add albert and deberta

modeling_albert.py

@@ -0,0 +1,1363 @@
+# 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 Optional, Tuple
+
+import torch
+from packaging import version
+from torch import nn
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+
+from transformers.activations import ACT2FN
+from transformers.file_utils import (
+    ModelOutput,
+    add_code_sample_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    replace_return_docstrings,
+)
+from transformers.modeling_outputs import (
+    BaseModelOutput,
+    BaseModelOutputWithPooling,
+    MaskedLMOutput,
+    MultipleChoiceModelOutput,
+    QuestionAnsweringModelOutput,
+    SequenceClassifierOutput,
+    TokenClassifierOutput,
+)
+from transformers.modeling_utils import (
+    PreTrainedModel,
+    apply_chunking_to_forward,
+    find_pruneable_heads_and_indices,
+    prune_linear_layer,
+)
+from transformers.utils import logging
+from transformers import AlbertConfig
+
+
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "albert-base-v2"
+_CONFIG_FOR_DOC = "AlbertConfig"
+_TOKENIZER_FOR_DOC = "AlbertTokenizer"
+
+
+ALBERT_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "albert-base-v1",
+    "albert-large-v1",
+    "albert-xlarge-v1",
+    "albert-xxlarge-v1",
+    "albert-base-v2",
+    "albert-large-v2",
+    "albert-xlarge-v2",
+    "albert-xxlarge-v2",
+    # See all ALBERT models at https://huggingface.co/models?filter=albert
+]
+
+
+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 AssertionError 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):
+        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)))
+        self.position_embedding_type = getattr(config, "position_embedding_type", "absolute")
+        if version.parse(torch.__version__) > version.parse("1.6.0"):
+            self.register_buffer(
+                "token_type_ids",
+                torch.zeros(self.position_ids.size(), dtype=torch.long, device=self.position_ids.device),
+                persistent=False,
+            )
+
+    # Copied from transformers.models.bert.modeling_bert.BertEmbeddings.forward
+    def forward(
+        self, input_ids=None, token_type_ids=None, position_ids=None, inputs_embeds=None, past_key_values_length=0
+    ):
+        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):
+        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):
+        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):
+        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, attention_mask=None, head_mask=None, output_attentions=False):
+        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.Softmax(dim=-1)(attention_scores)
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        attention_probs = self.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 AlbertLayer(nn.Module):
+    def __init__(self, config):
+        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 = AlbertAttention(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, attention_mask=None, head_mask=None, output_attentions=False, output_hidden_states=False
+    ):
+        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):
+        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):
+        super().__init__()
+
+        self.albert_layers = nn.ModuleList([AlbertLayer(config) for _ in range(config.inner_group_num)])
+
+    def forward(
+        self, hidden_states, attention_mask=None, head_mask=None, output_attentions=False, output_hidden_states=False
+    ):
+        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):
+        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,
+        attention_mask=None,
+        head_mask=None,
+        output_attentions=False,
+        output_hidden_states=False,
+        return_dict=True,
+    ):
+        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"
+    _keys_to_ignore_on_load_missing = [r"position_ids"]
+
+    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)
+
+
+@dataclass
+class AlbertForPreTrainingOutput(ModelOutput):
+    """
+    Output type of :class:`~transformers.AlbertForPreTraining`.
+
+    Args:
+        loss (`optional`, returned when ``labels`` is provided, ``torch.FloatTensor`` of shape :obj:`(1,)`):
+            Total loss as the sum of the masked language modeling loss and the next sequence prediction
+            (classification) loss.
+        prediction_logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`):
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        sop_logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, 2)`):
+            Prediction scores of the next sequence prediction (classification) head (scores of True/False continuation
+            before SoftMax).
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(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: torch.FloatTensor = None
+    sop_logits: torch.FloatTensor = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+ALBERT_START_DOCSTRING = r"""
+
+    This model inherits from :class:`~transformers.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 (:class:`~transformers.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 :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model
+            weights.
+"""
+
+ALBERT_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (:obj:`torch.LongTensor` of shape :obj:`({0})`):
+            Indices of input sequence tokens in the vocabulary.
+
+            Indices can be obtained using :class:`~transformers.AlbertTokenizer`. See
+            :meth:`transformers.PreTrainedTokenizer.__call__` and :meth:`transformers.PreTrainedTokenizer.encode` for
+            details.
+
+            `What are input IDs? <../glossary.html#input-ids>`__
+        attention_mask (:obj:`torch.FloatTensor` of shape :obj:`({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.html#attention-mask>`__
+        token_type_ids (:obj:`torch.LongTensor` of shape :obj:`({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.html#token-type-ids>`_
+        position_ids (:obj:`torch.LongTensor` of shape :obj:`({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.html#position-ids>`_
+        head_mask (:obj:`torch.FloatTensor` of shape :obj:`(num_heads,)` or :obj:`(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 (:obj:`torch.FloatTensor` of shape :obj:`({0}, hidden_size)`, `optional`):
+            Optionally, instead of passing :obj:`input_ids` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert :obj:`input_ids` indices into associated
+            vectors than the model's internal embedding lookup matrix.
+        output_attentions (:obj:`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 (:obj:`bool`, `optional`):
+            Whether or not to return the hidden states of all layers. See ``hidden_states`` under returned tensors for
+            more detail.
