[PyTorch Bart] Split Bart into different models (#9343)

* first try

* remove old template

* finish bart

* finish mbart

* delete unnecessary line

* init pegasus

* save intermediate

* correct pegasus

* finish pegasus

* remove cookie cutter leftover

* add marian

* finish blenderbot

* replace in file

* correctly split blenderbot

* delete "old" folder

* correct "add statement"

* adapt config for tf comp

* correct configs for tf

* remove ipdb

* fix more stuff

* fix mbart

* push pegasus fix

* fix mbart

* more fixes

* fix research projects code

* finish docs for bart, mbart, and marian

* delete unnecessary file

* correct attn typo

* correct configs

* remove pegasus for seq class

* correct peg docs

* correct peg docs

* finish configs

* further improve docs

* add copied from statements to mbart

* fix copied from in mbart

* add copy statements to marian

* add copied from to marian

* add pegasus copied from

* finish pegasus

* finish copied from

* Apply suggestions from code review

* make style

* backward comp blenderbot

* apply lysandres and sylvains suggestions

* apply suggestions

* push last fixes

* fix docs

* fix tok tests

* fix imports code style

* fix doc

src/transformers/models/blenderbot/modeling_blenderbot.py

-#!/usr/bin/env python3
 # coding=utf-8
-# Copyright (c) Facebook, Inc. and its affiliates.
+# Copyright 2021 The Facebook, Inc. and The HuggingFace Inc. team. All rights reserved.
 #
-# This source code is licensed under the MIT license found in the;
+# 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
 #
@@ -13,18 +12,444 @@
 # 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.
-# LICENSE file in the root directory of this source tree.
-""""BlenderbotForConditionalGeneration which inherits from BART"""
+""" PyTorch Blenderbot model. """
+
+
+import math
+import os
+import random
+import warnings
+from typing import Optional, Tuple, Union
 
 import torch
+import torch.nn.functional as F
+from torch import nn
+from torch.nn import CrossEntropyLoss
 
-from ...file_utils import add_start_docstrings
-from ..bart.modeling_bart import BartForConditionalGeneration, BartModel
+from ...activations import ACT2FN
+from ...file_utils import (
+    add_end_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    replace_return_docstrings,
+)
+from ...modeling_outputs import (
+    BaseModelOutput,
+    BaseModelOutputWithPastAndCrossAttentions,
+    Seq2SeqLMOutput,
+    Seq2SeqModelOutput,
+)
+from ...modeling_utils import PreTrainedModel
+from ...utils import logging
+from ..blenderbot_small import BlenderbotSmallForConditionalGeneration, BlenderbotSmallModel
 from .configuration_blenderbot import BlenderbotConfig
 
 
-BLENDER_START_DOCSTRING = r"""
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = "BlenderbotConfig"
+_TOKENIZER_FOR_DOC = "BlenderbotTokenizer"
+
+
+BLENDERBOT_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "facebook/blenderbot-3B",
+    # See all Blenderbot models at https://huggingface.co/models?filter=blenderbot
+]
+
+
+# Copied from transformers.models.bart.modeling_bart.shift_tokens_right
+def shift_tokens_right(input_ids: torch.Tensor, pad_token_id: int, decoder_start_token_id: int):
+    """
+    Shift input ids one token to the right.
+    """
+    shifted_input_ids = input_ids.new_zeros(input_ids.shape)
+    shifted_input_ids[:, 1:] = input_ids[:, :-1].clone()
+    shifted_input_ids[:, 0] = decoder_start_token_id
+
+    assert pad_token_id is not None, "self.model.config.pad_token_id has to be defined."
+    # replace possible -100 values in labels by `pad_token_id`
+    shifted_input_ids.masked_fill_(shifted_input_ids == -100, pad_token_id)
+
+    return shifted_input_ids
+
+
+# Copied from transformers.models.bart.modeling_bart._make_causal_mask
+def _make_causal_mask(input_ids_shape: torch.Size, dtype: torch.dtype, past_key_values_length: int = 0):
+    """
+    Make causal mask used for bi-directional self-attention.
+    """
+    bsz, tgt_len = input_ids_shape
+    mask = torch.full((tgt_len, tgt_len), float("-inf"))
+    mask_cond = torch.arange(mask.size(-1))
+    mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)
+    mask = mask.to(dtype)
+
+    if past_key_values_length > 0:
+        mask = torch.cat([torch.zeros(tgt_len, past_key_values_length, dtype=dtype), mask], dim=-1)
+    return mask[None, None, :, :].expand(bsz, 1, tgt_len, tgt_len + past_key_values_length)
+
+
+# Copied from transformers.models.bart.modeling_bart._expand_mask
+def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
+    """
+    Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
+    """
+    bsz, src_len = mask.size()
+    tgt_len = tgt_len if tgt_len is not None else src_len
+
+    expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)
+
+    inverted_mask = 1.0 - expanded_mask
+
+    return inverted_mask.masked_fill(inverted_mask.bool(), torch.finfo(dtype).min)
+
+
+class BlenderbotLearnedPositionalEmbedding(nn.Embedding):
+    """
+    This module learns positional embeddings up to a fixed maximum size.
+    """
+
+    def __init__(self, num_embeddings: int, embedding_dim: int, padding_idx: int):
+        assert padding_idx is not None, "`padding_idx` should not be None, but of type int"
+        super().__init__(num_embeddings, embedding_dim, padding_idx=padding_idx)
+
+    def forward(self, input_ids_shape: torch.Size, past_key_values_length: int = 0):
+        """`input_ids_shape` is expected to be [bsz x seqlen]."""
+        bsz, seq_len = input_ids_shape[:2]
+        positions = torch.arange(
+            past_key_values_length, past_key_values_length + seq_len, dtype=torch.long, device=self.weight.device
+        )
+        return super().forward(positions)
+
+
+# Copied from transformers.models.bart.modeling_bart.BartAttention with Bart->Blenderbot
+class BlenderbotAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(
+        self,
+        embed_dim: int,
+        num_heads: int,
+        dropout: float = 0.0,
+        is_decoder: bool = False,
+        bias: bool = True,
+    ):
+        super().__init__()
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.dropout = dropout
+        self.head_dim = embed_dim // num_heads
+        assert (
+            self.head_dim * num_heads == self.embed_dim
+        ), f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`: {num_heads})."
+        self.scaling = self.head_dim ** -0.5
+        self.is_decoder = is_decoder
+
+        self.k_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+        self.v_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+        self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+        self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+
+    def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
+        return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        key_value_states: Optional[torch.Tensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        output_attentions: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        """Input shape: Batch x Time x Channel"""
+
+        # if key_value_states are provided this layer is used as a cross-attention layer
+        # for the decoder
+        is_cross_attention = key_value_states is not None
+        bsz, tgt_len, embed_dim = hidden_states.size()
+
+        # get query proj
+        query_states = self.q_proj(hidden_states) * self.scaling
+        # get key, value proj
+        if is_cross_attention and past_key_value is not None:
+            # reuse k,v, cross_attentions
+            key_states = past_key_value[0]
+            value_states = past_key_value[1]
+        elif is_cross_attention:
+            # cross_attentions
+            key_states = self._shape(self.k_proj(key_value_states), -1, bsz)
+            value_states = self._shape(self.v_proj(key_value_states), -1, bsz)
+        elif past_key_value is not None:
+            # reuse k, v, self_attention
+            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
+            key_states = torch.cat([past_key_value[0], key_states], dim=2)
+            value_states = torch.cat([past_key_value[1], value_states], dim=2)
+        else:
+            # self_attention
+            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
+
+        if self.is_decoder:
+            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
+            # Further calls to cross_attention layer can then reuse all cross-attention
+            # key/value_states (first "if" case)
+            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
+            # all previous decoder key/value_states. Further calls to uni-directional self-attention
+            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
+            # if encoder bi-directional self-attention `past_key_value` is always `None`
+            past_key_value = (key_states, value_states)
+
+        proj_shape = (bsz * self.num_heads, -1, self.head_dim)
+        query_states = self._shape(query_states, tgt_len, bsz).view(*proj_shape)
+        key_states = key_states.view(*proj_shape)
+        value_states = value_states.view(*proj_shape)
+
+        src_len = key_states.size(1)
+        attn_weights = torch.bmm(query_states, key_states.transpose(1, 2))
+
+        assert attn_weights.size() == (
+            bsz * self.num_heads,
+            tgt_len,
+            src_len,
+        ), f"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is {attn_weights.size()}"
+
+        if attention_mask is not None:
+            assert attention_mask.size() == (
+                bsz,
+                1,
+                tgt_len,
+                src_len,
+            ), f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.size()}"
+            attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + attention_mask
+            attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
+
+        attn_weights = F.softmax(attn_weights, dim=-1)
+
+        if output_attentions:
+            # this operation is a bit akward, but it's required to
+            # make sure that attn_weights keeps its gradient.
+            # In order to do so, attn_weights have to reshaped
+            # twice and have to be reused in the following
+            attn_weights_reshaped = attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
+            attn_weights = attn_weights_reshaped.view(bsz * self.num_heads, tgt_len, src_len)
+        else:
+            attn_weights_reshaped = None
+
+        attn_probs = F.dropout(attn_weights, p=self.dropout, training=self.training)
+
+        attn_output = torch.bmm(attn_probs, value_states)
+
+        assert attn_output.size() == (
+            bsz * self.num_heads,
+            tgt_len,
+            self.head_dim,
+        ), f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is {attn_output.size()}"
+
+        attn_output = (
+            attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
+            .transpose(1, 2)
+            .reshape(bsz, tgt_len, embed_dim)
+        )
+
+        attn_output = self.out_proj(attn_output)
+
+        return attn_output, attn_weights_reshaped, past_key_value
 
+
+# Copied from transformers.models.mbart.modeling_mbart.MBartEncoderLayer with MBart->Blenderbot
+class BlenderbotEncoderLayer(nn.Module):
+    def __init__(self, config: BlenderbotConfig):
+        super().__init__()
+        self.embed_dim = config.d_model
+        self.self_attn = BlenderbotAttention(
+            embed_dim=self.embed_dim,
+            num_heads=config.encoder_attention_heads,
+            dropout=config.attention_dropout,
+        )
+        self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
+        self.dropout = config.dropout
+        self.activation_fn = ACT2FN[config.activation_function]
+        self.activation_dropout = config.activation_dropout
+        self.fc1 = nn.Linear(self.embed_dim, config.encoder_ffn_dim)
+        self.fc2 = nn.Linear(config.encoder_ffn_dim, self.embed_dim)
+        self.final_layer_norm = nn.LayerNorm(self.embed_dim)
+
+    def forward(self, hidden_states: torch.Tensor, attention_mask: torch.Tensor, output_attentions: bool = False):
+        """
+        Args:
+            hidden_states (:obj:`torch.FloatTensor`): input to the layer of shape `(seq_len, batch, embed_dim)`
+            attention_mask (:obj:`torch.FloatTensor`): attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            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.
+        """
+        residual = hidden_states
+        hidden_states = self.self_attn_layer_norm(hidden_states)
+        hidden_states, attn_weights, _ = self.self_attn(
+            hidden_states=hidden_states, attention_mask=attention_mask, output_attentions=output_attentions
+        )
+        hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
+        hidden_states = residual + hidden_states
+
+        residual = hidden_states
+        hidden_states = self.final_layer_norm(hidden_states)
+        hidden_states = self.activation_fn(self.fc1(hidden_states))
+        hidden_states = F.dropout(hidden_states, p=self.activation_dropout, training=self.training)
+        hidden_states = self.fc2(hidden_states)
+        hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
+        hidden_states = residual + hidden_states
+
+        if torch.isinf(hidden_states).any() or torch.isnan(hidden_states).any():
+            clamp_value = torch.finfo(hidden_states.dtype).max - 1000
+            hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (attn_weights,)
+
+        return outputs
+
+
+# Copied from transformers.models.mbart.modeling_mbart.MBartDecoderLayer with MBart->Blenderbot
+class BlenderbotDecoderLayer(nn.Module):
+    def __init__(self, config: BlenderbotConfig):
+        super().__init__()
+        self.embed_dim = config.d_model
+
+        self.self_attn = BlenderbotAttention(
+            embed_dim=self.embed_dim,
+            num_heads=config.decoder_attention_heads,
+            dropout=config.attention_dropout,
+            is_decoder=True,
+        )
+        self.dropout = config.dropout
+        self.activation_fn = ACT2FN[config.activation_function]
+        self.activation_dropout = config.activation_dropout
+
+        self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
+        self.encoder_attn = BlenderbotAttention(
+            self.embed_dim,
+            config.decoder_attention_heads,
+            dropout=config.attention_dropout,
+            is_decoder=True,
+        )
+        self.encoder_attn_layer_norm = nn.LayerNorm(self.embed_dim)
+        self.fc1 = nn.Linear(self.embed_dim, config.decoder_ffn_dim)
+        self.fc2 = nn.Linear(config.decoder_ffn_dim, self.embed_dim)
+        self.final_layer_norm = nn.LayerNorm(self.embed_dim)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = True,
+    ):
+        """
+        Args:
+            hidden_states (:obj:`torch.FloatTensor`): input to the layer of shape `(seq_len, batch, embed_dim)`
+            attention_mask (:obj:`torch.FloatTensor`): attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            encoder_hidden_states (:obj:`torch.FloatTensor`): cross attention input to the layer of shape `(seq_len, batch, embed_dim)`
+            encoder_attention_mask (:obj:`torch.FloatTensor`): encoder attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            past_key_value (:obj:`Tuple(torch.FloatTensor)`): cached past key and value projection states
+            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.
+        """
+        residual = hidden_states
+        hidden_states = self.self_attn_layer_norm(hidden_states)
+
+        # Self Attention
+        # decoder uni-directional self-attention cached key/values tuple is at positions 1,2
+        self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None
+        # add present self-attn cache to positions 1,2 of present_key_value tuple
+        hidden_states, self_attn_weights, present_key_value = self.self_attn(
+            hidden_states=hidden_states,
+            past_key_value=self_attn_past_key_value,
+            attention_mask=attention_mask,
+            output_attentions=output_attentions,
+        )
+        hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
+        hidden_states = residual + hidden_states
+
+        # Cross-Attention Block
+        cross_attn_present_key_value = None
+        cross_attn_weights = None
+        if encoder_hidden_states is not None:
+            residual = hidden_states
+            hidden_states = self.encoder_attn_layer_norm(hidden_states)
+
+            # cross_attn cached key/values tuple is at positions 3,4 of present_key_value tuple
+            cross_attn_past_key_value = past_key_value[-2:] if past_key_value is not None else None
+            hidden_states, cross_attn_weights, cross_attn_present_key_value = self.encoder_attn(
+                hidden_states=hidden_states,
+                key_value_states=encoder_hidden_states,
+                attention_mask=encoder_attention_mask,
+                past_key_value=cross_attn_past_key_value,
+                output_attentions=output_attentions,
+            )
+            hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
+            hidden_states = residual + hidden_states
+
+            # add cross-attn to positions 3,4 of present_key_value tuple
+            present_key_value = present_key_value + cross_attn_present_key_value
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.final_layer_norm(hidden_states)
+        hidden_states = self.activation_fn(self.fc1(hidden_states))
+        hidden_states = F.dropout(hidden_states, p=self.activation_dropout, training=self.training)
+        hidden_states = self.fc2(hidden_states)
+        hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights, cross_attn_weights)
+
+        if use_cache:
+            outputs += (present_key_value,)
+
+        return outputs
+
+
+class BlenderbotPreTrainedModel(PreTrainedModel):
+    config_class = BlenderbotConfig
+    base_model_prefix = "model"
+
+    def _init_weights(self, module):
+        std = self.config.init_std
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+    @property
+    def dummy_inputs(self):
+        pad_token = self.config.pad_token_id
+        input_ids = torch.tensor([[0, 6, 10, 4, 2], [0, 8, 12, 2, pad_token]], device=self.device)
+        dummy_inputs = {
+            "attention_mask": input_ids.ne(pad_token),
+            "input_ids": input_ids,
+            "decoder_input_ids": input_ids,
+        }
+        return dummy_inputs
+
+
+BLENDERBOT_START_DOCSTRING = r"""
     This model inherits from :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.)
@@ -33,37 +458,777 @@ BLENDER_START_DOCSTRING = r"""
     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 (:class:`~transformers.BlenderbotConfig`):
+            Model configuration class with all the parameters of the model. Initializing with a config file does not
+            load the weights associated with the model, only the configuration. Check out the
+            :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model weights.
 """
 
-BLENDERBOT_PRETRAINED_MODEL_ARCHIVE_LIST = ["facebook/blenderbot-3B", "facebook/blenderbot-90M"]
+BLENDERBOT_GENERATION_EXAMPLE = r"""
+    Conversation example::
+
+        >>> from transformers import BlenderbotTokenizer, BlenderbotForConditionalGeneration
+        >>> mname = 'facebook/blenderbot-400M-distill'
+        >>> model = BlenderbotForConditionalGeneration.from_pretrained(mname)
+        >>> tokenizer = BlenderbotTokenizer.from_pretrained(mname)
+        >>> UTTERANCE = "My friends are cool but they eat too many carbs."
+        >>> print("Human: ", UTTERANCE)
+        >>> inputs = tokenizer([UTTERANCE], return_tensors='pt')
+        >>> reply_ids = model.generate(**inputs)
+        >>> print("Bot: ", tokenizer.batch_decode(reply_ids, skip_special_tokens=True)[0])
+
+        >>> REPLY = "I'm not sure"
+        >>> print("Human: ", REPLY)
+        >>> NEXT_UTTERANCE = (
+        ... "My friends are cool but they eat too many carbs.</s> <s>That's unfortunate. "
+        ... "Are they trying to lose weight or are they just trying to be healthier?</s> "
+        ... "<s> I'm not sure."
+        ... )
+        >>> inputs = tokenizer([NEXT_UTTERANCE], return_tensors='pt')
+        >>> next_reply_ids = model.generate(**inputs)
+        >>> print("Bot: ", tokenizer.batch_decode(next_reply_ids, skip_special_tokens=True)[0])
+"""
+
+BLENDERBOT_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+
+            Indices can be obtained using :class:`~transformers.BlenderbotTokenizer`. See
+            :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__` for
+            details.
+
+            `What are input IDs? <../glossary.html#input-ids>`__
+        attention_mask (:obj:`torch.Tensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Mask to avoid performing attention on padding token indices. Mask values selected in ``[0, 1]``:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            `What are attention masks? <../glossary.html#attention-mask>`__
+        decoder_input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, target_sequence_length)`, `optional`):
+            Indices of decoder input sequence tokens in the vocabulary.
+
+            Indices can be obtained using :class:`~transformers.BlenderbotTokenizer`. See
+            :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__` for
+            details.
+
+            `What are input IDs? <../glossary.html#input-ids>`__
+
+            Blenderbot uses the :obj:`bos_token_id` as the starting token for :obj:`decoder_input_ids` generation. If
+            :obj:`past_key_values` is used, optionally only the last :obj:`decoder_input_ids` have to be input (see
+            :obj:`past_key_values`).
+        decoder_attention_mask (:obj:`torch.LongTensor` of shape :obj:`(batch_size, target_sequence_length)`, `optional`):
+            Default behavior: generate a tensor that ignores pad tokens in :obj:`decoder_input_ids`. Causal mask will
+            also be used by default.
+
+            If you want to change padding behavior, you should read :func:`modeling_blenderbot._prepare_decoder_inputs`
+            and modify to your needs. See diagram 1 in `the paper <https://arxiv.org/abs/1910.13461>`__ for more
+            information on the default strategy.
+        encoder_outputs (:obj:`tuple(tuple(torch.FloatTensor)`, `optional`):
+            Tuple consists of (:obj:`last_hidden_state`, `optional`: :obj:`hidden_states`, `optional`:
+            :obj:`attentions`) :obj:`last_hidden_state` of shape :obj:`(batch_size, sequence_length, hidden_size)`,
+            `optional`) is a sequence of hidden-states at the output of the last layer of the encoder. Used in the
+            cross-attention of the decoder.
+        past_key_values (:obj:`Tuple[Tuple[torch.Tensor]]` of length :obj:`config.n_layers` with each tuple having 2 tuples each of which has 2 tensors of shape :obj:`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
+            Contains precomputed key and value hidden-states of the attention blocks. Can be used to speed up decoding.
+
+            If :obj:`past_key_values` are used, the user can optionally input only the last :obj:`decoder_input_ids`
+            (those that don't have their past key value states given to this model) of shape :obj:`(batch_size, 1)`
+            instead of all :obj:`decoder_input_ids`` of shape :obj:`(batch_size, sequence_length)`.
+        inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, 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.
+        decoder_inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, target_sequence_length, hidden_size)`, `optional`):
+            Optionally, instead of passing :obj:`decoder_input_ids` you can choose to directly pass an embedded
+            representation. If :obj:`past_key_values` is used, optionally only the last :obj:`decoder_inputs_embeds`
+            have to be input (see :obj:`past_key_values`). This is useful if you want more control over how to convert
+            :obj:`decoder_input_ids` indices into associated vectors than the model's internal embedding lookup matrix.
+
+            If :obj:`decoder_input_ids` and :obj:`decoder_inputs_embeds` are both unset, :obj:`decoder_inputs_embeds`
+            takes the value of :obj:`inputs_embeds`.
+        use_cache (:obj:`bool`, `optional`):
+            If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up
+            decoding (see :obj:`past_key_values`).
+        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.
+"""
+
+
+class BlenderbotEncoder(BlenderbotPreTrainedModel):
+    """
+    Transformer encoder consisting of *config.encoder_layers* self attention layers. Each layer is a
+    :class:`BlenderbotEncoderLayer`.
+
+    Args:
+        config: BlenderbotConfig
+        embed_tokens (torch.nn.Embedding): output embedding
+    """
+
+    def __init__(self, config: BlenderbotConfig, embed_tokens: Optional[nn.Embedding] = None):
+        super().__init__(config)
+
+        self.dropout = config.dropout
+        self.layerdrop = config.encoder_layerdrop
+
+        embed_dim = config.d_model
+        self.padding_idx = config.pad_token_id
+        self.max_source_positions = config.max_position_embeddings
+        self.embed_scale = math.sqrt(embed_dim) if config.scale_embedding else 1.0
+
+        if embed_tokens is not None:
+            self.embed_tokens = embed_tokens
+        else:
+            self.embed_tokens = nn.Embedding(config.vocab_size, embed_dim, self.padding_idx)
+
+        self.embed_positions = BlenderbotLearnedPositionalEmbedding(
+            config.max_position_embeddings,
+            embed_dim,
+            self.padding_idx,
+        )
+        self.layers = nn.ModuleList([BlenderbotEncoderLayer(config) for _ in range(config.encoder_layers)])
+        self.layer_norm = nn.LayerNorm(config.d_model)
+
+        self.init_weights()
+
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        inputs_embeds=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        Args:
+            input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`):
+                Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you
+                provide it.
+
+                Indices can be obtained using :class:`~transformers.BlenderbotTokenizer`. See
+                :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__`
+                for details.
+
+                `What are input IDs? <../glossary.html#input-ids>`__
+            attention_mask (:obj:`torch.Tensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+                Mask to avoid performing attention on padding token indices. Mask values selected in ``[0, 1]``:
+
+                - 1 for tokens that are **not masked**,
+                - 0 for tokens that are **masked**.
+
+                `What are attention masks? <../glossary.html#attention-mask>`__
+            inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, 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.
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # retrieve input_ids and inputs_embeds
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+            input_ids = input_ids.view(-1, input_shape[-1])
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids) * self.embed_scale
+
+        embed_pos = self.embed_positions(input_shape)
+
+        hidden_states = inputs_embeds + embed_pos
+        hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
+
+        # expand attention_mask
+        if attention_mask is not None:
+            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+            attention_mask = _expand_mask(attention_mask, inputs_embeds.dtype)
+
+        encoder_states = () if output_hidden_states else None
+        all_attentions = () if output_attentions else None
+        for encoder_layer in self.layers:
+            if output_hidden_states:
+                encoder_states = encoder_states + (hidden_states,)
+            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
+            dropout_probability = random.uniform(0, 1)
+            if self.training and (dropout_probability < self.layerdrop):  # skip the layer
+                layer_outputs = (None, None)
+            else:
+                if getattr(self.config, "gradient_checkpointing", False):
+
+                    def create_custom_forward(module):
+                        def custom_forward(*inputs):
+                            return module(*inputs, output_attentions)
+
+                        return custom_forward
+
+                    layer_outputs = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(encoder_layer),
+                        hidden_states,
+                        attention_mask,
+                    )
+                else:
+                    layer_outputs = encoder_layer(hidden_states, attention_mask, output_attentions=output_attentions)
+
+                hidden_states = layer_outputs[0]
+
+            if output_attentions:
+                all_attentions = all_attentions + (layer_outputs[1],)
+
+        # add final layer norm
+        hidden_states = self.layer_norm(hidden_states)
+
+        if output_hidden_states:
+            encoder_states = encoder_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None)
+        return BaseModelOutput(
+            last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions
+        )
+
+
+class BlenderbotDecoder(BlenderbotPreTrainedModel):
+    """
+    Transformer decoder consisting of *config.decoder_layers* layers. Each layer is a :class:`BlenderbotDecoderLayer`
+
+    Args:
+        config: BlenderbotConfig
+        embed_tokens (torch.nn.Embedding): output embedding
+    """
+
+    def __init__(self, config: BlenderbotConfig, embed_tokens: Optional[nn.Embedding] = None):
+        super().__init__(config)
+        self.dropout = config.dropout
+        self.layerdrop = config.decoder_layerdrop
+        self.padding_idx = config.pad_token_id
+        self.max_target_positions = config.max_position_embeddings
+        self.embed_scale = math.sqrt(config.d_model) if config.scale_embedding else 1.0
+
+        if embed_tokens is not None:
+            self.embed_tokens = embed_tokens
+        else:
+            self.embed_tokens = nn.Embedding(config.vocab_size, config.d_model, self.padding_idx)
+
+        self.embed_positions = BlenderbotLearnedPositionalEmbedding(
+            config.max_position_embeddings,
+            config.d_model,
+            self.padding_idx,
+        )
+        self.layers = nn.ModuleList([BlenderbotDecoderLayer(config) for _ in range(config.decoder_layers)])
+        self.layer_norm = nn.LayerNorm(config.d_model)
+
+        self.init_weights()
+
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_values=None,
+        inputs_embeds=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        Args:
+            input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`):
+                Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you
+                provide it.
+
+                Indices can be obtained using :class:`~transformers.BlenderbotTokenizer`. See
+                :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__`
+                for details.
+
+                `What are input IDs? <../glossary.html#input-ids>`__
+            attention_mask (:obj:`torch.Tensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+                Mask to avoid performing attention on padding token indices. Mask values selected in ``[0, 1]``:
+
+                - 1 for tokens that are **not masked**,
+                - 0 for tokens that are **masked**.
+
+                `What are attention masks? <../glossary.html#attention-mask>`__
+            encoder_hidden_states (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, encoder_sequence_length, hidden_size)`, `optional`):
+                Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention
+                of the decoder.
+            encoder_attention_mask (:obj:`torch.LongTensor` of shape :obj:`(batch_size, encoder_sequence_length)`, `optional`):
+                Mask to avoid performing cross-attention on padding tokens indices of encoder input_ids. Mask values
+                selected in ``[0, 1]``:
+
+                - 1 for tokens that are **not masked**,
+                - 0 for tokens that are **masked**.
+
+                `What are attention masks? <../glossary.html#attention-mask>`__
+            past_key_values (:obj:`Tuple[Tuple[torch.Tensor]]` of length :obj:`config.n_layers` with each tuple having 2 tuples each of which has 2 tensors of shape :obj:`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
+                Contains precomputed key and value hidden-states of the attention blocks. Can be used to speed up
+                decoding.
+
+                If :obj:`past_key_values` are used, the user can optionally input only the last
+                :obj:`decoder_input_ids` (those that don't have their past key value states given to this model) of
+                shape :obj:`(batch_size, 1)` instead of all :obj:`decoder_input_ids`` of shape :obj:`(batch_size,
+                sequence_length)`.
+            inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, 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.
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # retrieve input_ids and inputs_embeds
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+            input_ids = input_ids.view(-1, input_shape[-1])
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+        else:
+            raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
+
+        # past_key_values_length
+        past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids) * self.embed_scale
+
+        # create causal mask
+        # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+        combined_attention_mask = None
+        if input_shape[-1] > 1:
+            combined_attention_mask = _make_causal_mask(
+                input_shape, inputs_embeds.dtype, past_key_values_length=past_key_values_length
+            ).to(self.device)
+
+        if attention_mask is not None and combined_attention_mask is not None:
+            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+            combined_attention_mask = combined_attention_mask + _expand_mask(
+                attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]
+            )
+
+        # expand encoder attention mask
+        if encoder_hidden_states is not None and encoder_attention_mask is not None:
+            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+            encoder_attention_mask = _expand_mask(encoder_attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1])
+
+        # embed positions
+        positions = self.embed_positions(input_shape, past_key_values_length)
+
+        # in constrast to Bart, Blenderbot applies layernorm on inputs_embeds
+        hidden_states = inputs_embeds + positions
+
+        hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        all_cross_attentions = () if output_attentions else None
+        next_decoder_cache = () if use_cache else None
+        for idx, decoder_layer in enumerate(self.layers):
+            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+            dropout_probability = random.uniform(0, 1)
+            if self.training and (dropout_probability < self.layerdrop):
+                continue
+
+            past_key_value = past_key_values[idx] if past_key_values is not None else None
+
+            if getattr(self.config, "gradient_checkpointing", False):
+                if use_cache:
+                    raise ValueError(
+                        "When using `gradient_checkpointing, make sure that `use_cache=False` and `config.use_cache=False`."
+                    )
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        # None for past_key_value
+                        return module(*inputs, output_attentions, use_cache)
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(decoder_layer),
+                    hidden_states,
+                    combined_attention_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    None,
+                )
+            else:
+
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask=combined_attention_mask,
+                    encoder_hidden_states=encoder_hidden_states,
+                    encoder_attention_mask=encoder_attention_mask,
+                    past_key_value=past_key_value,
+                    output_attentions=output_attentions,
+                    use_cache=use_cache,
+                )
+            hidden_states = layer_outputs[0]
+
+            if use_cache:
+                next_decoder_cache += (layer_outputs[3 if output_attentions else 1],)
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+                all_cross_attentions += (layer_outputs[2],)
+
+        # add final layer norm
+        hidden_states = self.layer_norm(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        next_cache = next_decoder_cache if use_cache else None
+        if not return_dict:
+            return tuple(
+                v
+                for v in [hidden_states, next_cache, all_hidden_states, all_self_attns, all_cross_attentions]
+                if v is not None
+            )
+        return BaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+            cross_attentions=all_cross_attentions,
+        )
 
 
 @add_start_docstrings(
-    "The bare BlenderBot Model transformer outputting raw hidden-states without any specific head on top.",
-    BLENDER_START_DOCSTRING,
+    "The bare Blenderbot Model outputting raw hidden-states without any specific head on top.",
+    BLENDERBOT_START_DOCSTRING,
 )
-class BlenderbotModel(BartModel):
-    r"""
-    This class overrides :class:`~transformers.BartModel`. Please check the superclass for the appropriate
-    documentation alongside usage examples.
-    """
+class BlenderbotModel(BlenderbotPreTrainedModel):
+    def __init__(self, config: BlenderbotConfig):
+        super().__init__(config)
 
-    config_class = BlenderbotConfig
+        padding_idx, vocab_size = config.pad_token_id, config.vocab_size
+        self.shared = nn.Embedding(vocab_size, config.d_model, padding_idx)
+
+        self.encoder = BlenderbotEncoder(config, self.shared)
+        self.decoder = BlenderbotDecoder(config, self.shared)
+
+        self.init_weights()
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], *model_args, **kwargs):
+        if pretrained_model_name_or_path == "facebook/blenderbot-90M":
+            warnings.warn(
+                "The checkpoint `facebook/blenderbot-90M` is deprecated. In the future, please use the identical checkpoint `facebook/small_blenderbot-90M` with `BlenderbotSmallModel.from_pretrained('facebook/small_blenderbot-90M')` instead.",
+                FutureWarning,
+            )
+            return BlenderbotSmallModel.from_pretrained(pretrained_model_name_or_path)
+
+        return super(BlenderbotModel, cls).from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+
+    def get_input_embeddings(self):
+        return self.shared
+
+    def set_input_embeddings(self, value):
+        self.shared = value
+        self.encoder.embed_tokens = self.shared
+        self.decoder.embed_tokens = self.shared
+
+    def get_encoder(self):
+        return self.encoder
+
+    def get_decoder(self):
+        return self.decoder
+
+    @add_start_docstrings_to_model_forward(BLENDERBOT_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=Seq2SeqModelOutput, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        decoder_input_ids=None,
+        decoder_attention_mask=None,
+        encoder_outputs=None,
+        past_key_values=None,
+        inputs_embeds=None,
+        decoder_inputs_embeds=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        Returns:
+
+        Example::
+
+            >>> from transformers import BlenderbotTokenizer, BlenderbotModel
+
+            >>> model = BlenderbotModel.from_pretrained("facebook/blenderbot-400M-distill")
+            >>> tokenizer = BlenderbotTokenizer.from_pretrained("facebook/blenderbot-400M-distill")
+
+            >>> input_ids = tokenizer("Studies have been shown that owning a dog is good for you", return_tensors="pt").input_ids  # Batch size 1
+            >>> decoder_input_ids = tokenizer("Studies show that", return_tensors="pt").input_ids  # Batch size 1
+            >>> outputs = model(input_ids=input_ids, decoder_input_ids=decoder_input_ids)
+
+            >>> last_hidden_states = outputs.last_hidden_state
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if encoder_outputs is None:
+            encoder_outputs = self.encoder(
+                input_ids=input_ids,
+                attention_mask=attention_mask,
+                inputs_embeds=inputs_embeds,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+            )
+        # If the user passed a tuple for encoder_outputs, we wrap it in a BaseModelOutput when return_dict=True
+        elif return_dict and not isinstance(encoder_outputs, BaseModelOutput):
+            encoder_outputs = BaseModelOutput(
+                last_hidden_state=encoder_outputs[0],
+                hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None,
+                attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None,
+            )
+
+        # decoder outputs consists of (dec_features, past_key_value, dec_hidden, dec_attn)
+        decoder_outputs = self.decoder(
+            input_ids=decoder_input_ids,
+            attention_mask=decoder_attention_mask,
+            encoder_hidden_states=encoder_outputs[0],
+            encoder_attention_mask=attention_mask,
+            past_key_values=past_key_values,
+            inputs_embeds=decoder_inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        if not return_dict:
+            return decoder_outputs + encoder_outputs
+
+        return Seq2SeqModelOutput(
+            last_hidden_state=decoder_outputs.last_hidden_state,
+            past_key_values=decoder_outputs.past_key_values,
+            decoder_hidden_states=decoder_outputs.hidden_states,
+            decoder_attentions=decoder_outputs.attentions,
+            cross_attentions=decoder_outputs.cross_attentions,
+            encoder_last_hidden_state=encoder_outputs.last_hidden_state,
+            encoder_hidden_states=encoder_outputs.hidden_states,
+            encoder_attentions=encoder_outputs.attentions,
+        )
 
