Upload 2 files

configuration_opt.py

@@ -0,0 +1,151 @@
+# coding=utf-8
+# Copyright 2022 The Metaseq Authors and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" OPT model configuration"""
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+OPT_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "facebook/opt-125m": "https://huggingface.co/facebook/opt-125m/blob/main/config.json",
+    "facebook/opt-350m": "https://huggingface.co/facebook/opt-350m/blob/main/config.json",
+    "facebook/opt-1.3b": "https://huggingface.co/facebook/opt-1.3b/blob/main/config.json",
+    "facebook/opt-2.7b": "https://huggingface.co/facebook/opt-2.7b/blob/main/config.json",
+    "facebook/opt-6.7b": "https://huggingface.co/facebook/opt-6.7b/blob/main/config.json",
+    "facebook/opt-13b": "https://huggingface.co/facebook/opt-13b/blob/main/config.json",
+}
+
+
+class OPTConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`OPTModel`]. It is used to instantiate a OPT 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 OPT
+    [facebook/opt-350m](https://huggingface.co/facebook/opt-350m) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 50272):
+            Vocabulary size of the OPT model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`OPTModel`]
+        hidden_size (`int`, *optional*, defaults to 768):
+            Dimensionality of the layers and the pooler layer.
+        num_hidden_layers (`int`, *optional*, defaults to 12):
+            Number of decoder layers.
+        ffn_dim (`int`, *optional*, defaults to 3072):
+            Dimensionality of the "intermediate" (often named feed-forward) layer in decoder.
+        num_attention_heads (`int`, *optional*, defaults to 12):
+            Number of attention heads for each attention layer in the Transformer decoder.
+        activation_function (`str` or `function`, *optional*, defaults to `"relu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"silu"` and `"gelu_new"` are supported.
+        max_position_embeddings (`int`, *optional*, defaults to 2048):
+            The maximum sequence length that this model might ever be used with. Typically set this to something large
+            just in case (e.g., 512 or 1024 or 2048).
+        do_layer_norm_before (`bool`, *optional*, defaults to `True`):
+            Whether to perform layer normalization before the attention block.
+        word_embed_proj_dim (`int`, *optional*):
+            `word_embed_proj_dim` can be set to down-project word embeddings, *e.g.* `opt-350m`. Defaults to
+            `hidden_size`.
+        dropout (`float`, *optional*, defaults to 0.1):
+            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        layerdrop (`float`, *optional*, defaults to 0.0):
+            The LayerDrop probability. See the [LayerDrop paper](see https://arxiv.org/abs/1909.11556) for more
+            details.
+        init_std (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models).
+        enable_bias (`bool`, *optional*, defaults to `True`):
+            Whether or not if the linear layers in the attention blocks should use the bias term.
+        layer_norm_elementwise_affine (`bool`, *optional*, defaults to `True`):
+            Whether or not if the layer norms should have learnable parameters.
+
+    Example:
+
+    ```python
+    >>> from transformers import OPTConfig, OPTModel
+
+    >>> # Initializing a OPT facebook/opt-large style configuration
+    >>> configuration = OPTConfig()
+
+    >>> # Initializing a model (with random weights) from the facebook/opt-large style configuration
+    >>> model = OPTModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "opt"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        vocab_size=50272,
+        hidden_size=768,
+        num_hidden_layers=12,
+        ffn_dim=3072,
+        max_position_embeddings=2048,
+        do_layer_norm_before=True,
+        _remove_final_layer_norm=False,
+        word_embed_proj_dim=None,
+        dropout=0.1,
+        attention_dropout=0.0,
+        num_attention_heads=12,
+        activation_function="relu",
+        layerdrop=0.0,
+        init_std=0.02,
+        use_cache=True,
+        pad_token_id=1,
+        bos_token_id=2,
+        eos_token_id=2,
+        enable_bias=True,
+        layer_norm_elementwise_affine=True,
+        **kwargs,
+    ):
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            **kwargs,
+        )
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.num_attention_heads = num_attention_heads
+        self.word_embed_proj_dim = word_embed_proj_dim if word_embed_proj_dim is not None else hidden_size
+        self.ffn_dim = ffn_dim
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.dropout = dropout
+        self.attention_dropout = attention_dropout
+        self.activation_function = activation_function
+        self.init_std = init_std
+        self.layerdrop = layerdrop
+        self.use_cache = use_cache
+        self.do_layer_norm_before = do_layer_norm_before
+        # We keep these variables at `True` for backward compatibility.