+        return_dict (:obj:`bool`, `optional`):
+            Whether or not to return a :class:`~transformers.file_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, add_pooling_layer=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.init_weights()
+
+    def get_input_embeddings(self):
+        return self.embeddings.word_embeddings
+
+    def set_input_embeddings(self, value):
+        self.embeddings.word_embeddings = value
+
+    def _prune_heads(self, heads_to_prune):
+        """
+        Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} 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(
+        processor_class=_TOKENIZER_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=BaseModelOutputWithPooling,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=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.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:
+            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)
+
+        # extended_attention_mask = attention_mask.unsqueeze(1).unsqueeze(2) # 
+        extended_attention_mask = attention_mask[:, None, :, :]
+        extended_attention_mask = extended_attention_mask.to(dtype=self.dtype)  # fp16 compatibility
+        extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0
+        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
+
+        embedding_output = self.embeddings(
+            input_ids, position_ids=position_ids, token_type_ids=token_type_ids, inputs_embeds=inputs_embeds
+        )
+        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):
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.albert = AlbertModel(config)
+        self.predictions = AlbertMLMHead(config)
+        self.sop_classifier = AlbertSOPHead(config)
+
+        self.init_weights()
+
+    def get_output_embeddings(self):
+        return self.predictions.decoder
+
+    def set_output_embeddings(self, new_embeddings):
+        self.predictions.decoder = new_embeddings
+
+    def get_input_embeddings(self):
+        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=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        labels=None,
+        sentence_order_label=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        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 :obj:`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::
+
+            >>> from transformers import AlbertTokenizer, AlbertForPreTraining
+            >>> import torch
+
+            >>> tokenizer = AlbertTokenizer.from_pretrained('albert-base-v2')
+            >>> model = AlbertForPreTraining.from_pretrained('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):
+        super().__init__()
+
+        self.LayerNorm = nn.LayerNorm(config.embedding_size)
+        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):
+        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):
+        # 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):
+        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):
+        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):
+
+    _keys_to_ignore_on_load_unexpected = [r"pooler"]
+
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.albert = AlbertModel(config, add_pooling_layer=False)
+        self.predictions = AlbertMLMHead(config)
+
+        self.init_weights()
+
+    def get_output_embeddings(self):
+        return self.predictions.decoder
+
+    def set_output_embeddings(self, new_embeddings):
+        self.predictions.decoder = new_embeddings
+
+    def get_input_embeddings(self):
+        return self.albert.embeddings.word_embeddings
+
+    @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        processor_class=_TOKENIZER_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=MaskedLMOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        labels=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        labels (:obj:`torch.LongTensor` of shape :obj:`(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]``
+        """
+        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):
+        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)
+
+        self.init_weights()
+
+    @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        processor_class=_TOKENIZER_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=SequenceClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        labels=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        labels (:obj:`torch.LongTensor` of shape :obj:`(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):
+
+    _keys_to_ignore_on_load_unexpected = [r"pooler"]
+
+    def __init__(self, config):
+        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)
+
+        self.init_weights()
+
+    @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        processor_class=_TOKENIZER_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=TokenClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        labels=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        labels (:obj:`torch.LongTensor` of shape :obj:`(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()
+            # 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,) + 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):
+
+    _keys_to_ignore_on_load_unexpected = [r"pooler"]
+
+    def __init__(self, config):
+        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)
+
+        self.init_weights()
+
+    @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        processor_class=_TOKENIZER_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=QuestionAnsweringModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        start_positions=None,
+        end_positions=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        start_positions (:obj:`torch.LongTensor` of shape :obj:`(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 (:obj:`sequence_length`). Position outside of the
+            sequence are not taken into account for computing the loss.
+        end_positions (:obj:`torch.LongTensor` of shape :obj:`(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 (:obj:`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 = 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):
+        super().__init__(config)
+
+        self.albert = AlbertModel(config)
+        self.dropout = nn.Dropout(config.classifier_dropout_prob)
+        self.classifier = nn.Linear(config.hidden_size, 1)
+
+        self.init_weights()
+
+    @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length"))
+    @add_code_sample_docstrings(
+        processor_class=_TOKENIZER_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=MultipleChoiceModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        labels=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        labels (:obj:`torch.LongTensor` of shape :obj:`(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 = 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,
+        )

modeling_deberta_v2.py

@@ -0,0 +1,1617 @@
+# coding=utf-8
+# Copyright 2020 Microsoft and the Hugging Face 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 DeBERTa-v2 model."""
+
+import math
+from collections.abc import Sequence
+from typing import Optional, Tuple, Union
+
+import numpy as np
+import torch
+from torch import nn
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, LayerNorm, MSELoss
+
+from transformers.activations import ACT2FN
+from transformers.modeling_outputs import (
+    BaseModelOutput,
+    MaskedLMOutput,
+    MultipleChoiceModelOutput,
+    QuestionAnsweringModelOutput,
+    SequenceClassifierOutput,
+    TokenClassifierOutput,
+)
+from transformers.modeling_utils import PreTrainedModel
+from transformers.pytorch_utils import softmax_backward_data
+from transformers.utils import add_code_sample_docstrings, add_start_docstrings, add_start_docstrings_to_model_forward, logging
+from transformers import DebertaV2Config
+
+
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = "DebertaV2Config"
+_TOKENIZER_FOR_DOC = "DebertaV2Tokenizer"
+_CHECKPOINT_FOR_DOC = "microsoft/deberta-v2-xlarge"
+
+DEBERTA_V2_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "microsoft/deberta-v2-xlarge",
+    "microsoft/deberta-v2-xxlarge",
+    "microsoft/deberta-v2-xlarge-mnli",
+    "microsoft/deberta-v2-xxlarge-mnli",
+]
+
+
+# Copied from transformers.models.deberta.modeling_deberta.ContextPooler
+class ContextPooler(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.pooler_hidden_size, config.pooler_hidden_size)
+        self.dropout = StableDropout(config.pooler_dropout)
+        self.config = config
+
+    def forward(self, hidden_states):
+        # We "pool" the model by simply taking the hidden state corresponding
+        # to the first token.
+
+        context_token = hidden_states[:, 0]
+        context_token = self.dropout(context_token)
+        pooled_output = self.dense(context_token)
+        pooled_output = ACT2FN[self.config.pooler_hidden_act](pooled_output)
+        return pooled_output
+
+    @property
+    def output_dim(self):
+        return self.config.hidden_size
+
+
+# Copied from transformers.models.deberta.modeling_deberta.XSoftmax with deberta->deberta_v2
+class XSoftmax(torch.autograd.Function):
+    """
+    Masked Softmax which is optimized for saving memory
+
+    Args:
+        input (`torch.tensor`): The input tensor that will apply softmax.
+        mask (`torch.IntTensor`):
+            The mask matrix where 0 indicate that element will be ignored in the softmax calculation.