 
 @add_start_docstrings(
-    "The BlenderBot Model with a language modeling head. Can be used for summarization.", BLENDER_START_DOCSTRING
+    "The Blenderbot Model with a language modeling head. Can be used for summarization.", BLENDERBOT_START_DOCSTRING
 )
-class BlenderbotForConditionalGeneration(BartForConditionalGeneration):
-    """
-    This class overrides :class:`~transformers.BartForConditionalGeneration`. Please check the superclass for the
-    appropriate documentation alongside usage examples.
-    """
+class BlenderbotForConditionalGeneration(BlenderbotPreTrainedModel):
+    base_model_prefix = "model"
+    _keys_to_ignore_on_load_missing = [
+        r"final_logits_bias",
+        r"encoder\.version",
+        r"decoder\.version",
+        r"lm_head\.weight",
+    ]
 
-    config_class = BlenderbotConfig
+    def __init__(self, config: BlenderbotConfig):
+        super().__init__(config)
+        self.model = BlenderbotModel(config)
+        self.register_buffer("final_logits_bias", torch.zeros((1, self.model.shared.num_embeddings)))
+        self.lm_head = nn.Linear(config.d_model, self.model.shared.num_embeddings, bias=False)
+
+        self.init_weights()
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], *model_args, **kwargs):
+        if pretrained_model_name_or_path == "facebook/blenderbot-90M":
+            warnings.warn(
+                "The checkpoint `facebook/blenderbot-90M` is deprecated. In the future, please use the identical checkpoint `facebook/small_blenderbot-90M` with `BlenderbotSmallForConditionalGeneration.from_pretrained('facebook/small_blenderbot-90M')` instead.",
+                FutureWarning,
+            )
+            return BlenderbotSmallForConditionalGeneration.from_pretrained(pretrained_model_name_or_path)
+
+        return super(BlenderbotForConditionalGeneration, cls).from_pretrained(
+            pretrained_model_name_or_path, *model_args, **kwargs
+        )
+
+    def get_encoder(self):
+        return self.model.get_encoder()
+
+    def get_decoder(self):
+        return self.model.get_decoder()
+
+    def resize_token_embeddings(self, new_num_tokens: int) -> nn.Embedding:
+        new_embeddings = super().resize_token_embeddings(new_num_tokens)
+        self._resize_final_logits_bias(new_num_tokens)
+        return new_embeddings
+
+    def _resize_final_logits_bias(self, new_num_tokens: int) -> None:
+        old_num_tokens = self.final_logits_bias.shape[-1]
+        if new_num_tokens <= old_num_tokens:
+            new_bias = self.final_logits_bias[:, :new_num_tokens]
+        else:
+            extra_bias = torch.zeros((1, new_num_tokens - old_num_tokens), device=self.final_logits_bias.device)
+            new_bias = torch.cat([self.final_logits_bias, extra_bias], dim=1)
+        self.register_buffer("final_logits_bias", new_bias)
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    @add_start_docstrings_to_model_forward(BLENDERBOT_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=Seq2SeqLMOutput, config_class=_CONFIG_FOR_DOC)
+    @add_end_docstrings(BLENDERBOT_GENERATION_EXAMPLE)
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        decoder_input_ids=None,
+        decoder_attention_mask=None,
+        encoder_outputs=None,
+        past_key_values=None,
+        inputs_embeds=None,
+        decoder_inputs_embeds=None,
+        labels=None,
+        use_cache=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 either be in ``[0, ...,
+            config.vocab_size]`` or -100 (see ``input_ids`` docstring). Tokens with indices set to ``-100`` are ignored
+            (masked), the loss is only computed for the tokens with labels in ``[0, ..., config.vocab_size]``.
+
+        Returns:
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if labels is not None:
+            if decoder_input_ids is None:
+                decoder_input_ids = shift_tokens_right(
+                    labels, self.config.pad_token_id, self.config.decoder_start_token_id
+                )
+
+        outputs = self.model(
+            input_ids,
+            attention_mask=attention_mask,
+            decoder_input_ids=decoder_input_ids,
+            encoder_outputs=encoder_outputs,
+            decoder_attention_mask=decoder_attention_mask,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            decoder_inputs_embeds=decoder_inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        lm_logits = self.lm_head(outputs[0]) + self.final_logits_bias
+
+        masked_lm_loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            masked_lm_loss = loss_fct(lm_logits.view(-1, self.config.vocab_size), labels.view(-1))
+
+        if not return_dict:
+            output = (lm_logits,) + outputs[1:]
+            return ((masked_lm_loss,) + output) if masked_lm_loss is not None else output
+
+        return Seq2SeqLMOutput(
+            loss=masked_lm_loss,
+            logits=lm_logits,
+            past_key_values=outputs.past_key_values,
+            decoder_hidden_states=outputs.decoder_hidden_states,
+            decoder_attentions=outputs.decoder_attentions,
+            cross_attentions=outputs.cross_attentions,
+            encoder_last_hidden_state=outputs.encoder_last_hidden_state,
+            encoder_hidden_states=outputs.encoder_hidden_states,
+            encoder_attentions=outputs.encoder_attentions,
+        )
+
+    def prepare_inputs_for_generation(
+        self, decoder_input_ids, past=None, attention_mask=None, use_cache=None, encoder_outputs=None, **kwargs
+    ):
+        # cut decoder_input_ids if past is used
+        if past is not None:
+            decoder_input_ids = decoder_input_ids[:, -1:]
+
+        return {
+            "input_ids": None,  # encoder_outputs is defined. input_ids not needed
+            "encoder_outputs": encoder_outputs,
+            "past_key_values": past,
+            "decoder_input_ids": decoder_input_ids,
+            "attention_mask": attention_mask,
+            "use_cache": use_cache,  # change this to avoid caching (presumably for debugging)
+        }
 
     def adjust_logits_during_generation(self, logits, cur_len, max_length):
-        logits[:, self.config.bos_token_id] = -torch.finfo(torch.float16).max  # near infinity fp16
         if cur_len == max_length - 1 and self.config.eos_token_id is not None:
             self._force_token_id_to_be_generated(logits, self.config.eos_token_id)
         return logits
+
+    @staticmethod
+    def _force_token_id_to_be_generated(scores, token_id) -> None:
+        """force one of token_ids to be generated by setting prob of all other tokens to 0 (logprob=-float("inf"))"""
+        scores[:, [x for x in range(scores.shape[1]) if x != token_id]] = -float("inf")
+
+    @staticmethod
+    def _reorder_cache(past, beam_idx):
+        reordered_past = ()
+        for layer_past in past:
+            # cached cross_attention states don't have to be reordered -> they are always the same
+            reordered_past += (
+                tuple(past_state.index_select(0, beam_idx) for past_state in layer_past[:2]) + layer_past[2:],
+            )
+        return reordered_past

src/transformers/models/blenderbot/tokenization_blenderbot.py

-#!/usr/bin/env python3
 # coding=utf-8
-# Copyright (c) Facebook, Inc. and its affiliates.
+# Copyright 2021 The Facebook Inc. and The HuggingFace Inc. team. All rights reserved.
 #
-# This source code is licensed under the MIT license found in the;
+# 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
 #
@@ -13,15 +12,10 @@
 # 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.
-# LICENSE file in the root directory of this source tree.
-""""BlenderbotTokenizer and BlenderbotSmallTokenizer"""
-import json
-import os
-from typing import Dict, List, Optional, Tuple
+"""Tokenization class for Blenderbot."""
 
-import regex as re
+from typing import List
 
-from ...tokenization_utils import PreTrainedTokenizer
 from ...utils import logging
 from ..roberta.tokenization_roberta import RobertaTokenizer
 
@@ -93,177 +87,3 @@ def get_pairs(word):
 
     pairs = set(pairs)
     return pairs
-
-
-class BlenderbotSmallTokenizer(PreTrainedTokenizer):
-    """
-    Constructs a Blenderbot-90M tokenizer based on BPE (Byte-Pair-Encoding)
-
-    This tokenizer inherits from :class:`~transformers.PreTrainedTokenizer` which contains most of the main methods.
-    Users should refer to the superclass for more information regarding methods.
-
-    Args:
-        vocab_file (:obj:`str`):
-            File containing the vocabulary.
-        merges_file (:obj:`str`):
-            Path to the merges file.
-        bos_token (:obj:`str`, `optional`, defaults to :obj:`"__start__"`):
-            The beginning of sentence token.
-        eos_token (:obj:`str`, `optional`, defaults to :obj:`"__end__"`):
-            The end of sentence token.
-        unk_token (:obj:`str`, `optional`, defaults to :obj:`"__unk__"`):
-            The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this
-            token instead.
-        pad_token (:obj:`str`, `optional`, defaults to :obj:`"__pad__"`):
-            The token used for padding, for example when batching sequences of different lengths.
-        **kwargs
-            Additional keyword arguments passed along to :class:`~transformers.PreTrainedTokenizer`
-    """
-
-    vocab_files_names = {"vocab_file": "vocab.json", "merges_file": "merges.txt"}
-    pretrained_vocab_files_map = {
-        "vocab_file": {"facebook/blenderbot-90M": "https://cdn.huggingface.co/facebook/blenderbot-90M/vocab.json"},
-        "merges_file": {"facebook/blenderbot-90M": "https://cdn.huggingface.co/facebook/blenderbot-90M/merges.txt"},
-    }
-    max_model_input_sizes = {"facebook/blenderbot-90M": 512}
-
-    def __init__(
-        self,
-        vocab_file,
-        merges_file,
-        bos_token="__start__",
-        eos_token="__end__",
-        unk_token="__unk__",
-        pad_token="__null__",
-        **kwargs
-    ):
-        super().__init__(unk_token=unk_token, bos_token=bos_token, eos_token=eos_token, pad_token=pad_token, **kwargs)
-
-        with open(vocab_file, encoding="utf-8") as vocab_handle:
-            self.encoder = json.load(vocab_handle)
-        self.decoder = {v: k for k, v in self.encoder.items()}
-        with open(merges_file, encoding="utf-8") as merges_handle:
-            merges = merges_handle.read().split("\n")[1:-1]
-        merges = [tuple(merge.split()) for merge in merges]
-        self.bpe_ranks = dict(zip(merges, range(len(merges))))
-        self.cache = {}
-
-    @property
-    def vocab_size(self) -> int:
-        return len(self.encoder)
-
-    def get_vocab(self) -> Dict:
-        return dict(self.encoder, **self.added_tokens_encoder)
-
-    def bpe(self, token: str) -> str:
-        if token in self.cache:
-            return self.cache[token]
-        token = re.sub("([.,!?()])", r" \1", token)
-        token = re.sub("(')", r" \1 ", token)
-        token = re.sub(r"\s{2,}", " ", token)
-        if "\n" in token:
-            token = token.replace("\n", " __newln__")
-
-        tokens = token.split(" ")
-        words = []
-        for token in tokens:
-            if not len(token):
-                continue
-
-            token = token.lower()
-            word = tuple(token)
-            word = tuple(list(word[:-1]) + [word[-1] + "</w>"])
-            pairs = get_pairs(word)
-
-            if not pairs:
-                words.append(token)
-                continue
-
-            while True:
-                bigram = min(pairs, key=lambda pair: self.bpe_ranks.get(pair, float("inf")))
-                if bigram not in self.bpe_ranks:
-                    break
-                first, second = bigram
-                new_word = []
-                i = 0
-
-                while i < len(word):
-                    try:
-                        j = word.index(first, i)
-                        new_word.extend(word[i:j])
-                        i = j
-                    except ValueError:
-                        new_word.extend(word[i:])
-                        break
-
-                    if word[i] == first and i < len(word) - 1 and word[i + 1] == second:
-                        new_word.append(first + second)
-                        i += 2
-                    else:
-                        new_word.append(word[i])
-                        i += 1
-                new_word = tuple(new_word)
-                word = new_word
-                if len(word) == 1:
-                    break
-                else:
-                    pairs = get_pairs(word)
-            word = "@@ ".join(word)
-            word = word[:-4]
-
-            self.cache[token] = word
-            words.append(word)
-        return " ".join(words)
-
-    def _tokenize(self, text: str) -> List[str]:
-        """ Split a string into tokens using BPE."""
-        split_tokens = []
-
-        words = re.findall(r"\S+\n?", text)
-
-        for token in words:
-            split_tokens.extend([t for t in self.bpe(token).split(" ")])
-        return split_tokens
-
-    def _convert_token_to_id(self, token: str) -> int:
-        """ Converts a token to an id using the vocab. """
-        token = token.lower()
-        return self.encoder.get(token, self.encoder.get(self.unk_token))
-
-    def _convert_id_to_token(self, index: int) -> str:
-        """Converts an index (integer) in a token (str) using the vocab."""
-        return self.decoder.get(index, self.unk_token)
-
-    def convert_tokens_to_string(self, tokens: List[str]) -> str:
-        """ Converts a sequence of tokens  in a single string. """
-        out_string = " ".join(tokens).replace("@@ ", "").strip()
-        return out_string
-
-    def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
-        if not os.path.isdir(save_directory):
-            logger.error("Vocabulary path ({}) should be a directory".format(save_directory))
-            return
-        vocab_file = os.path.join(
-            save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]
-        )
-        merge_file = os.path.join(
-            save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["merges_file"]
-        )
-
-        with open(vocab_file, "w", encoding="utf-8") as f:
-            f.write(json.dumps(self.encoder, ensure_ascii=False))
-
-        index = 0
-        with open(merge_file, "w", encoding="utf-8") as writer:
-            writer.write("#version: 0.2\n")
-            for bpe_tokens, token_index in sorted(self.bpe_ranks.items(), key=lambda kv: kv[1]):
-                if index != token_index:
-                    logger.warning(
-                        "Saving vocabulary to {}: BPE merge indices are not consecutive."
-                        " Please check that the tokenizer is not corrupted!".format(merge_file)
-                    )
-                    index = token_index
-                writer.write(" ".join(bpe_tokens) + "\n")
-                index += 1
-
-        return vocab_file, merge_file

src/transformers/models/blenderbot_small/__init__.py

+# flake8: noqa
+# There's no way to ignore "F401 '...' imported but unused" warnings in this
+# module, but to preserve other warnings. So, don't check this module at all.
+
+# Copyright 2020 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from ...file_utils import is_torch_available
+from .configuration_blenderbot_small import BLENDERBOT_SMALL_PRETRAINED_CONFIG_ARCHIVE_MAP, BlenderbotSmallConfig
+from .tokenization_blenderbot_small import BlenderbotSmallTokenizer
+
+
+if is_torch_available():
+    from .modeling_blenderbot_small import (
+        BLENDERBOT_SMALL_PRETRAINED_MODEL_ARCHIVE_LIST,
+        BlenderbotSmallForConditionalGeneration,
+        BlenderbotSmallModel,
+        BlenderbotSmallPreTrainedModel,
+    )

src/transformers/models/blenderbot_small/configuration_blenderbot_small.py

+# coding=utf-8
+# Copyright 2021 The Facebook, Inc. and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" BlenderbotSmall model configuration """
+
+from ...configuration_utils import PretrainedConfig
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+BLENDERBOT_SMALL_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "facebook/blenderbot_small-90M": "https://huggingface.co/facebook/blenderbot_small-90M/resolve/main/config.json",
+    # See all BlenderbotSmall models at https://huggingface.co/models?filter=blenderbot_small
+}
+
+
+class BlenderbotSmallConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a :class:`~transformers.BlenderbotSmallModel`. It is
+    used to instantiate an BlenderbotSmall model according to the specified arguments, defining the model architecture.
+    Instantiating a configuration with the defaults will yield a similar configuration to that of the BlenderbotSmall
+    `facebook/blenderbot_small-90M <https://huggingface.co/facebook/blenderbot_small-90M>`__ architecture.
+
+    Configuration objects inherit from :class:`~transformers.PretrainedConfig` and can be used to control the model
+    outputs. Read the documentation from :class:`~transformers.PretrainedConfig` for more information.
+
+
+    Args:
+        vocab_size (:obj:`int`, `optional`, defaults to 50265):
+            Vocabulary size of the BlenderbotSmall model. Defines the number of different tokens that can be
+            represented by the :obj:`inputs_ids` passed when calling :class:`~transformers.BlenderbotSmallModel` or
+            :class:`~transformers.TFBlenderbotSmallModel`.
+        d_model (:obj:`int`, `optional`, defaults to 512):
+            Dimensionality of the layers and the pooler layer.
+        encoder_layers (:obj:`int`, `optional`, defaults to 8):
+            Number of encoder layers.
+        decoder_layers (:obj:`int`, `optional`, defaults to 8):
+            Number of decoder layers.
+        encoder_attention_heads (:obj:`int`, `optional`, defaults to 16):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        decoder_attention_heads (:obj:`int`, `optional`, defaults to 16):
+            Number of attention heads for each attention layer in the Transformer decoder.
+        decoder_ffn_dim (:obj:`int`, `optional`, defaults to 2048):
+            Dimensionality of the "intermediate" (often named feed-forward) layer in decoder.
+        encoder_ffn_dim (:obj:`int`, `optional`, defaults to 2048):
+            Dimensionality of the "intermediate" (often named feed-forward) layer in decoder.
+        activation_function (:obj:`str` or :obj:`function`, `optional`, defaults to :obj:`"gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string,
+            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"silu"` and :obj:`"gelu_new"` are supported.
+        dropout (:obj:`float`, `optional`, defaults to 0.1):
+            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_dropout (:obj:`float`, `optional`, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        activation_dropout (:obj:`float`, `optional`, defaults to 0.0):
+            The dropout ratio for activations inside the fully connected layer.
+        classifier_dropout (:obj:`float`, `optional`, defaults to 0.0):
+            The dropout ratio for classifier.
+        max_position_embeddings (:obj:`int`, `optional`, defaults to 512):
+            The maximum sequence length that this model might ever be used with. Typically set this to something large
+            just in case (e.g., 512 or 1024 or 2048).
+        init_std (:obj:`float`, `optional`, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        encoder_layerdrop: (:obj:`float`, `optional`, defaults to 0.0):
+            The LayerDrop probability for the encoder. See the `LayerDrop paper <see
+            https://arxiv.org/abs/1909.11556>`__ for more details.
+        decoder_layerdrop: (:obj:`float`, `optional`, defaults to 0.0):
+            The LayerDrop probability for the decoder. See the `LayerDrop paper <see
+            https://arxiv.org/abs/1909.11556>`__ for more details.
+        gradient_checkpointing (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            If True, use gradient checkpointing to save memory at the expense of slower backward pass.
+        scale_embedding (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Scale embeddings by diving by sqrt(d_model).
+        use_cache (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Whether or not the model should return the last key/values attentions (not used by all models)
+
+    Example::
+
+        >>> from transformers import BlenderbotSmallModel, BlenderbotSmallConfig
+
+        >>> # Initializing a BlenderbotSmall facebook/blenderbot_small-90M style configuration
+        >>> configuration = BlenderbotSmallConfig()
+
+        >>> # Initializing a model from the facebook/blenderbot_small-90M style configuration
+        >>> model = BlenderbotSmallModel(configuration)
+
+        >>> # Accessing the model configuration
+        >>> configuration = model.config
+    """
+    model_type = "blenderbot-small"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        vocab_size=50265,
+        max_position_embeddings=512,
+        encoder_layers=8,
+        encoder_ffn_dim=2048,
+        encoder_attention_heads=16,
+        decoder_layers=8,
+        decoder_ffn_dim=2048,
+        decoder_attention_heads=16,
+        encoder_layerdrop=0.0,
+        decoder_layerdrop=0.0,
+        use_cache=True,
+        is_encoder_decoder=True,
+        activation_function="gelu",
+        d_model=512,
+        dropout=0.1,
+        attention_dropout=0.0,
+        activation_dropout=0.0,
+        init_std=0.02,
+        decoder_start_token_id=1,
+        classifier_dropout=0.0,
+        scale_embedding=False,
+        gradient_checkpointing=False,
+        pad_token_id=0,
+        bos_token_id=1,
+        eos_token_id=2,
+        **kwargs
+    ):
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            is_encoder_decoder=is_encoder_decoder,
+            decoder_start_token_id=decoder_start_token_id,
+            **kwargs,
+        )
+
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.d_model = d_model
+        self.encoder_ffn_dim = encoder_ffn_dim
+        self.encoder_layers = encoder_layers
+        self.encoder_attention_heads = encoder_attention_heads
+        self.decoder_ffn_dim = decoder_ffn_dim
+        self.decoder_layers = decoder_layers
+        self.decoder_attention_heads = decoder_attention_heads
+        self.dropout = dropout
+        self.attention_dropout = attention_dropout
+        self.activation_dropout = activation_dropout
+        self.activation_function = activation_function
+        self.init_std = init_std
+        self.encoder_layerdrop = encoder_layerdrop
+        self.decoder_layerdrop = decoder_layerdrop
+        self.classifier_dropout = classifier_dropout
+        self.use_cache = use_cache
+        self.num_hidden_layers = encoder_layers
+        self.gradient_checkpointing = gradient_checkpointing
+        self.scale_embedding = scale_embedding  # scale factor will be sqrt(d_model) if True
+
+    @property
+    def num_attention_heads(self) -> int:
+        return self.encoder_attention_heads
+
+    @property
+    def hidden_size(self) -> int:
+        return self.d_model