+        self.enable_bias = enable_bias
+        self.layer_norm_elementwise_affine = layer_norm_elementwise_affine
+
+        # Note that the only purpose of `_remove_final_layer_norm` is to keep backward compatibility
+        # with checkpoints that have been fine-tuned before transformers v4.20.1
+        # see https://github.com/facebookresearch/metaseq/pull/164
+        self._remove_final_layer_norm = _remove_final_layer_norm

modeling_opt.py

@@ -0,0 +1,1671 @@
+# coding=utf-8
+# Copyright 2022 The Fairseq Authors and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" PyTorch OPT model."""
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.nn.functional as F
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+
+from transformers.activations import ACT2FN
+from transformers.modeling_attn_mask_utils import _prepare_4d_causal_attention_mask
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPast,
+    CausalLMOutputWithPast,
+    QuestionAnsweringModelOutput,
+    SequenceClassifierOutputWithPast,
+)
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import (
+    add_code_sample_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    is_flash_attn_2_available,
+    is_flash_attn_greater_or_equal_2_10,
+    logging,
+    replace_return_docstrings,
+)
+from .configuration_opt import OPTConfig
+
+
+if is_flash_attn_2_available():
+    from flash_attn import flash_attn_func, flash_attn_varlen_func
+    from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input  # noqa
+
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "facebook/opt-350m"
+_CONFIG_FOR_DOC = "OPTConfig"
+
+# Base model docstring
+_EXPECTED_OUTPUT_SHAPE = [1, 8, 1024]
+
+# SequenceClassification docstring
+_CHECKPOINT_FOR_SEQUENCE_CLASSIFICATION = "ArthurZ/opt-350m-dummy-sc"
+_SEQ_CLASS_EXPECTED_LOSS = 1.71
+_SEQ_CLASS_EXPECTED_OUTPUT = "'LABEL_0'"
+
+OPT_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "facebook/opt-125m",
+    "facebook/opt-350m",
+    "facebook/opt-1.3b",
+    "facebook/opt-2.7b",
+    "facebook/opt-6.7b",
+    "facebook/opt-13b",
+    "facebook/opt-30b",
+    # See all OPT models at https://huggingface.co/models?filter=opt
+]
+
+
+# Copied from transformers.models.llama.modeling_llama._get_unpad_data
+def _get_unpad_data(attention_mask):
+    seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
+    indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
+    max_seqlen_in_batch = seqlens_in_batch.max().item()
+    cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.int32), (1, 0))
+    return (
+        indices,
+        cu_seqlens,
+        max_seqlen_in_batch,
+    )
+
+
+# class OPTLearnedPositionalEmbedding(nn.Embedding):
+#     """
+#     This module learns positional embeddings up to a fixed maximum size.
+#     """
+
+#     def __init__(self, num_embeddings: int, embedding_dim: int):
+#         # OPT is set up so that if padding_idx is specified then offset the embedding ids by 2
+#         # and adjust num_embeddings appropriately. Other models don't have this hack
+#         self.offset = 2
+#         super().__init__(num_embeddings + self.offset, embedding_dim)
+
+#     def forward(self, attention_mask: torch.LongTensor, past_key_values_length: int = 0):
+#         """`input_ids_shape` is expected to be [bsz x seqlen]."""
+#         attention_mask = attention_mask.long()
+
+#         # create positions depending on attention_mask
+#         positions = (torch.cumsum(attention_mask, dim=1).type_as(attention_mask) * attention_mask).long() - 1
+
+#         # cut positions if `past_key_values_length` is > 0
+#         positions = positions[:, past_key_values_length:]
+
+#         return super().forward(positions + self.offset)
+
+
+class OPTLearnedPositionalEmbedding(nn.Module):
+    """
+    This module learns positional embeddings up to a fixed maximum size.
+    """
+
+    def __init__(self, num_embeddings: int, embedding_dim: int):
+        # OPT is set up so that if padding_idx is specified then offset the embedding ids by 2
+        # and adjust num_embeddings appropriately. Other models don't have this hack
+        super().__init__()
+        self.offset = 2
+        self.embeddings = nn.Embedding(num_embeddings + self.offset, embedding_dim)
+
+    def forward(
+        self, attention_mask: torch.LongTensor, past_key_values_length: int = 0
+    ):
+        """`input_ids_shape` is expected to be [bsz x seqlen]."""
