, out_type=str)
+        if not text.startswith((SPIECE_UNDERLINE, " ")):
+            return tokens
+        # 1. Encode string + prefix ex: "<unk> Hey"
+        tokens = self.sp_model.encode(self.unk_token + text, out_type=str)
+        # 2. Remove self.unk_token from ['<','unk','>', '▁Hey']
+        return tokens[self.unk_token_length :] if len(tokens) >= self.unk_token_length else tokens
+
+    # Copied from transformers.models.llama.tokenization_llama.LlamaTokenizer._convert_token_to_id
+    def _convert_token_to_id(self, token):
+        """Converts a token (str) in an id using the vocab."""
+        return self.sp_model.piece_to_id(token)
+
+    # Copied from transformers.models.llama.tokenization_llama.LlamaTokenizer._convert_id_to_token
+    def _convert_id_to_token(self, index):
+        """Converts an index (integer) in a token (str) using the vocab."""
+        token = self.sp_model.IdToPiece(index)
+        return token
+
+    def convert_tokens_to_string(self, tokens):
+        """Converts a sequence of tokens (string) in a single string."""
+        # since we manually add the prefix space, we have to remove it when decoding
+        if tokens[0].startswith(SPIECE_UNDERLINE):
+            tokens[0] = tokens[0][1:]
+
+        current_sub_tokens = []
+        out_string = ""
+        for _, token in enumerate(tokens):
+            # make sure that special tokens are not decoded using sentencepiece model
+            if token in self.all_special_tokens:
+                out_string += self.sp_model.decode(current_sub_tokens) + token
+                current_sub_tokens = []
+            else:
+                current_sub_tokens.append(token)
+        out_string += self.sp_model.decode(current_sub_tokens)
+        return out_string
+
+    # Copied from transformers.models.llama.tokenization_llama.LlamaTokenizer.save_vocabulary
+    def save_vocabulary(self, save_directory, filename_prefix: Optional[str] = None) -> Tuple[str]:
+        """
+        Save the vocabulary and special tokens file to a directory.
+
+        Args:
+            save_directory (`str`):
+                The directory in which to save the vocabulary.
+
+        Returns:
+            `Tuple(str)`: Paths to the files saved.
+        """
+        if not os.path.isdir(save_directory):
+            logger.error(f"Vocabulary path ({save_directory}) should be a directory")
+            return
+        out_vocab_file = os.path.join(
+            save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]
+        )
+
+        if os.path.abspath(self.vocab_file) != os.path.abspath(out_vocab_file) and os.path.isfile(self.vocab_file):
+            copyfile(self.vocab_file, out_vocab_file)
+        elif not os.path.isfile(self.vocab_file):
+            with open(out_vocab_file, "wb") as fi:
+                content_spiece_model = self.sp_model.serialized_model_proto()
+                fi.write(content_spiece_model)
+
+        return (out_vocab_file,)
+
+    # Copied from transformers.models.llama.tokenization_llama.LlamaTokenizer.build_inputs_with_special_tokens
+    def build_inputs_with_special_tokens(self, token_ids_0, token_ids_1=None):
+        bos_token_id = [self.bos_token_id] if self.add_bos_token else []
+        eos_token_id = [self.eos_token_id] if self.add_eos_token else []
+
+        output = bos_token_id + token_ids_0 + eos_token_id
+
+        if token_ids_1 is not None:
+            output = output + bos_token_id + token_ids_1 + eos_token_id
+
+        return output
+
+    # Copied from transformers.models.llama.tokenization_llama.LlamaTokenizer.get_special_tokens_mask
+    def get_special_tokens_mask(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None, already_has_special_tokens: bool = False
+    ) -> List[int]:
+        """
+        Retrieve sequence ids from a token list that has no special tokens added. This method is called when adding
+        special tokens using the tokenizer `prepare_for_model` method.
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+            already_has_special_tokens (`bool`, *optional*, defaults to `False`):
+                Whether or not the token list is already formatted with special tokens for the model.