src/transformers/models/blenderbot_small/modeling_blenderbot_small.py

+# coding=utf-8
+# Copyright 2021 The Facebook, Inc. and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" PyTorch BlenderbotSmall model. """
+
+
+import math
+import random
+from typing import Optional, Tuple
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+from torch.nn import CrossEntropyLoss
+
+from ...activations import ACT2FN
+from ...file_utils import (
+    add_end_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    replace_return_docstrings,
+)
+from ...modeling_outputs import (
+    BaseModelOutput,
+    BaseModelOutputWithPastAndCrossAttentions,
+    Seq2SeqLMOutput,
+    Seq2SeqModelOutput,
+)
+from ...modeling_utils import PreTrainedModel
+from ...utils import logging
+from .configuration_blenderbot_small import BlenderbotSmallConfig
+
+
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = "BlenderbotSmallConfig"
+_TOKENIZER_FOR_DOC = "BlenderbotSmallTokenizer"
+
+
+BLENDERBOT_SMALL_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "facebook/blenderbot_small-90M",
+    # See all BlenderbotSmall models at https://huggingface.co/models?filter=blenderbot_small
+]
+
+
+# Copied from transformers.models.bart.modeling_bart.shift_tokens_right
+def shift_tokens_right(input_ids: torch.Tensor, pad_token_id: int, decoder_start_token_id: int):
+    """
+    Shift input ids one token to the right.
+    """
+    shifted_input_ids = input_ids.new_zeros(input_ids.shape)
+    shifted_input_ids[:, 1:] = input_ids[:, :-1].clone()
+    shifted_input_ids[:, 0] = decoder_start_token_id
+
+    assert pad_token_id is not None, "self.model.config.pad_token_id has to be defined."
+    # replace possible -100 values in labels by `pad_token_id`
+    shifted_input_ids.masked_fill_(shifted_input_ids == -100, pad_token_id)
+
+    return shifted_input_ids
+
+
+# Copied from transformers.models.bart.modeling_bart._make_causal_mask
+def _make_causal_mask(input_ids_shape: torch.Size, dtype: torch.dtype, past_key_values_length: int = 0):
+    """
+    Make causal mask used for bi-directional self-attention.
+    """
+    bsz, tgt_len = input_ids_shape
+    mask = torch.full((tgt_len, tgt_len), float("-inf"))
+    mask_cond = torch.arange(mask.size(-1))
+    mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)
+    mask = mask.to(dtype)
+
+    if past_key_values_length > 0:
+        mask = torch.cat([torch.zeros(tgt_len, past_key_values_length, dtype=dtype), mask], dim=-1)
+    return mask[None, None, :, :].expand(bsz, 1, tgt_len, tgt_len + past_key_values_length)
+
+
+# Copied from transformers.models.bart.modeling_bart._expand_mask
+def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
+    """
+    Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
+    """
+    bsz, src_len = mask.size()
+    tgt_len = tgt_len if tgt_len is not None else src_len
+
+    expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)
+
+    inverted_mask = 1.0 - expanded_mask
+
+    return inverted_mask.masked_fill(inverted_mask.bool(), torch.finfo(dtype).min)
+
+
+# Copied from transformers.models.blenderbot.modeling_blenderbot.BlenderbotLearnedPositionalEmbedding with Blenderbot->BlenderbotSmall
+class BlenderbotSmallLearnedPositionalEmbedding(nn.Embedding):
+    """
+    This module learns positional embeddings up to a fixed maximum size.
+    """
+
+    def __init__(self, num_embeddings: int, embedding_dim: int, padding_idx: int):
+        assert padding_idx is not None, "`padding_idx` should not be None, but of type int"
+        super().__init__(num_embeddings, embedding_dim, padding_idx=padding_idx)
+
+    def forward(self, input_ids_shape: torch.Size, past_key_values_length: int = 0):
+        """`input_ids_shape` is expected to be [bsz x seqlen]."""
+        bsz, seq_len = input_ids_shape[:2]
+        positions = torch.arange(
+            past_key_values_length, past_key_values_length + seq_len, dtype=torch.long, device=self.weight.device
+        )
+        return super().forward(positions)
+
+
+# Copied from transformers.models.bart.modeling_bart.BartAttention with Bart->BlenderbotSmall
+class BlenderbotSmallAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(
+        self,
+        embed_dim: int,
+        num_heads: int,
+        dropout: float = 0.0,
+        is_decoder: bool = False,
+        bias: bool = True,
+    ):
+        super().__init__()
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.dropout = dropout
+        self.head_dim = embed_dim // num_heads
+        assert (
+            self.head_dim * num_heads == self.embed_dim
+        ), f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`: {num_heads})."
+        self.scaling = self.head_dim ** -0.5
+        self.is_decoder = is_decoder
+
+        self.k_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+        self.v_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+        self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+        self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+
+    def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
+        return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        key_value_states: Optional[torch.Tensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        output_attentions: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        """Input shape: Batch x Time x Channel"""
+
+        # if key_value_states are provided this layer is used as a cross-attention layer
+        # for the decoder
+        is_cross_attention = key_value_states is not None
+        bsz, tgt_len, embed_dim = hidden_states.size()
+
+        # get query proj
+        query_states = self.q_proj(hidden_states) * self.scaling
+        # get key, value proj
+        if is_cross_attention and past_key_value is not None:
+            # reuse k,v, cross_attentions
+            key_states = past_key_value[0]
+            value_states = past_key_value[1]
+        elif is_cross_attention:
+            # cross_attentions
+            key_states = self._shape(self.k_proj(key_value_states), -1, bsz)
+            value_states = self._shape(self.v_proj(key_value_states), -1, bsz)
+        elif past_key_value is not None:
+            # reuse k, v, self_attention
+            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
+            key_states = torch.cat([past_key_value[0], key_states], dim=2)
+            value_states = torch.cat([past_key_value[1], value_states], dim=2)
+        else:
+            # self_attention
+            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
+
+        if self.is_decoder:
+            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
+            # Further calls to cross_attention layer can then reuse all cross-attention
+            # key/value_states (first "if" case)
+            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
+            # all previous decoder key/value_states. Further calls to uni-directional self-attention
+            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
+            # if encoder bi-directional self-attention `past_key_value` is always `None`
+            past_key_value = (key_states, value_states)
+
+        proj_shape = (bsz * self.num_heads, -1, self.head_dim)
+        query_states = self._shape(query_states, tgt_len, bsz).view(*proj_shape)
+        key_states = key_states.view(*proj_shape)
+        value_states = value_states.view(*proj_shape)
+
+        src_len = key_states.size(1)
+        attn_weights = torch.bmm(query_states, key_states.transpose(1, 2))
+
+        assert attn_weights.size() == (
+            bsz * self.num_heads,
+            tgt_len,
+            src_len,
+        ), f"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is {attn_weights.size()}"
+
+        if attention_mask is not None:
+            assert attention_mask.size() == (
+                bsz,
+                1,
+                tgt_len,
+                src_len,
+            ), f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.size()}"
+            attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + attention_mask
+            attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
+
+        attn_weights = F.softmax(attn_weights, dim=-1)
+
+        if output_attentions:
+            # this operation is a bit akward, but it's required to
+            # make sure that attn_weights keeps its gradient.
+            # In order to do so, attn_weights have to reshaped
+            # twice and have to be reused in the following
+            attn_weights_reshaped = attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
+            attn_weights = attn_weights_reshaped.view(bsz * self.num_heads, tgt_len, src_len)
+        else:
+            attn_weights_reshaped = None
+
+        attn_probs = F.dropout(attn_weights, p=self.dropout, training=self.training)
+
+        attn_output = torch.bmm(attn_probs, value_states)
+
+        assert attn_output.size() == (
+            bsz * self.num_heads,
+            tgt_len,
+            self.head_dim,
+        ), f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is {attn_output.size()}"
+
+        attn_output = (
+            attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
+            .transpose(1, 2)
+            .reshape(bsz, tgt_len, embed_dim)
+        )
+
+        attn_output = self.out_proj(attn_output)
+
+        return attn_output, attn_weights_reshaped, past_key_value
+
+
+# Copied from transformers.models.bart.modeling_bart.BartEncoderLayer with Bart->BlenderbotSmall
+class BlenderbotSmallEncoderLayer(nn.Module):
+    def __init__(self, config: BlenderbotSmallConfig):
+        super().__init__()
+        self.embed_dim = config.d_model
+        self.self_attn = BlenderbotSmallAttention(
+            embed_dim=self.embed_dim,
+            num_heads=config.encoder_attention_heads,
+            dropout=config.attention_dropout,
+        )
+        self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
+        self.dropout = config.dropout
+        self.activation_fn = ACT2FN[config.activation_function]
+        self.activation_dropout = config.activation_dropout
+        self.fc1 = nn.Linear(self.embed_dim, config.encoder_ffn_dim)
+        self.fc2 = nn.Linear(config.encoder_ffn_dim, self.embed_dim)
+        self.final_layer_norm = nn.LayerNorm(self.embed_dim)
+
+    def forward(self, hidden_states: torch.Tensor, attention_mask: torch.Tensor, output_attentions: bool = False):
+        """
+        Args:
+            hidden_states (:obj:`torch.FloatTensor`): input to the layer of shape `(seq_len, batch, embed_dim)`
+            attention_mask (:obj:`torch.FloatTensor`): attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            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.
+        """
+        residual = hidden_states
+        hidden_states, attn_weights, _ = self.self_attn(
+            hidden_states=hidden_states, attention_mask=attention_mask, output_attentions=output_attentions
+        )
+        hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
+        hidden_states = residual + hidden_states
+        hidden_states = self.self_attn_layer_norm(hidden_states)
+
+        residual = hidden_states
+        hidden_states = self.activation_fn(self.fc1(hidden_states))
+        hidden_states = F.dropout(hidden_states, p=self.activation_dropout, training=self.training)
+        hidden_states = self.fc2(hidden_states)
+        hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
+        hidden_states = residual + hidden_states
+        hidden_states = self.final_layer_norm(hidden_states)
+
+        if torch.isinf(hidden_states).any() or torch.isnan(hidden_states).any():
+            clamp_value = torch.finfo(hidden_states.dtype).max - 1000
+            hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (attn_weights,)
+
+        return outputs
+
+
+# Copied from transformers.models.bart.modeling_bart.BartDecoderLayer with Bart->BlenderbotSmall
+class BlenderbotSmallDecoderLayer(nn.Module):
+    def __init__(self, config: BlenderbotSmallConfig):
+        super().__init__()
+        self.embed_dim = config.d_model
+
+        self.self_attn = BlenderbotSmallAttention(
+            embed_dim=self.embed_dim,
+            num_heads=config.decoder_attention_heads,
+            dropout=config.attention_dropout,
+            is_decoder=True,
+        )
+        self.dropout = config.dropout
+        self.activation_fn = ACT2FN[config.activation_function]
+        self.activation_dropout = config.activation_dropout
+
+        self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
+        self.encoder_attn = BlenderbotSmallAttention(
+            self.embed_dim,
+            config.decoder_attention_heads,
+            dropout=config.attention_dropout,
+            is_decoder=True,
+        )
+        self.encoder_attn_layer_norm = nn.LayerNorm(self.embed_dim)
+        self.fc1 = nn.Linear(self.embed_dim, config.decoder_ffn_dim)
+        self.fc2 = nn.Linear(config.decoder_ffn_dim, self.embed_dim)
+        self.final_layer_norm = nn.LayerNorm(self.embed_dim)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = True,
+    ):
+        """
+        Args:
+            hidden_states (:obj:`torch.FloatTensor`): input to the layer of shape `(seq_len, batch, embed_dim)`
+            attention_mask (:obj:`torch.FloatTensor`): attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            encoder_hidden_states (:obj:`torch.FloatTensor`): cross attention input to the layer of shape `(seq_len, batch, embed_dim)`
+            encoder_attention_mask (:obj:`torch.FloatTensor`): encoder attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            past_key_value (:obj:`Tuple(torch.FloatTensor)`): cached past key and value projection states
+            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.
+        """
+        residual = hidden_states
+
+        # Self Attention
+        # decoder uni-directional self-attention cached key/values tuple is at positions 1,2
+        self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None
+        # add present self-attn cache to positions 1,2 of present_key_value tuple
+        hidden_states, self_attn_weights, present_key_value = self.self_attn(
+            hidden_states=hidden_states,
+            past_key_value=self_attn_past_key_value,
+            attention_mask=attention_mask,
+            output_attentions=output_attentions,
+        )
+        hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
+        hidden_states = residual + hidden_states
+        hidden_states = self.self_attn_layer_norm(hidden_states)
+
+        # Cross-Attention Block
+        cross_attn_present_key_value = None
+        cross_attn_weights = None
+        if encoder_hidden_states is not None:
+            residual = hidden_states
+
+            # cross_attn cached key/values tuple is at positions 3,4 of present_key_value tuple
+            cross_attn_past_key_value = past_key_value[-2:] if past_key_value is not None else None
+            hidden_states, cross_attn_weights, cross_attn_present_key_value = self.encoder_attn(
+                hidden_states=hidden_states,
+                key_value_states=encoder_hidden_states,
+                attention_mask=encoder_attention_mask,
+                past_key_value=cross_attn_past_key_value,
+                output_attentions=output_attentions,
+            )
+            hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
+            hidden_states = residual + hidden_states
+            hidden_states = self.encoder_attn_layer_norm(hidden_states)
+
+            # add cross-attn to positions 3,4 of present_key_value tuple
+            present_key_value = present_key_value + cross_attn_present_key_value
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.activation_fn(self.fc1(hidden_states))
+        hidden_states = F.dropout(hidden_states, p=self.activation_dropout, training=self.training)
+        hidden_states = self.fc2(hidden_states)
+        hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
+        hidden_states = residual + hidden_states
+        hidden_states = self.final_layer_norm(hidden_states)
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights, cross_attn_weights)
+
+        if use_cache:
+            outputs += (present_key_value,)
+
+        return outputs
+
+
+class BlenderbotSmallPreTrainedModel(PreTrainedModel):
+    config_class = BlenderbotSmallConfig
+    base_model_prefix = "model"
+
+    def _init_weights(self, module):
+        std = self.config.init_std
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+    @property
+    def dummy_inputs(self):
+        pad_token = self.config.pad_token_id
+        input_ids = torch.tensor([[0, 6, 10, 4, 2], [0, 8, 12, 2, pad_token]], device=self.device)
+        dummy_inputs = {
+            "attention_mask": input_ids.ne(pad_token),
+            "input_ids": input_ids,
+            "decoder_input_ids": input_ids,
+        }
+        return dummy_inputs
+
+
+BLENDERBOT_SMALL_START_DOCSTRING = r"""
+    This model inherits from :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.
+
+    Parameters:
+        config (:class:`~transformers.BlenderbotSmallConfig`):
+            Model configuration class with all the parameters of the model. Initializing with a config file does not
+            load the weights associated with the model, only the configuration. Check out the
+            :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model weights.
+"""
+
+BLENDERBOT_SMALL_GENERATION_EXAMPLE = r"""
+    Conversation example::
+
+        >>> from transformers import BlenderbotSmallTokenizer, BlenderbotSmallForConditionalGeneration
+        >>> mname = 'facebook/blenderbot_small-90M'
+        >>> model = BlenderbotSmallForConditionalGeneration.from_pretrained(mname)
+        >>> tokenizer = BlenderbotSmallTokenizer.from_pretrained(mname)
+        >>> UTTERANCE = "My friends are cool but they eat too many carbs."
+        >>> print("Human: ", UTTERANCE)
+        >>> inputs = tokenizer([UTTERANCE], return_tensors='pt')
+        >>> inputs.pop("token_type_ids")
+        >>> reply_ids = model.generate(**inputs)
+        >>> print("Bot: ", tokenizer.batch_decode(reply_ids, skip_special_tokens=True)[0])
+        what kind of carbs do they eat? i don't know much about carbs.
+
+        >>> REPLY = "I'm not sure"
+        >>> print("Human: ", REPLY)
+        >>> NEXT_UTTERANCE = (
+        ... "My friends are cool but they eat too many carbs.</s> "
+        ... "<s>what kind of carbs do they eat? i don't know much about carbs.</s> "
+        ... "<s>I'm not sure."
+        ... )
+        >>> inputs = tokenizer([NEXT_UTTERANCE], return_tensors='pt')
+        >>> inputs.pop("token_type_ids")
+        >>> next_reply_ids = model.generate(**inputs)
+        >>> print("Bot: ", tokenizer.batch_decode(next_reply_ids, skip_special_tokens=True)[0])
+"""
+
+BLENDERBOT_SMALL_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+
+            Indices can be obtained using :class:`~transformers.BlenderbotSmallTokenizer`. See
+            :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__` for
+            details.
+
+            `What are input IDs? <../glossary.html#input-ids>`__
+        attention_mask (:obj:`torch.Tensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Mask to avoid performing attention on padding token indices. Mask values selected in ``[0, 1]``:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            `What are attention masks? <../glossary.html#attention-mask>`__
+        decoder_input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, target_sequence_length)`, `optional`):
+            Indices of decoder input sequence tokens in the vocabulary.
+
+            Indices can be obtained using :class:`~transformers.BlenderbotSmallTokenizer`. See
+            :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__` for
+            details.
+
+            `What are input IDs? <../glossary.html#input-ids>`__
+
+            BlenderbotSmall uses the :obj:`bos_token_id` as the starting token for :obj:`decoder_input_ids` generation.
+            If :obj:`past_key_values` is used, optionally only the last :obj:`decoder_input_ids` have to be input (see
+            :obj:`past_key_values`).
+        decoder_attention_mask (:obj:`torch.LongTensor` of shape :obj:`(batch_size, target_sequence_length)`, `optional`):
+            Default behavior: generate a tensor that ignores pad tokens in :obj:`decoder_input_ids`. Causal mask will
+            also be used by default.
+
+            If you want to change padding behavior, you should read
+            :func:`modeling_blenderbot_small._prepare_decoder_inputs` and modify to your needs. See diagram 1 in `the
+            paper <https://arxiv.org/abs/1910.13461>`__ for more information on the default strategy.
+        encoder_outputs (:obj:`tuple(tuple(torch.FloatTensor)`, `optional`):
+            Tuple consists of (:obj:`last_hidden_state`, `optional`: :obj:`hidden_states`, `optional`:
+            :obj:`attentions`) :obj:`last_hidden_state` of shape :obj:`(batch_size, sequence_length, hidden_size)`,
+            `optional`) is a sequence of hidden-states at the output of the last layer of the encoder. Used in the
+            cross-attention of the decoder.
+        past_key_values (:obj:`Tuple[Tuple[torch.Tensor]]` of length :obj:`config.n_layers` with each tuple having 2 tuples each of which has 2 tensors of shape :obj:`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
+            Contains precomputed key and value hidden-states of the attention blocks. Can be used to speed up decoding.
+
+            If :obj:`past_key_values` are used, the user can optionally input only the last :obj:`decoder_input_ids`
+            (those that don't have their past key value states given to this model) of shape :obj:`(batch_size, 1)`
+            instead of all :obj:`decoder_input_ids`` of shape :obj:`(batch_size, sequence_length)`.
+        inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, 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.
+        decoder_inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, target_sequence_length, hidden_size)`, `optional`):
+            Optionally, instead of passing :obj:`decoder_input_ids` you can choose to directly pass an embedded
+            representation. If :obj:`past_key_values` is used, optionally only the last :obj:`decoder_inputs_embeds`
+            have to be input (see :obj:`past_key_values`). This is useful if you want more control over how to convert
+            :obj:`decoder_input_ids` indices into associated vectors than the model's internal embedding lookup matrix.
+
+            If :obj:`decoder_input_ids` and :obj:`decoder_inputs_embeds` are both unset, :obj:`decoder_inputs_embeds`
+            takes the value of :obj:`inputs_embeds`.
+        use_cache (:obj:`bool`, `optional`):
+            If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up
+            decoding (see :obj:`past_key_values`).
+        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.
+"""
+
+
+class BlenderbotSmallEncoder(BlenderbotSmallPreTrainedModel):
+    """
+    Transformer encoder consisting of *config.encoder_layers* self attention layers. Each layer is a
+    :class:`BlenderbotSmallEncoderLayer`.
+
+    Args:
+        config: BlenderbotSmallConfig
+        embed_tokens (torch.nn.Embedding): output embedding
+    """
+
+    def __init__(self, config: BlenderbotSmallConfig, embed_tokens: Optional[nn.Embedding] = None):
+        super().__init__(config)
+
+        self.dropout = config.dropout
+        self.layerdrop = config.encoder_layerdrop
+
+        embed_dim = config.d_model
+        self.padding_idx = config.pad_token_id
+        self.max_source_positions = config.max_position_embeddings
+        self.embed_scale = math.sqrt(embed_dim) if config.scale_embedding else 1.0
+
+        if embed_tokens is not None:
+            self.embed_tokens = embed_tokens
+        else:
+            self.embed_tokens = nn.Embedding(config.vocab_size, embed_dim, self.padding_idx)
+
+        self.embed_positions = BlenderbotSmallLearnedPositionalEmbedding(
+            config.max_position_embeddings,
+            embed_dim,
+            self.padding_idx,
+        )
+        self.layers = nn.ModuleList([BlenderbotSmallEncoderLayer(config) for _ in range(config.encoder_layers)])
+        self.layernorm_embedding = nn.LayerNorm(embed_dim)
+
+        self.init_weights()
+
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        inputs_embeds=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        Args:
+            input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`):
+                Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you
+                provide it.
+
+                Indices can be obtained using :class:`~transformers.BlenderbotSmallTokenizer`. See
+                :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__`
+                for details.
+
+                `What are input IDs? <../glossary.html#input-ids>`__
+            attention_mask (:obj:`torch.Tensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+                Mask to avoid performing attention on padding token indices. Mask values selected in ``[0, 1]``:
+
+                - 1 for tokens that are **not masked**,
+                - 0 for tokens that are **masked**.
+
+                `What are attention masks? <../glossary.html#attention-mask>`__
+            inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, 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.
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # retrieve input_ids and inputs_embeds
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+            input_ids = input_ids.view(-1, input_shape[-1])
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids) * self.embed_scale
+
+        embed_pos = self.embed_positions(input_shape)
+
+        hidden_states = inputs_embeds + embed_pos
+        hidden_states = self.layernorm_embedding(hidden_states)
+        hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
+
+        # expand attention_mask
+        if attention_mask is not None:
+            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+            attention_mask = _expand_mask(attention_mask, inputs_embeds.dtype)
+
+        encoder_states = () if output_hidden_states else None
+        all_attentions = () if output_attentions else None
+        for encoder_layer in self.layers:
+            if output_hidden_states:
+                encoder_states = encoder_states + (hidden_states,)
+            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
+            dropout_probability = random.uniform(0, 1)
+            if self.training and (dropout_probability < self.layerdrop):  # skip the layer
+                layer_outputs = (None, None)
+            else:
+                if getattr(self.config, "gradient_checkpointing", False):
+
+                    def create_custom_forward(module):
+                        def custom_forward(*inputs):
+                            return module(*inputs, output_attentions)
+
+                        return custom_forward
+
+                    layer_outputs = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(encoder_layer),
+                        hidden_states,
+                        attention_mask,
+                    )
+                else:
+                    layer_outputs = encoder_layer(hidden_states, attention_mask, output_attentions=output_attentions)
+
+                hidden_states = layer_outputs[0]
+
+            if output_attentions:
+                all_attentions = all_attentions + (layer_outputs[1],)
+
+        if output_hidden_states:
+            encoder_states = encoder_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None)
+        return BaseModelOutput(
+            last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions
+        )
+
+
+class BlenderbotSmallDecoder(BlenderbotSmallPreTrainedModel):
+    """
+    Transformer decoder consisting of *config.decoder_layers* layers. Each layer is a
+    :class:`BlenderbotSmallDecoderLayer`
+
+    Args:
+        config: BlenderbotSmallConfig
+        embed_tokens (torch.nn.Embedding): output embedding
+    """
+
+    def __init__(self, config: BlenderbotSmallConfig, embed_tokens: Optional[nn.Embedding] = None):
+        super().__init__(config)
+        self.dropout = config.dropout
+        self.layerdrop = config.decoder_layerdrop
+        self.padding_idx = config.pad_token_id
+        self.max_target_positions = config.max_position_embeddings
+        self.embed_scale = math.sqrt(config.d_model) if config.scale_embedding else 1.0
+
+        if embed_tokens is not None:
+            self.embed_tokens = embed_tokens
+        else:
+            self.embed_tokens = nn.Embedding(config.vocab_size, config.d_model, self.padding_idx)
+
+        self.embed_positions = BlenderbotSmallLearnedPositionalEmbedding(
+            config.max_position_embeddings,
+            config.d_model,
+            self.padding_idx,
+        )
+        self.layers = nn.ModuleList([BlenderbotSmallDecoderLayer(config) for _ in range(config.decoder_layers)])
+        self.layernorm_embedding = nn.LayerNorm(config.d_model)
+
+        self.init_weights()
+
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_values=None,
+        inputs_embeds=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        Args:
+            input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`):
+                Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you
+                provide it.
+
+                Indices can be obtained using :class:`~transformers.BlenderbotSmallTokenizer`. See
+                :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__`
+                for details.
+
+                `What are input IDs? <../glossary.html#input-ids>`__
+            attention_mask (:obj:`torch.Tensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+                Mask to avoid performing attention on padding token indices. Mask values selected in ``[0, 1]``:
+
+                - 1 for tokens that are **not masked**,
+                - 0 for tokens that are **masked**.
+
+                `What are attention masks? <../glossary.html#attention-mask>`__
+            encoder_hidden_states (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, encoder_sequence_length, hidden_size)`, `optional`):
+                Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention
+                of the decoder.
+            encoder_attention_mask (:obj:`torch.LongTensor` of shape :obj:`(batch_size, encoder_sequence_length)`, `optional`):
+                Mask to avoid performing cross-attention on padding tokens indices of encoder input_ids. Mask values
+                selected in ``[0, 1]``:
+
+                - 1 for tokens that are **not masked**,
+                - 0 for tokens that are **masked**.
+
+                `What are attention masks? <../glossary.html#attention-mask>`__
+            past_key_values (:obj:`Tuple[Tuple[torch.Tensor]]` of length :obj:`config.n_layers` with each tuple having 2 tuples each of which has 2 tensors of shape :obj:`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
+                Contains precomputed key and value hidden-states of the attention blocks. Can be used to speed up
+                decoding.
+
+                If :obj:`past_key_values` are used, the user can optionally input only the last
+                :obj:`decoder_input_ids` (those that don't have their past key value states given to this model) of
+                shape :obj:`(batch_size, 1)` instead of all :obj:`decoder_input_ids`` of shape :obj:`(batch_size,
+                sequence_length)`.
+            inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, 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.
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # retrieve input_ids and inputs_embeds
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+            input_ids = input_ids.view(-1, input_shape[-1])
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+        else:
+            raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
+
+        # past_key_values_length
+        past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids) * self.embed_scale
+
+        # create causal mask
+        # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+        combined_attention_mask = None
+        if input_shape[-1] > 1:
+            combined_attention_mask = _make_causal_mask(
+                input_shape, inputs_embeds.dtype, past_key_values_length=past_key_values_length
+            ).to(self.device)
+
+        if attention_mask is not None and combined_attention_mask is not None:
+            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+            combined_attention_mask = combined_attention_mask + _expand_mask(
+                attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]
+            )
+
+        # expand encoder attention mask
+        if encoder_hidden_states is not None and encoder_attention_mask is not None:
+            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+            encoder_attention_mask = _expand_mask(encoder_attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1])
+
+        # embed positions
+        positions = self.embed_positions(input_shape, past_key_values_length)
+
+        # BlenderbotSmall applies layer norm on hidden_states
+        inputs_embeds = self.layernorm_embedding(inputs_embeds)
+        hidden_states = inputs_embeds + positions
+
+        hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        all_cross_attentions = () if output_attentions else None
+        next_decoder_cache = () if use_cache else None
+        for idx, decoder_layer in enumerate(self.layers):
+            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+            dropout_probability = random.uniform(0, 1)
+            if self.training and (dropout_probability < self.layerdrop):
+                continue
+
+            past_key_value = past_key_values[idx] if past_key_values is not None else None
+
+            if getattr(self.config, "gradient_checkpointing", False):
+                if use_cache:
+                    raise ValueError(
+                        "When using `gradient_checkpointing, make sure that `use_cache=False` and `config.use_cache=False`."
+                    )
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        # None for past_key_value
+                        return module(*inputs, output_attentions, use_cache)
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(decoder_layer),
+                    hidden_states,
+                    combined_attention_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    None,
+                )
+            else:
+
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask=combined_attention_mask,
+                    encoder_hidden_states=encoder_hidden_states,
+                    encoder_attention_mask=encoder_attention_mask,
+                    past_key_value=past_key_value,
+                    output_attentions=output_attentions,
+                    use_cache=use_cache,
+                )
+            hidden_states = layer_outputs[0]
+
+            if use_cache:
+                next_decoder_cache += (layer_outputs[3 if output_attentions else 1],)
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+                all_cross_attentions += (layer_outputs[2],)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        next_cache = next_decoder_cache if use_cache else None
+        if not return_dict:
+            return tuple(
+                v
+                for v in [hidden_states, next_cache, all_hidden_states, all_self_attns, all_cross_attentions]
+                if v is not None
+            )
+        return BaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+            cross_attentions=all_cross_attentions,
+        )
+
+
+@add_start_docstrings(
+    "The bare BlenderbotSmall Model outputting raw hidden-states without any specific head on top.",
+    BLENDERBOT_SMALL_START_DOCSTRING,
+)
+class BlenderbotSmallModel(BlenderbotSmallPreTrainedModel):
+    def __init__(self, config: BlenderbotSmallConfig):
+        super().__init__(config)
+
+        padding_idx, vocab_size = config.pad_token_id, config.vocab_size
+        self.shared = nn.Embedding(vocab_size, config.d_model, padding_idx)
+
+        self.encoder = BlenderbotSmallEncoder(config, self.shared)
+        self.decoder = BlenderbotSmallDecoder(config, self.shared)
+
+        self.init_weights()
+
+    def get_input_embeddings(self):
+        return self.shared
+
+    def set_input_embeddings(self, value):
+        self.shared = value
+        self.encoder.embed_tokens = self.shared
+        self.decoder.embed_tokens = self.shared
+
+    def get_encoder(self):
+        return self.encoder
+
+    def get_decoder(self):
+        return self.decoder
+
+    @add_start_docstrings_to_model_forward(BLENDERBOT_SMALL_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=Seq2SeqModelOutput, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        decoder_input_ids=None,
+        decoder_attention_mask=None,
+        encoder_outputs=None,
+        past_key_values=None,
+        inputs_embeds=None,
+        decoder_inputs_embeds=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        Returns:
+
+        Example::
+
+            >>> from transformers import BlenderbotSmallTokenizer, BlenderbotSmallModel
+
+            >>> model = BlenderbotSmallModel.from_pretrained("facebook/blenderbot_small-90M")
+            >>> tokenizer = BlenderbotSmallTokenizer.from_pretrained("facebook/blenderbot_small-90M")
+
+            >>> input_ids = tokenizer("Studies have been shown that owning a dog is good for you", return_tensors="pt").input_ids  # Batch size 1
+            >>> decoder_input_ids = tokenizer("Studies show that", return_tensors="pt").input_ids  # Batch size 1
+            >>> outputs = model(input_ids=input_ids, decoder_input_ids=decoder_input_ids)
+
+            >>> last_hidden_states = outputs.last_hidden_state
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if encoder_outputs is None:
+            encoder_outputs = self.encoder(
+                input_ids=input_ids,
+                attention_mask=attention_mask,
+                inputs_embeds=inputs_embeds,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+            )
+        # If the user passed a tuple for encoder_outputs, we wrap it in a BaseModelOutput when return_dict=True
+        elif return_dict and not isinstance(encoder_outputs, BaseModelOutput):
+            encoder_outputs = BaseModelOutput(
+                last_hidden_state=encoder_outputs[0],
+                hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None,
+                attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None,
+            )
+
+        # decoder outputs consists of (dec_features, past_key_value, dec_hidden, dec_attn)
+        decoder_outputs = self.decoder(
+            input_ids=decoder_input_ids,
+            attention_mask=decoder_attention_mask,
+            encoder_hidden_states=encoder_outputs[0],
+            encoder_attention_mask=attention_mask,
+            past_key_values=past_key_values,
+            inputs_embeds=decoder_inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        if not return_dict:
+            return decoder_outputs + encoder_outputs
+
+        return Seq2SeqModelOutput(
+            last_hidden_state=decoder_outputs.last_hidden_state,
+            past_key_values=decoder_outputs.past_key_values,
+            decoder_hidden_states=decoder_outputs.hidden_states,
+            decoder_attentions=decoder_outputs.attentions,
+            cross_attentions=decoder_outputs.cross_attentions,
+            encoder_last_hidden_state=encoder_outputs.last_hidden_state,
+            encoder_hidden_states=encoder_outputs.hidden_states,
+            encoder_attentions=encoder_outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    "The BlenderbotSmall Model with a language modeling head. Can be used for summarization.",
+    BLENDERBOT_SMALL_START_DOCSTRING,
+)
+class BlenderbotSmallForConditionalGeneration(BlenderbotSmallPreTrainedModel):
+    base_model_prefix = "model"
+    _keys_to_ignore_on_load_missing = [
+        r"final_logits_bias",
+        r"encoder\.version",
+        r"decoder\.version",
+        r"lm_head\.weight",
+    ]
+
+    def __init__(self, config: BlenderbotSmallConfig):
+        super().__init__(config)
+        self.model = BlenderbotSmallModel(config)
+        self.register_buffer("final_logits_bias", torch.zeros((1, self.model.shared.num_embeddings)))
+        self.lm_head = nn.Linear(config.d_model, self.model.shared.num_embeddings, bias=False)
+
+        self.init_weights()
+
+    def get_encoder(self):
+        return self.model.get_encoder()
+
+    def get_decoder(self):
+        return self.model.get_decoder()
+
+    def resize_token_embeddings(self, new_num_tokens: int) -> nn.Embedding:
+        new_embeddings = super().resize_token_embeddings(new_num_tokens)
+        self._resize_final_logits_bias(new_num_tokens)
+        return new_embeddings
+
+    def _resize_final_logits_bias(self, new_num_tokens: int) -> None:
+        old_num_tokens = self.final_logits_bias.shape[-1]
+        if new_num_tokens <= old_num_tokens:
+            new_bias = self.final_logits_bias[:, :new_num_tokens]
+        else:
+            extra_bias = torch.zeros((1, new_num_tokens - old_num_tokens), device=self.final_logits_bias.device)
+            new_bias = torch.cat([self.final_logits_bias, extra_bias], dim=1)
+        self.register_buffer("final_logits_bias", new_bias)
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    @add_start_docstrings_to_model_forward(BLENDERBOT_SMALL_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=Seq2SeqLMOutput, config_class=_CONFIG_FOR_DOC)
+    @add_end_docstrings(BLENDERBOT_SMALL_GENERATION_EXAMPLE)
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        decoder_input_ids=None,
+        decoder_attention_mask=None,
+        encoder_outputs=None,
+        past_key_values=None,
+        inputs_embeds=None,
+        decoder_inputs_embeds=None,
+        labels=None,
+        use_cache=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 either be in ``[0, ...,
+            config.vocab_size]`` or -100 (see ``input_ids`` docstring). Tokens with indices set to ``-100`` are ignored
+            (masked), the loss is only computed for the tokens with labels in ``[0, ..., config.vocab_size]``.
+
+        Returns:
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if labels is not None:
+            if decoder_input_ids is None:
+                decoder_input_ids = shift_tokens_right(
+                    labels, self.config.pad_token_id, self.config.decoder_start_token_id
+                )
+
+        outputs = self.model(
+            input_ids,
+            attention_mask=attention_mask,
+            decoder_input_ids=decoder_input_ids,
+            encoder_outputs=encoder_outputs,
+            decoder_attention_mask=decoder_attention_mask,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            decoder_inputs_embeds=decoder_inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        lm_logits = self.lm_head(outputs[0]) + self.final_logits_bias
+
+        masked_lm_loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            masked_lm_loss = loss_fct(lm_logits.view(-1, self.config.vocab_size), labels.view(-1))
+
+        if not return_dict:
+            output = (lm_logits,) + outputs[1:]
+            return ((masked_lm_loss,) + output) if masked_lm_loss is not None else output
+
+        return Seq2SeqLMOutput(
+            loss=masked_lm_loss,
+            logits=lm_logits,
+            past_key_values=outputs.past_key_values,
+            decoder_hidden_states=outputs.decoder_hidden_states,
+            decoder_attentions=outputs.decoder_attentions,
+            cross_attentions=outputs.cross_attentions,
+            encoder_last_hidden_state=outputs.encoder_last_hidden_state,
+            encoder_hidden_states=outputs.encoder_hidden_states,
+            encoder_attentions=outputs.encoder_attentions,
+        )
+
+    def prepare_inputs_for_generation(
+        self, decoder_input_ids, past=None, attention_mask=None, use_cache=None, encoder_outputs=None, **kwargs
+    ):
+        # cut decoder_input_ids if past is used
+        if past is not None:
+            decoder_input_ids = decoder_input_ids[:, -1:]
+
+        return {
+            "input_ids": None,  # encoder_outputs is defined. input_ids not needed
+            "encoder_outputs": encoder_outputs,
+            "past_key_values": past,
+            "decoder_input_ids": decoder_input_ids,
+            "attention_mask": attention_mask,
+            "use_cache": use_cache,  # change this to avoid caching (presumably for debugging)
+        }
+
+    def adjust_logits_during_generation(self, logits, cur_len, max_length):
+        if cur_len == max_length - 1 and self.config.eos_token_id is not None:
+            self._force_token_id_to_be_generated(logits, self.config.eos_token_id)
+        return logits
+
+    @staticmethod
+    def _force_token_id_to_be_generated(scores, token_id) -> None:
+        """force one of token_ids to be generated by setting prob of all other tokens to 0 (logprob=-float("inf"))"""
+        scores[:, [x for x in range(scores.shape[1]) if x != token_id]] = -float("inf")
+
+    @staticmethod
+    def _reorder_cache(past, beam_idx):
+        reordered_past = ()
+        for layer_past in past:
+            # cached cross_attention states don't have to be reordered -> they are always the same
+            reordered_past += (
+                tuple(past_state.index_select(0, beam_idx) for past_state in layer_past[:2]) + layer_past[2:],
+            )
+        return reordered_past

src/transformers/models/blenderbot_small/tokenization_blenderbot_small.py

+# coding=utf-8
+# Copyright 2021 The Facebook Inc. and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Tokenization class for BlenderbotSmall."""
+
+import json
+import os
+from typing import Dict, List, Optional, Tuple
+
+import regex as re
+
+from ...tokenization_utils import PreTrainedTokenizer
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+VOCAB_FILES_NAMES = {
+    "vocab_file": "vocab.json",
+    "merges_file": "merges.txt",
+    # "tokenizer_config_file": "tokenizer_config.json",
+}
+
+
+def get_pairs(word):
+    """
+    Return set of symbol pairs in a word.
+
+    Word is represented as tuple of symbols (symbols being variable-length strings).
+    """
+    pairs = set()
+    prev_char = word[0]
+    for char in word[1:]:
+        pairs.add((prev_char, char))
+        prev_char = char
+
+    pairs = set(pairs)
+    return pairs
+
+
+class BlenderbotSmallTokenizer(PreTrainedTokenizer):
+    """
+    Constructs a Blenderbot-90M tokenizer based on BPE (Byte-Pair-Encoding)
+
+    This tokenizer inherits from :class:`~transformers.PreTrainedTokenizer` which contains most of the main methods.
+    Users should refer to the superclass for more information regarding methods.
+
+    Args:
+        vocab_file (:obj:`str`):
+            File containing the vocabulary.
+        merges_file (:obj:`str`):
+            Path to the merges file.
+        bos_token (:obj:`str`, `optional`, defaults to :obj:`"__start__"`):
+            The beginning of sentence token.
+        eos_token (:obj:`str`, `optional`, defaults to :obj:`"__end__"`):
+            The end of sentence token.
+        unk_token (:obj:`str`, `optional`, defaults to :obj:`"__unk__"`):
+            The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this
+            token instead.
+        pad_token (:obj:`str`, `optional`, defaults to :obj:`"__pad__"`):
+            The token used for padding, for example when batching sequences of different lengths.
+        **kwargs
+            Additional keyword arguments passed along to :class:`~transformers.PreTrainedTokenizer`
+    """
+
+    vocab_files_names = {"vocab_file": "vocab.json", "merges_file": "merges.txt"}
+    pretrained_vocab_files_map = {
+        "vocab_file": {
+            "facebook/blenderbot_small-90M": "https://cdn.huggingface.co/facebook/blenderbot_small-90M/vocab.json"
+        },
+        "merges_file": {
+            "facebook/blenderbot_small-90M": "https://cdn.huggingface.co/facebook/blenderbot_small-90M/merges.txt"
+        },
+    }
+    max_model_input_sizes = {"facebook/blenderbot_small-90M": 512}
+
+    def __init__(
+        self,
+        vocab_file,
+        merges_file,
+        bos_token="__start__",
+        eos_token="__end__",
+        unk_token="__unk__",
+        pad_token="__null__",
+        **kwargs
+    ):
+        super().__init__(unk_token=unk_token, bos_token=bos_token, eos_token=eos_token, pad_token=pad_token, **kwargs)
+
+        with open(vocab_file, encoding="utf-8") as vocab_handle:
+            self.encoder = json.load(vocab_handle)
+        self.decoder = {v: k for k, v in self.encoder.items()}
+        with open(merges_file, encoding="utf-8") as merges_handle:
+            merges = merges_handle.read().split("\n")[1:-1]
+        merges = [tuple(merge.split()) for merge in merges]
+        self.bpe_ranks = dict(zip(merges, range(len(merges))))
+        self.cache = {}
+
+    @property
+    def vocab_size(self) -> int:
+        return len(self.encoder)
+
+    def get_vocab(self) -> Dict:
+        return dict(self.encoder, **self.added_tokens_encoder)
+
+    def bpe(self, token: str) -> str:
+        if token in self.cache:
+            return self.cache[token]
+        token = re.sub("([.,!?()])", r" \1", token)
+        token = re.sub("(')", r" \1 ", token)
+        token = re.sub(r"\s{2,}", " ", token)
+        if "\n" in token:
+            token = token.replace("\n", " __newln__")
+
+        tokens = token.split(" ")
+        words = []
+        for token in tokens:
+            if not len(token):
+                continue
+
+            token = token.lower()
+            word = tuple(token)
+            word = tuple(list(word[:-1]) + [word[-1] + "</w>"])
+            pairs = get_pairs(word)
+
+            if not pairs:
+                words.append(token)
+                continue
+
+            while True:
+                bigram = min(pairs, key=lambda pair: self.bpe_ranks.get(pair, float("inf")))
+                if bigram not in self.bpe_ranks:
+                    break
+                first, second = bigram
+                new_word = []
+                i = 0
+
+                while i < len(word):
+                    try:
+                        j = word.index(first, i)
+                        new_word.extend(word[i:j])
+                        i = j
+                    except ValueError:
+                        new_word.extend(word[i:])
+                        break
+
+                    if word[i] == first and i < len(word) - 1 and word[i + 1] == second:
+                        new_word.append(first + second)
+                        i += 2
+                    else:
+                        new_word.append(word[i])
+                        i += 1
+                new_word = tuple(new_word)
+                word = new_word
+                if len(word) == 1:
+                    break
+                else:
+                    pairs = get_pairs(word)
+            word = "@@ ".join(word)
+            word = word[:-4]
+
+            self.cache[token] = word
+            words.append(word)
+        return " ".join(words)
+
+    def _tokenize(self, text: str) -> List[str]:
+        """ Split a string into tokens using BPE."""
+        split_tokens = []
+
+        words = re.findall(r"\S+\n?", text)
+
+        for token in words:
+            split_tokens.extend([t for t in self.bpe(token).split(" ")])
+        return split_tokens
+
+    def _convert_token_to_id(self, token: str) -> int:
+        """ Converts a token to an id using the vocab. """
+        token = token.lower()
+        return self.encoder.get(token, self.encoder.get(self.unk_token))
+
+    def _convert_id_to_token(self, index: int) -> str:
+        """Converts an index (integer) in a token (str) using the vocab."""
+        return self.decoder.get(index, self.unk_token)
+
+    def convert_tokens_to_string(self, tokens: List[str]) -> str:
+        """ Converts a sequence of tokens  in a single string. """
+        out_string = " ".join(tokens).replace("@@ ", "").strip()
+        return out_string
+
+    def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
+        if not os.path.isdir(save_directory):
+            logger.error("Vocabulary path ({}) should be a directory".format(save_directory))
+            return
+        vocab_file = os.path.join(
+            save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]
+        )
+        merge_file = os.path.join(
+            save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["merges_file"]
+        )
+
+        with open(vocab_file, "w", encoding="utf-8") as f:
+            f.write(json.dumps(self.encoder, ensure_ascii=False))
+
+        index = 0
+        with open(merge_file, "w", encoding="utf-8") as writer:
+            writer.write("#version: 0.2\n")
+            for bpe_tokens, token_index in sorted(self.bpe_ranks.items(), key=lambda kv: kv[1]):
+                if index != token_index:
+                    logger.warning(
+                        "Saving vocabulary to {}: BPE merge indices are not consecutive."
+                        " Please check that the tokenizer is not corrupted!".format(merge_file)
+                    )
+                    index = token_index
+                writer.write(" ".join(bpe_tokens) + "\n")
+                index += 1
+
+        return vocab_file, merge_file

src/transformers/models/blenderbot_small/tokenization_blenderbot_small_fast.py

+# coding=utf-8
+# Copyright 2021, The Facebook, Inc. and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Fast tokenization class for BlenderbotSmall."""
+from typing import List, Optional
+
+from tokenizers import ByteLevelBPETokenizer
+
+from ...tokenization_utils_fast import PreTrainedTokenizerFast
+from ...utils import logging
+from .tokenization_blenderbot_small import BlenderbotSmallTokenizer
+
+
+logger = logging.get_logger(__name__)
+
+VOCAB_FILES_NAMES = {}
+
+PRETRAINED_VOCAB_FILES_MAP = {}
+
+PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {
+    "facebook/blenderbot_small-90M": 512,
+}
+
+
+class BlenderbotSmallTokenizerFast(PreTrainedTokenizerFast):
+    """
+    Construct a "fast" BlenderbotSmall tokenizer (backed by HuggingFace's `tokenizers` library).
+
+    Args:
+        vocab_file (:obj:`str`):
+            Path to the vocabulary file.
+    """
+
+    vocab_files_names = VOCAB_FILES_NAMES
+    pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
+    max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
+    slow_tokenizer_class = BlenderbotSmallTokenizer
+
+    def __init__(
+        self,
+        vocab_file,
+        merges_file,
+        unk_token="<|endoftext|>",
+        bos_token="<|endoftext|>",
+        eos_token="<|endoftext|>",
+        add_prefix_space=False,
+        trim_offsets=True,
+        **kwargs
+    ):
+        super().__init__(
+            ByteLevelBPETokenizer(
+                vocab_file=vocab_file,
+                merges_file=merges_file,
+                add_prefix_space=add_prefix_space,
+                trim_offsets=trim_offsets,
+            ),
+            bos_token=bos_token,
+            eos_token=eos_token,
+            unk_token=unk_token,
+            **kwargs,
+        )
+        self.add_prefix_space = add_prefix_space
+
+    def build_inputs_with_special_tokens(self, token_ids_0, token_ids_1=None):
+        output = [self.bos_token_id] + token_ids_0 + [self.eos_token_id]
+        if token_ids_1 is None:
+            return output
+
+        return output + [self.eos_token_id] + token_ids_1 + [self.eos_token_id]
+
+    def create_token_type_ids_from_sequences(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        Create a mask from the two sequences passed to be used in a sequence-pair classification task. BlenderbotSmall
+        does not make use of token type ids, therefore a list of zeros is returned.
+
+        Args:
+            token_ids_0 (:obj:`List[int]`):
+                List of IDs.
+            token_ids_1 (:obj:`List[int]`, `optional`):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            :obj:`List[int]`: List of zeros.
+        """
+        sep = [self.sep_token_id]
+        cls = [self.cls_token_id]
+
+        if token_ids_1 is None:
+            return len(cls + token_ids_0 + sep) * [0]
+        return len(cls + token_ids_0 + sep + sep + token_ids_1 + sep) * [0]

src/transformers/models/encoder_decoder/modeling_encoder_decoder.py

@@ -77,10 +77,21 @@ ENCODER_DECODER_INPUTS_DOCSTRING = r"""
 