+        attention_mask = attention_mask.long()
+
+        # create positions depending on attention_mask
+        positions = (
+            torch.cumsum(attention_mask, dim=1).type_as(attention_mask) * attention_mask
+        ).long() - 1
+
+        # cut positions if `past_key_values_length` is > 0
+        positions = positions[:, past_key_values_length:]
+
+        return self.embeddings(positions + self.offset)
+
+
+class OPTAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(
+        self,
+        config: OPTConfig,
+        is_decoder: bool = False,
+        **kwargs,
+    ):
+        super().__init__()
+        self.config = config
+
+        def _handle_deprecated_argument(config_arg_name, config, fn_arg_name, kwargs):
+            """
+            If a the deprecated argument `fn_arg_name` is passed, raise a deprecation
+            warning and return that value, otherwise take the equivalent config.config_arg_name
+            """
+            val = None
+            if fn_arg_name in kwargs:
+                logging.warning(
+                    "Passing in {fn_arg_name} to {self.__class__.__name__} is deprecated and won't be supported from "
+                    "v4.39. Please set it in the config instead"
+                )
+                val = kwargs.pop(fn_arg_name)
+            else:
+                val = getattr(config, config_arg_name)
+            return val
+
+        self.embed_dim = _handle_deprecated_argument(
+            "hidden_size", config, "embed_dim", kwargs
+        )
+        self.num_heads = _handle_deprecated_argument(
+            "num_attention_heads", config, "num_heads", kwargs
+        )
+        self.dropout = _handle_deprecated_argument(
+            "attention_dropout", config, "dropout", kwargs
+        )
+        self.enable_bias = _handle_deprecated_argument(
+            "enable_bias", config, "bias", kwargs
+        )
+
+        self.head_dim = self.embed_dim // self.num_heads
+        self.is_causal = True
+
+        if (self.head_dim * self.num_heads) != self.embed_dim:
+            raise ValueError(
+                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim}"
+                f" and `num_heads`: {self.num_heads})."
+            )
+        self.scaling = self.head_dim**-0.5
+        self.is_decoder = is_decoder
+
+        self.k_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=self.enable_bias)
+        self.v_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=self.enable_bias)
+        self.q_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=self.enable_bias)
+        self.out_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=self.enable_bias)
+
+    def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
+        return (
+            tensor.view(bsz, seq_len, self.num_heads, self.head_dim)
+            .transpose(1, 2)
+            .contiguous()
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        key_value_states: Optional[torch.Tensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        layer_head_mask: Optional[torch.Tensor] = None,
+        output_attentions: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        """Input shape: Batch x Time x Channel"""
+
+        # if key_value_states are provided this layer is used as a cross-attention layer
+        # for the decoder
+        is_cross_attention = key_value_states is not None
+
+        bsz, tgt_len, _ = hidden_states.size()
+
+        # get query proj
+        query_states = self.q_proj(hidden_states) * self.scaling
+        # get key, value proj
+        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))
+
+        if attn_weights.size() != (bsz * self.num_heads, tgt_len, src_len):
+            raise ValueError(
+                f"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is"
+                f" {attn_weights.size()}"
+            )
+
+        if attention_mask is not None:
+            if attention_mask.size() != (bsz, 1, tgt_len, src_len):
+                raise ValueError(
+                    f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.size()}"
+                )
+            attn_weights = (
+                attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
+                + attention_mask
+            )
+            attn_weights = torch.max(
+                attn_weights,
+                torch.tensor(
+                    torch.finfo(attn_weights.dtype).min, device=attn_weights.device
+                ),
+            )
+            attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
+
+        # upcast to fp32 if the weights are in fp16. Please see https://github.com/huggingface/transformers/pull/17437
+        if attn_weights.dtype == torch.float16:
+            attn_weights = nn.functional.softmax(
+                attn_weights, dim=-1, dtype=torch.float32
+            ).to(torch.float16)
+        else:
+            attn_weights = nn.functional.softmax(attn_weights, dim=-1)
+
+        if layer_head_mask is not None:
+            if layer_head_mask.size() != (self.num_heads,):
+                raise ValueError(
+                    f"Head mask for a single layer should be of size {(self.num_heads,)}, but is"
+                    f" {layer_head_mask.size()}"
+                )
+            attn_weights = layer_head_mask.view(1, -1, 1, 1) * attn_weights.view(
+                bsz, self.num_heads, tgt_len, src_len
+            )
+            attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
+
+        if output_attentions:
+            # this operation is a bit awkward, but it's required to
+            # make sure that attn_weights keeps its gradient.