+
+        Returns:
+            `List[int]`: A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
+        """
+        if already_has_special_tokens:
+            return super().get_special_tokens_mask(
+                token_ids_0=token_ids_0, token_ids_1=token_ids_1, already_has_special_tokens=True
+            )
+
+        bos_token_id = [1] if self.add_bos_token else []
+        eos_token_id = [1] if self.add_eos_token else []
+
+        if token_ids_1 is None:
+            return bos_token_id + ([0] * len(token_ids_0)) + eos_token_id
+        return (
+            bos_token_id
+            + ([0] * len(token_ids_0))
+            + eos_token_id
+            + bos_token_id
+            + ([0] * len(token_ids_1))
+            + eos_token_id
+        )
+
+    # Copied from transformers.models.llama.tokenization_llama.LlamaTokenizer.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]:
+        """
+        Creates a mask from the two sequences passed to be used in a sequence-pair classification task. An ALBERT
+        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, 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).
+        """
+        bos_token_id = [self.bos_token_id] if self.add_bos_token else []
+        eos_token_id = [self.eos_token_id] if self.add_eos_token else []
+
+        output = [0] * len(bos_token_id + token_ids_0 + eos_token_id)
+
+        if token_ids_1 is not None:
+            output += [1] * len(bos_token_id + token_ids_1 + eos_token_id)
+
+        return output
+
+    def __getstate__(self):
+        state = self.__dict__.copy()
+        state["sp_model"] = None
+        state["sp_model_proto"] = self.sp_model.serialized_model_proto()
+        return state
+
+    def __setstate__(self, d):
+        self.__dict__ = d
+        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
+        self.sp_model.LoadFromSerializedProto(self.sp_model_proto)
+
+
+__all__ = ["CodeLlamaTokenizer"]

lib/python3.12/site-packages/transformers/models/code_llama/tokenization_code_llama_fast.py

@@ -0,0 +1,381 @@
+# coding=utf-8
+# Copyright 2023 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import os
+from shutil import copyfile
+from typing import List, Optional, Tuple
+
+from tokenizers import normalizers, processors
+
+from ...tokenization_utils_fast import PreTrainedTokenizerFast
+from ...utils import is_sentencepiece_available, logging
+from ...utils.versions import require_version
+
+
+require_version("tokenizers>=0.13.3")
+
+if is_sentencepiece_available():
+    from .tokenization_code_llama import CodeLlamaTokenizer
+else:
+    CodeLlamaTokenizer = None
+
+logger = logging.get_logger(__name__)
+VOCAB_FILES_NAMES = {"vocab_file": "tokenizer.model", "tokenizer_file": "tokenizer.json"}
+
+SPIECE_UNDERLINE = "▁"
+
+
+B_INST, E_INST = "[INST]", "[/INST]"
+B_SYS, E_SYS = "<<SYS>>\n", "\n<</SYS>>\n\n"
+
+# fmt: off
+DEFAULT_SYSTEM_PROMPT = """You are a helpful, respectful and honest assistant. Always answer as helpfully as possible, while being safe. Your \
+answers should not include any harmful, unethical, racist, sexist, toxic, dangerous, or illegal content. Please ensure\
+ that your responses are socially unbiased and positive in nature.
+
+If a question does not make any sense, or is not factually coherent, explain why instead of answering something not \
+correct. If you don't know the answer to a question, please don't share false information."""
+# fmt: on
+
+
+class CodeLlamaTokenizerFast(PreTrainedTokenizerFast):
+    """
+    Construct a Llama tokenizer. Based on byte-level Byte-Pair-Encoding.
+
+    This uses notably ByteFallback and no normalization.
+
+    ```python
+    >>> from transformers import CodeLlamaTokenizerFast
+
+    >>> tokenizer = CodeLlamaTokenizerFast.from_pretrained("hf-internal-testing/llama-tokenizer")
+    >>> tokenizer.encode("Hello this is a test")
+    [1, 15043, 445, 338, 263, 1243]
+    ```
+
+    If you want to change the `bos_token` or the `eos_token`, make sure to specify them when initializing the model, or
+    call `tokenizer.update_post_processor()` to make sure that the post-processing is correctly done (otherwise the
+    values of the first token and final token of an encoded sequence will not be correct). For more details, checkout
+    [post-processors] (https://huggingface.co/docs/tokenizers/api/post-processors) documentation.