             `What are attention masks? <../glossary.html#attention-mask>`__
         decoder_input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, target_sequence_length)`, `optional`):
+            Indices of decoder input sequence tokens in the vocabulary.
+
+            Indices can be obtained using :class:`~transformers.PreTrainedTokenizer`. See
+            :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__` for
+            details.
+
+            `What are input IDs? <../glossary.html#input-ids>`__
+
+            If :obj:`past_key_values` is used, optionally only the last :obj:`decoder_input_ids` have to be input (see
+            :obj:`past_key_values`).
+
             Provide for sequence to sequence training to the decoder. Indices can be obtained using
             :class:`~transformers.PretrainedTokenizer`. See :meth:`transformers.PreTrainedTokenizer.encode` and
             :meth:`transformers.PreTrainedTokenizer.__call__` for details.
-        decoder_attention_mask (:obj:`torch.BoolTensor` of shape :obj:`(batch_size, tgt_seq_len)`, `optional`):
+        decoder_attention_mask (:obj:`torch.BoolTensor` of shape :obj:`(batch_size, target_sequence_length)`, `optional`):
             Default behavior: generate a tensor that ignores pad tokens in :obj:`decoder_input_ids`. Causal mask will
             also be used by default.
         encoder_outputs (:obj:`tuple(torch.FloatTensor)`, `optional`):

src/transformers/models/fsmt/modeling_fsmt.py

@@ -235,7 +235,7 @@ FSMT_INPUTS_DOCSTRING = r"""
         decoder_input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, target_sequence_length)`, `optional`):
             Provide for translation and summarization training. By default, the model will create this tensor by
             shifting the input_ids right, following the paper.
-        decoder_attention_mask (:obj:`torch.BoolTensor` of shape :obj:`(batch_size, tgt_seq_len)`, `optional`):
+        decoder_attention_mask (:obj:`torch.BoolTensor` of shape :obj:`(batch_size, target_sequence_length)`, `optional`):
             Default behavior: generate a tensor that ignores pad tokens in :obj:`decoder_input_ids`. Causal mask will
             also be used by default. If you want to change padding behavior, you should read
             :func:`modeling_fstm._prepare_fstm_decoder_inputs` and modify. See diagram 1 in the paper for more info on

src/transformers/models/marian/__init__.py

@@ -15,16 +15,20 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-
-from ...file_utils import is_sentencepiece_available, is_tf_available, is_torch_available
-from .configuration_marian import MarianConfig
+from ...file_utils import is_sentencepiece_available, is_tf_available, is_tokenizers_available, is_torch_available
+from .configuration_marian import MARIAN_PRETRAINED_CONFIG_ARCHIVE_MAP, MarianConfig
 
 
 if is_sentencepiece_available():
     from .tokenization_marian import MarianTokenizer
 
 if is_torch_available():
-    from .modeling_marian import MarianMTModel
+    from .modeling_marian import (
+        MARIAN_PRETRAINED_MODEL_ARCHIVE_LIST,
+        MarianModel,
+        MarianMTModel,
+        MarianPreTrainedModel,
+    )
 
 if is_tf_available():
     from .modeling_tf_marian import TFMarianMTModel

src/transformers/models/marian/configuration_marian.py

 # coding=utf-8
-# Copyright 2020 The OPUS-NMT Team, Marian team, and The HuggingFace Inc. team.
+# Copyright 2021 The Marian Team Authors and The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -14,40 +14,48 @@
 # limitations under the License.
 """ Marian model configuration """
 
-from ..bart.configuration_bart import BartConfig
+from ...configuration_utils import PretrainedConfig
+from ...utils import logging
 
 
-PRETRAINED_CONFIG_ARCHIVE_MAP = {
+logger = logging.get_logger(__name__)
+
+MARIAN_PRETRAINED_CONFIG_ARCHIVE_MAP = {
     "Helsinki-NLP/opus-mt-en-de": "https://huggingface.co/Helsinki-NLP/opus-mt-en-de/resolve/main/config.json",
+    # See all Marian models at https://huggingface.co/models?filter=marian
 }
 
 
-class MarianConfig(BartConfig):
-    """
-    This is the configuration class to store the configuration of a :class:`~transformers.MarianMTModel`. It is used to
-    instantiate a Marian model according to the specified arguments, defining the model architecture.
+class MarianConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a :class:`~transformers.MarianModel`. It is used to
+    instantiate an Marian model according to the specified arguments, defining the model architecture. Instantiating a
+    configuration with the defaults will yield a similar configuration to that of the Marian
+    `Helsinki-NLP/opus-mt-en-de <https://huggingface.co/Helsinki-NLP/opus-mt-en-de>`__ architecture.
 
     Configuration objects inherit from :class:`~transformers.PretrainedConfig` and can be used to control the model
     outputs. Read the documentation from :class:`~transformers.PretrainedConfig` for more information.
 
+
     Args:
-        vocab_size (:obj:`int`, `optional`, defaults to 58101):
+        vocab_size (:obj:`int`, `optional`, defaults to 50265):
             Vocabulary size of the Marian model. Defines the number of different tokens that can be represented by the
-            :obj:`inputs_ids` passed when calling :class:`~transformers.MarianMTModel`.
-        d_model (:obj:`int`, `optional`, defaults to 512):
+            :obj:`inputs_ids` passed when calling :class:`~transformers.MarianModel` or
+            :class:`~transformers.TFMarianModel`.
+        d_model (:obj:`int`, `optional`, defaults to 1024):
             Dimensionality of the layers and the pooler layer.
-        encoder_layers (:obj:`int`, `optional`, defaults to 6):
+        encoder_layers (:obj:`int`, `optional`, defaults to 12):
             Number of encoder layers.
-        decoder_layers (:obj:`int`, `optional`, defaults to 6):
+        decoder_layers (:obj:`int`, `optional`, defaults to 12):
             Number of decoder layers.
-        encoder_attention_heads (:obj:`int`, `optional`, defaults to 8):
+        encoder_attention_heads (:obj:`int`, `optional`, defaults to 16):
             Number of attention heads for each attention layer in the Transformer encoder.
-        decoder_attention_heads (:obj:`int`, `optional`, defaults to 8):
+        decoder_attention_heads (:obj:`int`, `optional`, defaults to 16):
             Number of attention heads for each attention layer in the Transformer decoder.
-        decoder_ffn_dim (:obj:`int`, `optional`, defaults to 2048):
-            Dimensionality of the "intermediate" (i.e., feed-forward) layer in decoder.
-        encoder_ffn_dim (:obj:`int`, `optional`, defaults to 2048):
-            Dimensionality of the "intermediate" (i.e., feed-forward) layer in decoder.
+        decoder_ffn_dim (:obj:`int`, `optional`, defaults to 4096):
+            Dimensionality of the "intermediate" (often named feed-forward) layer in decoder.
+        encoder_ffn_dim (:obj:`int`, `optional`, defaults to 4096):
+            Dimensionality of the "intermediate" (often named feed-forward) layer in decoder.
         activation_function (:obj:`str` or :obj:`function`, `optional`, defaults to :obj:`"gelu"`):
             The non-linear activation function (function or string) in the encoder and pooler. If string,
             :obj:`"gelu"`, :obj:`"relu"`, :obj:`"silu"` and :obj:`"gelu_new"` are supported.
@@ -59,42 +67,113 @@ class MarianConfig(BartConfig):
             The dropout ratio for activations inside the fully connected layer.
         classifier_dropout (:obj:`float`, `optional`, defaults to 0.0):
             The dropout ratio for classifier.
-        max_position_embeddings (:obj:`int`, `optional`, defaults to 512):
+        max_position_embeddings (:obj:`int`, `optional`, defaults to 1024):
             The maximum sequence length that this model might ever be used with. Typically set this to something large
             just in case (e.g., 512 or 1024 or 2048).
         init_std (:obj:`float`, `optional`, defaults to 0.02):
             The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
-        add_bias_logits (:obj:`bool`, `optional`, defaults to :obj:`False`):
-            This should be completed, specific to marian.
-        normalize_before (:obj:`bool`, `optional`, defaults to :obj:`False`):
-            Call layernorm before attention ops.
-        normalize_embedding (:obj:`bool`, `optional`, defaults to :obj:`False`):
-            Call layernorm after embeddings.
-        static_position_embeddings (:obj:`bool`, `optional`, defaults to :obj:`True`):
-            Don't learn positional embeddings, use sinusoidal.
-        add_final_layer_norm (:obj:`bool`, `optional`, defaults to :obj:`False`):
-            Why not add another layernorm?
-        scale_embedding (:obj:`bool`, `optional`, defaults to :obj:`False`):
-            Scale embeddings by diving by sqrt(d_model).
-        eos_token_id (:obj:`int`, `optional`, defaults to 2)
-            End of stream token id.
-        pad_token_id (:obj:`int`, `optional`, defaults to 1)
-            Padding token id.
-        bos_token_id (:obj:`int`, `optional`, defaults to 0)
-            Beginning of stream token id.
         encoder_layerdrop: (:obj:`float`, `optional`, defaults to 0.0):
             The LayerDrop probability for the encoder. See the `LayerDrop paper <see
             https://arxiv.org/abs/1909.11556>`__ for more details.
         decoder_layerdrop: (:obj:`float`, `optional`, defaults to 0.0):
             The LayerDrop probability for the decoder. See the `LayerDrop paper <see
             https://arxiv.org/abs/1909.11556>`__ for more details.
-        extra_pos_embeddings: (:obj:`int`, `optional`, defaults to 2):
-            How many extra learned positional embeddings to use.
-        is_encoder_decoder (:obj:`bool`, `optional`, defaults to :obj:`True`):
-            Whether this is an encoder/decoder model
-        force_bos_token_to_be_generated (:obj:`bool`, `optional`, defaults to :obj:`False`):
-            Whether or not to force BOS token to be generated at step 1 (after ``decoder_start_token_id``).
-    """
+        gradient_checkpointing (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            If True, use gradient checkpointing to save memory at the expense of slower backward pass.
+        scale_embedding (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Scale embeddings by diving by sqrt(d_model).
+        use_cache (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Whether or not the model should return the last key/values attentions (not used by all models)
+
+    Examples::
 
+        >>> from transformers import MarianModel, MarianConfig
+
+        >>> # Initializing a Marian Helsinki-NLP/opus-mt-en-de style configuration
+        >>> configuration = MarianConfig()
+
+        >>> # Initializing a model from the Helsinki-NLP/opus-mt-en-de style configuration
+        >>> model = MarianModel(configuration)
+
+        >>> # Accessing the model configuration
+        >>> configuration = model.config
+    """
     model_type = "marian"
     keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        vocab_size=50265,
+        max_position_embeddings=1024,
+        encoder_layers=12,
+        encoder_ffn_dim=4096,
+        encoder_attention_heads=16,
+        decoder_layers=12,
+        decoder_ffn_dim=4096,
+        decoder_attention_heads=16,
+        encoder_layerdrop=0.0,
+        decoder_layerdrop=0.0,
+        use_cache=True,
+        is_encoder_decoder=True,
+        activation_function="gelu",
+        d_model=1024,
+        dropout=0.1,
+        attention_dropout=0.0,
+        activation_dropout=0.0,
+        init_std=0.02,
+        decoder_start_token_id=58100,
+        classifier_dropout=0.0,
+        scale_embedding=False,
+        gradient_checkpointing=False,
+        pad_token_id=58100,
+        eos_token_id=0,
+        **kwargs
+    ):
+        super().__init__(
+            pad_token_id=pad_token_id,
+            eos_token_id=eos_token_id,
+            is_encoder_decoder=is_encoder_decoder,
+            decoder_start_token_id=decoder_start_token_id,
+            **kwargs,
+        )
+
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.d_model = d_model
+        self.encoder_ffn_dim = encoder_ffn_dim
+        self.encoder_layers = encoder_layers
+        self.encoder_attention_heads = encoder_attention_heads
+        self.decoder_ffn_dim = decoder_ffn_dim
+        self.decoder_layers = decoder_layers
+        self.decoder_attention_heads = decoder_attention_heads
+        self.dropout = dropout
+        self.attention_dropout = attention_dropout
+        self.activation_dropout = activation_dropout
+        self.activation_function = activation_function
+        self.init_std = init_std
+        self.encoder_layerdrop = encoder_layerdrop
+        self.decoder_layerdrop = decoder_layerdrop
+        self.classifier_dropout = classifier_dropout
+        self.use_cache = use_cache
+        self.num_hidden_layers = encoder_layers
+        self.gradient_checkpointing = gradient_checkpointing
+        self.scale_embedding = scale_embedding  # scale factor will be sqrt(d_model) if True
+
+        # IMPORTANT
+        # DELETE ALL OF THE FOLLOWING LINES AS SOON AS TF IS READY
+        self.extra_pos_embeddings = 0
+        self.normalize_before = False
+        self.add_final_layer_norm = False
+        self.do_blenderbot_90_layernorm = False
+        self.normalize_embedding = False
+        self.static_position_embeddings = True
+        self.add_bias_logits = False
+        self.force_bos_token_to_be_generated = False
+
+    @property
+    def num_attention_heads(self) -> int:
+        return self.encoder_attention_heads
+
+    @property
+    def hidden_size(self) -> int:
+        return self.d_model

src/transformers/models/marian/modeling_marian.py

 # coding=utf-8
-# Copyright 2020 Marian Team Authors and The HuggingFace Inc. team.
+# Copyright 2021 The Marian Team Authors and The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -15,49 +15,1205 @@
 """PyTorch MarianMTModel model, ported from the Marian C++ repo."""
 
 
-from ..bart.modeling_bart import BartForConditionalGeneration
+import math
+import random
+from typing import Optional, Tuple
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from torch import nn
+from torch.nn import CrossEntropyLoss
+
+from ...activations import ACT2FN
+from ...file_utils import (
+    add_end_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    replace_return_docstrings,
+)
+from ...modeling_outputs import (
+    BaseModelOutput,
+    BaseModelOutputWithPastAndCrossAttentions,
+    Seq2SeqLMOutput,
+    Seq2SeqModelOutput,
+)
+from ...modeling_utils import PreTrainedModel
+from ...utils import logging
 from .configuration_marian import MarianConfig
 
 
-# See all Marian models at https://huggingface.co/models?search=Helsinki-NLP
+logger = logging.get_logger(__name__)
 
+_CONFIG_FOR_DOC = "MarianConfig"
+_TOKENIZER_FOR_DOC = "MarianTokenizer"
 
-class MarianMTModel(BartForConditionalGeneration):
-    r"""
-    Pytorch version of marian-nmt's transformer.h (c++). Designed for the OPUS-NMT translation checkpoints. Available
-    models are listed `here <https://huggingface.co/models?search=Helsinki-NLP>`__.
 
-    This class overrides :class:`~transformers.BartForConditionalGeneration`. Please check the superclass for the
-    appropriate documentation alongside usage examples.
+MARIAN_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "Helsinki-NLP/opus-mt-en-de",
+    # See all Marian models at https://huggingface.co/models?filter=marian
+]
 
-    Examples::
 
-        >>> from transformers import MarianTokenizer, MarianMTModel
-        >>> from typing import List
-        >>> src = 'fr'  # source language
-        >>> trg = 'en'  # target language
-        >>> sample_text = "où est l'arrêt de bus ?"
-        >>> mname = f'Helsinki-NLP/opus-mt-{src}-{trg}'
+# Copied from transformers.models.bart.modeling_bart.shift_tokens_right
+def shift_tokens_right(input_ids: torch.Tensor, pad_token_id: int, decoder_start_token_id: int):
+    """
+    Shift input ids one token to the right.
+    """
+    shifted_input_ids = input_ids.new_zeros(input_ids.shape)
+    shifted_input_ids[:, 1:] = input_ids[:, :-1].clone()
+    shifted_input_ids[:, 0] = decoder_start_token_id
 
-        >>> model = MarianMTModel.from_pretrained(mname)
-        >>> tok = MarianTokenizer.from_pretrained(mname)
-        >>> batch = tok.prepare_seq2seq_batch(src_texts=[sample_text], return_tensors="pt")  # don't need tgt_text for inference
-        >>> gen = model.generate(**batch)  # for forward pass: model(**batch)
-        >>> words: List[str] = tok.batch_decode(gen, skip_special_tokens=True)  # returns "Where is the bus stop ?"
+    assert pad_token_id is not None, "self.model.config.pad_token_id has to be defined."
+    # replace possible -100 values in labels by `pad_token_id`
+    shifted_input_ids.masked_fill_(shifted_input_ids == -100, pad_token_id)
 