+        dim (int): The dimension that will apply softmax
+
+    Example:
+
+    ```python
+    >>> import torch
+    >>> from transformers.models.deberta_v2.modeling_deberta_v2 import XSoftmax
+
+    >>> # Make a tensor
+    >>> x = torch.randn([4, 20, 100])
+
+    >>> # Create a mask
+    >>> mask = (x > 0).int()
+
+    >>> # Specify the dimension to apply softmax
+    >>> dim = -1
+
+    >>> y = XSoftmax.apply(x, mask, dim)
+    ```"""
+
+    @staticmethod
+    def forward(self, input, mask, dim):
+        self.dim = dim
+        rmask = ~(mask.to(torch.bool))
+
+        output = input.masked_fill(rmask, torch.tensor(torch.finfo(input.dtype).min))
+        output = torch.softmax(output, self.dim)
+        output.masked_fill_(rmask, 0)
+        self.save_for_backward(output)
+        return output
+
+    @staticmethod
+    def backward(self, grad_output):
+        (output,) = self.saved_tensors
+        inputGrad = softmax_backward_data(self, grad_output, output, self.dim, output)
+        return inputGrad, None, None
+
+    @staticmethod
+    def symbolic(g, self, mask, dim):
+        import torch.onnx.symbolic_helper as sym_help
+        from torch.onnx.symbolic_opset9 import masked_fill, softmax
+
+        mask_cast_value = g.op("Cast", mask, to_i=sym_help.cast_pytorch_to_onnx["Long"])
+        r_mask = g.op(
+            "Cast",
+            g.op("Sub", g.op("Constant", value_t=torch.tensor(1, dtype=torch.int64)), mask_cast_value),
+            to_i=sym_help.cast_pytorch_to_onnx["Byte"],
+        )
+        output = masked_fill(g, self, r_mask, g.op("Constant", value_t=torch.tensor(torch.finfo(self.dtype).min)))
+        output = softmax(g, output, dim)
+        return masked_fill(g, output, r_mask, g.op("Constant", value_t=torch.tensor(0, dtype=torch.uint8)))
+
+
+# Copied from transformers.models.deberta.modeling_deberta.DropoutContext
+class DropoutContext(object):
+    def __init__(self):
+        self.dropout = 0
+        self.mask = None
+        self.scale = 1
+        self.reuse_mask = True
+
+
+# Copied from transformers.models.deberta.modeling_deberta.get_mask
+def get_mask(input, local_context):
+    if not isinstance(local_context, DropoutContext):
+        dropout = local_context
+        mask = None
+    else:
+        dropout = local_context.dropout
+        dropout *= local_context.scale
+        mask = local_context.mask if local_context.reuse_mask else None
+
+    if dropout > 0 and mask is None:
+        mask = (1 - torch.empty_like(input).bernoulli_(1 - dropout)).to(torch.bool)
+
+    if isinstance(local_context, DropoutContext):
+        if local_context.mask is None:
+            local_context.mask = mask
+
+    return mask, dropout
+
+
+# Copied from transformers.models.deberta.modeling_deberta.XDropout
+class XDropout(torch.autograd.Function):
+    """Optimized dropout function to save computation and memory by using mask operation instead of multiplication."""
+
+    @staticmethod
+    def forward(ctx, input, local_ctx):
+        mask, dropout = get_mask(input, local_ctx)
+        ctx.scale = 1.0 / (1 - dropout)
+        if dropout > 0:
+            ctx.save_for_backward(mask)
+            return input.masked_fill(mask, 0) * ctx.scale
+        else:
+            return input
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        if ctx.scale > 1:
+            (mask,) = ctx.saved_tensors
+            return grad_output.masked_fill(mask, 0) * ctx.scale, None
+        else:
+            return grad_output, None
+
+
+# Copied from transformers.models.deberta.modeling_deberta.StableDropout
+class StableDropout(nn.Module):
+    """
+    Optimized dropout module for stabilizing the training
+
+    Args:
+        drop_prob (float): the dropout probabilities
+    """
+
+    def __init__(self, drop_prob):
+        super().__init__()
+        self.drop_prob = drop_prob
+        self.count = 0
+        self.context_stack = None
+
+    def forward(self, x):
+        """
+        Call the module
+
+        Args:
+            x (`torch.tensor`): The input tensor to apply dropout
+        """
+        if self.training and self.drop_prob > 0:
+            return XDropout.apply(x, self.get_context())
+        return x
+
+    def clear_context(self):
+        self.count = 0
+        self.context_stack = None
+
+    def init_context(self, reuse_mask=True, scale=1):
+        if self.context_stack is None:
+            self.context_stack = []
+        self.count = 0
+        for c in self.context_stack:
+            c.reuse_mask = reuse_mask
+            c.scale = scale
+
+    def get_context(self):
+        if self.context_stack is not None:
+            if self.count >= len(self.context_stack):
+                self.context_stack.append(DropoutContext())
+            ctx = self.context_stack[self.count]
+            ctx.dropout = self.drop_prob
+            self.count += 1
+            return ctx
+        else:
+            return self.drop_prob
+
+
+# Copied from transformers.models.deberta.modeling_deberta.DebertaSelfOutput with DebertaLayerNorm->LayerNorm
+class DebertaV2SelfOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.LayerNorm = LayerNorm(config.hidden_size, config.layer_norm_eps)
+        self.dropout = StableDropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states, input_tensor):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+# Copied from transformers.models.deberta.modeling_deberta.DebertaAttention with Deberta->DebertaV2
+class DebertaV2Attention(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.self = DisentangledSelfAttention(config)
+        self.output = DebertaV2SelfOutput(config)
+        self.config = config
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask,
+        output_attentions=False,
+        query_states=None,
+        relative_pos=None,
+        rel_embeddings=None,
+    ):
+        self_output = self.self(
+            hidden_states,
+            attention_mask,
+            output_attentions,
+            query_states=query_states,
+            relative_pos=relative_pos,
+            rel_embeddings=rel_embeddings,
+        )
+        if output_attentions:
+            self_output, att_matrix = self_output
+        if query_states is None:
+            query_states = hidden_states
+        attention_output = self.output(self_output, query_states)
+
+        if output_attentions:
+            return (attention_output, att_matrix)
+        else:
+            return attention_output
+
+
+# Copied from transformers.models.bert.modeling_bert.BertIntermediate with Bert->DebertaV2
+class DebertaV2Intermediate(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.intermediate_size)
+        if isinstance(config.hidden_act, str):
+            self.intermediate_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.intermediate_act_fn = config.hidden_act
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.intermediate_act_fn(hidden_states)
+        return hidden_states
+
+
+# Copied from transformers.models.deberta.modeling_deberta.DebertaOutput with DebertaLayerNorm->LayerNorm
+class DebertaV2Output(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
+        self.LayerNorm = LayerNorm(config.hidden_size, config.layer_norm_eps)
+        self.dropout = StableDropout(config.hidden_dropout_prob)
+        self.config = config