+
+
+    This tokenizer inherits from [`PreTrainedTokenizerFast`] which contains most of the main methods. Users should
+    refer to this superclass for more information regarding those methods. The default configuration match that of
+    [meta-llama/CodeLlama-7b-Instruct-hf](https://huggingface.co/meta-llama/CodeLlama-7b-Instruct-hf/blob/main/tokenizer_config.json)
+    which supports prompt infilling.
+
+    Args:
+        vocab_file (`str`, *optional*):
+            [SentencePiece](https://github.com/google/sentencepiece) file (generally has a .model extension) that
+            contains the vocabulary necessary to instantiate a tokenizer.
+        tokenizer_file (`str`, *optional*):
+            [tokenizers](https://github.com/huggingface/tokenizers) file (generally has a .json extension) that
+            contains everything needed to load the tokenizer.
+        clean_up_tokenization_spaces (`str`, *optional*, defaults to `False`):
+            Whether to cleanup spaces after decoding, cleanup consists in removing potential artifacts like extra
+            spaces.
+        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.
+        bos_token (`str`, *optional*, defaults to `"<s>"`):
+            The beginning of sequence token that was used during pretraining. Can be used a sequence classifier token.
+        eos_token (`str`, *optional*, defaults to `"</s>"`):
+            The end of sequence token.
+        prefix_token (`str`, *optional*, defaults to `"▁<PRE>"`):
+            Prefix token used for infilling.
+        middle_token (`str`, *optional*, defaults to `"▁<MID>"`):
+            Middle token used for infilling.
+        suffix_token (`str`, *optional*, defaults to `"▁<SUF>"`):
+            Suffix token used for infilling.
+        eot_token (`str`, *optional*, defaults to `"▁<EOT>"`):
+            End of text token used for infilling.
+        fill_token (`str`, *optional*, defaults to `"<FILL_ME>"`):
+            The token used to split the input between the prefix and suffix.
+        additional_special_tokens (`List[str]`, *optional*):
+            Additional special tokens used by the tokenizer.
+        add_bos_token (`bool`, *optional*, defaults to `True`):
+            Whether to add a beginning of sequence token at the start of sequences.
+        add_eos_token (`bool`, *optional*, defaults to `False`):
+            Whether to add an end of sequence token at the end of sequences.
+        use_default_system_prompt (`bool`, *optional*, defaults to `False`):
+            Whether or not the default system prompt for Llama should be used.
+    """
+
+    vocab_files_names = VOCAB_FILES_NAMES
+    slow_tokenizer_class = CodeLlamaTokenizer
+    padding_side = "left"
+    model_input_names = ["input_ids", "attention_mask"]
+
+    def __init__(
+        self,
+        vocab_file=None,
+        tokenizer_file=None,
+        clean_up_tokenization_spaces=False,
+        unk_token="<unk>",
+        bos_token="<s>",
+        eos_token="</s>",
+        prefix_token="▁<PRE>",
+        middle_token="▁<MID>",
+        suffix_token="▁<SUF>",
+        eot_token="▁<EOT>",
+        fill_token="<FILL_ME>",
+        additional_special_tokens=None,
+        add_bos_token=True,
+        add_eos_token=False,
+        use_default_system_prompt=False,
+        **kwargs,
+    ):
+        # mark tokens special to skip them
+        additional_special_tokens = additional_special_tokens or []
+        for token in [prefix_token, middle_token, suffix_token, eot_token]:
+            additional_special_tokens += [token] if token is not None else []
+        self.use_default_system_prompt = use_default_system_prompt
+
+        super().__init__(
+            vocab_file=vocab_file,
+            tokenizer_file=tokenizer_file,
+            clean_up_tokenization_spaces=clean_up_tokenization_spaces,
+            additional_special_tokens=additional_special_tokens,
+            unk_token=unk_token,
+            bos_token=bos_token,
+            eos_token=eos_token,
+            add_bos_token=add_bos_token,
+            add_eos_token=add_eos_token,
+            prefix_token=prefix_token,
+            middle_token=middle_token,
+            suffix_token=suffix_token,
+            eot_token=eot_token,
+            fill_token=fill_token,
+            use_default_system_prompt=use_default_system_prompt,
+            **kwargs,
+        )
+        self._add_bos_token = add_bos_token
+        self._add_eos_token = add_eos_token
+        self.update_post_processor()
+
+        self.vocab_file = vocab_file
+
+        self._prefix_token = prefix_token
+        self._middle_token = middle_token
+        self._suffix_token = suffix_token
+        self._eot_token = eot_token
+        self.fill_token = fill_token
+
+    @property
+    def can_save_slow_tokenizer(self) -> bool:
+        return os.path.isfile(self.vocab_file) if self.vocab_file else False
+
+    # Copied from transformers.models.llama.tokenization_llama_fast.LlamaTokenizerFast.update_post_processor
+    def update_post_processor(self):
+        """
+        Updates the underlying post processor with the current `bos_token` and `eos_token`.