+    return shifted_input_ids
+
+
+# Copied from transformers.models.bart.modeling_bart._make_causal_mask
+def _make_causal_mask(input_ids_shape: torch.Size, dtype: torch.dtype, past_key_values_length: int = 0):
+    """
+    Make causal mask used for bi-directional self-attention.
     """
+    bsz, tgt_len = input_ids_shape
+    mask = torch.full((tgt_len, tgt_len), float("-inf"))
+    mask_cond = torch.arange(mask.size(-1))
+    mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)
+    mask = mask.to(dtype)
+
+    if past_key_values_length > 0:
+        mask = torch.cat([torch.zeros(tgt_len, past_key_values_length, dtype=dtype), mask], dim=-1)
+    return mask[None, None, :, :].expand(bsz, 1, tgt_len, tgt_len + past_key_values_length)
+
+
+# Copied from transformers.models.bart.modeling_bart._expand_mask
+def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
+    """
+    Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
+    """
+    bsz, src_len = mask.size()
+    tgt_len = tgt_len if tgt_len is not None else src_len
+
+    expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)
+
+    inverted_mask = 1.0 - expanded_mask
+
+    return inverted_mask.masked_fill(inverted_mask.bool(), torch.finfo(dtype).min)
+
+
+class MarianSinusoidalPositionalEmbedding(nn.Embedding):
+    """This module produces sinusoidal positional embeddings of any length."""
+
+    def __init__(self, num_positions: int, embedding_dim: int, padding_idx: Optional[int] = None):
+        super().__init__(num_positions, embedding_dim)
+        self.weight = self._init_weight(self.weight)
+
+    @staticmethod
+    def _init_weight(out: nn.Parameter):
+        """
+        Identical to the XLM create_sinusoidal_embeddings except features are not interleaved. The cos features are in
+        the 2nd half of the vector. [dim // 2:]
+        """
+        n_pos, dim = out.shape
+        position_enc = np.array(
+            [[pos / np.power(10000, 2 * (j // 2) / dim) for j in range(dim)] for pos in range(n_pos)]
+        )
+        out.requires_grad = False  # set early to avoid an error in pytorch-1.8+
+        sentinel = dim // 2 if dim % 2 == 0 else (dim // 2) + 1
+        out[:, 0:sentinel] = torch.FloatTensor(np.sin(position_enc[:, 0::2]))
+        out[:, sentinel:] = torch.FloatTensor(np.cos(position_enc[:, 1::2]))
+        out.detach_()
+        return out
+
+    @torch.no_grad()
+    def forward(self, input_ids_shape: torch.Size, past_key_values_length: int = 0):
+        """`input_ids_shape` is expected to be [bsz x seqlen]."""
+        bsz, seq_len = input_ids_shape[:2]
+        positions = torch.arange(
+            past_key_values_length, past_key_values_length + seq_len, dtype=torch.long, device=self.weight.device
+        )
+        return super().forward(positions)
+
+
+# Copied from transformers.models.bart.modeling_bart.BartAttention with Bart->Marian
+class MarianAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(
+        self,
+        embed_dim: int,
+        num_heads: int,
+        dropout: float = 0.0,
+        is_decoder: bool = False,
+        bias: bool = True,
+    ):
+        super().__init__()
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.dropout = dropout
+        self.head_dim = embed_dim // num_heads
+        assert (
+            self.head_dim * num_heads == self.embed_dim
+        ), f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`: {num_heads})."
+        self.scaling = self.head_dim ** -0.5
+        self.is_decoder = is_decoder
+
+        self.k_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+        self.v_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+        self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+        self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+
+    def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
+        return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        key_value_states: Optional[torch.Tensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        output_attentions: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        """Input shape: Batch x Time x Channel"""
+
+        # if key_value_states are provided this layer is used as a cross-attention layer
+        # for the decoder
+        is_cross_attention = key_value_states is not None
+        bsz, tgt_len, embed_dim = hidden_states.size()
+
+        # get query proj
+        query_states = self.q_proj(hidden_states) * self.scaling
+        # get key, value proj
+        if is_cross_attention and past_key_value is not None:
+            # reuse k,v, cross_attentions
+            key_states = past_key_value[0]
+            value_states = past_key_value[1]
+        elif is_cross_attention:
+            # cross_attentions
+            key_states = self._shape(self.k_proj(key_value_states), -1, bsz)
+            value_states = self._shape(self.v_proj(key_value_states), -1, bsz)
+        elif past_key_value is not None:
+            # reuse k, v, self_attention
+            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
+            key_states = torch.cat([past_key_value[0], key_states], dim=2)
+            value_states = torch.cat([past_key_value[1], value_states], dim=2)
+        else:
+            # self_attention
+            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
+
+        if self.is_decoder:
+            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
+            # Further calls to cross_attention layer can then reuse all cross-attention
+            # key/value_states (first "if" case)
+            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
+            # all previous decoder key/value_states. Further calls to uni-directional self-attention
+            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
+            # if encoder bi-directional self-attention `past_key_value` is always `None`
+            past_key_value = (key_states, value_states)
+
+        proj_shape = (bsz * self.num_heads, -1, self.head_dim)
+        query_states = self._shape(query_states, tgt_len, bsz).view(*proj_shape)
+        key_states = key_states.view(*proj_shape)
+        value_states = value_states.view(*proj_shape)
+
+        src_len = key_states.size(1)
+        attn_weights = torch.bmm(query_states, key_states.transpose(1, 2))
+
+        assert attn_weights.size() == (
+            bsz * self.num_heads,
+            tgt_len,
+            src_len,
+        ), f"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is {attn_weights.size()}"
+
+        if attention_mask is not None:
+            assert attention_mask.size() == (
+                bsz,
+                1,
+                tgt_len,
+                src_len,
+            ), f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.size()}"
+            attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + attention_mask
+            attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
+
+        attn_weights = F.softmax(attn_weights, dim=-1)
+
+        if output_attentions:
+            # this operation is a bit akward, but it's required to
+            # make sure that attn_weights keeps its gradient.
+            # In order to do so, attn_weights have to reshaped
+            # twice and have to be reused in the following
+            attn_weights_reshaped = attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
+            attn_weights = attn_weights_reshaped.view(bsz * self.num_heads, tgt_len, src_len)
+        else:
+            attn_weights_reshaped = None
+
+        attn_probs = F.dropout(attn_weights, p=self.dropout, training=self.training)
+
+        attn_output = torch.bmm(attn_probs, value_states)
+
+        assert attn_output.size() == (
+            bsz * self.num_heads,
+            tgt_len,
+            self.head_dim,
+        ), f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is {attn_output.size()}"
+
+        attn_output = (
+            attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
+            .transpose(1, 2)
+            .reshape(bsz, tgt_len, embed_dim)
+        )
+
+        attn_output = self.out_proj(attn_output)
+
+        return attn_output, attn_weights_reshaped, past_key_value
+
+
+# Copied from transformers.models.bart.modeling_bart.BartEncoderLayer with Bart->Marian
+class MarianEncoderLayer(nn.Module):
+    def __init__(self, config: MarianConfig):
+        super().__init__()
+        self.embed_dim = config.d_model
+        self.self_attn = MarianAttention(
+            embed_dim=self.embed_dim,
+            num_heads=config.encoder_attention_heads,
+            dropout=config.attention_dropout,
+        )
+        self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
+        self.dropout = config.dropout
+        self.activation_fn = ACT2FN[config.activation_function]
+        self.activation_dropout = config.activation_dropout
+        self.fc1 = nn.Linear(self.embed_dim, config.encoder_ffn_dim)
+        self.fc2 = nn.Linear(config.encoder_ffn_dim, self.embed_dim)
+        self.final_layer_norm = nn.LayerNorm(self.embed_dim)
+
+    def forward(self, hidden_states: torch.Tensor, attention_mask: torch.Tensor, output_attentions: bool = False):
+        """
+        Args:
+            hidden_states (:obj:`torch.FloatTensor`): input to the layer of shape `(seq_len, batch, embed_dim)`
+            attention_mask (:obj:`torch.FloatTensor`): attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            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.
+        """
+        residual = hidden_states
+        hidden_states, attn_weights, _ = self.self_attn(
+            hidden_states=hidden_states, attention_mask=attention_mask, output_attentions=output_attentions
+        )
+        hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
+        hidden_states = residual + hidden_states
+        hidden_states = self.self_attn_layer_norm(hidden_states)
+
+        residual = hidden_states
+        hidden_states = self.activation_fn(self.fc1(hidden_states))
+        hidden_states = F.dropout(hidden_states, p=self.activation_dropout, training=self.training)
+        hidden_states = self.fc2(hidden_states)
+        hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
+        hidden_states = residual + hidden_states
+        hidden_states = self.final_layer_norm(hidden_states)
+
+        if torch.isinf(hidden_states).any() or torch.isnan(hidden_states).any():
+            clamp_value = torch.finfo(hidden_states.dtype).max - 1000
+            hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (attn_weights,)
+
+        return outputs
+
+
+# Copied from transformers.models.bart.modeling_bart.BartDecoderLayer with Bart->Marian
+class MarianDecoderLayer(nn.Module):
+    def __init__(self, config: MarianConfig):
+        super().__init__()
+        self.embed_dim = config.d_model
+
+        self.self_attn = MarianAttention(
+            embed_dim=self.embed_dim,
+            num_heads=config.decoder_attention_heads,
+            dropout=config.attention_dropout,
+            is_decoder=True,
+        )
+        self.dropout = config.dropout
+        self.activation_fn = ACT2FN[config.activation_function]
+        self.activation_dropout = config.activation_dropout
+
+        self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
+        self.encoder_attn = MarianAttention(
+            self.embed_dim,
+            config.decoder_attention_heads,
+            dropout=config.attention_dropout,
+            is_decoder=True,
+        )
+        self.encoder_attn_layer_norm = nn.LayerNorm(self.embed_dim)
+        self.fc1 = nn.Linear(self.embed_dim, config.decoder_ffn_dim)
+        self.fc2 = nn.Linear(config.decoder_ffn_dim, self.embed_dim)
+        self.final_layer_norm = nn.LayerNorm(self.embed_dim)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = True,
+    ):
+        """
+        Args:
+            hidden_states (:obj:`torch.FloatTensor`): input to the layer of shape `(seq_len, batch, embed_dim)`
+            attention_mask (:obj:`torch.FloatTensor`): attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            encoder_hidden_states (:obj:`torch.FloatTensor`): cross attention input to the layer of shape `(seq_len, batch, embed_dim)`
+            encoder_attention_mask (:obj:`torch.FloatTensor`): encoder attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            past_key_value (:obj:`Tuple(torch.FloatTensor)`): cached past key and value projection states
+            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.
+        """
+        residual = hidden_states
+
+        # Self Attention
+        # decoder uni-directional self-attention cached key/values tuple is at positions 1,2
+        self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None
+        # add present self-attn cache to positions 1,2 of present_key_value tuple
+        hidden_states, self_attn_weights, present_key_value = self.self_attn(
+            hidden_states=hidden_states,
+            past_key_value=self_attn_past_key_value,
+            attention_mask=attention_mask,
+            output_attentions=output_attentions,
+        )
+        hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
+        hidden_states = residual + hidden_states
+        hidden_states = self.self_attn_layer_norm(hidden_states)
+
+        # Cross-Attention Block
+        cross_attn_present_key_value = None
+        cross_attn_weights = None
+        if encoder_hidden_states is not None:
+            residual = hidden_states
+
+            # cross_attn cached key/values tuple is at positions 3,4 of present_key_value tuple
+            cross_attn_past_key_value = past_key_value[-2:] if past_key_value is not None else None
+            hidden_states, cross_attn_weights, cross_attn_present_key_value = self.encoder_attn(
+                hidden_states=hidden_states,
+                key_value_states=encoder_hidden_states,
+                attention_mask=encoder_attention_mask,
+                past_key_value=cross_attn_past_key_value,
+                output_attentions=output_attentions,
+            )
+            hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
+            hidden_states = residual + hidden_states
+            hidden_states = self.encoder_attn_layer_norm(hidden_states)
+
+            # add cross-attn to positions 3,4 of present_key_value tuple
+            present_key_value = present_key_value + cross_attn_present_key_value
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.activation_fn(self.fc1(hidden_states))
+        hidden_states = F.dropout(hidden_states, p=self.activation_dropout, training=self.training)
+        hidden_states = self.fc2(hidden_states)
+        hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
+        hidden_states = residual + hidden_states
+        hidden_states = self.final_layer_norm(hidden_states)
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights, cross_attn_weights)
+
+        if use_cache:
+            outputs += (present_key_value,)
+
+        return outputs
+
+
+class MarianPreTrainedModel(PreTrainedModel):
     config_class = MarianConfig
+    base_model_prefix = "model"
+
+    def _init_weights(self, module):
+        std = self.config.init_std
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, MarianSinusoidalPositionalEmbedding):
+            pass
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+    @property
+    def dummy_inputs(self):
+        pad_token = self.config.pad_token_id
+        input_ids = torch.tensor([[0, 6, 10, 4, 2], [0, 8, 12, 2, pad_token]], device=self.device)
+        dummy_inputs = {
+            "attention_mask": input_ids.ne(pad_token),
+            "input_ids": input_ids,
+            "decoder_input_ids": input_ids,
+        }
+        return dummy_inputs
+
+
+MARIAN_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.
+
+    Parameters:
+        config (:class:`~transformers.MarianConfig`):
+            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.
+"""
+
+MARIAN_GENERATION_EXAMPLE = r"""
+        Pytorch version of marian-nmt's transformer.h (c++). Designed for the OPUS-NMT translation checkpoints.
+        Available models are listed `here <https://huggingface.co/models?search=Helsinki-NLP>`__.
+
+        Examples::
+
+            >>> from transformers import MarianTokenizer, MarianMTModel
+            >>> from typing import List
+            >>> src = 'fr'  # source language
+            >>> trg = 'en'  # target language
+            >>> sample_text = "où est l'arrêt de bus ?"
+            >>> mname = f'Helsinki-NLP/opus-mt-{src}-{trg}'
+
+            >>> model = MarianMTModel.from_pretrained(mname)
+            >>> tok = MarianTokenizer.from_pretrained(mname)
+            >>> batch = tok.prepare_seq2seq_batch(src_texts=[sample_text], return_tensors="pt")  # don't need tgt_text for inference
+            >>> gen = model.generate(**batch)
+            >>> words: List[str] = tok.batch_decode(gen, skip_special_tokens=True)  # returns "Where is the bus stop ?"
+"""
+
+MARIAN_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+
+            Indices can be obtained using :class:`~transformers.MarianTokenizer`. See
+            :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__` for
+            details.
+
+            `What are input IDs? <../glossary.html#input-ids>`__
+        attention_mask (:obj:`torch.Tensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Mask to avoid performing attention on padding token indices. Mask values selected in ``[0, 1]``:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            `What are attention masks? <../glossary.html#attention-mask>`__
+        decoder_input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, target_sequence_length)`, `optional`):
+            Indices of decoder input sequence tokens in the vocabulary.
+
+            Indices can be obtained using :class:`~transformers.MarianTokenizer`. See
+            :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__` for
+            details.
+
+            `What are input IDs? <../glossary.html#input-ids>`__
+
+            Marian uses the :obj:`pad_token_id` as the starting token for :obj:`decoder_input_ids` generation. If
+            :obj:`past_key_values` is used, optionally only the last :obj:`decoder_input_ids` have to be input (see
+            :obj:`past_key_values`).
+        decoder_attention_mask (:obj:`torch.LongTensor` of shape :obj:`(batch_size, target_sequence_length)`, `optional`):
+            Default behavior: generate a tensor that ignores pad tokens in :obj:`decoder_input_ids`. Causal mask will
+            also be used by default.
+
+            If you want to change padding behavior, you should read :func:`modeling_marian._prepare_decoder_inputs` and
+            modify to your needs. See diagram 1 in `the paper <https://arxiv.org/abs/1910.13461>`__ for more
+            information on the default strategy.
+        encoder_outputs (:obj:`tuple(tuple(torch.FloatTensor)`, `optional`):
+            Tuple consists of (:obj:`last_hidden_state`, `optional`: :obj:`hidden_states`, `optional`:
+            :obj:`attentions`) :obj:`last_hidden_state` of shape :obj:`(batch_size, sequence_length, hidden_size)`,
+            `optional`) is a sequence of hidden-states at the output of the last layer of the encoder. Used in the
+            cross-attention of the decoder.
+        past_key_values (:obj:`Tuple[Tuple[torch.Tensor]]` of length :obj:`config.n_layers` with each tuple having 2 tuples each of which has 2 tensors of shape :obj:`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
+            Contains precomputed key and value hidden-states of the attention blocks. Can be used to speed up decoding.
+
+            If :obj:`past_key_values` are used, the user can optionally input only the last :obj:`decoder_input_ids`
+            (those that don't have their past key value states given to this model) of shape :obj:`(batch_size, 1)`
+            instead of all :obj:`decoder_input_ids`` of shape :obj:`(batch_size, sequence_length)`.
+        inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, 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.
+        decoder_inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, target_sequence_length, hidden_size)`, `optional`):
+            Optionally, instead of passing :obj:`decoder_input_ids` you can choose to directly pass an embedded
+            representation. If :obj:`past_key_values` is used, optionally only the last :obj:`decoder_inputs_embeds`
+            have to be input (see :obj:`past_key_values`). This is useful if you want more control over how to convert
+            :obj:`decoder_input_ids` indices into associated vectors than the model's internal embedding lookup matrix.
+
+            If :obj:`decoder_input_ids` and :obj:`decoder_inputs_embeds` are both unset, :obj:`decoder_inputs_embeds`
+            takes the value of :obj:`inputs_embeds`.
+        use_cache (:obj:`bool`, `optional`):
+            If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up
+            decoding (see :obj:`past_key_values`).
+        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.
+"""
+
+
+class MarianEncoder(MarianPreTrainedModel):
+    """
+    Transformer encoder consisting of *config.encoder_layers* self attention layers. Each layer is a
+    :class:`MarianEncoderLayer`.
+
+    Args:
+        config: MarianConfig
+        embed_tokens (torch.nn.Embedding): output embedding
+    """
+
+    def __init__(self, config: MarianConfig, embed_tokens: Optional[nn.Embedding] = None):
+        super().__init__(config)
+
+        self.dropout = config.dropout
+        self.layerdrop = config.encoder_layerdrop
+
+        embed_dim = config.d_model
+        self.padding_idx = config.pad_token_id
+        self.max_source_positions = config.max_position_embeddings
+        self.embed_scale = math.sqrt(embed_dim) if config.scale_embedding else 1.0
+
+        if embed_tokens is not None:
+            self.embed_tokens = embed_tokens
+        else:
+            self.embed_tokens = nn.Embedding(config.vocab_size, embed_dim, self.padding_idx)
+
+        self.embed_positions = MarianSinusoidalPositionalEmbedding(
+            config.max_position_embeddings,
+            embed_dim,
+            self.padding_idx,
+        )
+        self.layers = nn.ModuleList([MarianEncoderLayer(config) for _ in range(config.encoder_layers)])
+        self.init_weights()
+
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        inputs_embeds=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        Args:
+            input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`):
+                Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you
+                provide it.
+
+                Indices can be obtained using :class:`~transformers.MarianTokenizer`. See
+                :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__`
+                for details.
+
+                `What are input IDs? <../glossary.html#input-ids>`__
+            attention_mask (:obj:`torch.Tensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+                Mask to avoid performing attention on padding token indices. Mask values selected in ``[0, 1]``:
+
+                - 1 for tokens that are **not masked**,
+                - 0 for tokens that are **masked**.
+
+                `What are attention masks? <../glossary.html#attention-mask>`__
+            inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, 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.
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # retrieve input_ids and inputs_embeds
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+            input_ids = input_ids.view(-1, input_shape[-1])
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids) * self.embed_scale
+
+        embed_pos = self.embed_positions(input_shape)
+
+        hidden_states = inputs_embeds + embed_pos
+        hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
+
+        # expand attention_mask
+        if attention_mask is not None:
+            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+            attention_mask = _expand_mask(attention_mask, inputs_embeds.dtype)
+
+        encoder_states = () if output_hidden_states else None
+        all_attentions = () if output_attentions else None
+        for encoder_layer in self.layers:
+            if output_hidden_states:
+                encoder_states = encoder_states + (hidden_states,)
+            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
+            dropout_probability = random.uniform(0, 1)
+            if self.training and (dropout_probability < self.layerdrop):  # skip the layer
+                layer_outputs = (None, None)
+            else:
+                if getattr(self.config, "gradient_checkpointing", False):
+
+                    def create_custom_forward(module):
+                        def custom_forward(*inputs):
+                            return module(*inputs, output_attentions)
+
+                        return custom_forward
+
+                    layer_outputs = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(encoder_layer),
+                        hidden_states,
+                        attention_mask,
+                    )
+                else:
+                    layer_outputs = encoder_layer(hidden_states, attention_mask, output_attentions=output_attentions)
+
+                hidden_states = layer_outputs[0]
+
+            if output_attentions:
+                all_attentions = all_attentions + (layer_outputs[1],)
+
+        if output_hidden_states:
+            encoder_states = encoder_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None)
+        return BaseModelOutput(
+            last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions
+        )
+
+
+class MarianDecoder(MarianPreTrainedModel):
+    """
+    Transformer decoder consisting of *config.decoder_layers* layers. Each layer is a :class:`MarianDecoderLayer`
+
+    Args:
+        config: MarianConfig
+        embed_tokens (torch.nn.Embedding): output embedding
+    """
+
+    def __init__(self, config: MarianConfig, embed_tokens: Optional[nn.Embedding] = None):
+        super().__init__(config)
+        self.dropout = config.dropout
+        self.layerdrop = config.decoder_layerdrop
+        self.padding_idx = config.pad_token_id
+        self.max_target_positions = config.max_position_embeddings
+        self.embed_scale = math.sqrt(config.d_model) if config.scale_embedding else 1.0
+
+        if embed_tokens is not None:
+            self.embed_tokens = embed_tokens
+        else:
+            self.embed_tokens = nn.Embedding(config.vocab_size, config.d_model, self.padding_idx)
+
+        self.embed_positions = MarianSinusoidalPositionalEmbedding(
+            config.max_position_embeddings,
+            config.d_model,
+            self.padding_idx,
+        )
+        self.layers = nn.ModuleList([MarianDecoderLayer(config) for _ in range(config.decoder_layers)])
+        self.init_weights()
+
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_values=None,
+        inputs_embeds=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        Args:
+            input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`):
+                Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you
+                provide it.
+
+                Indices can be obtained using :class:`~transformers.MarianTokenizer`. See
+                :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__`
+                for details.
+
+                `What are input IDs? <../glossary.html#input-ids>`__
+            attention_mask (:obj:`torch.Tensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+                Mask to avoid performing attention on padding token indices. Mask values selected in ``[0, 1]``:
+
+                - 1 for tokens that are **not masked**,
+                - 0 for tokens that are **masked**.
+
+                `What are attention masks? <../glossary.html#attention-mask>`__
+            encoder_hidden_states (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, encoder_sequence_length, hidden_size)`, `optional`):
+                Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention
+                of the decoder.
+            encoder_attention_mask (:obj:`torch.LongTensor` of shape :obj:`(batch_size, encoder_sequence_length)`, `optional`):
+                Mask to avoid performing cross-attention on padding tokens indices of encoder input_ids. Mask values
+                selected in ``[0, 1]``:
+
+                - 1 for tokens that are **not masked**,
+                - 0 for tokens that are **masked**.
+
+                `What are attention masks? <../glossary.html#attention-mask>`__
+            past_key_values (:obj:`Tuple[Tuple[torch.Tensor]]` of length :obj:`config.n_layers` with each tuple having 2 tuples each of which has 2 tensors of shape :obj:`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
+                Contains precomputed key and value hidden-states of the attention blocks. Can be used to speed up
+                decoding.
+
+                If :obj:`past_key_values` are used, the user can optionally input only the last
+                :obj:`decoder_input_ids` (those that don't have their past key value states given to this model) of
+                shape :obj:`(batch_size, 1)` instead of all :obj:`decoder_input_ids`` of shape :obj:`(batch_size,
+                sequence_length)`.
+            inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, 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.
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # retrieve input_ids and inputs_embeds
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+            input_ids = input_ids.view(-1, input_shape[-1])
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+        else:
+            raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
+
+        # past_key_values_length
+        past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids) * self.embed_scale
+
+        # create causal mask
+        # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+        combined_attention_mask = None
+        if input_shape[-1] > 1:
+            combined_attention_mask = _make_causal_mask(
+                input_shape, inputs_embeds.dtype, past_key_values_length=past_key_values_length
+            ).to(self.device)
+
+        if attention_mask is not None and combined_attention_mask is not None:
+            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+            combined_attention_mask = combined_attention_mask + _expand_mask(
+                attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]
+            )
+
+        # expand encoder attention mask
+        if encoder_hidden_states is not None and encoder_attention_mask is not None:
+            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+            encoder_attention_mask = _expand_mask(encoder_attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1])
+
+        # embed positions
+        positions = self.embed_positions(input_shape, past_key_values_length)
+
+        hidden_states = inputs_embeds + positions
+
+        hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        all_cross_attentions = () if output_attentions else None
+        next_decoder_cache = () if use_cache else None
+        for idx, decoder_layer in enumerate(self.layers):
+            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+            dropout_probability = random.uniform(0, 1)
+            if self.training and (dropout_probability < self.layerdrop):
+                continue
+
+            past_key_value = past_key_values[idx] if past_key_values is not None else None
+
+            if getattr(self.config, "gradient_checkpointing", False):
+                if use_cache:
+                    raise ValueError(
+                        "When using `gradient_checkpointing, make sure that `use_cache=False` and `config.use_cache=False`."
+                    )
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        # None for past_key_value
+                        return module(*inputs, output_attentions, use_cache)
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(decoder_layer),
+                    hidden_states,
+                    combined_attention_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    None,
+                )
+            else:
+
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask=combined_attention_mask,
+                    encoder_hidden_states=encoder_hidden_states,
+                    encoder_attention_mask=encoder_attention_mask,
+                    past_key_value=past_key_value,
+                    output_attentions=output_attentions,
+                    use_cache=use_cache,
+                )
+            hidden_states = layer_outputs[0]
+
+            if use_cache:
+                next_decoder_cache += (layer_outputs[3 if output_attentions else 1],)
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+                all_cross_attentions += (layer_outputs[2],)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        next_cache = next_decoder_cache if use_cache else None
+        if not return_dict:
+            return tuple(
+                v
+                for v in [hidden_states, next_cache, all_hidden_states, all_self_attns, all_cross_attentions]
+                if v is not None
+            )
+        return BaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+            cross_attentions=all_cross_attentions,
+        )
+
+
+@add_start_docstrings(
+    "The bare Marian Model outputting raw hidden-states without any specific head on top.",
+    MARIAN_START_DOCSTRING,
+)
+class MarianModel(MarianPreTrainedModel):
+    def __init__(self, config: MarianConfig):
+        super().__init__(config)
+
+        padding_idx, vocab_size = config.pad_token_id, config.vocab_size
+        self.shared = nn.Embedding(vocab_size, config.d_model, padding_idx)
+
+        self.encoder = MarianEncoder(config, self.shared)
+        self.decoder = MarianDecoder(config, self.shared)
+
+        self.init_weights()
+
+    def get_input_embeddings(self):
+        return self.shared
+
+    def set_input_embeddings(self, value):
+        self.shared = value
+        self.encoder.embed_tokens = self.shared
+        self.decoder.embed_tokens = self.shared
+
+    def get_encoder(self):
+        return self.encoder
+
+    def get_decoder(self):
+        return self.decoder
+
+    @add_start_docstrings_to_model_forward(MARIAN_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=Seq2SeqModelOutput, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        decoder_input_ids=None,
+        decoder_attention_mask=None,
+        encoder_outputs=None,
+        past_key_values=None,
+        inputs_embeds=None,
+        decoder_inputs_embeds=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        Returns:
+
+        Example::
+
+            >>> from transformers import MarianTokenizer, MarianModel
+
+            >>> tokenizer = MarianTokenizer.from_pretrained('Helsinki-NLP/opus-mt-en-de')
+            >>> model = MarianModel.from_pretrained('Helsinki-NLP/opus-mt-en-de')
+
+            >>> input_ids = tokenizer("Studies have been shown that owning a dog is good for you", return_tensors="pt").input_ids  # Batch size 1
+            >>> decoder_input_ids = tokenizer("<pad> Studien haben gezeigt dass es hilfreich ist einen Hund zu besitzen",
+            ... return_tensors="pt", add_special_tokens=False).input_ids  # Batch size 1
+            >>> outputs = model(input_ids=input_ids, decoder_input_ids=decoder_input_ids)
+
+            >>> last_hidden_states = outputs.last_hidden_state
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if encoder_outputs is None:
+            encoder_outputs = self.encoder(
+                input_ids=input_ids,
+                attention_mask=attention_mask,
+                inputs_embeds=inputs_embeds,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+            )
+        # If the user passed a tuple for encoder_outputs, we wrap it in a BaseModelOutput when return_dict=True
+        elif return_dict and not isinstance(encoder_outputs, BaseModelOutput):
+            encoder_outputs = BaseModelOutput(
+                last_hidden_state=encoder_outputs[0],
+                hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None,
+                attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None,
+            )
+
+        # decoder outputs consists of (dec_features, past_key_value, dec_hidden, dec_attn)
+        decoder_outputs = self.decoder(
+            input_ids=decoder_input_ids,
+            attention_mask=decoder_attention_mask,
+            encoder_hidden_states=encoder_outputs[0],
+            encoder_attention_mask=attention_mask,
+            past_key_values=past_key_values,
+            inputs_embeds=decoder_inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        if not return_dict:
+            return decoder_outputs + encoder_outputs
+
+        return Seq2SeqModelOutput(
+            last_hidden_state=decoder_outputs.last_hidden_state,
+            past_key_values=decoder_outputs.past_key_values,
+            decoder_hidden_states=decoder_outputs.hidden_states,
+            decoder_attentions=decoder_outputs.attentions,
+            cross_attentions=decoder_outputs.cross_attentions,
+            encoder_last_hidden_state=encoder_outputs.last_hidden_state,
+            encoder_hidden_states=encoder_outputs.hidden_states,
+            encoder_attentions=encoder_outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    "The Marian Model with a language modeling head. Can be used for summarization.", MARIAN_START_DOCSTRING
+)
+class MarianMTModel(MarianPreTrainedModel):
+    base_model_prefix = "model"
     _keys_to_ignore_on_load_missing = [
-        "model.encoder.embed_positions.weight",
-        "model.decoder.embed_positions.weight",
+        r"final_logits_bias",
+        r"encoder\.version",
+        r"decoder\.version",
+        r"lm_head\.weight",
+        r"embed_positions",
     ]
+
     _keys_to_ignore_on_save = [
         "model.encoder.embed_positions.weight",
         "model.decoder.embed_positions.weight",
     ]
 
+    def __init__(self, config: MarianConfig):
+        super().__init__(config)
+        self.model = MarianModel(config)
+        self.register_buffer("final_logits_bias", torch.zeros((1, self.model.shared.num_embeddings)))
+        self.lm_head = nn.Linear(config.d_model, self.model.shared.num_embeddings, bias=False)
+
+        self.init_weights()
+
+    def get_encoder(self):
+        return self.model.get_encoder()
+
+    def get_decoder(self):
+        return self.model.get_decoder()
+
+    def resize_token_embeddings(self, new_num_tokens: int) -> nn.Embedding:
+        new_embeddings = super().resize_token_embeddings(new_num_tokens)
+        self._resize_final_logits_bias(new_num_tokens)
+        return new_embeddings
+
+    def _resize_final_logits_bias(self, new_num_tokens: int) -> None:
+        old_num_tokens = self.final_logits_bias.shape[-1]
+        if new_num_tokens <= old_num_tokens:
+            new_bias = self.final_logits_bias[:, :new_num_tokens]
+        else:
+            extra_bias = torch.zeros((1, new_num_tokens - old_num_tokens), device=self.final_logits_bias.device)
+            new_bias = torch.cat([self.final_logits_bias, extra_bias], dim=1)
+        self.register_buffer("final_logits_bias", new_bias)
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    @add_start_docstrings_to_model_forward(MARIAN_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=Seq2SeqLMOutput, config_class=_CONFIG_FOR_DOC)
+    @add_end_docstrings(MARIAN_GENERATION_EXAMPLE)
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        decoder_input_ids=None,
+        decoder_attention_mask=None,
+        encoder_outputs=None,
+        past_key_values=None,
+        inputs_embeds=None,
+        decoder_inputs_embeds=None,
+        labels=None,
+        use_cache=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 either be in ``[0, ...,
+            config.vocab_size]`` or -100 (see ``input_ids`` docstring). Tokens with indices set to ``-100`` are ignored
+            (masked), the loss is only computed for the tokens with labels in ``[0, ..., config.vocab_size]``.
+
+        Returns:
+
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if labels is not None:
+            if decoder_input_ids is None:
+                decoder_input_ids = shift_tokens_right(
+                    labels, self.config.pad_token_id, self.config.decoder_start_token_id
+                )
+
+        outputs = self.model(
+            input_ids,
+            attention_mask=attention_mask,
+            decoder_input_ids=decoder_input_ids,
+            encoder_outputs=encoder_outputs,
+            decoder_attention_mask=decoder_attention_mask,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            decoder_inputs_embeds=decoder_inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        lm_logits = self.lm_head(outputs[0]) + self.final_logits_bias
+
+        masked_lm_loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            masked_lm_loss = loss_fct(lm_logits.view(-1, self.config.vocab_size), labels.view(-1))
+
+        if not return_dict:
+            output = (lm_logits,) + outputs[1:]
+            return ((masked_lm_loss,) + output) if masked_lm_loss is not None else output
+
+        return Seq2SeqLMOutput(
+            loss=masked_lm_loss,
+            logits=lm_logits,
+            past_key_values=outputs.past_key_values,
+            decoder_hidden_states=outputs.decoder_hidden_states,
+            decoder_attentions=outputs.decoder_attentions,
+            cross_attentions=outputs.cross_attentions,
+            encoder_last_hidden_state=outputs.encoder_last_hidden_state,
+            encoder_hidden_states=outputs.encoder_hidden_states,
+            encoder_attentions=outputs.encoder_attentions,
+        )
+
+    def prepare_inputs_for_generation(
+        self, decoder_input_ids, past=None, attention_mask=None, use_cache=None, encoder_outputs=None, **kwargs
+    ):
+        # cut decoder_input_ids if past is used
+        if past is not None:
+            decoder_input_ids = decoder_input_ids[:, -1:]
+
+        return {
+            "input_ids": None,  # encoder_outputs is defined. input_ids not needed
+            "encoder_outputs": encoder_outputs,
+            "past_key_values": past,
+            "decoder_input_ids": decoder_input_ids,
+            "attention_mask": attention_mask,
+            "use_cache": use_cache,  # change this to avoid caching (presumably for debugging)
+        }
+
     def adjust_logits_during_generation(self, logits, cur_len, max_length):
         logits[:, self.config.pad_token_id] = float("-inf")  # never predict pad token.
         if cur_len == max_length - 1 and self.config.eos_token_id is not None:
             self._force_token_id_to_be_generated(logits, self.config.eos_token_id)
         return logits
+
+    @staticmethod
+    def _force_token_id_to_be_generated(scores, token_id) -> None:
+        """force one of token_ids to be generated by setting prob of all other tokens to 0 (logprob=-float("inf"))"""
+        scores[:, [x for x in range(scores.shape[1]) if x != token_id]] = -float("inf")
+
+    @staticmethod
+    def _reorder_cache(past, beam_idx):
+        reordered_past = ()
+        for layer_past in past:
+            # cached cross_attention states don't have to be reordered -> they are always the same
+            reordered_past += (
+                tuple(past_state.index_select(0, beam_idx) for past_state in layer_past[:2]) + layer_past[2:],
+            )
+        return reordered_past

src/transformers/models/marian/tokenization_marian.py

@@ -84,7 +84,7 @@ class MarianTokenizer(PreTrainedTokenizer):
         >>> tok = MarianTokenizer.from_pretrained('Helsinki-NLP/opus-mt-en-de')
         >>> src_texts = [ "I am a small frog.", "Tom asked his teacher for advice."]
         >>> tgt_texts = ["Ich bin ein kleiner Frosch.", "Tom bat seinen Lehrer um Rat."]  # optional
-        >>> batch_enc: BatchEncoding = tok.prepare_seq2seq_batch(src_texts, tgt_texts=tgt_texts, return_tensors="pt")
+        >>> batch_enc = tok.prepare_seq2seq_batch(src_texts, tgt_texts=tgt_texts, return_tensors="pt")
         >>> # keys  [input_ids, attention_mask, labels].
         >>> # model(**batch) should work
     """

src/transformers/models/mbart/__init__.py

@@ -15,9 +15,8 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-
 from ...file_utils import is_sentencepiece_available, is_tf_available, is_tokenizers_available, is_torch_available
-from .configuration_mbart import MBartConfig
+from .configuration_mbart import MBART_PRETRAINED_CONFIG_ARCHIVE_MAP, MBartConfig
 
 
 if is_sentencepiece_available():
@@ -27,7 +26,14 @@ if is_tokenizers_available():
     from .tokenization_mbart_fast import MBartTokenizerFast
 
 if is_torch_available():
-    from .modeling_mbart import MBartForConditionalGeneration, MBartModel
+    from .modeling_mbart import (
+        MBART_PRETRAINED_MODEL_ARCHIVE_LIST,
+        MBartForConditionalGeneration,
+        MBartForQuestionAnswering,
+        MBartForSequenceClassification,
+        MBartModel,
+        MBartPreTrainedModel,
+    )
 
 if is_tf_available():
     from .modeling_tf_mbart import TFMBartForConditionalGeneration

src/transformers/models/mbart/configuration_mbart.py

 # coding=utf-8
-# Copyright 2020 The Fairseq Authors and The HuggingFace Inc. team.
+# Copyright 2021, The Facebook AI Research Team and The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -12,33 +12,36 @@
 # 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.
-""" MBART configuration """
+""" MBART model configuration """
 
+from ...configuration_utils import PretrainedConfig
 from ...utils import logging
-from ..bart.configuration_bart import BartConfig
 
 
 logger = logging.get_logger(__name__)
 
 MBART_PRETRAINED_CONFIG_ARCHIVE_MAP = {
-    "facebook/mbart-large-en-ro": "https://huggingface.co/facebook/mbart-large-en-ro/resolve/main/config.json",
     "facebook/mbart-large-cc25": "https://huggingface.co/facebook/mbart-large-cc25/resolve/main/config.json",
+    # See all MBART models at https://huggingface.co/models?filter=mbart
 }
 
 
-class MBartConfig(BartConfig):
-    """
-    This is the configuration class to store the configuration of a
-    :class:`~transformers.MBartForConditionalGeneration`. It is used to instantiate a BART model according to the
-    specified arguments, defining the model architecture.
+class MBartConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a :class:`~transformers.MBartModel`. It is used to
+    instantiate an MBART model according to the specified arguments, defining the model architecture. Instantiating a
+    configuration with the defaults will yield a similar configuration to that of the MBART `facebook/mbart-large-cc25
+    <https://huggingface.co/facebook/mbart-large-cc25>`__ architecture.
 
     Configuration objects inherit from :class:`~transformers.PretrainedConfig` and can be used to control the model
     outputs. Read the documentation from :class:`~transformers.PretrainedConfig` for more information.
 
+
     Args:
-        vocab_size (:obj:`int`, `optional`, defaults to 250027):
+        vocab_size (:obj:`int`, `optional`, defaults to 50265):
             Vocabulary size of the MBART model. Defines the number of different tokens that can be represented by the
-            :obj:`inputs_ids` passed when calling :class:`~transformers.MBartForConditionalGeneration`.
+            :obj:`inputs_ids` passed when calling :class:`~transformers.MBartModel` or
+            :class:`~transformers.TFMBartModel`.
         d_model (:obj:`int`, `optional`, defaults to 1024):
             Dimensionality of the layers and the pooler layer.
         encoder_layers (:obj:`int`, `optional`, defaults to 12):
@@ -50,9 +53,9 @@ class MBartConfig(BartConfig):
         decoder_attention_heads (:obj:`int`, `optional`, defaults to 16):
             Number of attention heads for each attention layer in the Transformer decoder.
         decoder_ffn_dim (:obj:`int`, `optional`, defaults to 4096):
-            Dimensionality of the "intermediate" (i.e., feed-forward) layer in decoder.
+            Dimensionality of the "intermediate" (often named feed-forward) layer in decoder.
         encoder_ffn_dim (:obj:`int`, `optional`, defaults to 4096):
-            Dimensionality of the "intermediate" (i.e., feed-forward) layer in decoder.
+            Dimensionality of the "intermediate" (often named feed-forward) layer in decoder.
         activation_function (:obj:`str` or :obj:`function`, `optional`, defaults to :obj:`"gelu"`):
             The non-linear activation function (function or string) in the encoder and pooler. If string,
             :obj:`"gelu"`, :obj:`"relu"`, :obj:`"silu"` and :obj:`"gelu_new"` are supported.
@@ -69,37 +72,108 @@ class MBartConfig(BartConfig):
             just in case (e.g., 512 or 1024 or 2048).
         init_std (:obj:`float`, `optional`, defaults to 0.02):
             The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
-        add_bias_logits (:obj:`bool`, `optional`, defaults to :obj:`False`):
-            This should be completed, specific to marian.
-        normalize_before (:obj:`bool`, `optional`, defaults to :obj:`True`):
-            Call layernorm before attention ops.
-        normalize_embedding (:obj:`bool`, `optional`, defaults to :obj:`True`):
-            Call layernorm after embeddings. Only True for Bart.
-        static_position_embeddings (:obj:`bool`, `optional`, defaults to :obj:`False`):
-            Don't learn positional embeddings, use sinusoidal.
-        add_final_layer_norm (:obj:`bool`, `optional`, defaults to :obj:`True`):
-            Why not add another layernorm?
-        scale_embedding (:obj:`bool`, `optional`, defaults to :obj:`False`):
-            Scale embeddings by diving by sqrt(d_model).
-        eos_token_id (:obj:`int`, `optional`, defaults to 2)
-            End of stream token id.
-        pad_token_id (:obj:`int`, `optional`, defaults to 1)
-            Padding token id.
-        bos_token_id (:obj:`int`, `optional`, defaults to 0)
-            Beginning of stream token id.
         encoder_layerdrop: (:obj:`float`, `optional`, defaults to 0.0):
             The LayerDrop probability for the encoder. See the `LayerDrop paper <see
             https://arxiv.org/abs/1909.11556>`__ for more details.
         decoder_layerdrop: (:obj:`float`, `optional`, defaults to 0.0):
             The LayerDrop probability for the decoder. See the `LayerDrop paper <see
             https://arxiv.org/abs/1909.11556>`__ for more details.
-        extra_pos_embeddings: (:obj:`int`, `optional`, defaults to 2):
-            How many extra learned positional embeddings to use. Should be equal to :obj:`pad_token_id+1`.
-        is_encoder_decoder (:obj:`bool`, `optional`, defaults to :obj:`True`):
-            Whether this is an encoder/decoder model
-        force_bos_token_to_be_generated (:obj:`bool`, `optional`, defaults to :obj:`False`):
-            Whether or not to force BOS token to be generated at step 1 (after ``decoder_start_token_id``).
-    """
+        gradient_checkpointing (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            If True, use gradient checkpointing to save memory at the expense of slower backward pass.
+        scale_embedding (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Scale embeddings by diving by sqrt(d_model).
+        use_cache (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Whether or not the model should return the last key/values attentions (not used by all models)
+
+    Example::
+
+        >>> from transformers import MBartModel, MBartConfig
 
+        >>> # Initializing a MBART facebook/mbart-large-cc25 style configuration
+        >>> configuration = MBartConfig()
+
+        >>> # Initializing a model from the facebook/mbart-large-cc25 style configuration
+        >>> model = MBartModel(configuration)
+
+        >>> # Accessing the model configuration
+        >>> configuration = model.config
+    """
     model_type = "mbart"
     keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        vocab_size=50265,
+        max_position_embeddings=1024,
+        encoder_layers=12,
+        encoder_ffn_dim=4096,
+        encoder_attention_heads=16,
+        decoder_layers=12,
+        decoder_ffn_dim=4096,
+        decoder_attention_heads=16,
+        encoder_layerdrop=0.0,
+        decoder_layerdrop=0.0,
+        use_cache=True,
+        is_encoder_decoder=True,
+        activation_function="gelu",
+        d_model=1024,
+        dropout=0.1,
+        attention_dropout=0.0,
+        activation_dropout=0.0,
+        init_std=0.02,
+        classifier_dropout=0.0,
+        scale_embedding=False,
+        gradient_checkpointing=False,
+        pad_token_id=1,
+        bos_token_id=0,
+        eos_token_id=2,
+        **kwargs
+    ):
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            is_encoder_decoder=is_encoder_decoder,
+            **kwargs,
+        )
+
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.d_model = d_model
+        self.encoder_ffn_dim = encoder_ffn_dim
+        self.encoder_layers = encoder_layers
+        self.encoder_attention_heads = encoder_attention_heads
+        self.decoder_ffn_dim = decoder_ffn_dim
+        self.decoder_layers = decoder_layers
+        self.decoder_attention_heads = decoder_attention_heads
+        self.dropout = dropout
+        self.attention_dropout = attention_dropout
+        self.activation_dropout = activation_dropout
+        self.activation_function = activation_function
+        self.init_std = init_std
+        self.encoder_layerdrop = encoder_layerdrop
+        self.decoder_layerdrop = decoder_layerdrop
+        self.classifier_dropout = classifier_dropout
+        self.use_cache = use_cache
+        self.num_hidden_layers = encoder_layers
+        self.gradient_checkpointing = gradient_checkpointing
+        self.scale_embedding = scale_embedding  # scale factor will be sqrt(d_model) if True
+
+        # IMPORTANT
+        # DELETE ALL OF THE FOLLOWING LINES AS SOON AS TF IS READY
+        self.extra_pos_embeddings = 2
+        self.normalize_before = True
+        self.add_final_layer_norm = True
+        self.do_blenderbot_90_layernorm = False
+        self.normalize_embedding = True
+        self.static_position_embeddings = False
+        self.add_bias_logits = False
+        self.force_bos_token_to_be_generated = False
+
+    @property
+    def num_attention_heads(self) -> int:
+        return self.encoder_attention_heads
+
+    @property
+    def hidden_size(self) -> int:
+        return self.d_model

src/transformers/models/mbart/modeling_mbart.py

-# Copyright 2020 The Facebook AI Research Team Authors and The HuggingFace Inc. team.
+# coding=utf-8
+# Copyright 2021, The Facebook AI Research Team and The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -11,60 +12,1457 @@
 # 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 MBART model. """
 