+
+    def forward(self, hidden_states, input_tensor):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+# Copied from transformers.models.deberta.modeling_deberta.DebertaLayer with Deberta->DebertaV2
+class DebertaV2Layer(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.attention = DebertaV2Attention(config)
+        self.intermediate = DebertaV2Intermediate(config)
+        self.output = DebertaV2Output(config)
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask,
+        query_states=None,
+        relative_pos=None,
+        rel_embeddings=None,
+        output_attentions=False,
+    ):
+        attention_output = self.attention(
+            hidden_states,
+            attention_mask,
+            output_attentions=output_attentions,
+            query_states=query_states,
+            relative_pos=relative_pos,
+            rel_embeddings=rel_embeddings,
+        )
+        if output_attentions:
+            attention_output, att_matrix = attention_output
+        intermediate_output = self.intermediate(attention_output)
+        layer_output = self.output(intermediate_output, attention_output)
+        if output_attentions:
+            return (layer_output, att_matrix)
+        else:
+            return layer_output
+
+
+class ConvLayer(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        kernel_size = getattr(config, "conv_kernel_size", 3)
+        groups = getattr(config, "conv_groups", 1)
+        self.conv_act = getattr(config, "conv_act", "tanh")
+        self.conv = nn.Conv1d(
+            config.hidden_size, config.hidden_size, kernel_size, padding=(kernel_size - 1) // 2, groups=groups
+        )
+        self.LayerNorm = LayerNorm(config.hidden_size, config.layer_norm_eps)
+        self.dropout = StableDropout(config.hidden_dropout_prob)
+        self.config = config
+
+    def forward(self, hidden_states, residual_states, input_mask):
+        out = self.conv(hidden_states.permute(0, 2, 1).contiguous()).permute(0, 2, 1).contiguous()
+        rmask = (1 - input_mask).bool()
+        out.masked_fill_(rmask.unsqueeze(-1).expand(out.size()), 0)
+        out = ACT2FN[self.conv_act](self.dropout(out))
+
+        layer_norm_input = residual_states + out
+        output = self.LayerNorm(layer_norm_input).to(layer_norm_input)
+
+        if input_mask is None:
+            output_states = output
+        else:
+            if input_mask.dim() != layer_norm_input.dim():
+                if input_mask.dim() == 4:
+                    input_mask = input_mask.squeeze(1).squeeze(1)
+                input_mask = input_mask.unsqueeze(2)
+
+            input_mask = input_mask.to(output.dtype)
+            output_states = output * input_mask
+
+        return output_states
+
+
+class DebertaV2Encoder(nn.Module):
+    """Modified BertEncoder with relative position bias support"""
+
+    def __init__(self, config):
+        super().__init__()
+
+        self.layer = nn.ModuleList([DebertaV2Layer(config) for _ in range(config.num_hidden_layers)])
+        self.relative_attention = getattr(config, "relative_attention", False)
+
+        if self.relative_attention:
+            self.max_relative_positions = getattr(config, "max_relative_positions", -1)
+            if self.max_relative_positions < 1:
+                self.max_relative_positions = config.max_position_embeddings
+
+            self.position_buckets = getattr(config, "position_buckets", -1)
+            pos_ebd_size = self.max_relative_positions * 2
+
+            if self.position_buckets > 0:
+                pos_ebd_size = self.position_buckets * 2
+
+            self.rel_embeddings = nn.Embedding(pos_ebd_size, config.hidden_size)
+
+        self.norm_rel_ebd = [x.strip() for x in getattr(config, "norm_rel_ebd", "none").lower().split("|")]
+
+        if "layer_norm" in self.norm_rel_ebd:
+            self.LayerNorm = LayerNorm(config.hidden_size, config.layer_norm_eps, elementwise_affine=True)
+
+        self.conv = ConvLayer(config) if getattr(config, "conv_kernel_size", 0) > 0 else None
+        self.gradient_checkpointing = False
+
+    def get_rel_embedding(self):
+        rel_embeddings = self.rel_embeddings.weight if self.relative_attention else None
+        if rel_embeddings is not None and ("layer_norm" in self.norm_rel_ebd):
+            rel_embeddings = self.LayerNorm(rel_embeddings)
+        return rel_embeddings
+
+    def get_attention_mask(self, attention_mask):
+        if attention_mask.dim() <= 2:
+            extended_attention_mask = attention_mask.unsqueeze(1).unsqueeze(2)
+            attention_mask = extended_attention_mask * extended_attention_mask.squeeze(-2).unsqueeze(-1)
+            attention_mask = attention_mask.byte()
+        elif attention_mask.dim() == 3:
+            attention_mask = attention_mask.unsqueeze(1)
+
+        return attention_mask
+
+    def get_rel_pos(self, hidden_states, query_states=None, relative_pos=None):
+        if self.relative_attention and relative_pos is None:
+            q = query_states.size(-2) if query_states is not None else hidden_states.size(-2)
+            relative_pos = build_relative_position(
+                q, hidden_states.size(-2), bucket_size=self.position_buckets, max_position=self.max_relative_positions
+            )
+        return relative_pos
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask,
+        output_hidden_states=True,
+        output_attentions=False,
+        query_states=None,
+        relative_pos=None,
+        return_dict=True,
+    ):
+        if attention_mask.dim() <= 2:
+            input_mask = attention_mask
+        else:
+            input_mask = (attention_mask.sum(-2) > 0).byte()
+        attention_mask = self.get_attention_mask(attention_mask)
+        relative_pos = self.get_rel_pos(hidden_states, query_states, relative_pos)
+
+        all_hidden_states = () if output_hidden_states else None
+        all_attentions = () if output_attentions else None
+
+        if isinstance(hidden_states, Sequence):
+            next_kv = hidden_states[0]
+        else:
+            next_kv = hidden_states
+        rel_embeddings = self.get_rel_embedding()
+        output_states = next_kv
+        for i, layer_module in enumerate(self.layer):
+
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (output_states,)
+
+            if self.gradient_checkpointing and self.training:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs, output_attentions)
+
+                    return custom_forward
+
+                output_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(layer_module),
+                    next_kv,
+                    attention_mask,
+                    query_states,
+                    relative_pos,
+                    rel_embeddings,
+                )
+            else:
+                output_states = layer_module(
+                    next_kv,
+                    attention_mask,
+                    query_states=query_states,
+                    relative_pos=relative_pos,
+                    rel_embeddings=rel_embeddings,
+                    output_attentions=output_attentions,
+                )
+
+            if output_attentions:
+                output_states, att_m = output_states
+
+            if i == 0 and self.conv is not None:
+                output_states = self.conv(hidden_states, output_states, input_mask)
+
+            if query_states is not None:
+                query_states = output_states
+                if isinstance(hidden_states, Sequence):
+                    next_kv = hidden_states[i + 1] if i + 1 < len(self.layer) else None
+            else:
+                next_kv = output_states
+
+            if output_attentions:
+                all_attentions = all_attentions + (att_m,)
+
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (output_states,)
+
+        if not return_dict:
+            return tuple(v for v in [output_states, all_hidden_states, all_attentions] if v is not None)
+        return BaseModelOutput(
+            last_hidden_state=output_states, hidden_states=all_hidden_states, attentions=all_attentions
+        )
+
+
+def make_log_bucket_position(relative_pos, bucket_size, max_position):
+    sign = np.sign(relative_pos)
+    mid = bucket_size // 2
+    abs_pos = np.where((relative_pos < mid) & (relative_pos > -mid), mid - 1, np.abs(relative_pos))
+    log_pos = np.ceil(np.log(abs_pos / mid) / np.log((max_position - 1) / mid) * (mid - 1)) + mid
+    bucket_pos = np.where(abs_pos <= mid, relative_pos, log_pos * sign).astype(np.int)
+    return bucket_pos
+
+
+def build_relative_position(query_size, key_size, bucket_size=-1, max_position=-1):
+    """
+    Build relative position according to the query and key
+
+    We assume the absolute position of query \\(P_q\\) is range from (0, query_size) and the absolute position of key
+    \\(P_k\\) is range from (0, key_size), The relative positions from query to key is \\(R_{q \\rightarrow k} = P_q -
+    P_k\\)
+
+    Args:
+        query_size (int): the length of query
+        key_size (int): the length of key
+        bucket_size (int): the size of position bucket
+        max_position (int): the maximum allowed absolute position
+
+    Return:
+        `torch.LongTensor`: A tensor with shape [1, query_size, key_size]
+
+    """
+    q_ids = np.arange(0, query_size)
+    k_ids = np.arange(0, key_size)
+    rel_pos_ids = q_ids[:, None] - np.tile(k_ids, (q_ids.shape[0], 1))
+    if bucket_size > 0 and max_position > 0:
+        rel_pos_ids = make_log_bucket_position(rel_pos_ids, bucket_size, max_position)
+    rel_pos_ids = torch.tensor(rel_pos_ids, dtype=torch.long)
+    rel_pos_ids = rel_pos_ids[:query_size, :]
+    rel_pos_ids = rel_pos_ids.unsqueeze(0)
+    return rel_pos_ids
+
+
+@torch.jit.script
+# Copied from transformers.models.deberta.modeling_deberta.c2p_dynamic_expand
+def c2p_dynamic_expand(c2p_pos, query_layer, relative_pos):
+    return c2p_pos.expand([query_layer.size(0), query_layer.size(1), query_layer.size(2), relative_pos.size(-1)])
+
+
+@torch.jit.script
+# Copied from transformers.models.deberta.modeling_deberta.p2c_dynamic_expand
+def p2c_dynamic_expand(c2p_pos, query_layer, key_layer):
+    return c2p_pos.expand([query_layer.size(0), query_layer.size(1), key_layer.size(-2), key_layer.size(-2)])
+
+
+@torch.jit.script
+# Copied from transformers.models.deberta.modeling_deberta.pos_dynamic_expand
+def pos_dynamic_expand(pos_index, p2c_att, key_layer):
+    return pos_index.expand(p2c_att.size()[:2] + (pos_index.size(-2), key_layer.size(-2)))
+
+
+class DisentangledSelfAttention(nn.Module):
+    """
+    Disentangled self-attention module
+
+    Parameters:
+        config (`DebertaV2Config`):
+            A model config class instance with the configuration to build a new model. The schema is similar to
+            *BertConfig*, for more details, please refer [`DebertaV2Config`]
+
+    """
+
+    def __init__(self, config):
+        super().__init__()
+        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 of attention "
+                f"heads ({config.num_attention_heads})"
+            )
+        self.num_attention_heads = config.num_attention_heads
+        _attention_head_size = config.hidden_size // config.num_attention_heads
+        self.attention_head_size = getattr(config, "attention_head_size", _attention_head_size)
+        self.all_head_size = self.num_attention_heads * self.attention_head_size
+        self.query_proj = nn.Linear(config.hidden_size, self.all_head_size, bias=True)
+        self.key_proj = nn.Linear(config.hidden_size, self.all_head_size, bias=True)
+        self.value_proj = nn.Linear(config.hidden_size, self.all_head_size, bias=True)
+
+        self.share_att_key = getattr(config, "share_att_key", False)
+        self.pos_att_type = config.pos_att_type if config.pos_att_type is not None else []
+        self.relative_attention = getattr(config, "relative_attention", False)
+
+        if self.relative_attention:
+            self.position_buckets = getattr(config, "position_buckets", -1)
+            self.max_relative_positions = getattr(config, "max_relative_positions", -1)
+            if self.max_relative_positions < 1:
+                self.max_relative_positions = config.max_position_embeddings
+            self.pos_ebd_size = self.max_relative_positions
+            if self.position_buckets > 0:
+                self.pos_ebd_size = self.position_buckets
+
+            self.pos_dropout = StableDropout(config.hidden_dropout_prob)
+
+            if not self.share_att_key:
+                if "c2p" in self.pos_att_type:
+                    self.pos_key_proj = nn.Linear(config.hidden_size, self.all_head_size, bias=True)
+                if "p2c" in self.pos_att_type:
+                    self.pos_query_proj = nn.Linear(config.hidden_size, self.all_head_size)
+
+        self.dropout = StableDropout(config.attention_probs_dropout_prob)
+
+    def transpose_for_scores(self, x, attention_heads):
+        new_x_shape = x.size()[:-1] + (attention_heads, -1)
+        x = x.view(new_x_shape)
+        return x.permute(0, 2, 1, 3).contiguous().view(-1, x.size(1), x.size(-1))
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask,
+        output_attentions=False,
+        query_states=None,
+        relative_pos=None,
+        rel_embeddings=None,
+    ):
+        """
+        Call the module
+
+        Args:
+            hidden_states (`torch.FloatTensor`):
+                Input states to the module usually the output from previous layer, it will be the Q,K and V in
+                *Attention(Q,K,V)*
+
+            attention_mask (`torch.ByteTensor`):
+                An attention mask matrix of shape [*B*, *N*, *N*] where *B* is the batch size, *N* is the maximum
+                sequence length in which element [i,j] = *1* means the *i* th token in the input can attend to the *j*
+                th token.
+
+            output_attentions (`bool`, optional):
+                Whether return the attention matrix.
+
+            query_states (`torch.FloatTensor`, optional):
+                The *Q* state in *Attention(Q,K,V)*.
+
+            relative_pos (`torch.LongTensor`):
+                The relative position encoding between the tokens in the sequence. It's of shape [*B*, *N*, *N*] with
+                values ranging in [*-max_relative_positions*, *max_relative_positions*].
+
+            rel_embeddings (`torch.FloatTensor`):
+                The embedding of relative distances. It's a tensor of shape [\\(2 \\times
+                \\text{max_relative_positions}\\), *hidden_size*].