+        """
+        bos = self.bos_token
+        bos_token_id = self.bos_token_id
+        if bos is None and self.add_bos_token:
+            raise ValueError("add_bos_token = True but bos_token = None")
+
+        eos = self.eos_token
+        eos_token_id = self.eos_token_id
+        if eos is None and self.add_eos_token:
+            raise ValueError("add_eos_token = True but eos_token = None")
+
+        single = f"{(bos + ':0 ') if self.add_bos_token else ''}$A:0{(' ' + eos + ':0') if self.add_eos_token else ''}"
+        pair = f"{single}{(' ' + bos + ':1') if self.add_bos_token else ''} $B:1{(' ' + eos + ':1') if self.add_eos_token else ''}"
+
+        special_tokens = []
+        if self.add_bos_token:
+            special_tokens.append((bos, bos_token_id))
+        if self.add_eos_token:
+            special_tokens.append((eos, eos_token_id))
+        self._tokenizer.post_processor = processors.TemplateProcessing(
+            single=single, pair=pair, special_tokens=special_tokens
+        )
+
+    @property
+    def prefix_token(self):
+        return self._prefix_token
+
+    @property
+    def prefix_id(self):
+        if self._prefix_token is None:
+            return None
+        return self.convert_tokens_to_ids(self.prefix_token)
+
+    @property
+    def middle_token(self):
+        return self._middle_token
+
+    @property
+    def middle_id(self):
+        if self._middle_token is None:
+            return None
+        return self.convert_tokens_to_ids(self.middle_token)
+
+    @property
+    def suffix_token(self):
+        return self._suffix_token
+
+    @property
+    def suffix_id(self):
+        if self._suffix_token is None:
+            return None
+        return self.convert_tokens_to_ids(self.suffix_token)
+
+    @property
+    def eot_id(self):
+        if self._eot_token is None:
+            return None
+        return self.convert_tokens_to_ids(self.eot_token)
+
+    @property
+    def eot_token(self):
+        return self._eot_token
+
+    @property
+    def add_eos_token(self):
+        return self._add_eos_token
+
+    @property
+    def add_bos_token(self):
+        return self._add_bos_token
+
+    @add_eos_token.setter
+    def add_eos_token(self, value):
+        self._add_eos_token = value
+        self.update_post_processor()
+
+    @add_bos_token.setter
+    def add_bos_token(self, value):
+        self._add_bos_token = value
+        self.update_post_processor()
+
+    def set_infilling_processor(self, reset, suffix_first=False, add_special_tokens=True):
+        """
+        Updates the normalizer to make sure the prompt format for `infilling` is respected. The infilling format is the
+        following: if suffix_first
+            " <PRE> <SUF>{suf} <MID> {pre}"
+        else:
+            " <PRE> {pre} <SUF>{suf} <MID>"
+
+        If `reset` is set to `True`, the `normalizer` and `post_processor` are reset to their "normal" behaviour, which
+        is to add a prefix space for the normalizer, and add a `bos_token` to the input text for the `post_processor`.