-from ..bart.modeling_bart import BartForConditionalGeneration, BartModel
+
+import math
+import random
+from typing import Optional, Tuple
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+from torch.nn import CrossEntropyLoss
+
+from ...activations import ACT2FN
+from ...file_utils import (
+    add_code_sample_docstrings,
+    add_end_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    replace_return_docstrings,
+)
+from ...modeling_outputs import (
+    BaseModelOutput,
+    BaseModelOutputWithPastAndCrossAttentions,
+    Seq2SeqLMOutput,
+    Seq2SeqModelOutput,
+    Seq2SeqQuestionAnsweringModelOutput,
+    Seq2SeqSequenceClassifierOutput,
+)
+from ...modeling_utils import PreTrainedModel
+from ...utils import logging
 from .configuration_mbart import MBartConfig
 
 
+logger = logging.get_logger(__name__)
+
 _CONFIG_FOR_DOC = "MBartConfig"
 _TOKENIZER_FOR_DOC = "MBartTokenizer"
 
+
 MBART_PRETRAINED_MODEL_ARCHIVE_LIST = [
     "facebook/mbart-large-cc25",
-    "facebook/mbart-large-en-ro",
-    # See all multilingual BART models at https://huggingface.co/models?filter=mbart
+    # See all MBART models at https://huggingface.co/models?filter=mbart
 ]
 
 
-class MBartModel(BartModel):
-    r"""
-    This class overrides :class:`~transformers.BartModel`. Please check the superclass for the appropriate
-    documentation alongside usage examples.
+def shift_tokens_right(input_ids: torch.Tensor, pad_token_id: int):
+    """
+    Shift input ids one token to the right, and wrap the last non pad token (the <LID> token) Note that MBart does not
+    have a single `decoder_start_token_id` in contrast to other Bart-like models.
     """
+    prev_output_tokens = input_ids.clone()
 
-    config_class = MBartConfig
-    _keys_to_ignore_on_load_missing = [
-        "encoder.embed_positions.weight",
-        "decoder.embed_positions.weight",
-    ]
-    _keys_to_ignore_on_save = [
-        "encoder.embed_positions.weight",
-        "decoder.embed_positions.weight",
-    ]
+    assert pad_token_id is not None, "self.model.config.pad_token_id has to be defined."
+    # replace possible -100 values in labels by `pad_token_id`
+    prev_output_tokens.masked_fill_(prev_output_tokens == -100, pad_token_id)
+
+    index_of_eos = (prev_output_tokens.ne(pad_token_id).sum(dim=1) - 1).unsqueeze(-1)
+    decoder_start_tokens = prev_output_tokens.gather(1, index_of_eos).squeeze()
+    prev_output_tokens[:, 1:] = prev_output_tokens[:, :-1].clone()
+    prev_output_tokens[:, 0] = decoder_start_tokens
+
+    return prev_output_tokens
+
+
+# Copied from transformers.models.bart.modeling_bart._make_causal_mask
+def _make_causal_mask(input_ids_shape: torch.Size, dtype: torch.dtype, past_key_values_length: int = 0):
+    """
+    Make causal mask used for bi-directional self-attention.
+    """
+    bsz, tgt_len = input_ids_shape
+    mask = torch.full((tgt_len, tgt_len), float("-inf"))
+    mask_cond = torch.arange(mask.size(-1))
+    mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)
+    mask = mask.to(dtype)
+
+    if past_key_values_length > 0:
+        mask = torch.cat([torch.zeros(tgt_len, past_key_values_length, dtype=dtype), mask], dim=-1)
+    return mask[None, None, :, :].expand(bsz, 1, tgt_len, tgt_len + past_key_values_length)
 
 
-class MBartForConditionalGeneration(BartForConditionalGeneration):
-    r"""
-    This class overrides :class:`~transformers.BartForConditionalGeneration`. Please check the superclass for the
-    appropriate documentation alongside usage examples.
-
-    Examples::
-        >>> from transformers import MBartForConditionalGeneration, MBartTokenizer
-        >>> model = MBartForConditionalGeneration.from_pretrained("facebook/mbart-large-en-ro")
-        >>> tokenizer = MBartTokenizer.from_pretrained("facebook/mbart-large-en-ro")
-        >>> article = "UN Chief Says There Is No Military Solution in Syria"
-        >>> batch = tokenizer.prepare_seq2seq_batch(src_texts=[article], return_tensors="pt")
-        >>> translated_tokens = model.generate(**batch)
-        >>> translation = tokenizer.batch_decode(translated_tokens, skip_special_tokens=True)[0]
-        >>> assert translation == "Şeful ONU declară că nu există o soluţie militară în Siria"
+# Copied from transformers.models.bart.modeling_bart._expand_mask
+def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
     """
-    model_type = "mbart"
+    Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
+    """
+    bsz, src_len = mask.size()
+    tgt_len = tgt_len if tgt_len is not None else src_len
+
+    expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)
+
+    inverted_mask = 1.0 - expanded_mask
+
+    return inverted_mask.masked_fill(inverted_mask.bool(), torch.finfo(dtype).min)
+
+
+# Copied from transformers.models.bart.modeling_bart.BartLearnedPositionalEmbedding with Bart->MBart
+class MBartLearnedPositionalEmbedding(nn.Embedding):
+    """
+    This module learns positional embeddings up to a fixed maximum size.
+    """
+
+    def __init__(self, num_embeddings: int, embedding_dim: int, padding_idx: int):
+        assert padding_idx is not None, "`padding_idx` should not be None, but of type int"
+        # MBart is set up so that if padding_idx is specified then offset the embedding ids by 2
+        # and adjust num_embeddings appropriately. Other models dont have this hack
+        self.offset = 2
+        super().__init__(num_embeddings + self.offset, embedding_dim, padding_idx=padding_idx)
+
+    def forward(self, input_ids_shape: torch.Size, past_key_values_length: int = 0):
+        """`input_ids_shape` is expected to be [bsz x seqlen]."""
+        bsz, seq_len = input_ids_shape[:2]
+        positions = torch.arange(
+            past_key_values_length, past_key_values_length + seq_len, dtype=torch.long, device=self.weight.device
+        )
+        return super().forward(positions + self.offset)
+
+
+# Copied from transformers.models.bart.modeling_bart.BartAttention with Bart->MBart
+class MBartAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(
+        self,
+        embed_dim: int,
+        num_heads: int,
+        dropout: float = 0.0,
+        is_decoder: bool = False,
+        bias: bool = True,
+    ):
+        super().__init__()
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.dropout = dropout
+        self.head_dim = embed_dim // num_heads
+        assert (
+            self.head_dim * num_heads == self.embed_dim
+        ), f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`: {num_heads})."
+        self.scaling = self.head_dim ** -0.5
+        self.is_decoder = is_decoder
+
+        self.k_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+        self.v_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+        self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+        self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+
+    def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
+        return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        key_value_states: Optional[torch.Tensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        output_attentions: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        """Input shape: Batch x Time x Channel"""
+
+        # if key_value_states are provided this layer is used as a cross-attention layer
+        # for the decoder
+        is_cross_attention = key_value_states is not None
+        bsz, tgt_len, embed_dim = hidden_states.size()
+
+        # get query proj
+        query_states = self.q_proj(hidden_states) * self.scaling
+        # get key, value proj
+        if is_cross_attention and past_key_value is not None:
+            # reuse k,v, cross_attentions
+            key_states = past_key_value[0]
+            value_states = past_key_value[1]
+        elif is_cross_attention:
+            # cross_attentions
+            key_states = self._shape(self.k_proj(key_value_states), -1, bsz)
+            value_states = self._shape(self.v_proj(key_value_states), -1, bsz)
+        elif past_key_value is not None:
+            # reuse k, v, self_attention
+            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
+            key_states = torch.cat([past_key_value[0], key_states], dim=2)
+            value_states = torch.cat([past_key_value[1], value_states], dim=2)
+        else:
+            # self_attention
+            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
+
+        if self.is_decoder:
+            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
+            # Further calls to cross_attention layer can then reuse all cross-attention
+            # key/value_states (first "if" case)
+            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
+            # all previous decoder key/value_states. Further calls to uni-directional self-attention
+            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
+            # if encoder bi-directional self-attention `past_key_value` is always `None`
+            past_key_value = (key_states, value_states)
+
+        proj_shape = (bsz * self.num_heads, -1, self.head_dim)
+        query_states = self._shape(query_states, tgt_len, bsz).view(*proj_shape)
+        key_states = key_states.view(*proj_shape)
+        value_states = value_states.view(*proj_shape)
+
+        src_len = key_states.size(1)
+        attn_weights = torch.bmm(query_states, key_states.transpose(1, 2))
+
+        assert attn_weights.size() == (
+            bsz * self.num_heads,
+            tgt_len,
+            src_len,
+        ), f"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is {attn_weights.size()}"
+
+        if attention_mask is not None:
+            assert attention_mask.size() == (
+                bsz,
+                1,
+                tgt_len,
+                src_len,
+            ), f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.size()}"
+            attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + attention_mask
+            attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
+
+        attn_weights = F.softmax(attn_weights, dim=-1)
+
+        if output_attentions:
+            # this operation is a bit akward, but it's required to
+            # make sure that attn_weights keeps its gradient.
+            # In order to do so, attn_weights have to reshaped
+            # twice and have to be reused in the following
+            attn_weights_reshaped = attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
+            attn_weights = attn_weights_reshaped.view(bsz * self.num_heads, tgt_len, src_len)
+        else:
+            attn_weights_reshaped = None
+
+        attn_probs = F.dropout(attn_weights, p=self.dropout, training=self.training)
+
+        attn_output = torch.bmm(attn_probs, value_states)
+
+        assert attn_output.size() == (
+            bsz * self.num_heads,
+            tgt_len,
+            self.head_dim,
+        ), f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is {attn_output.size()}"
+
+        attn_output = (
+            attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
+            .transpose(1, 2)
+            .reshape(bsz, tgt_len, embed_dim)
+        )
+
+        attn_output = self.out_proj(attn_output)
+
+        return attn_output, attn_weights_reshaped, past_key_value
+
+
+class MBartEncoderLayer(nn.Module):
+    def __init__(self, config: MBartConfig):
+        super().__init__()
+        self.embed_dim = config.d_model
+        self.self_attn = MBartAttention(
+            embed_dim=self.embed_dim,
+            num_heads=config.encoder_attention_heads,
+            dropout=config.attention_dropout,
+        )
+        self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
+        self.dropout = config.dropout
+        self.activation_fn = ACT2FN[config.activation_function]
+        self.activation_dropout = config.activation_dropout
+        self.fc1 = nn.Linear(self.embed_dim, config.encoder_ffn_dim)
+        self.fc2 = nn.Linear(config.encoder_ffn_dim, self.embed_dim)
+        self.final_layer_norm = nn.LayerNorm(self.embed_dim)
+
+    def forward(self, hidden_states: torch.Tensor, attention_mask: torch.Tensor, output_attentions: bool = False):
+        """
+        Args:
+            hidden_states (:obj:`torch.FloatTensor`): input to the layer of shape `(seq_len, batch, embed_dim)`
+            attention_mask (:obj:`torch.FloatTensor`): attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            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.
+        """
+        residual = hidden_states
+        hidden_states = self.self_attn_layer_norm(hidden_states)
+        hidden_states, attn_weights, _ = self.self_attn(
+            hidden_states=hidden_states, attention_mask=attention_mask, output_attentions=output_attentions
+        )
+        hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
+        hidden_states = residual + hidden_states
+
+        residual = hidden_states
+        hidden_states = self.final_layer_norm(hidden_states)
+        hidden_states = self.activation_fn(self.fc1(hidden_states))
+        hidden_states = F.dropout(hidden_states, p=self.activation_dropout, training=self.training)
+        hidden_states = self.fc2(hidden_states)
+        hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
+        hidden_states = residual + hidden_states
+
+        if torch.isinf(hidden_states).any() or torch.isnan(hidden_states).any():
+            clamp_value = torch.finfo(hidden_states.dtype).max - 1000
+            hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (attn_weights,)
+
+        return outputs
+
+
+class MBartDecoderLayer(nn.Module):
+    def __init__(self, config: MBartConfig):
+        super().__init__()
+        self.embed_dim = config.d_model
+
+        self.self_attn = MBartAttention(
+            embed_dim=self.embed_dim,
+            num_heads=config.decoder_attention_heads,
+            dropout=config.attention_dropout,
+            is_decoder=True,
+        )
+        self.dropout = config.dropout
+        self.activation_fn = ACT2FN[config.activation_function]
+        self.activation_dropout = config.activation_dropout
+
+        self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
+        self.encoder_attn = MBartAttention(
+            self.embed_dim,
+            config.decoder_attention_heads,
+            dropout=config.attention_dropout,
+            is_decoder=True,
+        )
+        self.encoder_attn_layer_norm = nn.LayerNorm(self.embed_dim)
+        self.fc1 = nn.Linear(self.embed_dim, config.decoder_ffn_dim)
+        self.fc2 = nn.Linear(config.decoder_ffn_dim, self.embed_dim)
+        self.final_layer_norm = nn.LayerNorm(self.embed_dim)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = True,
+    ):
+        """
+        Args:
+            hidden_states (:obj:`torch.FloatTensor`): input to the layer of shape `(seq_len, batch, embed_dim)`
+            attention_mask (:obj:`torch.FloatTensor`): attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            encoder_hidden_states (:obj:`torch.FloatTensor`): cross attention input to the layer of shape `(seq_len, batch, embed_dim)`
+            encoder_attention_mask (:obj:`torch.FloatTensor`): encoder attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            past_key_value (:obj:`Tuple(torch.FloatTensor)`): cached past key and value projection states
+            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.
+        """
+        residual = hidden_states
+        hidden_states = self.self_attn_layer_norm(hidden_states)
+
+        # Self Attention
+        # decoder uni-directional self-attention cached key/values tuple is at positions 1,2
+        self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None
+        # add present self-attn cache to positions 1,2 of present_key_value tuple
+        hidden_states, self_attn_weights, present_key_value = self.self_attn(
+            hidden_states=hidden_states,
+            past_key_value=self_attn_past_key_value,
+            attention_mask=attention_mask,
+            output_attentions=output_attentions,
+        )
+        hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
+        hidden_states = residual + hidden_states
+
+        # Cross-Attention Block
+        cross_attn_present_key_value = None
+        cross_attn_weights = None
+        if encoder_hidden_states is not None:
+            residual = hidden_states
+            hidden_states = self.encoder_attn_layer_norm(hidden_states)
+
+            # cross_attn cached key/values tuple is at positions 3,4 of present_key_value tuple
+            cross_attn_past_key_value = past_key_value[-2:] if past_key_value is not None else None
+            hidden_states, cross_attn_weights, cross_attn_present_key_value = self.encoder_attn(
+                hidden_states=hidden_states,
+                key_value_states=encoder_hidden_states,
+                attention_mask=encoder_attention_mask,
+                past_key_value=cross_attn_past_key_value,
+                output_attentions=output_attentions,
+            )
+            hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
+            hidden_states = residual + hidden_states
+
+            # add cross-attn to positions 3,4 of present_key_value tuple
+            present_key_value = present_key_value + cross_attn_present_key_value
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.final_layer_norm(hidden_states)
+        hidden_states = self.activation_fn(self.fc1(hidden_states))
+        hidden_states = F.dropout(hidden_states, p=self.activation_dropout, training=self.training)
+        hidden_states = self.fc2(hidden_states)
+        hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights, cross_attn_weights)
+
+        if use_cache:
+            outputs += (present_key_value,)
+
+        return outputs
+
+
+# Copied from transformers.models.bart.modeling_bart.BartClassificationHead with Bart->MBart
+class MBartClassificationHead(nn.Module):
+    """Head for sentence-level classification tasks."""
+
+    def __init__(
+        self,
+        input_dim: int,
+        inner_dim: int,
+        num_classes: int,
+        pooler_dropout: float,
+    ):
+        super().__init__()
+        self.dense = nn.Linear(input_dim, inner_dim)
+        self.dropout = nn.Dropout(p=pooler_dropout)
+        self.out_proj = nn.Linear(inner_dim, num_classes)
+
+    def forward(self, hidden_states: torch.Tensor):
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.dense(hidden_states)
+        hidden_states = torch.tanh(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.out_proj(hidden_states)
+        return hidden_states
+
+
+class MBartPreTrainedModel(PreTrainedModel):
     config_class = MBartConfig
+    base_model_prefix = "model"
+
+    def _init_weights(self, module):
+        std = self.config.init_std
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+    @property
+    def dummy_inputs(self):
+        pad_token = self.config.pad_token_id
+        input_ids = torch.tensor([[0, 6, 10, 4, 2], [0, 8, 12, 2, pad_token]], device=self.device)
+        dummy_inputs = {
+            "attention_mask": input_ids.ne(pad_token),
+            "input_ids": input_ids,
+        }
+        return dummy_inputs
+
+
+MBART_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.
+
+    Parameters:
+        config (:class:`~transformers.MBartConfig`):
+            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.
+"""
+
+MBART_GENERATION_EXAMPLE = r"""
+    Summarization example::
+
+        >>> from transformers import MBartTokenizer, MBartForConditionalGeneration, MBartConfig
+
+        >>> model = MBartForConditionalGeneration.from_pretrained('facebook/mbart-large-cc25')
+        >>> tokenizer = MBartTokenizer.from_pretrained('facebook/mbart-large-cc25')
+
+        >>> ARTICLE_TO_SUMMARIZE = "Meine Freunde sind cool, aber sie essen zu viel Kuchen."
+        >>> inputs = tokenizer([ARTICLE_TO_SUMMARIZE], max_length=1024, return_tensors='pt')
+
+        >>> # Generate Summary
+        >>> summary_ids = model.generate(inputs['input_ids'], num_beams=4, max_length=5, early_stopping=True)
+        >>> print([tokenizer.decode(g, skip_special_tokens=True, clean_up_tokenization_spaces=False) for g in summary_ids])
+
+    Mask filling example::
+
+        >>> from transformers import MBartTokenizer, MBartForConditionalGeneration
+        >>> tokenizer = MBartTokenizer.from_pretrained('facebook/mbart-large-cc25')
+        >>> # de_DE is the language symbol id <LID> for German
+        >>> TXT = "</s> Meine Freunde sind <mask> nett aber sie essen zu viel Kuchen. </s> de_DE"
+
+        >>> model = MBartForConditionalGeneration.from_pretrained('facebook/mbart-large-cc25')
+        >>> input_ids = tokenizer([TXT], add_special_tokens=False, return_tensors='pt')['input_ids']
+        >>> logits = model(input_ids).logits
+
+        >>> masked_index = (input_ids[0] == tokenizer.mask_token_id).nonzero().item()
+        >>> probs = logits[0, masked_index].softmax(dim=0)
+        >>> values, predictions = probs.topk(5)
+
+        >>> tokenizer.decode(predictions).split()
+"""
+
+MBART_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+
+            Indices can be obtained using :class:`~transformers.MBartTokenizer`. See
+            :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__` for
+            details.
+
+            `What are input IDs? <../glossary.html#input-ids>`__
+        attention_mask (:obj:`torch.Tensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Mask to avoid performing attention on padding token indices. Mask values selected in ``[0, 1]``:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            `What are attention masks? <../glossary.html#attention-mask>`__
+        decoder_input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, target_sequence_length)`, `optional`):
+            Provide for translation and summarization training. By default, the model will create this tensor by
+            shifting the :obj:`input_ids` to the right, following the paper.
+        decoder_input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, target_sequence_length)`, `optional`):
+            Indices of decoder input sequence tokens in the vocabulary.
+
+            Indices can be obtained using :class:`~transformers.MBartTokenizer`. See
+            :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__` for
+            details.
+
+            `What are input IDs? <../glossary.html#input-ids>`__
+
+            MBart uses a specific language id token as the starting token for :obj:`decoder_input_ids` generation that
+            varies according to source and target language, *e.g.* 25004 for `en_XX`, and 25003 for `de_DE`. If
+            :obj:`past_key_values` is used, optionally only the last :obj:`decoder_input_ids` have to be input (see
+            :obj:`past_key_values`).
+
+            For translation and summarization training, :obj:`decoder_input_ids` should be provided. If no
+            :obj:`decoder_input_ids` is provided, the model will create this tensor by shifting the :obj:`input_ids` to
+            the right for denoising pre-training following the paper.
+        decoder_attention_mask (:obj:`torch.LongTensor` of shape :obj:`(batch_size, target_sequence_length)`, `optional`):
+            Default behavior: generate a tensor that ignores pad tokens in :obj:`decoder_input_ids`. Causal mask will
+            also be used by default.
+
+            If you want to change padding behavior, you should read :func:`modeling_mbart._prepare_decoder_inputs` and
+            modify to your needs. See diagram 1 in `the paper <https://arxiv.org/abs/1910.13461>`__ for more
+            information on the default strategy.
+        encoder_outputs (:obj:`tuple(tuple(torch.FloatTensor)`, `optional`):
+            Tuple consists of (:obj:`last_hidden_state`, `optional`: :obj:`hidden_states`, `optional`:
+            :obj:`attentions`) :obj:`last_hidden_state` of shape :obj:`(batch_size, sequence_length, hidden_size)`,
+            `optional`) is a sequence of hidden-states at the output of the last layer of the encoder. Used in the
+            cross-attention of the decoder.
+        past_key_values (:obj:`Tuple[Tuple[torch.Tensor]]` of length :obj:`config.n_layers` with each tuple having 2 tuples each of which has 2 tensors of shape :obj:`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
+            Contains precomputed key and value hidden-states of the attention blocks. Can be used to speed up decoding.
+
+            If :obj:`past_key_values` are used, the user can optionally input only the last :obj:`decoder_input_ids`
+            (those that don't have their past key value states given to this model) of shape :obj:`(batch_size, 1)`
+            instead of all :obj:`decoder_input_ids`` of shape :obj:`(batch_size, sequence_length)`.
+        inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, 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.
+        decoder_inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, target_sequence_length, hidden_size)`, `optional`):
+            Optionally, instead of passing :obj:`decoder_input_ids` you can choose to directly pass an embedded
+            representation. If :obj:`past_key_values` is used, optionally only the last :obj:`decoder_inputs_embeds`
+            have to be input (see :obj:`past_key_values`). This is useful if you want more control over how to convert
+            :obj:`decoder_input_ids` indices into associated vectors than the model's internal embedding lookup matrix.
+
+            If :obj:`decoder_input_ids` and :obj:`decoder_inputs_embeds` are both unset, :obj:`decoder_inputs_embeds`
+            takes the value of :obj:`inputs_embeds`.
+        use_cache (:obj:`bool`, `optional`):
+            If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up
+            decoding (see :obj:`past_key_values`).
+        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.
+"""
+
+
+class MBartEncoder(MBartPreTrainedModel):
+    """
+    Transformer encoder consisting of *config.encoder_layers* self attention layers. Each layer is a
+    :class:`MBartEncoderLayer`.
+
+    Args:
+        config: MBartConfig
+        embed_tokens (torch.nn.Embedding): output embedding
+    """
+
+    def __init__(self, config: MBartConfig, embed_tokens: Optional[nn.Embedding] = None):
+        super().__init__(config)
+
+        self.dropout = config.dropout
+        self.layerdrop = config.encoder_layerdrop
+
+        embed_dim = config.d_model
+        self.padding_idx = config.pad_token_id
+        self.max_source_positions = config.max_position_embeddings
+        self.embed_scale = math.sqrt(embed_dim) if config.scale_embedding else 1.0
+
+        if embed_tokens is not None:
+            self.embed_tokens = embed_tokens
+        else:
+            self.embed_tokens = nn.Embedding(config.vocab_size, embed_dim, self.padding_idx)
+
+        self.embed_positions = MBartLearnedPositionalEmbedding(
+            config.max_position_embeddings,
+            embed_dim,
+            self.padding_idx,
+        )
+        self.layers = nn.ModuleList([MBartEncoderLayer(config) for _ in range(config.encoder_layers)])
+        self.layernorm_embedding = nn.LayerNorm(embed_dim)
+        self.layer_norm = nn.LayerNorm(config.d_model)
+
+        self.init_weights()
+
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        inputs_embeds=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        Args:
+            input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`):
+                Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you
+                provide it.
+
+                Indices can be obtained using :class:`~transformers.MBartTokenizer`. See
+                :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__`
+                for details.
+
+                `What are input IDs? <../glossary.html#input-ids>`__
+            attention_mask (:obj:`torch.Tensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+                Mask to avoid performing attention on padding token indices. Mask values selected in ``[0, 1]``:
+
+                - 1 for tokens that are **not masked**,
+                - 0 for tokens that are **masked**.
+
+                `What are attention masks? <../glossary.html#attention-mask>`__
+            inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, 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.
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # retrieve input_ids and inputs_embeds
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+            input_ids = input_ids.view(-1, input_shape[-1])
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids) * self.embed_scale
+
+        embed_pos = self.embed_positions(input_shape)
+
+        hidden_states = inputs_embeds + embed_pos
+        hidden_states = self.layernorm_embedding(hidden_states)
+        hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
+
+        # expand attention_mask
+        if attention_mask is not None:
+            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+            attention_mask = _expand_mask(attention_mask, inputs_embeds.dtype)
+
+        encoder_states = () if output_hidden_states else None
+        all_attentions = () if output_attentions else None
+        for encoder_layer in self.layers:
+            if output_hidden_states:
+                encoder_states = encoder_states + (hidden_states,)
+            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
+            dropout_probability = random.uniform(0, 1)
+            if self.training and (dropout_probability < self.layerdrop):  # skip the layer
+                layer_outputs = (None, None)
+            else:
+                if getattr(self.config, "gradient_checkpointing", False):
+
+                    def create_custom_forward(module):
+                        def custom_forward(*inputs):
+                            return module(*inputs, output_attentions)
+
+                        return custom_forward
+
+                    layer_outputs = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(encoder_layer),
+                        hidden_states,
+                        attention_mask,
+                    )
+                else:
+                    layer_outputs = encoder_layer(hidden_states, attention_mask, output_attentions=output_attentions)
+
+                hidden_states = layer_outputs[0]
+
+            if output_attentions:
+                all_attentions = all_attentions + (layer_outputs[1],)
+
+        hidden_states = self.layer_norm(hidden_states)
+
+        if output_hidden_states:
+            encoder_states = encoder_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None)
+        return BaseModelOutput(
+            last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions
+        )
+
+
+class MBartDecoder(MBartPreTrainedModel):
+    """
+    Transformer decoder consisting of *config.decoder_layers* layers. Each layer is a :class:`MBartDecoderLayer`
+
+    Args:
+        config: MBartConfig
+        embed_tokens (torch.nn.Embedding): output embedding
+    """
+
+    def __init__(self, config: MBartConfig, embed_tokens: Optional[nn.Embedding] = None):
+        super().__init__(config)
+        self.dropout = config.dropout
+        self.layerdrop = config.decoder_layerdrop
+        self.padding_idx = config.pad_token_id
+        self.max_target_positions = config.max_position_embeddings
+        self.embed_scale = math.sqrt(config.d_model) if config.scale_embedding else 1.0
+
+        if embed_tokens is not None:
+            self.embed_tokens = embed_tokens
+        else:
+            self.embed_tokens = nn.Embedding(config.vocab_size, config.d_model, self.padding_idx)
+
+        self.embed_positions = MBartLearnedPositionalEmbedding(
+            config.max_position_embeddings,
+            config.d_model,
+            self.padding_idx,
+        )
+        self.layers = nn.ModuleList([MBartDecoderLayer(config) for _ in range(config.decoder_layers)])
+        self.layernorm_embedding = nn.LayerNorm(config.d_model)
+        self.layer_norm = nn.LayerNorm(config.d_model)
+
+        self.init_weights()
+
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_values=None,
+        inputs_embeds=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        Args:
+            input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`):
+                Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you
+                provide it.
+
+                Indices can be obtained using :class:`~transformers.MBartTokenizer`. See
+                :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__`
+                for details.
+
+                `What are input IDs? <../glossary.html#input-ids>`__
+            attention_mask (:obj:`torch.Tensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+                Mask to avoid performing attention on padding token indices. Mask values selected in ``[0, 1]``:
+
+                - 1 for tokens that are **not masked**,
+                - 0 for tokens that are **masked**.
+
+                `What are attention masks? <../glossary.html#attention-mask>`__
+            encoder_hidden_states (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, encoder_sequence_length, hidden_size)`, `optional`):
+                Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention
+                of the decoder.
+            encoder_attention_mask (:obj:`torch.LongTensor` of shape :obj:`(batch_size, encoder_sequence_length)`, `optional`):
+                Mask to avoid performing cross-attention on padding tokens indices of encoder input_ids. Mask values
+                selected in ``[0, 1]``:
+
+                - 1 for tokens that are **not masked**,
+                - 0 for tokens that are **masked**.
+
+                `What are attention masks? <../glossary.html#attention-mask>`__
+            past_key_values (:obj:`Tuple[Tuple[torch.Tensor]]` of length :obj:`config.n_layers` with each tuple having 2 tuples each of which has 2 tensors of shape :obj:`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
+                Contains precomputed key and value hidden-states of the attention blocks. Can be used to speed up
+                decoding.
+
+                If :obj:`past_key_values` are used, the user can optionally input only the last
+                :obj:`decoder_input_ids` (those that don't have their past key value states given to this model) of
+                shape :obj:`(batch_size, 1)` instead of all :obj:`decoder_input_ids`` of shape :obj:`(batch_size,
+                sequence_length)`.
+            inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, 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.
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # retrieve input_ids and inputs_embeds
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+            input_ids = input_ids.view(-1, input_shape[-1])
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+        else:
+            raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
+
+        # past_key_values_length
+        past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids) * self.embed_scale
+
+        # create causal mask
+        # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+        combined_attention_mask = None
+        if input_shape[-1] > 1:
+            combined_attention_mask = _make_causal_mask(
+                input_shape, inputs_embeds.dtype, past_key_values_length=past_key_values_length
+            ).to(self.device)
+
+        if attention_mask is not None and combined_attention_mask is not None:
+            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+            combined_attention_mask = combined_attention_mask + _expand_mask(
+                attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]
+            )
+
+        # expand encoder attention mask
+        if encoder_hidden_states is not None and encoder_attention_mask is not None:
+            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+            encoder_attention_mask = _expand_mask(encoder_attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1])
+
+        # embed positions
+        positions = self.embed_positions(input_shape, past_key_values_length)
+
+        hidden_states = inputs_embeds + positions
+        hidden_states = self.layernorm_embedding(hidden_states)
+
+        hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        all_cross_attentions = () if output_attentions else None
+        next_decoder_cache = () if use_cache else None
+        for idx, decoder_layer in enumerate(self.layers):
+            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+            dropout_probability = random.uniform(0, 1)
+            if self.training and (dropout_probability < self.layerdrop):
+                continue
+
+            past_key_value = past_key_values[idx] if past_key_values is not None else None
+
+            if getattr(self.config, "gradient_checkpointing", False):
+                if use_cache:
+                    raise ValueError(
+                        "When using `gradient_checkpointing, make sure that `use_cache=False` and `config.use_cache=False`."
+                    )
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        # None for past_key_value
+                        return module(*inputs, output_attentions, use_cache)
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(decoder_layer),
+                    hidden_states,
+                    combined_attention_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    None,
+                )
+            else:
+
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask=combined_attention_mask,
+                    encoder_hidden_states=encoder_hidden_states,
+                    encoder_attention_mask=encoder_attention_mask,
+                    past_key_value=past_key_value,
+                    output_attentions=output_attentions,
+                    use_cache=use_cache,
+                )
+            hidden_states = layer_outputs[0]
+
+            if use_cache:
+                next_decoder_cache += (layer_outputs[3 if output_attentions else 1],)
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+                all_cross_attentions += (layer_outputs[2],)
+
+        hidden_states = self.layer_norm(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        next_cache = next_decoder_cache if use_cache else None
+        if not return_dict:
+            return tuple(
+                v
+                for v in [hidden_states, next_cache, all_hidden_states, all_self_attns, all_cross_attentions]
+                if v is not None
+            )
+        return BaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+            cross_attentions=all_cross_attentions,
+        )
+
+
+@add_start_docstrings(
+    "The bare MBART Model outputting raw hidden-states without any specific head on top.",
+    MBART_START_DOCSTRING,
+)
+class MBartModel(MBartPreTrainedModel):
+    def __init__(self, config: MBartConfig):
+        super().__init__(config)
+
+        padding_idx, vocab_size = config.pad_token_id, config.vocab_size
+        self.shared = nn.Embedding(vocab_size, config.d_model, padding_idx)
+
+        self.encoder = MBartEncoder(config, self.shared)
+        self.decoder = MBartDecoder(config, self.shared)
+
+        self.init_weights()
+
+    def get_input_embeddings(self):
+        return self.shared
+
+    def set_input_embeddings(self, value):
+        self.shared = value
+        self.encoder.embed_tokens = self.shared
+        self.decoder.embed_tokens = self.shared
+
+    def get_encoder(self):
+        return self.encoder
+
+    def get_decoder(self):
+        return self.decoder
+
+    @add_start_docstrings_to_model_forward(MBART_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="facebook/mbart-large-cc25",
+        output_type=Seq2SeqModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        decoder_input_ids=None,
+        decoder_attention_mask=None,
+        encoder_outputs=None,
+        past_key_values=None,
+        inputs_embeds=None,
+        decoder_inputs_embeds=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # different to other models, MBart automatically creates decoder_input_ids from
+        # input_ids if no decoder_input_ids are provided
+        if decoder_input_ids is None and decoder_inputs_embeds is None:
+            decoder_input_ids = shift_tokens_right(input_ids, self.config.pad_token_id)
+
+        if encoder_outputs is None:
+            encoder_outputs = self.encoder(
+                input_ids=input_ids,
+                attention_mask=attention_mask,
+                inputs_embeds=inputs_embeds,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+            )
+        # If the user passed a tuple for encoder_outputs, we wrap it in a BaseModelOutput when return_dict=True
+        elif return_dict and not isinstance(encoder_outputs, BaseModelOutput):
+            encoder_outputs = BaseModelOutput(
+                last_hidden_state=encoder_outputs[0],
+                hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None,
+                attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None,
+            )
+
+        # decoder outputs consists of (dec_features, past_key_value, dec_hidden, dec_attn)
+        decoder_outputs = self.decoder(
+            input_ids=decoder_input_ids,
+            attention_mask=decoder_attention_mask,
+            encoder_hidden_states=encoder_outputs[0],
+            encoder_attention_mask=attention_mask,
+            past_key_values=past_key_values,
+            inputs_embeds=decoder_inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        if not return_dict:
+            return decoder_outputs + encoder_outputs
+
+        return Seq2SeqModelOutput(
+            last_hidden_state=decoder_outputs.last_hidden_state,
+            past_key_values=decoder_outputs.past_key_values,
+            decoder_hidden_states=decoder_outputs.hidden_states,
+            decoder_attentions=decoder_outputs.attentions,
+            cross_attentions=decoder_outputs.cross_attentions,
+            encoder_last_hidden_state=encoder_outputs.last_hidden_state,
+            encoder_hidden_states=encoder_outputs.hidden_states,
+            encoder_attentions=encoder_outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    "The MBART Model with a language modeling head. Can be used for summarization.", MBART_START_DOCSTRING
+)
+class MBartForConditionalGeneration(MBartPreTrainedModel):
+    base_model_prefix = "model"
     _keys_to_ignore_on_load_missing = [
-        "model.encoder.embed_positions.weight",
-        "model.decoder.embed_positions.weight",
-    ]
-    _keys_to_ignore_on_save = [
-        "model.encoder.embed_positions.weight",
-        "model.decoder.embed_positions.weight",
+        r"final_logits_bias",
+        r"encoder\.version",
+        r"decoder\.version",
+        r"lm_head\.weight",
     ]
+
+    def __init__(self, config: MBartConfig):
+        super().__init__(config)
+        self.model = MBartModel(config)
+        self.register_buffer("final_logits_bias", torch.zeros((1, self.model.shared.num_embeddings)))
+        self.lm_head = nn.Linear(config.d_model, self.model.shared.num_embeddings, bias=False)
+
+        self.init_weights()
+
+    def get_encoder(self):
+        return self.model.get_encoder()
+
+    def get_decoder(self):
+        return self.model.get_decoder()
+
+    def resize_token_embeddings(self, new_num_tokens: int) -> nn.Embedding:
+        new_embeddings = super().resize_token_embeddings(new_num_tokens)
+        self._resize_final_logits_bias(new_num_tokens)
+        return new_embeddings
+
+    def _resize_final_logits_bias(self, new_num_tokens: int) -> None:
+        old_num_tokens = self.final_logits_bias.shape[-1]
+        if new_num_tokens <= old_num_tokens:
+            new_bias = self.final_logits_bias[:, :new_num_tokens]
+        else:
+            extra_bias = torch.zeros((1, new_num_tokens - old_num_tokens), device=self.final_logits_bias.device)
+            new_bias = torch.cat([self.final_logits_bias, extra_bias], dim=1)
+        self.register_buffer("final_logits_bias", new_bias)
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    @add_start_docstrings_to_model_forward(MBART_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=Seq2SeqLMOutput, config_class=_CONFIG_FOR_DOC)
+    @add_end_docstrings(MBART_GENERATION_EXAMPLE)
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        decoder_input_ids=None,
+        decoder_attention_mask=None,
+        encoder_outputs=None,
+        past_key_values=None,
+        inputs_embeds=None,
+        decoder_inputs_embeds=None,
+        labels=None,
+        use_cache=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 either be in ``[0, ...,
+            config.vocab_size]`` or -100 (see ``input_ids`` docstring). Tokens with indices set to ``-100`` are ignored
+            (masked), the loss is only computed for the tokens with labels in ``[0, ..., config.vocab_size]``.
+
+        Returns:
+
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if labels is not None:
+            if decoder_input_ids is None:
+                decoder_input_ids = shift_tokens_right(labels, self.config.pad_token_id)
+
+        outputs = self.model(
+            input_ids,
+            attention_mask=attention_mask,
+            decoder_input_ids=decoder_input_ids,
+            encoder_outputs=encoder_outputs,
+            decoder_attention_mask=decoder_attention_mask,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            decoder_inputs_embeds=decoder_inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        lm_logits = self.lm_head(outputs[0]) + self.final_logits_bias
+
+        masked_lm_loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            masked_lm_loss = loss_fct(lm_logits.view(-1, self.config.vocab_size), labels.view(-1))
+
+        if not return_dict:
+            output = (lm_logits,) + outputs[1:]
+            return ((masked_lm_loss,) + output) if masked_lm_loss is not None else output
+
+        return Seq2SeqLMOutput(
+            loss=masked_lm_loss,
+            logits=lm_logits,
+            past_key_values=outputs.past_key_values,
+            decoder_hidden_states=outputs.decoder_hidden_states,
+            decoder_attentions=outputs.decoder_attentions,
+            cross_attentions=outputs.cross_attentions,
+            encoder_last_hidden_state=outputs.encoder_last_hidden_state,
+            encoder_hidden_states=outputs.encoder_hidden_states,
+            encoder_attentions=outputs.encoder_attentions,
+        )
+
+    def prepare_inputs_for_generation(
+        self, decoder_input_ids, past=None, attention_mask=None, use_cache=None, encoder_outputs=None, **kwargs
+    ):
+        # cut decoder_input_ids if past is used
+        if past is not None:
+            decoder_input_ids = decoder_input_ids[:, -1:]
+
+        return {
+            "input_ids": None,  # encoder_outputs is defined. input_ids not needed
+            "encoder_outputs": encoder_outputs,
+            "past_key_values": past,
+            "decoder_input_ids": decoder_input_ids,
+            "attention_mask": attention_mask,
+            "use_cache": use_cache,  # change this to avoid caching (presumably for debugging)
+        }
+
+    def adjust_logits_during_generation(self, logits, cur_len, max_length):
+        if cur_len == max_length - 1 and self.config.eos_token_id is not None:
+            self._force_token_id_to_be_generated(logits, self.config.eos_token_id)
+        return logits
+
+    @staticmethod
+    def _force_token_id_to_be_generated(scores, token_id) -> None:
+        """force one of token_ids to be generated by setting prob of all other tokens to 0 (logprob=-float("inf"))"""
+        scores[:, [x for x in range(scores.shape[1]) if x != token_id]] = -float("inf")
+
+    @staticmethod
+    def _reorder_cache(past, beam_idx):
+        reordered_past = ()
+        for layer_past in past:
+            # cached cross_attention states don't have to be reordered -> they are always the same
+            reordered_past += (
+                tuple(past_state.index_select(0, beam_idx) for past_state in layer_past[:2]) + layer_past[2:],
+            )
+        return reordered_past
+
+
+@add_start_docstrings(
+    """
+    MBart model with a sequence classification/head on top (a linear layer on top of the pooled output) e.g. for GLUE
+    tasks.
+    """,
+    MBART_START_DOCSTRING,
+)
+class MBartForSequenceClassification(MBartPreTrainedModel):
+    def __init__(self, config: MBartConfig, **kwargs):
+        super().__init__(config, **kwargs)
+        self.model = MBartModel(config)
+        self.classification_head = MBartClassificationHead(
+            config.d_model,
+            config.d_model,
+            config.num_labels,
+            config.classifier_dropout,
+        )
+        self.model._init_weights(self.classification_head.dense)
+        self.model._init_weights(self.classification_head.out_proj)
+
+    @add_start_docstrings_to_model_forward(MBART_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="facebook/mbart-large-cc25",
+        output_type=Seq2SeqSequenceClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    # Copied from transformers.models.bart.modeling_bart.BartForSequenceClassification.forward
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        decoder_input_ids=None,
+        decoder_attention_mask=None,
+        encoder_outputs=None,
+        inputs_embeds=None,
+        decoder_inputs_embeds=None,
+        labels=None,
+        use_cache=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 :obj:`[0, ...,
+            config.num_labels - 1]`. If :obj:`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
+        if labels is not None:
+            use_cache = False
+
+        if input_ids is None and inputs_embeds is not None:
+            raise NotImplementedError(
+                f"Passing input embeddings is currently not supported for {self.__class__.__name__}"
+            )
+
+        outputs = self.model(
+            input_ids,
+            attention_mask=attention_mask,
+            decoder_input_ids=decoder_input_ids,
+            decoder_attention_mask=decoder_attention_mask,
+            encoder_outputs=encoder_outputs,
+            inputs_embeds=inputs_embeds,
+            decoder_inputs_embeds=decoder_inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        hidden_states = outputs[0]  # last hidden state
+
+        eos_mask = input_ids.eq(self.config.eos_token_id)
+
+        if len(torch.unique(eos_mask.sum(1))) > 1:
+            raise ValueError("All examples must have the same number of <eos> tokens.")
+        sentence_representation = hidden_states[eos_mask, :].view(hidden_states.size(0), -1, hidden_states.size(-1))[
+            :, -1, :
+        ]
+        logits = self.classification_head(sentence_representation)
+
+        loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(logits.view(-1, self.config.num_labels), labels.view(-1))
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return Seq2SeqSequenceClassifierOutput(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            decoder_hidden_states=outputs.decoder_hidden_states,
+            decoder_attentions=outputs.decoder_attentions,
+            cross_attentions=outputs.cross_attentions,
+            encoder_last_hidden_state=outputs.encoder_last_hidden_state,
+            encoder_hidden_states=outputs.encoder_hidden_states,
+            encoder_attentions=outputs.encoder_attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    MBART Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear
+    layer on top of the hidden-states output to compute `span start logits` and `span end logits`).
+    """,
+    MBART_START_DOCSTRING,
+)
+class MBartForQuestionAnswering(MBartPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+
+        config.num_labels = 2
+        self.num_labels = config.num_labels
+
+        self.model = MBartModel(config)
+        self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)
+
+        self.model._init_weights(self.qa_outputs)
+
+    @add_start_docstrings_to_model_forward(MBART_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="facebook/mbart-large-cc25",
+        output_type=Seq2SeqQuestionAnsweringModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    # Copied from transformers.models.bart.modeling_bart.BartForQuestionAnswering.forward
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        decoder_input_ids=None,
+        decoder_attention_mask=None,
+        encoder_outputs=None,
+        start_positions=None,
+        end_positions=None,
+        inputs_embeds=None,
+        decoder_inputs_embeds=None,
+        use_cache=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 (`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 (`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
+        if start_positions is not None and end_positions is not None:
+            use_cache = False
+
+        outputs = self.model(
+            input_ids,
+            attention_mask=attention_mask,
+            decoder_input_ids=decoder_input_ids,
+            decoder_attention_mask=decoder_attention_mask,
+            encoder_outputs=encoder_outputs,
+            inputs_embeds=inputs_embeds,
+            decoder_inputs_embeds=decoder_inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        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)
+        end_logits = end_logits.squeeze(-1)
+
+        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.clamp_(0, ignored_index)
+            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 Seq2SeqQuestionAnsweringModelOutput(
+            loss=total_loss,
+            start_logits=start_logits,
+            end_logits=end_logits,
+            past_key_values=outputs.past_key_values,
+            decoder_hidden_states=outputs.decoder_hidden_states,
+            decoder_attentions=outputs.decoder_attentions,
+            cross_attentions=outputs.cross_attentions,
+            encoder_last_hidden_state=outputs.encoder_last_hidden_state,
+            encoder_hidden_states=outputs.encoder_hidden_states,
+            encoder_attentions=outputs.encoder_attentions,
+        )