+
+
+        """
+        if query_states is None:
+            query_states = hidden_states
+        query_layer = self.transpose_for_scores(self.query_proj(query_states), self.num_attention_heads)
+        key_layer = self.transpose_for_scores(self.key_proj(hidden_states), self.num_attention_heads)
+        value_layer = self.transpose_for_scores(self.value_proj(hidden_states), self.num_attention_heads)
+
+        rel_att = None
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        scale_factor = 1
+        if "c2p" in self.pos_att_type:
+            scale_factor += 1
+        if "p2c" in self.pos_att_type:
+            scale_factor += 1
+        scale = math.sqrt(query_layer.size(-1) * scale_factor)
+        attention_scores = torch.bmm(query_layer, key_layer.transpose(-1, -2)) / scale
+        if self.relative_attention:
+            rel_embeddings = self.pos_dropout(rel_embeddings)
+            rel_att = self.disentangled_attention_bias(
+                query_layer, key_layer, relative_pos, rel_embeddings, scale_factor
+            )
+
+        if rel_att is not None:
+            attention_scores = attention_scores + rel_att
+        attention_scores = attention_scores
+        attention_scores = attention_scores.view(
+            -1, self.num_attention_heads, attention_scores.size(-2), attention_scores.size(-1)
+        )
+
+        # bsz x height x length x dimension
+        attention_probs = XSoftmax.apply(attention_scores, attention_mask, -1)
+        attention_probs = self.dropout(attention_probs)
+        context_layer = torch.bmm(
+            attention_probs.view(-1, attention_probs.size(-2), attention_probs.size(-1)), value_layer
+        )
+        context_layer = (
+            context_layer.view(-1, self.num_attention_heads, context_layer.size(-2), context_layer.size(-1))
+            .permute(0, 2, 1, 3)
+            .contiguous()
+        )
+        new_context_layer_shape = context_layer.size()[:-2] + (-1,)
+        context_layer = context_layer.view(new_context_layer_shape)
+        if output_attentions:
+            return (context_layer, attention_probs)
+        else:
+            return context_layer
+
+    def disentangled_attention_bias(self, query_layer, key_layer, relative_pos, rel_embeddings, scale_factor):
+        if relative_pos is None:
+            q = query_layer.size(-2)
+            relative_pos = build_relative_position(
+                q, key_layer.size(-2), bucket_size=self.position_buckets, max_position=self.max_relative_positions
+            )
+        if relative_pos.dim() == 2:
+            relative_pos = relative_pos.unsqueeze(0).unsqueeze(0)
+        elif relative_pos.dim() == 3:
+            relative_pos = relative_pos.unsqueeze(1)
+        # bsz x height x query x key
+        elif relative_pos.dim() != 4:
+            raise ValueError(f"Relative position ids must be of dim 2 or 3 or 4. {relative_pos.dim()}")
+
+        att_span = self.pos_ebd_size
+        relative_pos = relative_pos.long().to(query_layer.device)
+
+        rel_embeddings = rel_embeddings[0 : att_span * 2, :].unsqueeze(0)
+        if self.share_att_key:
+            pos_query_layer = self.transpose_for_scores(
+                self.query_proj(rel_embeddings), self.num_attention_heads
+            ).repeat(query_layer.size(0) // self.num_attention_heads, 1, 1)
+            pos_key_layer = self.transpose_for_scores(self.key_proj(rel_embeddings), self.num_attention_heads).repeat(
+                query_layer.size(0) // self.num_attention_heads, 1, 1
+            )
+        else:
+            if "c2p" in self.pos_att_type:
+                pos_key_layer = self.transpose_for_scores(
+                    self.pos_key_proj(rel_embeddings), self.num_attention_heads
+                ).repeat(
+                    query_layer.size(0) // self.num_attention_heads, 1, 1
+                )  # .split(self.all_head_size, dim=-1)
+            if "p2c" in self.pos_att_type:
+                pos_query_layer = self.transpose_for_scores(
+                    self.pos_query_proj(rel_embeddings), self.num_attention_heads
+                ).repeat(
+                    query_layer.size(0) // self.num_attention_heads, 1, 1
+                )  # .split(self.all_head_size, dim=-1)
+
+        score = 0
+        # content->position
+        if "c2p" in self.pos_att_type:
+            scale = math.sqrt(pos_key_layer.size(-1) * scale_factor)
+            c2p_att = torch.bmm(query_layer, pos_key_layer.transpose(-1, -2))
+            c2p_pos = torch.clamp(relative_pos + att_span, 0, att_span * 2 - 1)
+            c2p_att = torch.gather(
+                c2p_att,
+                dim=-1,
+                index=c2p_pos.squeeze(0).expand([query_layer.size(0), query_layer.size(1), relative_pos.size(-1)]),
+            )
+            score += c2p_att / scale
+
+        # position->content
+        if "p2c" in self.pos_att_type:
+            scale = math.sqrt(pos_query_layer.size(-1) * scale_factor)
+            if key_layer.size(-2) != query_layer.size(-2):
+                r_pos = build_relative_position(
+                    key_layer.size(-2),
+                    key_layer.size(-2),
+                    bucket_size=self.position_buckets,
+                    max_position=self.max_relative_positions,
+                ).to(query_layer.device)
+                r_pos = r_pos.unsqueeze(0)
+            else:
+                r_pos = relative_pos
+
+            p2c_pos = torch.clamp(-r_pos + att_span, 0, att_span * 2 - 1)
+            p2c_att = torch.bmm(key_layer, pos_query_layer.transpose(-1, -2))
+            p2c_att = torch.gather(
+                p2c_att,
+                dim=-1,
+                index=p2c_pos.squeeze(0).expand([query_layer.size(0), key_layer.size(-2), key_layer.size(-2)]),
+            ).transpose(-1, -2)
+            score += p2c_att / scale
+
+        return score
+
+
+# Copied from transformers.models.deberta.modeling_deberta.DebertaEmbeddings with DebertaLayerNorm->LayerNorm
+class DebertaV2Embeddings(nn.Module):
+    """Construct the embeddings from word, position and token_type embeddings."""