+        """
+        if reset:
+            self._tokenizer.normalizer = normalizers.Sequence(
+                [
+                    normalizers.Prepend(prepend="▁"),
+                    normalizers.Replace(pattern=" ", content="▁"),
+                ]
+            )
+            self.update_post_processor()
+            return
+
+        self._tokenizer.normalizer = normalizers.Replace(pattern=" ", content="▁")
+        pair = [self.bos_token] if self.add_bos_token and add_special_tokens else []
+        special_tokens = [(self.bos_token, self.bos_token_id)] if self.add_bos_token and add_special_tokens else []
+        if suffix_first:
+            # format as " <PRE> <SUF>{suf} <MID> {pre}"
+            pair += [self.prefix_token, self.suffix_token, "$B", self.middle_token, "$A"]
+            special_tokens += [
+                (self.prefix_token, self.prefix_id),
+                (self.suffix_token, self.suffix_id),
+                (self.middle_token, self.middle_id),
+            ]
+        else:
+            # format as " <PRE> {pre} <SUF>{suf} <MID>"
+            pair += [self.prefix_token, "$A", self.suffix_token, "$B", self.middle_token]
+            special_tokens += [
+                (self.prefix_token, self.prefix_id),
+                (self.suffix_token, self.suffix_id),
+                (self.middle_token, self.middle_id),
+            ]
+
+        if self.add_eos_token and add_special_tokens:
+            pair += [self.eos_token]
+            special_tokens += [(self.eos_token, self.eos_token_id)]
+        self._tokenizer.post_processor = processors.TemplateProcessing(
+            single="$A", pair=pair, special_tokens=special_tokens
+        )
+
+    def encode_plus(self, text, text_pair=None, suffix_first=False, add_special_tokens=True, **kwargs):
+        # hack to make sure the input is pre-process but outside rust
+        text_pair = kwargs.pop("suffix", text_pair)
+        if self.fill_token is not None and self.fill_token in text and text_pair is None:
+            text, text_pair = text.split(self.fill_token)
+
+        if text_pair is None or len(text_pair) < 1:
+            return super().encode_plus(text, text_pair, add_special_tokens=add_special_tokens, **kwargs)
+
+        if None in (self.prefix_id, self.middle_id, self.suffix_id):
+            raise ValueError(
+                "Then input includes a `prefix` and a `suffix` used for the infilling task,"
+                " the `prefix_id, middle_id, suffix_id` must all be initialized. Current"
+                f" values : {self.prefix_id, self.middle_id, self.suffix_id}"
+            )
+
+        self.set_infilling_processor(False, suffix_first=suffix_first, add_special_tokens=add_special_tokens)
+        tokens = super().encode_plus(" " + text, text_pair=text_pair, add_special_tokens=True, **kwargs)
+        self.set_infilling_processor(True)
+        return tokens
+
+    # Copied from transformers.models.llama.tokenization_llama_fast.LlamaTokenizerFast.save_vocabulary
+    def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
+        if not self.can_save_slow_tokenizer:
+            raise ValueError(
+                "Your fast tokenizer does not have the necessary information to save the vocabulary for a slow "
+                "tokenizer."
+            )
+
+        if not os.path.isdir(save_directory):
+            logger.error(f"Vocabulary path ({save_directory}) should be a directory")
+            return
+        out_vocab_file = os.path.join(
+            save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]
+        )
+
+        if os.path.abspath(self.vocab_file) != os.path.abspath(out_vocab_file):
+            copyfile(self.vocab_file, out_vocab_file)
+
+        return (out_vocab_file,)
+
+    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. The special tokens depend on calling set_lang.
+
+        An NLLB sequence has the following format, where `X` represents the sequence:
+
+        - `input_ids` (for encoder) `X [eos, src_lang_code]`
+        - `decoder_input_ids`: (for decoder) `X [eos, tgt_lang_code]`
+
+        BOS is never used. Pairs of sequences are not the expected use case, but they will be handled without a
+        separator.
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs to which the special tokens will be added.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            `List[int]`: list of [input IDs](../glossary#input-ids) with the appropriate special tokens.
+        """
+        if token_ids_1 is None:
+            return self.bos_token_id + token_ids_0 + self.eos_token_id
+        return self.bos_token_id + token_ids_0 + token_ids_1 + self.eos_token_id
+
+
+__all__ = ["CodeLlamaTokenizerFast"]

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

@@ -0,0 +1,27 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import TYPE_CHECKING
+
+from ...utils import _LazyModule
+from ...utils.import_utils import define_import_structure
+
+
+if TYPE_CHECKING:
+    from .configuration_dbrx import *
+    from .modeling_dbrx import *
+else:
+    import sys
+
+    _file = globals()["__file__"]
+    sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)