src/transformers/models/pegasus/__init__.py

@@ -15,9 +15,8 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-
 from ...file_utils import is_sentencepiece_available, is_tf_available, is_tokenizers_available, is_torch_available
-from .configuration_pegasus import PegasusConfig
+from .configuration_pegasus import PEGASUS_PRETRAINED_CONFIG_ARCHIVE_MAP, PegasusConfig
 
 
 if is_sentencepiece_available():
@@ -27,7 +26,12 @@ if is_tokenizers_available():
     from .tokenization_pegasus_fast import PegasusTokenizerFast
 
 if is_torch_available():
-    from .modeling_pegasus import PegasusForConditionalGeneration, PegasusModel
+    from .modeling_pegasus import (
+        PEGASUS_PRETRAINED_MODEL_ARCHIVE_LIST,
+        PegasusForConditionalGeneration,
+        PegasusModel,
+        PegasusPreTrainedModel,
+    )
 
 if is_tf_available():
     from .modeling_tf_pegasus import TFPegasusForConditionalGeneration

src/transformers/models/pegasus/configuration_pegasus.py

 # coding=utf-8
-# Copyright 2020 Google and The HuggingFace Inc. team.
+# Copyright 2021, Google and The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -14,84 +14,48 @@
 # limitations under the License.
 """ PEGASUS model configuration """
 
+from ...configuration_utils import PretrainedConfig
 from ...utils import logging
-from ..bart.configuration_bart import BartConfig
 
 
 logger = logging.get_logger(__name__)
 
-# These config values do not vary between checkpoints
-DEFAULTS = dict(
-    vocab_size=96103,
-    max_position_embeddings=512,
-    d_model=1024,
-    encoder_ffn_dim=4096,
-    decoder_ffn_dim=4096,
-    encoder_attention_heads=16,
-    decoder_attention_heads=16,
-    encoder_layers=16,
-    decoder_layers=16,
-    dropout=0.1,
-    attention_dropout=0.1,
-    activation_dropout=0.1,
-    pad_token_id=0,
-    eos_token_id=1,
-    is_encoder_decoder=True,
-    normalize_before=True,
-    scale_embedding=True,
-    normalize_embedding=False,
-    add_final_layer_norm=True,
-    static_position_embeddings=True,
-    num_beams=8,
-    activation_function="relu",
-)
-# Config values that vary between checkpoints: for testing and conversion
-task_specific_params = {
-    # These are task specific params for pegasus-large and normal params for finetuned checkpoints
-    "summarization_xsum": {"length_penalty": 0.6, "max_length": 64, "max_position_embeddings": 512},
-    "summarization_cnn_dailymail": {"length_penalty": 0.8, "max_length": 128, "max_position_embeddings": 1024},
-    "summarization_newsroom": {"length_penalty": 0.8, "max_length": 128, "max_position_embeddings": 512},
-    "summarization_wikihow": {"length_penalty": 0.6, "max_length": 256, "max_position_embeddings": 512},
-    "summarization_multi_news": {"length_penalty": 0.8, "max_length": 256, "max_position_embeddings": 1024},
-    "summarization_reddit_tifu": {"length_penalty": 0.6, "max_length": 128, "max_position_embeddings": 512},
-    "summarization_big_patent": {"length_penalty": 0.7, "max_length": 256, "max_position_embeddings": 1024},
-    "summarization_arxiv": {"length_penalty": 0.8, "max_length": 256, "max_position_embeddings": 1024},
-    "summarization_pubmed": {"length_penalty": 0.8, "max_length": 256, "max_position_embeddings": 1024},
-    "summarization_gigaword": {"length_penalty": 0.6, "max_length": 32, "max_position_embeddings": 128},
-    "summarization_aeslc": {"length_penalty": 0.6, "max_length": 32, "max_position_embeddings": 512},
-    "summarization_billsum": {"length_penalty": 0.6, "max_length": 256, "max_position_embeddings": 1024},
-    # this last entry is useless -- just for consistency
-    "summarization_large": {"length_penalty": 0.8, "max_length": 256, "max_position_embeddings": 1024},
+PEGASUS_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "google/pegasus-large": "https://huggingface.co/google/pegasus-large/resolve/main/config.json",
+    # See all PEGASUS models at https://huggingface.co/models?filter=pegasus
 }
 
 
-class PegasusConfig(BartConfig):
-    """
-    This is the configuration class to store the configuration of a
-    :class:`~transformers.PegasusForConditionalGeneration`. It is used to instantiate a Pegasus model according to the
-    specified arguments, defining the model architecture.
+class PegasusConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a :class:`~transformers.PegasusModel`. It is used to
+    instantiate an PEGASUS model according to the specified arguments, defining the model architecture. Instantiating a
+    configuration with the defaults will yield a similar configuration to that of the PEGASUS `google/pegasus-large
+    <https://huggingface.co/google/pegasus-large>`__ architecture.
 
     Configuration objects inherit from :class:`~transformers.PretrainedConfig` and can be used to control the model
     outputs. Read the documentation from :class:`~transformers.PretrainedConfig` for more information.
 
+
     Args:
-        vocab_size (:obj:`int`, `optional`, defaults to 96103):
-            Vocabulary size of the Pegasus model. Defines the number of different tokens that can be represented by the
-            :obj:`inputs_ids` passed when calling :class:`~transformers.PegasusForConditionalGeneration`.
+        vocab_size (:obj:`int`, `optional`, defaults to 50265):
+            Vocabulary size of the PEGASUS model. Defines the number of different tokens that can be represented by the
+            :obj:`inputs_ids` passed when calling :class:`~transformers.PegasusModel` or
+            :class:`~transformers.TFPegasusModel`.
         d_model (:obj:`int`, `optional`, defaults to 1024):
             Dimensionality of the layers and the pooler layer.
-        encoder_layers (:obj:`int`, `optional`, defaults to 16):
+        encoder_layers (:obj:`int`, `optional`, defaults to 12):
             Number of encoder layers.
-        decoder_layers (:obj:`int`, `optional`, defaults to 16):
+        decoder_layers (:obj:`int`, `optional`, defaults to 12):
             Number of decoder layers.
         encoder_attention_heads (:obj:`int`, `optional`, defaults to 16):
             Number of attention heads for each attention layer in the Transformer encoder.
         decoder_attention_heads (:obj:`int`, `optional`, defaults to 16):
             Number of attention heads for each attention layer in the Transformer decoder.
         decoder_ffn_dim (:obj:`int`, `optional`, defaults to 4096):
-            Dimensionality of the "intermediate" (i.e., feed-forward) layer in decoder.
+            Dimensionality of the "intermediate" (often named feed-forward) layer in decoder.
         encoder_ffn_dim (:obj:`int`, `optional`, defaults to 4096):
-            Dimensionality of the "intermediate" (i.e., feed-forward) layer in decoder.
+            Dimensionality of the "intermediate" (often named feed-forward) layer in decoder.
         activation_function (:obj:`str` or :obj:`function`, `optional`, defaults to :obj:`"gelu"`):
             The non-linear activation function (function or string) in the encoder and pooler. If string,
             :obj:`"gelu"`, :obj:`"relu"`, :obj:`"silu"` and :obj:`"gelu_new"` are supported.
@@ -108,38 +72,108 @@ class PegasusConfig(BartConfig):
             just in case (e.g., 512 or 1024 or 2048).
         init_std (:obj:`float`, `optional`, defaults to 0.02):
             The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
-        add_bias_logits (:obj:`bool`, `optional`, defaults to :obj:`False`):
-            This should be completed, specific to marian.
-        normalize_before (:obj:`bool`, `optional`, defaults to :obj:`True`):
-            Call layernorm before attention ops.
-        normalize_embedding (:obj:`bool`, `optional`, defaults to :obj:`False`):
-            Call layernorm after embeddings.
-        static_position_embeddings (:obj:`bool`, `optional`, defaults to :obj:`True`):
-            Don't learn positional embeddings, use sinusoidal.
-        add_final_layer_norm (:obj:`bool`, `optional`, defaults to :obj:`True`):
-            Why not add another layernorm?
-        scale_embedding (:obj:`bool`, `optional`, defaults to :obj:`True`):
-            Scale embeddings by diving by sqrt(d_model).
-        eos_token_id (:obj:`int`, `optional`, defaults to 2)
-            End of stream token id.
-        pad_token_id (:obj:`int`, `optional`, defaults to 1)
-            Padding token id.
-        bos_token_id (:obj:`int`, `optional`, defaults to 0)
-            Beginning of stream token id.
         encoder_layerdrop: (:obj:`float`, `optional`, defaults to 0.0):
             The LayerDrop probability for the encoder. See the `LayerDrop paper <see
             https://arxiv.org/abs/1909.11556>`__ for more details.
         decoder_layerdrop: (:obj:`float`, `optional`, defaults to 0.0):
             The LayerDrop probability for the decoder. See the `LayerDrop paper <see
             https://arxiv.org/abs/1909.11556>`__ for more details.
-        extra_pos_embeddings: (:obj:`int`, `optional`, defaults to 2):
-            How many extra learned positional embeddings to use. Should be pad_token_id+1 for bart.
-        is_encoder_decoder (:obj:`bool`, `optional`, defaults to :obj:`True`):
-            Whether this is an encoder/decoder model
-        force_bos_token_to_be_generated (:obj:`bool`, `optional`, defaults to :obj:`False`):
-            Whether or not to force BOS token to be generated at step 1 (after ``decoder_start_token_id``).
-    """
+        gradient_checkpointing (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            If True, use gradient checkpointing to save memory at the expense of slower backward pass.
+        scale_embedding (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Scale embeddings by diving by sqrt(d_model).
+        use_cache (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Whether or not the model should return the last key/values attentions (not used by all models)
+
+    Example::
+
+        >>> from transformers import PegasusModel, PegasusConfig
+
+        >>> # Initializing a PEGASUS google/pegasus-large style configuration
+        >>> configuration = PegasusConfig()
 
+        >>> # Initializing a model from the google/pegasus-large style configuration
+        >>> model = PegasusModel(configuration)
+
+        >>> # Accessing the model configuration
+        >>> configuration = model.config
+    """
     model_type = "pegasus"
     keys_to_ignore_at_inference = ["past_key_values"]
-    # The implementation of the config object is in BartConfig
+
+    def __init__(
+        self,
+        vocab_size=50265,
+        max_position_embeddings=1024,
+        encoder_layers=12,
+        encoder_ffn_dim=4096,
+        encoder_attention_heads=16,
+        decoder_layers=12,
+        decoder_ffn_dim=4096,
+        decoder_attention_heads=16,
+        encoder_layerdrop=0.0,
+        decoder_layerdrop=0.0,
+        use_cache=True,
+        is_encoder_decoder=True,
+        activation_function="gelu",
+        d_model=1024,
+        dropout=0.1,
+        attention_dropout=0.0,
+        activation_dropout=0.0,
+        init_std=0.02,
+        decoder_start_token_id=0,
+        classifier_dropout=0.0,
+        scale_embedding=False,
+        gradient_checkpointing=False,
+        pad_token_id=0,
+        eos_token_id=1,
+        **kwargs
+    ):
+        super().__init__(
+            pad_token_id=pad_token_id,
+            eos_token_id=eos_token_id,
+            is_encoder_decoder=is_encoder_decoder,
+            decoder_start_token_id=decoder_start_token_id,
+            **kwargs,
+        )
+
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.d_model = d_model
+        self.encoder_ffn_dim = encoder_ffn_dim
+        self.encoder_layers = encoder_layers
+        self.encoder_attention_heads = encoder_attention_heads
+        self.decoder_ffn_dim = decoder_ffn_dim
+        self.decoder_layers = decoder_layers
+        self.decoder_attention_heads = decoder_attention_heads
+        self.dropout = dropout
+        self.attention_dropout = attention_dropout
+        self.activation_dropout = activation_dropout
+        self.activation_function = activation_function
+        self.init_std = init_std
+        self.encoder_layerdrop = encoder_layerdrop
+        self.decoder_layerdrop = decoder_layerdrop
+        self.classifier_dropout = classifier_dropout
+        self.use_cache = use_cache
+        self.num_hidden_layers = encoder_layers
+        self.gradient_checkpointing = gradient_checkpointing
+        self.scale_embedding = scale_embedding  # scale factor will be sqrt(d_model) if True
+
+        # IMPORTANT
+        # DELETE ALL OF THE FOLLOWING LINES AS SOON AS TF IS READY
+        self.extra_pos_embeddings = 0
+        self.normalize_before = True
+        self.add_final_layer_norm = True
+        self.do_blenderbot_90_layernorm = False
+        self.normalize_embedding = False
+        self.static_position_embeddings = True
+        self.add_bias_logits = False
+        self.force_bos_token_to_be_generated = False
+
+    @property
+    def num_attention_heads(self) -> int:
+        return self.encoder_attention_heads
+
+    @property
+    def hidden_size(self) -> int:
+        return self.d_model

src/transformers/models/pegasus/modeling_pegasus.py

 # coding=utf-8
-# Copyright 2020 Google and The HuggingFace Inc. team.
+# Copyright 2021, Google and The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -12,72 +12,1212 @@
 # 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 Pegasus model, ported from https://github.com/google-research/pegasus"""
+""" PyTorch PEGASUS model. """
 
 
-from ...file_utils import add_start_docstrings
-from ..bart.modeling_bart import BART_START_DOCSTRING, BartForConditionalGeneration, BartModel
-from .configuration_pegasus import PegasusConfig
+import math
+import random
+from typing import Optional, Tuple
 
+import numpy as np
+import torch
+import torch.nn.functional as F
+from torch import nn
+from torch.nn import CrossEntropyLoss
 
-@add_start_docstrings(
-    "The bare Pegasus Model transformer outputting raw hidden-states without any specific head on top.",
-    BART_START_DOCSTRING,
+from ...activations import ACT2FN
+from ...file_utils import (
+    add_end_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    replace_return_docstrings,
 )
-class PegasusModel(BartModel):
-    r"""
-    This class overrides :class:`~transformers.BartModel`. Please check the superclass for the appropriate
-    documentation alongside usage examples.
+from ...modeling_outputs import (
+    BaseModelOutput,
+    BaseModelOutputWithPastAndCrossAttentions,
+    Seq2SeqLMOutput,
+    Seq2SeqModelOutput,
+)
+from ...modeling_utils import PreTrainedModel
+from ...utils import logging
+from .configuration_pegasus import PegasusConfig
+
+
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = "PegasusConfig"
+_TOKENIZER_FOR_DOC = "PegasusTokenizer"
+
+
+PEGASUS_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "google/pegasus-large",
+    # See all PEGASUS models at https://huggingface.co/models?filter=pegasus
+]
+
+
+# Copied from transformers.models.bart.modeling_bart.shift_tokens_right
+def shift_tokens_right(input_ids: torch.Tensor, pad_token_id: int, decoder_start_token_id: int):
     """
+    Shift input ids one token to the right.
+    """
+    shifted_input_ids = input_ids.new_zeros(input_ids.shape)
+    shifted_input_ids[:, 1:] = input_ids[:, :-1].clone()
+    shifted_input_ids[:, 0] = decoder_start_token_id
+
+    assert pad_token_id is not None, "self.model.config.pad_token_id has to be defined."
+    # replace possible -100 values in labels by `pad_token_id`
+    shifted_input_ids.masked_fill_(shifted_input_ids == -100, pad_token_id)
+
+    return shifted_input_ids
+
+
+# Copied from transformers.models.bart.modeling_bart._make_causal_mask
+def _make_causal_mask(input_ids_shape: torch.Size, dtype: torch.dtype, past_key_values_length: int = 0):
+    """
+    Make causal mask used for bi-directional self-attention.
+    """
+    bsz, tgt_len = input_ids_shape
+    mask = torch.full((tgt_len, tgt_len), float("-inf"))
+    mask_cond = torch.arange(mask.size(-1))
+    mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)
+    mask = mask.to(dtype)
+
+    if past_key_values_length > 0:
+        mask = torch.cat([torch.zeros(tgt_len, past_key_values_length, dtype=dtype), mask], dim=-1)
+    return mask[None, None, :, :].expand(bsz, 1, tgt_len, tgt_len + past_key_values_length)
+
+
+# Copied from transformers.models.bart.modeling_bart._expand_mask
+def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
+    """
+    Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
+    """
+    bsz, src_len = mask.size()
+    tgt_len = tgt_len if tgt_len is not None else src_len
+
+    expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)
+
+    inverted_mask = 1.0 - expanded_mask
+
+    return inverted_mask.masked_fill(inverted_mask.bool(), torch.finfo(dtype).min)
+
+
+# Copied from transformers.models.marian.modeling_marian.MarianSinusoidalPositionalEmbedding with Marian->Pegasus
+class PegasusSinusoidalPositionalEmbedding(nn.Embedding):
+    """This module produces sinusoidal positional embeddings of any length."""
+
+    def __init__(self, num_positions: int, embedding_dim: int, padding_idx: Optional[int] = None):
+        super().__init__(num_positions, embedding_dim)
+        self.weight = self._init_weight(self.weight)
+
+    @staticmethod
+    def _init_weight(out: nn.Parameter):
+        """
+        Identical to the XLM create_sinusoidal_embeddings except features are not interleaved. The cos features are in
+        the 2nd half of the vector. [dim // 2:]
+        """
+        n_pos, dim = out.shape
+        position_enc = np.array(
+            [[pos / np.power(10000, 2 * (j // 2) / dim) for j in range(dim)] for pos in range(n_pos)]
+        )
+        out.requires_grad = False  # set early to avoid an error in pytorch-1.8+
+        sentinel = dim // 2 if dim % 2 == 0 else (dim // 2) + 1
+        out[:, 0:sentinel] = torch.FloatTensor(np.sin(position_enc[:, 0::2]))
+        out[:, sentinel:] = torch.FloatTensor(np.cos(position_enc[:, 1::2]))
+        out.detach_()
+        return out
+
+    @torch.no_grad()
+    def forward(self, input_ids_shape: torch.Size, past_key_values_length: int = 0):
+        """`input_ids_shape` is expected to be [bsz x seqlen]."""
+        bsz, seq_len = input_ids_shape[:2]
+        positions = torch.arange(
+            past_key_values_length, past_key_values_length + seq_len, dtype=torch.long, device=self.weight.device
+        )
+        return super().forward(positions)
+
+
+# Copied from transformers.models.bart.modeling_bart.BartAttention with Bart->Pegasus
+class PegasusAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(
+        self,
+        embed_dim: int,
+        num_heads: int,
+        dropout: float = 0.0,
+        is_decoder: bool = False,
+        bias: bool = True,
+    ):
+        super().__init__()
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.dropout = dropout
+        self.head_dim = embed_dim // num_heads
+        assert (
+            self.head_dim * num_heads == self.embed_dim
+        ), f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`: {num_heads})."
+        self.scaling = self.head_dim ** -0.5
+        self.is_decoder = is_decoder
+
+        self.k_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+        self.v_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+        self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+        self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+
+    def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
+        return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        key_value_states: Optional[torch.Tensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        output_attentions: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        """Input shape: Batch x Time x Channel"""
+
+        # if key_value_states are provided this layer is used as a cross-attention layer
+        # for the decoder
+        is_cross_attention = key_value_states is not None
+        bsz, tgt_len, embed_dim = hidden_states.size()
+
+        # get query proj
+        query_states = self.q_proj(hidden_states) * self.scaling
+        # get key, value proj
+        if is_cross_attention and past_key_value is not None:
+            # reuse k,v, cross_attentions
+            key_states = past_key_value[0]
+            value_states = past_key_value[1]
+        elif is_cross_attention:
+            # cross_attentions
+            key_states = self._shape(self.k_proj(key_value_states), -1, bsz)
+            value_states = self._shape(self.v_proj(key_value_states), -1, bsz)
+        elif past_key_value is not None:
+            # reuse k, v, self_attention
+            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
+            key_states = torch.cat([past_key_value[0], key_states], dim=2)
+            value_states = torch.cat([past_key_value[1], value_states], dim=2)
+        else:
+            # self_attention
+            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
+
+        if self.is_decoder:
+            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
+            # Further calls to cross_attention layer can then reuse all cross-attention
+            # key/value_states (first "if" case)
+            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
+            # all previous decoder key/value_states. Further calls to uni-directional self-attention
+            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
+            # if encoder bi-directional self-attention `past_key_value` is always `None`
+            past_key_value = (key_states, value_states)
+
+        proj_shape = (bsz * self.num_heads, -1, self.head_dim)
+        query_states = self._shape(query_states, tgt_len, bsz).view(*proj_shape)
+        key_states = key_states.view(*proj_shape)
+        value_states = value_states.view(*proj_shape)
+
+        src_len = key_states.size(1)
+        attn_weights = torch.bmm(query_states, key_states.transpose(1, 2))
+
+        assert attn_weights.size() == (
+            bsz * self.num_heads,
+            tgt_len,
+            src_len,
+        ), f"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is {attn_weights.size()}"
+
+        if attention_mask is not None:
+            assert attention_mask.size() == (
+                bsz,
+                1,
+                tgt_len,
+                src_len,
+            ), f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.size()}"
+            attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + attention_mask
+            attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
+
+        attn_weights = F.softmax(attn_weights, dim=-1)
+
+        if output_attentions:
+            # this operation is a bit akward, but it's required to
+            # make sure that attn_weights keeps its gradient.
+            # In order to do so, attn_weights have to reshaped
+            # twice and have to be reused in the following
+            attn_weights_reshaped = attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
+            attn_weights = attn_weights_reshaped.view(bsz * self.num_heads, tgt_len, src_len)
+        else:
+            attn_weights_reshaped = None
+
+        attn_probs = F.dropout(attn_weights, p=self.dropout, training=self.training)
+
+        attn_output = torch.bmm(attn_probs, value_states)
+
+        assert attn_output.size() == (
+            bsz * self.num_heads,
+            tgt_len,
+            self.head_dim,
+        ), f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is {attn_output.size()}"
+
+        attn_output = (
+            attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
+            .transpose(1, 2)
+            .reshape(bsz, tgt_len, embed_dim)
+        )
+
+        attn_output = self.out_proj(attn_output)
+
+        return attn_output, attn_weights_reshaped, past_key_value
+
+
+# Copied from transformers.models.mbart.modeling_mbart.MBartEncoderLayer with MBart->Pegasus
+class PegasusEncoderLayer(nn.Module):
+    def __init__(self, config: PegasusConfig):
+        super().__init__()
+        self.embed_dim = config.d_model
+        self.self_attn = PegasusAttention(
+            embed_dim=self.embed_dim,
+            num_heads=config.encoder_attention_heads,
+            dropout=config.attention_dropout,
+        )
+        self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
+        self.dropout = config.dropout
+        self.activation_fn = ACT2FN[config.activation_function]
+        self.activation_dropout = config.activation_dropout
+        self.fc1 = nn.Linear(self.embed_dim, config.encoder_ffn_dim)
+        self.fc2 = nn.Linear(config.encoder_ffn_dim, self.embed_dim)
+        self.final_layer_norm = nn.LayerNorm(self.embed_dim)
 