+
+    def __init__(self, config):
+        super().__init__()
+        pad_token_id = getattr(config, "pad_token_id", 0)
+        self.embedding_size = getattr(config, "embedding_size", config.hidden_size)
+        self.word_embeddings = nn.Embedding(config.vocab_size, self.embedding_size, padding_idx=pad_token_id)
+
+        self.position_biased_input = getattr(config, "position_biased_input", True)
+        if not self.position_biased_input:
+            self.position_embeddings = None
+        else:
+            self.position_embeddings = nn.Embedding(config.max_position_embeddings, self.embedding_size)
+
+        if config.type_vocab_size > 0:
+            self.token_type_embeddings = nn.Embedding(config.type_vocab_size, self.embedding_size)
+
+        if self.embedding_size != config.hidden_size:
+            self.embed_proj = nn.Linear(self.embedding_size, config.hidden_size, bias=False)
+        self.LayerNorm = LayerNorm(config.hidden_size, config.layer_norm_eps)
+        self.dropout = StableDropout(config.hidden_dropout_prob)
+        self.config = config
+
+        # 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)))
+
+    def forward(self, input_ids=None, token_type_ids=None, position_ids=None, mask=None, inputs_embeds=None):
+        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[:, :seq_length]
+
+        if token_type_ids is None:
+            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)
+
+        if self.position_embeddings is not None:
+            position_embeddings = self.position_embeddings(position_ids.long())
+        else:
+            position_embeddings = torch.zeros_like(inputs_embeds)
+
+        embeddings = inputs_embeds
+        if self.position_biased_input:
+            embeddings += position_embeddings
+        if self.config.type_vocab_size > 0:
+            token_type_embeddings = self.token_type_embeddings(token_type_ids)
+            embeddings += token_type_embeddings
+
+        if self.embedding_size != self.config.hidden_size:
+            embeddings = self.embed_proj(embeddings)
+
+        embeddings = self.LayerNorm(embeddings)
+
+        # if mask is not None:
+        #     if mask.dim() != embeddings.dim():
+        #         if mask.dim() == 4:
+        #             mask = mask.squeeze(1).squeeze(1)
+        #         mask = mask.unsqueeze(2)
+        #     mask = mask.to(embeddings.dtype)
+
+        #     embeddings = embeddings * mask
+
+        embeddings = self.dropout(embeddings)
+        return embeddings
+
+
+# Copied from transformers.models.deberta.modeling_deberta.DebertaPreTrainedModel with Deberta->DebertaV2
+class DebertaV2PreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = DebertaV2Config
+    base_model_prefix = "deberta"
+    _keys_to_ignore_on_load_missing = ["position_ids"]
+    _keys_to_ignore_on_load_unexpected = ["position_embeddings"]
+    supports_gradient_checkpointing = 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_()
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, DebertaV2Encoder):
+            module.gradient_checkpointing = value
+
+
+DEBERTA_START_DOCSTRING = r"""
+    The DeBERTa model was proposed in [DeBERTa: Decoding-enhanced BERT with Disentangled
+    Attention](https://arxiv.org/abs/2006.03654) by Pengcheng He, Xiaodong Liu, Jianfeng Gao, Weizhu Chen. It's build
+    on top of BERT/RoBERTa with two improvements, i.e. disentangled attention and enhanced mask decoder. With those two
+    improvements, it out perform BERT/RoBERTa on a majority of tasks with 80GB pretraining data.
+
+    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 ([`DebertaV2Config`]): 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.
+"""
+
+DEBERTA_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `({0})`):
+            Indices of input sequence tokens in the vocabulary.
+
+            Indices can be obtained using [`DebertaV2Tokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.FloatTensor` of shape `({0})`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+        token_type_ids (`torch.LongTensor` of shape `({0})`, *optional*):
+            Segment token indices to indicate first and second portions of the inputs. Indices are selected in `[0,
+            1]`:
+
+            - 0 corresponds to a *sentence A* token,
+            - 1 corresponds to a *sentence B* token.
+
+            [What are token type IDs?](../glossary#token-type-ids)
+        position_ids (`torch.LongTensor` of shape `({0})`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.max_position_embeddings - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        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 DeBERTa Model transformer outputting raw hidden-states without any specific head on top.",
+    DEBERTA_START_DOCSTRING,
+)
+# Copied from transformers.models.deberta.modeling_deberta.DebertaModel with Deberta->DebertaV2
+class DebertaV2Model(DebertaV2PreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.embeddings = DebertaV2Embeddings(config)
+        self.encoder = DebertaV2Encoder(config)
+        self.z_steps = 2
+        self.config = config
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embeddings.word_embeddings
+
+    def set_input_embeddings(self, new_embeddings):
+        self.embeddings.word_embeddings = new_embeddings
+
+    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("The prune function is not implemented in DeBERTa model.")
+
+    @add_start_docstrings_to_model_forward(DEBERTA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        processor_class=_TOKENIZER_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=BaseModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutput]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        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:
+            token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
+
+        embedding_output = self.embeddings(
+            input_ids=input_ids,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            mask=attention_mask,
+            inputs_embeds=inputs_embeds,
+        )
+
+        encoder_outputs = self.encoder(
+            embedding_output,
+            attention_mask,
+            output_hidden_states=True,
+            output_attentions=output_attentions,
+            return_dict=return_dict,
+        )
+        encoded_layers = encoder_outputs[1]
+
+        if self.z_steps > 1:
+            hidden_states = encoded_layers[-2]
+            layers = [self.encoder.layer[-1] for _ in range(self.z_steps)]
+            query_states = encoded_layers[-1]
+            rel_embeddings = self.encoder.get_rel_embedding()
+            attention_mask = self.encoder.get_attention_mask(attention_mask)
+            rel_pos = self.encoder.get_rel_pos(embedding_output)
+            for layer in layers[1:]:
+                query_states = layer(
+                    hidden_states,
+                    attention_mask,
+                    output_attentions=False,
+                    query_states=query_states,
+                    relative_pos=rel_pos,
+                    rel_embeddings=rel_embeddings,
+                )
+                encoded_layers = encoded_layers + (query_states,)
+
+        sequence_output = encoded_layers[-1]
+
+        if not return_dict:
+            return (sequence_output,) + encoder_outputs[(1 if output_hidden_states else 2) :]
+
+        return BaseModelOutput(
+            last_hidden_state=sequence_output,
+            hidden_states=encoder_outputs.hidden_states if output_hidden_states else None,
+            attentions=encoder_outputs.attentions,
+        )
+
+
+@add_start_docstrings("""DeBERTa Model with a `language modeling` head on top.""", DEBERTA_START_DOCSTRING)
+# Copied from transformers.models.deberta.modeling_deberta.DebertaForMaskedLM with Deberta->DebertaV2
+class DebertaV2ForMaskedLM(DebertaV2PreTrainedModel):
+    _keys_to_ignore_on_load_unexpected = [r"pooler"]
+    _keys_to_ignore_on_load_missing = [r"position_ids", r"predictions.decoder.bias"]
+
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.deberta = DebertaV2Model(config)
+        self.cls = DebertaV2OnlyMLMHead(config)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_output_embeddings(self):
+        return self.cls.predictions.decoder
+
+    def set_output_embeddings(self, new_embeddings):
+        self.cls.predictions.decoder = new_embeddings
+
+    @add_start_docstrings_to_model_forward(DEBERTA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        processor_class=_TOKENIZER_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=MaskedLMOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        labels: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, MaskedLMOutput]:
+        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]`
+        """
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.deberta(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        sequence_output = outputs[0]
+        prediction_scores = self.cls(sequence_output)
+
+        masked_lm_loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()  # -100 index = padding token
+            masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), labels.view(-1))
+
+        if not return_dict:
+            output = (prediction_scores,) + outputs[1:]
+            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,
+        )
+
+
+# copied from transformers.models.bert.BertPredictionHeadTransform with bert -> deberta
+class DebertaV2PredictionHeadTransform(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        if isinstance(config.hidden_act, str):
+            self.transform_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.transform_act_fn = config.hidden_act
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+
+    def forward(self, hidden_states):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.transform_act_fn(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states)
+        return hidden_states
+
+
+# copied from transformers.models.bert.BertLMPredictionHead with bert -> deberta
+class DebertaV2LMPredictionHead(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.transform = DebertaV2PredictionHeadTransform(config)
+
+        # The output weights are the same as the input embeddings, but there is
+        # an output-only bias for each token.