+    def forward(self, hidden_states: torch.Tensor, attention_mask: torch.Tensor, output_attentions: bool = False):
+        """
+        Args:
+            hidden_states (:obj:`torch.FloatTensor`): input to the layer of shape `(seq_len, batch, embed_dim)`
+            attention_mask (:obj:`torch.FloatTensor`): attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            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.
+        """
+        residual = hidden_states
+        hidden_states = self.self_attn_layer_norm(hidden_states)
+        hidden_states, attn_weights, _ = self.self_attn(
+            hidden_states=hidden_states, attention_mask=attention_mask, output_attentions=output_attentions
+        )
+        hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
+        hidden_states = residual + hidden_states
+
+        residual = hidden_states
+        hidden_states = self.final_layer_norm(hidden_states)
+        hidden_states = self.activation_fn(self.fc1(hidden_states))
+        hidden_states = F.dropout(hidden_states, p=self.activation_dropout, training=self.training)
+        hidden_states = self.fc2(hidden_states)
+        hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
+        hidden_states = residual + hidden_states
+
+        if torch.isinf(hidden_states).any() or torch.isnan(hidden_states).any():
+            clamp_value = torch.finfo(hidden_states.dtype).max - 1000
+            hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (attn_weights,)
+
+        return outputs
+
+
+# Copied from transformers.models.mbart.modeling_mbart.MBartDecoderLayer with MBart->Pegasus
+class PegasusDecoderLayer(nn.Module):
+    def __init__(self, config: PegasusConfig):
+        super().__init__()
+        self.embed_dim = config.d_model
+
+        self.self_attn = PegasusAttention(
+            embed_dim=self.embed_dim,
+            num_heads=config.decoder_attention_heads,
+            dropout=config.attention_dropout,
+            is_decoder=True,
+        )
+        self.dropout = config.dropout
+        self.activation_fn = ACT2FN[config.activation_function]
+        self.activation_dropout = config.activation_dropout
+
+        self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
+        self.encoder_attn = PegasusAttention(
+            self.embed_dim,
+            config.decoder_attention_heads,
+            dropout=config.attention_dropout,
+            is_decoder=True,
+        )
+        self.encoder_attn_layer_norm = nn.LayerNorm(self.embed_dim)
+        self.fc1 = nn.Linear(self.embed_dim, config.decoder_ffn_dim)
+        self.fc2 = nn.Linear(config.decoder_ffn_dim, self.embed_dim)
+        self.final_layer_norm = nn.LayerNorm(self.embed_dim)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = True,
+    ):
+        """
+        Args:
+            hidden_states (:obj:`torch.FloatTensor`): input to the layer of shape `(seq_len, batch, embed_dim)`
+            attention_mask (:obj:`torch.FloatTensor`): attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            encoder_hidden_states (:obj:`torch.FloatTensor`): cross attention input to the layer of shape `(seq_len, batch, embed_dim)`
+            encoder_attention_mask (:obj:`torch.FloatTensor`): encoder attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            past_key_value (:obj:`Tuple(torch.FloatTensor)`): cached past key and value projection states
+            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.
+        """
+        residual = hidden_states
+        hidden_states = self.self_attn_layer_norm(hidden_states)
+
+        # Self Attention
+        # decoder uni-directional self-attention cached key/values tuple is at positions 1,2
+        self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None
+        # add present self-attn cache to positions 1,2 of present_key_value tuple
+        hidden_states, self_attn_weights, present_key_value = self.self_attn(
+            hidden_states=hidden_states,
+            past_key_value=self_attn_past_key_value,
+            attention_mask=attention_mask,
+            output_attentions=output_attentions,
+        )
+        hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
+        hidden_states = residual + hidden_states
+
+        # Cross-Attention Block
+        cross_attn_present_key_value = None
+        cross_attn_weights = None
+        if encoder_hidden_states is not None:
+            residual = hidden_states
+            hidden_states = self.encoder_attn_layer_norm(hidden_states)
+
+            # cross_attn cached key/values tuple is at positions 3,4 of present_key_value tuple
+            cross_attn_past_key_value = past_key_value[-2:] if past_key_value is not None else None
+            hidden_states, cross_attn_weights, cross_attn_present_key_value = self.encoder_attn(
+                hidden_states=hidden_states,
+                key_value_states=encoder_hidden_states,
+                attention_mask=encoder_attention_mask,
+                past_key_value=cross_attn_past_key_value,
+                output_attentions=output_attentions,
+            )
+            hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
+            hidden_states = residual + hidden_states
+
+            # add cross-attn to positions 3,4 of present_key_value tuple
+            present_key_value = present_key_value + cross_attn_present_key_value
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.final_layer_norm(hidden_states)
+        hidden_states = self.activation_fn(self.fc1(hidden_states))
+        hidden_states = F.dropout(hidden_states, p=self.activation_dropout, training=self.training)
+        hidden_states = self.fc2(hidden_states)
+        hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights, cross_attn_weights)
+
+        if use_cache:
+            outputs += (present_key_value,)
+
+        return outputs
+
+
+class PegasusPreTrainedModel(PreTrainedModel):
     config_class = PegasusConfig
-    _keys_to_ignore_on_load_missing = [
-        r"final_logits_bias",
-        r"encoder\.version",
-        r"decoder\.version",
-        "encoder.embed_positions",
-        "decoder.embed_positions",
-    ]
-    _keys_to_ignore_on_save = [
-        "encoder.embed_positions.weight",
-        "decoder.embed_positions.weight",
-    ]
+    base_model_prefix = "model"
 
+    def _init_weights(self, module):
+        std = self.config.init_std
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, PegasusSinusoidalPositionalEmbedding):
+            pass
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
 
-@add_start_docstrings("The Pegasus Model for summarization ", BART_START_DOCSTRING)
-class PegasusForConditionalGeneration(BartForConditionalGeneration):
-    r"""
-    Pytorch version of google's pegasus model for summarization. Available models are listed `here
-    <https://huggingface.co/models?search=pegasus>`__.
+    @property
+    def dummy_inputs(self):
+        pad_token = self.config.pad_token_id
+        input_ids = torch.tensor([[0, 6, 10, 4, 2], [0, 8, 12, 2, pad_token]], device=self.device)
+        dummy_inputs = {
+            "attention_mask": input_ids.ne(pad_token),
+            "input_ids": input_ids,
+            "decoder_input_ids": input_ids,
+        }
+        return dummy_inputs
 
-    This class overrides :class:`~transformers.BartForConditionalGeneration`. Please check the superclass for the
-    appropriate documentation alongside usage examples.
 
-    Examples::
+PEGASUS_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.
+
+    Parameters:
+        config (:class:`~transformers.PegasusConfig`):
+            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.
+"""
+
+PEGASUS_GENERATION_EXAMPLE = r"""
+    Summarization example::
 
         >>> from transformers import PegasusTokenizer, PegasusForConditionalGeneration
-        >>> from typing import List
-        >>> PGE_ARTICLE = "PG&E stated it scheduled the blackouts in response to forecasts for high winds amid dry conditions. The aim is to reduce the risk of wildfires. Nearly 800 thousand customers were scheduled to be affected by the shutoffs which were expected to last through at least midday tomorrow."
-        >>> mname = "google/pegasus-xsum"
 
-        >>> model = PegasusForConditionalGeneration.from_pretrained(mname)
-        >>> tok = PegasusTokenizer.from_pretrained(mname)
-        >>> batch = tok.prepare_seq2seq_batch(src_texts=[PGE_ARTICLE], return_tensors="pt")  # don't need tgt_text for inference
-        >>> gen = model.generate(**batch)  # for forward pass: model(**batch)
-        >>> summary: List[str] = tok.batch_decode(gen, skip_special_tokens=True)
-        >>> assert summary == "California's largest electricity provider has turned off power to tens of thousands of customers."
+        >>> model = PegasusForConditionalGeneration.from_pretrained('google/pegasus-xsum')
+        >>> tokenizer = PegasusTokenizer.from_pretrained('google/pegasus-xsum')
+
+        >>> ARTICLE_TO_SUMMARIZE = (
+        ... "PG&E stated it scheduled the blackouts in response to forecasts for high winds "
+        ... "amid dry conditions. The aim is to reduce the risk of wildfires. Nearly 800 thousand customers were "
+        ... "scheduled to be affected by the shutoffs which were expected to last through at least midday tomorrow."
+        ... )
+        >>> inputs = tokenizer([ARTICLE_TO_SUMMARIZE], max_length=1024, return_tensors='pt')
+
+        >>> # Generate Summary
+        >>> summary_ids = model.generate(inputs['input_ids'])
+        >>> print([tokenizer.decode(g, skip_special_tokens=True, clean_up_tokenization_spaces=False) for g in summary_ids])
+"""
+
+PEGASUS_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+
+            Indices can be obtained using :class:`~transformers.PegasusTokenizer`. See
+            :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__` for
+            details.
+
+            `What are input IDs? <../glossary.html#input-ids>`__
+        attention_mask (:obj:`torch.Tensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Mask to avoid performing attention on padding token indices. Mask values selected in ``[0, 1]``:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            `What are attention masks? <../glossary.html#attention-mask>`__
+        decoder_input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, target_sequence_length)`, `optional`):
+            Indices of decoder input sequence tokens in the vocabulary.
+
+            Indices can be obtained using :class:`~transformers.PegasusTokenizer`. See
+            :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__` for
+            details.
+
+            `What are input IDs? <../glossary.html#input-ids>`__
+
+            Pegasus uses the :obj:`pad_token_id` as the starting token for :obj:`decoder_input_ids` generation. If
+            :obj:`past_key_values` is used, optionally only the last :obj:`decoder_input_ids` have to be input (see
+            :obj:`past_key_values`).
+        decoder_attention_mask (:obj:`torch.LongTensor` of shape :obj:`(batch_size, target_sequence_length)`, `optional`):
+            Default behavior: generate a tensor that ignores pad tokens in :obj:`decoder_input_ids`. Causal mask will
+            also be used by default.
 
+            If you want to change padding behavior, you should read :func:`modeling_pegasus._prepare_decoder_inputs`
+            and modify to your needs. See diagram 1 in `the paper <https://arxiv.org/abs/1910.13461>`__ for more
+            information on the default strategy.
+        encoder_outputs (:obj:`tuple(tuple(torch.FloatTensor)`, `optional`):
+            Tuple consists of (:obj:`last_hidden_state`, `optional`: :obj:`hidden_states`, `optional`:
+            :obj:`attentions`) :obj:`last_hidden_state` of shape :obj:`(batch_size, sequence_length, hidden_size)`,
+            `optional`) is a sequence of hidden-states at the output of the last layer of the encoder. Used in the
+            cross-attention of the decoder.
+        past_key_values (:obj:`Tuple[Tuple[torch.Tensor]]` of length :obj:`config.n_layers` with each tuple having 2 tuples each of which has 2 tensors of shape :obj:`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
+            Contains precomputed key and value hidden-states of the attention blocks. Can be used to speed up decoding.
+
+            If :obj:`past_key_values` are used, the user can optionally input only the last :obj:`decoder_input_ids`
+            (those that don't have their past key value states given to this model) of shape :obj:`(batch_size, 1)`
+            instead of all :obj:`decoder_input_ids`` of shape :obj:`(batch_size, sequence_length)`.
+        inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, 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.
+        decoder_inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, target_sequence_length, hidden_size)`, `optional`):
+            Optionally, instead of passing :obj:`decoder_input_ids` you can choose to directly pass an embedded
+            representation. If :obj:`past_key_values` is used, optionally only the last :obj:`decoder_inputs_embeds`
+            have to be input (see :obj:`past_key_values`). This is useful if you want more control over how to convert
+            :obj:`decoder_input_ids` indices into associated vectors than the model's internal embedding lookup matrix.
+
+            If :obj:`decoder_input_ids` and :obj:`decoder_inputs_embeds` are both unset, :obj:`decoder_inputs_embeds`
+            takes the value of :obj:`inputs_embeds`.
+        use_cache (:obj:`bool`, `optional`):
+            If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up
+            decoding (see :obj:`past_key_values`).
+        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.
+"""
+
+
+class PegasusEncoder(PegasusPreTrainedModel):
     """
-    # All the code is in src/transformers/models/bart/modeling_bart.py
-    config_class = PegasusConfig
+    Transformer encoder consisting of *config.encoder_layers* self attention layers. Each layer is a
+    :class:`PegasusEncoderLayer`.
+
+    Args:
+        config: PegasusConfig
+        embed_tokens (torch.nn.Embedding): output embedding
+    """
+
+    def __init__(self, config: PegasusConfig, embed_tokens: Optional[nn.Embedding] = None):
+        super().__init__(config)
+
+        self.dropout = config.dropout
+        self.layerdrop = config.encoder_layerdrop
+
+        embed_dim = config.d_model
+        self.padding_idx = config.pad_token_id
+        self.max_source_positions = config.max_position_embeddings
+        self.embed_scale = math.sqrt(embed_dim) if config.scale_embedding else 1.0
+
+        if embed_tokens is not None:
+            self.embed_tokens = embed_tokens
+        else:
+            self.embed_tokens = nn.Embedding(config.vocab_size, embed_dim, self.padding_idx)
+
+        self.embed_positions = PegasusSinusoidalPositionalEmbedding(
+            config.max_position_embeddings,
+            embed_dim,
+            self.padding_idx,
+        )
+        self.layers = nn.ModuleList([PegasusEncoderLayer(config) for _ in range(config.encoder_layers)])
+        self.layer_norm = nn.LayerNorm(config.d_model)
+
+        self.init_weights()
+
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        inputs_embeds=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        Args:
+            input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`):
+                Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you
+                provide it.
+
+                Indices can be obtained using :class:`~transformers.PegasusTokenizer`. See
+                :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__`
+                for details.
+
+                `What are input IDs? <../glossary.html#input-ids>`__
+            attention_mask (:obj:`torch.Tensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+                Mask to avoid performing attention on padding token indices. Mask values selected in ``[0, 1]``:
+
+                - 1 for tokens that are **not masked**,
+                - 0 for tokens that are **masked**.
+
+                `What are attention masks? <../glossary.html#attention-mask>`__
+            inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, 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.
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # retrieve input_ids and inputs_embeds
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+            input_ids = input_ids.view(-1, input_shape[-1])
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids) * self.embed_scale
+
+        embed_pos = self.embed_positions(input_shape)
+
+        hidden_states = inputs_embeds + embed_pos
+
+        hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
+
+        # expand attention_mask
+        if attention_mask is not None:
+            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+            attention_mask = _expand_mask(attention_mask, inputs_embeds.dtype)
+
+        encoder_states = () if output_hidden_states else None
+        all_attentions = () if output_attentions else None
+        for encoder_layer in self.layers:
+            if output_hidden_states:
+                encoder_states = encoder_states + (hidden_states,)
+            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
+            dropout_probability = random.uniform(0, 1)
+            if self.training and (dropout_probability < self.layerdrop):  # skip the layer
+                layer_outputs = (None, None)
+            else:
+                if getattr(self.config, "gradient_checkpointing", False):
+
+                    def create_custom_forward(module):
+                        def custom_forward(*inputs):
+                            return module(*inputs, output_attentions)
+
+                        return custom_forward
+
+                    layer_outputs = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(encoder_layer),
+                        hidden_states,
+                        attention_mask,
+                    )
+                else:
+                    layer_outputs = encoder_layer(hidden_states, attention_mask, output_attentions=output_attentions)
+
+                hidden_states = layer_outputs[0]
+
+            if output_attentions:
+                all_attentions = all_attentions + (layer_outputs[1],)
+
+        hidden_states = self.layer_norm(hidden_states)
+
+        if output_hidden_states:
+            encoder_states = encoder_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None)
+        return BaseModelOutput(
+            last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions
+        )
+
+
+class PegasusDecoder(PegasusPreTrainedModel):
+    """
+    Transformer decoder consisting of *config.decoder_layers* layers. Each layer is a :class:`PegasusDecoderLayer`
+
+    Args:
+        config: PegasusConfig
+        embed_tokens (torch.nn.Embedding): output embedding
+    """
+
+    def __init__(self, config: PegasusConfig, embed_tokens: Optional[nn.Embedding] = None):
+        super().__init__(config)
+        self.dropout = config.dropout
+        self.layerdrop = config.decoder_layerdrop
+        self.padding_idx = config.pad_token_id
+        self.max_target_positions = config.max_position_embeddings
+        self.embed_scale = math.sqrt(config.d_model) if config.scale_embedding else 1.0
+
+        if embed_tokens is not None:
+            self.embed_tokens = embed_tokens
+        else:
+            self.embed_tokens = nn.Embedding(config.vocab_size, config.d_model, self.padding_idx)
+
+        self.embed_positions = PegasusSinusoidalPositionalEmbedding(
+            config.max_position_embeddings,
+            config.d_model,
+            self.padding_idx,
+        )
+        self.layers = nn.ModuleList([PegasusDecoderLayer(config) for _ in range(config.decoder_layers)])
+        self.layer_norm = nn.LayerNorm(config.d_model)
+
+        self.init_weights()
+
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_values=None,
+        inputs_embeds=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        Args:
+            input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`):
+                Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you
+                provide it.
+
+                Indices can be obtained using :class:`~transformers.PegasusTokenizer`. See
+                :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__`
+                for details.
+
+                `What are input IDs? <../glossary.html#input-ids>`__
+            attention_mask (:obj:`torch.Tensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+                Mask to avoid performing attention on padding token indices. Mask values selected in ``[0, 1]``:
+
+                - 1 for tokens that are **not masked**,
+                - 0 for tokens that are **masked**.
+
+                `What are attention masks? <../glossary.html#attention-mask>`__
+            encoder_hidden_states (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, encoder_sequence_length, hidden_size)`, `optional`):
+                Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention
+                of the decoder.
+            encoder_attention_mask (:obj:`torch.LongTensor` of shape :obj:`(batch_size, encoder_sequence_length)`, `optional`):
+                Mask to avoid performing cross-attention on padding tokens indices of encoder input_ids. Mask values
+                selected in ``[0, 1]``:
+
+                - 1 for tokens that are **not masked**,
+                - 0 for tokens that are **masked**.
+
+                `What are attention masks? <../glossary.html#attention-mask>`__
+            past_key_values (:obj:`Tuple[Tuple[torch.Tensor]]` of length :obj:`config.n_layers` with each tuple having 2 tuples each of which has 2 tensors of shape :obj:`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
+                Contains precomputed key and value hidden-states of the attention blocks. Can be used to speed up
+                decoding.
+
+                If :obj:`past_key_values` are used, the user can optionally input only the last
+                :obj:`decoder_input_ids` (those that don't have their past key value states given to this model) of
+                shape :obj:`(batch_size, 1)` instead of all :obj:`decoder_input_ids`` of shape :obj:`(batch_size,
+                sequence_length)`.
+            inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, 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.
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # retrieve input_ids and inputs_embeds
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+            input_ids = input_ids.view(-1, input_shape[-1])
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+        else:
+            raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
+
+        # past_key_values_length
+        past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids) * self.embed_scale
+
+        # create causal mask
+        # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+        combined_attention_mask = None
+        if input_shape[-1] > 1:
+            combined_attention_mask = _make_causal_mask(
+                input_shape, inputs_embeds.dtype, past_key_values_length=past_key_values_length
+            ).to(self.device)
+
+        if attention_mask is not None and combined_attention_mask is not None:
+            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+            combined_attention_mask = combined_attention_mask + _expand_mask(
+                attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]
+            )
+
+        # expand encoder attention mask
+        if encoder_hidden_states is not None and encoder_attention_mask is not None:
+            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+            encoder_attention_mask = _expand_mask(encoder_attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1])
+
+        # embed positions
+        positions = self.embed_positions(input_shape, past_key_values_length)
+
+        hidden_states = inputs_embeds + positions
+
+        hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        all_cross_attentions = () if output_attentions else None
+        next_decoder_cache = () if use_cache else None
+        for idx, decoder_layer in enumerate(self.layers):
+            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+            dropout_probability = random.uniform(0, 1)
+            if self.training and (dropout_probability < self.layerdrop):
+                continue
+
+            past_key_value = past_key_values[idx] if past_key_values is not None else None
+
+            if getattr(self.config, "gradient_checkpointing", False):
+                if use_cache:
+                    raise ValueError(
+                        "When using `gradient_checkpointing, make sure that `use_cache=False` and `config.use_cache=False`."
+                    )
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        # None for past_key_value
+                        return module(*inputs, output_attentions, use_cache)
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(decoder_layer),
+                    hidden_states,
+                    combined_attention_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    None,
+                )
+            else:
+
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask=combined_attention_mask,
+                    encoder_hidden_states=encoder_hidden_states,
+                    encoder_attention_mask=encoder_attention_mask,
+                    past_key_value=past_key_value,
+                    output_attentions=output_attentions,
+                    use_cache=use_cache,
+                )
+            hidden_states = layer_outputs[0]
+
+            if use_cache:
+                next_decoder_cache += (layer_outputs[3 if output_attentions else 1],)
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+                all_cross_attentions += (layer_outputs[2],)
+
+        hidden_states = self.layer_norm(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        next_cache = next_decoder_cache if use_cache else None
+        if not return_dict:
+            return tuple(
+                v
+                for v in [hidden_states, next_cache, all_hidden_states, all_self_attns, all_cross_attentions]
+                if v is not None
+            )
+        return BaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+            cross_attentions=all_cross_attentions,
+        )
+
+
+@add_start_docstrings(
+    "The bare PEGASUS Model outputting raw hidden-states without any specific head on top.",
+    PEGASUS_START_DOCSTRING,
+)
+class PegasusModel(PegasusPreTrainedModel):
+    def __init__(self, config: PegasusConfig):
+        super().__init__(config)
+
+        padding_idx, vocab_size = config.pad_token_id, config.vocab_size
+        self.shared = nn.Embedding(vocab_size, config.d_model, padding_idx)
+
+        self.encoder = PegasusEncoder(config, self.shared)
+        self.decoder = PegasusDecoder(config, self.shared)
+
+        self.init_weights()
+
+    def get_input_embeddings(self):
+        return self.shared
+
+    def set_input_embeddings(self, value):
+        self.shared = value
+        self.encoder.embed_tokens = self.shared
+        self.decoder.embed_tokens = self.shared
+
+    def get_encoder(self):
+        return self.encoder
+
+    def get_decoder(self):
+        return self.decoder
+
+    @add_start_docstrings_to_model_forward(PEGASUS_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=Seq2SeqModelOutput, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        decoder_input_ids=None,
+        decoder_attention_mask=None,
+        encoder_outputs=None,
+        past_key_values=None,
+        inputs_embeds=None,
+        decoder_inputs_embeds=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        Returns:
+
+        Example::
+
+            >>> from transformers import PegasusTokenizer, PegasusModel
+
+            >>> tokenizer = PegasusTokenizer.from_pretrained("google/pegasus-large")
+            >>> model = PegasusModel.from_pretrained("google/pegasus-large")
+
+            >>> input_ids = tokenizer("Studies have been shown that owning a dog is good for you", return_tensors="pt").input_ids  # Batch size 1
+            >>> decoder_input_ids = tokenizer("Studies show that", return_tensors="pt").input_ids  # Batch size 1
+            >>> outputs = model(input_ids=input_ids, decoder_input_ids=decoder_input_ids)
+
+            >>> last_hidden_states = outputs.last_hidden_state
+        """
+
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if encoder_outputs is None:
+            encoder_outputs = self.encoder(
+                input_ids=input_ids,
+                attention_mask=attention_mask,
+                inputs_embeds=inputs_embeds,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+            )
+        # If the user passed a tuple for encoder_outputs, we wrap it in a BaseModelOutput when return_dict=True
+        elif return_dict and not isinstance(encoder_outputs, BaseModelOutput):
+            encoder_outputs = BaseModelOutput(
+                last_hidden_state=encoder_outputs[0],
+                hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None,
+                attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None,
+            )
+
+        # decoder outputs consists of (dec_features, past_key_value, dec_hidden, dec_attn)
+        decoder_outputs = self.decoder(
+            input_ids=decoder_input_ids,
+            attention_mask=decoder_attention_mask,
+            encoder_hidden_states=encoder_outputs[0],
+            encoder_attention_mask=attention_mask,
+            past_key_values=past_key_values,
+            inputs_embeds=decoder_inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        if not return_dict:
+            return decoder_outputs + encoder_outputs
+
+        return Seq2SeqModelOutput(
+            last_hidden_state=decoder_outputs.last_hidden_state,
+            past_key_values=decoder_outputs.past_key_values,
+            decoder_hidden_states=decoder_outputs.hidden_states,
+            decoder_attentions=decoder_outputs.attentions,
+            cross_attentions=decoder_outputs.cross_attentions,
+            encoder_last_hidden_state=encoder_outputs.last_hidden_state,
+            encoder_hidden_states=encoder_outputs.hidden_states,
+            encoder_attentions=encoder_outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    "The PEGASUS Model with a language modeling head. Can be used for summarization.", PEGASUS_START_DOCSTRING
+)
+class PegasusForConditionalGeneration(PegasusPreTrainedModel):
+    base_model_prefix = "model"
     _keys_to_ignore_on_load_missing = [
         r"final_logits_bias",
         r"encoder\.version",
         r"decoder\.version",
-        "model.encoder.embed_positions",
-        "model.decoder.embed_positions",
-    ]
-    _keys_to_ignore_on_save = [
-        "model.encoder.embed_positions.weight",
-        "model.decoder.embed_positions.weight",
+        r"lm_head\.weight",
+        r"embed_positions\.weight",
     ]
+
+    def __init__(self, config: PegasusConfig):
+        super().__init__(config)
+        self.model = PegasusModel(config)
+        self.register_buffer("final_logits_bias", torch.zeros((1, self.model.shared.num_embeddings)))
+        self.lm_head = nn.Linear(config.d_model, self.model.shared.num_embeddings, bias=False)
+
+        self.init_weights()
+
+    def get_encoder(self):
+        return self.model.get_encoder()
+
+    def get_decoder(self):
+        return self.model.get_decoder()
+
+    def resize_token_embeddings(self, new_num_tokens: int) -> nn.Embedding:
+        new_embeddings = super().resize_token_embeddings(new_num_tokens)
+        self._resize_final_logits_bias(new_num_tokens)
+        return new_embeddings
+
+    def _resize_final_logits_bias(self, new_num_tokens: int) -> None:
+        old_num_tokens = self.final_logits_bias.shape[-1]
+        if new_num_tokens <= old_num_tokens:
+            new_bias = self.final_logits_bias[:, :new_num_tokens]
+        else:
+            extra_bias = torch.zeros((1, new_num_tokens - old_num_tokens), device=self.final_logits_bias.device)
+            new_bias = torch.cat([self.final_logits_bias, extra_bias], dim=1)
+        self.register_buffer("final_logits_bias", new_bias)
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    @add_start_docstrings_to_model_forward(PEGASUS_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=Seq2SeqLMOutput, config_class=_CONFIG_FOR_DOC)
+    @add_end_docstrings(PEGASUS_GENERATION_EXAMPLE)
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        decoder_input_ids=None,
+        decoder_attention_mask=None,
+        encoder_outputs=None,
+        past_key_values=None,
+        inputs_embeds=None,
+        decoder_inputs_embeds=None,
+        labels=None,
+        use_cache=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 either be in ``[0, ...,
+            config.vocab_size]`` or -100 (see ``input_ids`` docstring). Tokens with indices set to ``-100`` are ignored
+            (masked), the loss is only computed for the tokens with labels in ``[0, ..., config.vocab_size]``.
+
+        Returns:
+
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if labels is not None:
+            if decoder_input_ids is None:
+                decoder_input_ids = shift_tokens_right(
+                    labels, self.config.pad_token_id, self.config.decoder_start_token_id
+                )
+
+        outputs = self.model(
+            input_ids,
+            attention_mask=attention_mask,
+            decoder_input_ids=decoder_input_ids,
+            encoder_outputs=encoder_outputs,
+            decoder_attention_mask=decoder_attention_mask,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            decoder_inputs_embeds=decoder_inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        lm_logits = self.lm_head(outputs[0]) + self.final_logits_bias
+
+        masked_lm_loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            masked_lm_loss = loss_fct(lm_logits.view(-1, self.config.vocab_size), labels.view(-1))
+
+        if not return_dict:
+            output = (lm_logits,) + outputs[1:]
+            return ((masked_lm_loss,) + output) if masked_lm_loss is not None else output
+
+        return Seq2SeqLMOutput(
+            loss=masked_lm_loss,
+            logits=lm_logits,
+            past_key_values=outputs.past_key_values,
+            decoder_hidden_states=outputs.decoder_hidden_states,
+            decoder_attentions=outputs.decoder_attentions,
+            cross_attentions=outputs.cross_attentions,
+            encoder_last_hidden_state=outputs.encoder_last_hidden_state,
+            encoder_hidden_states=outputs.encoder_hidden_states,
+            encoder_attentions=outputs.encoder_attentions,
+        )
+
+    def prepare_inputs_for_generation(
+        self, decoder_input_ids, past=None, attention_mask=None, use_cache=None, encoder_outputs=None, **kwargs
+    ):
+        # cut decoder_input_ids if past is used
+        if past is not None:
+            decoder_input_ids = decoder_input_ids[:, -1:]
+
+        return {
+            "input_ids": None,  # encoder_outputs is defined. input_ids not needed
+            "encoder_outputs": encoder_outputs,
+            "past_key_values": past,
+            "decoder_input_ids": decoder_input_ids,
+            "attention_mask": attention_mask,
+            "use_cache": use_cache,  # change this to avoid caching (presumably for debugging)
+        }
+
+    def adjust_logits_during_generation(self, logits, cur_len, max_length):
+        if cur_len == max_length - 1 and self.config.eos_token_id is not None:
+            self._force_token_id_to_be_generated(logits, self.config.eos_token_id)
+        return logits
+
+    @staticmethod
+    def _force_token_id_to_be_generated(scores, token_id) -> None:
+        """force one of token_ids to be generated by setting prob of all other tokens to 0 (logprob=-float("inf"))"""
+        scores[:, [x for x in range(scores.shape[1]) if x != token_id]] = -float("inf")
+
+    @staticmethod
+    def _reorder_cache(past, beam_idx):
+        reordered_past = ()
+        for layer_past in past:
+            # cached cross_attention states don't have to be reordered -> they are always the same
+            reordered_past += (
+                tuple(past_state.index_select(0, beam_idx) for past_state in layer_past[:2]) + layer_past[2:],
+            )
+        return reordered_past

src/transformers/models/prophetnet/modeling_prophetnet.py

@@ -88,9 +88,19 @@ PROPHETNET_INPUTS_DOCSTRING = r"""
 
             `What are attention masks? <../glossary.html#attention-mask>`__
         decoder_input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, target_sequence_length)`, `optional`):
-            Provide for translation and summarization training. By default, the model will create this tensor by
-            shifting the :obj:`input_ids` to the right, following the paper.
-        decoder_attention_mask (:obj:`torch.BoolTensor` of shape :obj:`(batch_size, tgt_seq_len)`, `optional`):
+            Indices of decoder input sequence tokens in the vocabulary.
+
+            Indices can be obtained using :class:`~transformers.PreTrainedTokenizer`. See
+            :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__` for
+            details.
+
+            `What are input IDs? <../glossary.html#input-ids>`__
+
+            ProphetNet uses the :obj:`eos_token_id` as the starting token for :obj:`decoder_input_ids` generation. If
+            :obj:`past_key_values` is used, optionally only the last :obj:`decoder_input_ids` have to be input (see
+            :obj:`past_key_values`).
+
+        decoder_attention_mask (:obj:`torch.BoolTensor` of shape :obj:`(batch_size, target_sequence_length)`, `optional`):
             Default behavior: generate a tensor that ignores pad tokens in :obj:`decoder_input_ids`. Causal mask will
             also be used by default.