+        self.decoder = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        self.bias = nn.Parameter(torch.zeros(config.vocab_size))
+
+        # Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`
+        self.decoder.bias = self.bias
+
+    def forward(self, hidden_states):
+        hidden_states = self.transform(hidden_states)
+        hidden_states = self.decoder(hidden_states)
+        return hidden_states
+
+
+# copied from transformers.models.bert.BertOnlyMLMHead with bert -> deberta
+class DebertaV2OnlyMLMHead(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.predictions = DebertaV2LMPredictionHead(config)
+
+    def forward(self, sequence_output):
+        prediction_scores = self.predictions(sequence_output)
+        return prediction_scores
+
+
+@add_start_docstrings(
+    """
+    DeBERTa Model transformer with a sequence classification/regression head on top (a linear layer on top of the
+    pooled output) e.g. for GLUE tasks.
+    """,
+    DEBERTA_START_DOCSTRING,
+)
+# Copied from transformers.models.deberta.modeling_deberta.DebertaForSequenceClassification with Deberta->DebertaV2
+class DebertaV2ForSequenceClassification(DebertaV2PreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+
+        num_labels = getattr(config, "num_labels", 2)
+        self.num_labels = num_labels
+
+        self.deberta = DebertaV2Model(config)
+        self.pooler = ContextPooler(config)
+        output_dim = self.pooler.output_dim
+
+        self.classifier = nn.Linear(output_dim, num_labels)
+        drop_out = getattr(config, "cls_dropout", None)
+        drop_out = self.config.hidden_dropout_prob if drop_out is None else drop_out
+        self.dropout = StableDropout(drop_out)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.deberta.get_input_embeddings()
+
+    def set_input_embeddings(self, new_embeddings):
+        self.deberta.set_input_embeddings(new_embeddings)
+
+    @add_start_docstrings_to_model_forward(DEBERTA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        processor_class=_TOKENIZER_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=SequenceClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        labels: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, SequenceClassifierOutput]:
+        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.deberta(
+            input_ids,
+            token_type_ids=token_type_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        encoder_layer = outputs[0]
+        pooled_output = self.pooler(encoder_layer)
+        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:
+                    # regression task
+                    loss_fn = nn.MSELoss()
+                    logits = logits.view(-1).to(labels.dtype)
+                    loss = loss_fn(logits, labels.view(-1))
+                elif labels.dim() == 1 or labels.size(-1) == 1:
+                    label_index = (labels >= 0).nonzero()
+                    labels = labels.long()
+                    if label_index.size(0) > 0:
+                        labeled_logits = torch.gather(
+                            logits, 0, label_index.expand(label_index.size(0), logits.size(1))
+                        )
+                        labels = torch.gather(labels, 0, label_index.view(-1))
+                        loss_fct = CrossEntropyLoss()
+                        loss = loss_fct(labeled_logits.view(-1, self.num_labels).float(), labels.view(-1))
+                    else:
+                        loss = torch.tensor(0).to(logits)
+                else:
+                    log_softmax = nn.LogSoftmax(-1)
+                    loss = -((log_softmax(logits) * labels).sum(-1)).mean()
+            elif 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[1:]
+            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(
+    """
+    DeBERTa 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.
+    """,
+    DEBERTA_START_DOCSTRING,
+)
+# Copied from transformers.models.deberta.modeling_deberta.DebertaForTokenClassification with Deberta->DebertaV2
+class DebertaV2ForTokenClassification(DebertaV2PreTrainedModel):
+    _keys_to_ignore_on_load_unexpected = [r"pooler"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.deberta = DebertaV2Model(config)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.classifier = nn.Linear(config.hidden_size, config.num_labels)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(DEBERTA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        processor_class=_TOKENIZER_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=TokenClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        labels: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, TokenClassifierOutput]:
+        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.deberta(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            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[1:]
+            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(
+    """
+    DeBERTa 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`).
+    """,
+    DEBERTA_START_DOCSTRING,
+)
+# Copied from transformers.models.deberta.modeling_deberta.DebertaForQuestionAnswering with Deberta->DebertaV2
+class DebertaV2ForQuestionAnswering(DebertaV2PreTrainedModel):
+    _keys_to_ignore_on_load_unexpected = [r"pooler"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.deberta = DebertaV2Model(config)
+        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(DEBERTA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        processor_class=_TOKENIZER_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=QuestionAnsweringModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        start_positions: Optional[torch.Tensor] = None,
+        end_positions: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, QuestionAnsweringModelOutput]:
+        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.deberta(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        sequence_output = outputs[0]
+
+        logits = 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[1:]
+            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(
+    """
+    DeBERTa 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.
+    """,
+    DEBERTA_START_DOCSTRING,
+)
+class DebertaV2ForMultipleChoice(DebertaV2PreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+
+        num_labels = getattr(config, "num_labels", 2)
+        self.num_labels = num_labels
+
+        self.deberta = DebertaV2Model(config)
+        self.pooler = ContextPooler(config)
+        output_dim = self.pooler.output_dim
+
+        self.classifier = nn.Linear(output_dim, 1)
+        drop_out = getattr(config, "cls_dropout", None)
+        drop_out = self.config.hidden_dropout_prob if drop_out is None else drop_out
+        self.dropout = StableDropout(drop_out)
+
+        self.init_weights()
+
+    def get_input_embeddings(self):
+        return self.deberta.get_input_embeddings()
+
+    def set_input_embeddings(self, new_embeddings):
+        self.deberta.set_input_embeddings(new_embeddings)
+
+    @add_start_docstrings_to_model_forward(DEBERTA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        processor_class=_TOKENIZER_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=MultipleChoiceModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        inputs_embeds=None,
+        labels=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        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]
+
+        flat_input_ids = input_ids.view(-1, input_ids.size(-1)) if input_ids is not None else None
+        flat_position_ids = position_ids.view(-1, position_ids.size(-1)) if position_ids is not None else None
+        flat_token_type_ids = token_type_ids.view(-1, token_type_ids.size(-1)) if token_type_ids is not None else None
+        flat_attention_mask = attention_mask.view(-1, attention_mask.size(-1)) if attention_mask is not None else None
+        flat_inputs_embeds = (
+            inputs_embeds.view(-1, inputs_embeds.size(-2), inputs_embeds.size(-1))
+            if inputs_embeds is not None
+            else None
+        )
+
+        outputs = self.deberta(
+            flat_input_ids,
+            position_ids=flat_position_ids,
+            token_type_ids=flat_token_type_ids,
+            attention_mask=flat_attention_mask,
+            inputs_embeds=flat_inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        encoder_layer = outputs[0]
+        pooled_output = self.pooler(encoder_layer)
+        pooled_output = self.dropout(pooled_output)
+        logits = 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[1:]
+            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,
+        )