+            # In order to do so, attn_weights have to be reshaped
+            # twice and have to be reused in the following
+            attn_weights_reshaped = attn_weights.view(
+                bsz, self.num_heads, tgt_len, src_len
+            )
+            attn_weights = attn_weights_reshaped.view(
+                bsz * self.num_heads, tgt_len, src_len
+            )
+        else:
+            attn_weights_reshaped = None
+
+        attn_probs = nn.functional.dropout(
+            attn_weights, p=self.dropout, training=self.training
+        )
+
+        attn_output = torch.bmm(attn_probs, value_states)
+
+        if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
+                f" {attn_output.size()}"
+            )
+
+        attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
+        attn_output = attn_output.transpose(1, 2)
+
+        # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
+        # partitioned aross GPUs when using tensor-parallelism.
+        attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)
+
+        attn_output = self.out_proj(attn_output)
+
+        return attn_output, attn_weights_reshaped, past_key_value
+
+
+class OptFlashAttention2(OPTAttention):
+    """
+    OPT flash attention module. This module inherits from `OPTAttention` as the weights of the module stays untouched.
+    The only required change would be on the forward pass where it needs to correctly call the public API of flash
+    attention and deal with padding tokens in case the input contains any of them.
+    """
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2.__init__
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
+        # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
+        # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
+        self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        key_value_states: Optional[torch.Tensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        layer_head_mask: Optional[torch.Tensor] = None,
+        output_attentions: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        """Input shape: Batch x Time x Channel"""
+
+        # if key_value_states are provided this layer is used as a cross-attention layer
+        # for the decoder
+        is_cross_attention = key_value_states is not None
+
+        bsz, _, _ = hidden_states.size()
+
+        # get query proj
+        query_states = self.q_proj(hidden_states)
+        # 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)
+
+        query_length = query_states.shape[1]
+        tgt_len = key_states.shape[-2]
+
+        # Flash attention requires the input to have the shape
+        # batch_size x seq_length x head_dim x hidden_dim
+        query_states = query_states.view(
+            bsz, query_length, self.num_heads, self.head_dim
+        )
+        key_states = key_states.transpose(1, 2).view(
+            bsz, tgt_len, self.num_heads, self.head_dim
+        )
+        value_states = value_states.transpose(1, 2).view(
+            bsz, tgt_len, self.num_heads, self.head_dim
+        )
+
+        attn_dropout = self.dropout if self.training else 0.0
+
+        # In PEFT, usually we cast the layer norms in float32 for training stability reasons
+        # therefore the input hidden states gets silently casted in float32. Hence, we need
+        # cast them back in float16 just to be sure everything works as expected.
+        input_dtype = query_states.dtype
+        if input_dtype == torch.float32:
+            if torch.is_autocast_enabled():
+                target_dtype = torch.get_autocast_gpu_dtype()
+            # Handle the case where the model is quantized
+            elif hasattr(self.config, "_pre_quantization_dtype"):
+                target_dtype = self.config._pre_quantization_dtype
+            else:
+                target_dtype = self.q_proj.weight.dtype
+
+            logger.warning_once(
+                f"The input hidden states seems to be silently casted in float32, this might be related to"
+                f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
+                f" {target_dtype}."
+            )
+
+            query_states = query_states.to(target_dtype)
+            key_states = key_states.to(target_dtype)
+            value_states = value_states.to(target_dtype)
+
+        attn_output = self._flash_attention_forward(
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            query_length,
+            dropout=attn_dropout,
+        )
+
+        attn_weights_reshaped = attn_output.reshape(
+            bsz, query_length, self.num_heads * self.head_dim
+        )
+        attn_output = self.out_proj(attn_weights_reshaped)
+
+        if not output_attentions:
+            attn_weights_reshaped = None
+
+        return attn_output, attn_weights_reshaped, past_key_value
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2._flash_attention_forward
+    def _flash_attention_forward(
+        self,
+        query_states,
+        key_states,
+        value_states,
+        attention_mask,
+        query_length,
+        dropout=0.0,
+        softmax_scale=None,
+    ):
+        """
+        Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token
+        first unpad the input, then computes the attention scores and pad the final attention scores.
+
+        Args:
+            query_states (`torch.Tensor`):
+                Input query states to be passed to Flash Attention API
+            key_states (`torch.Tensor`):
+                Input key states to be passed to Flash Attention API
+            value_states (`torch.Tensor`):
+                Input value states to be passed to Flash Attention API
+            attention_mask (`torch.Tensor`):
+                The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
+                position of padding tokens and 1 for the position of non-padding tokens.
+            dropout (`int`, *optional*):
+                Attention dropout
+            softmax_scale (`float`, *optional*):
+                The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
+        """
+        if not self._flash_attn_uses_top_left_mask:
+            causal = self.is_causal
+        else:
+            # TODO: Remove the `query_length != 1` check once Flash Attention for RoCm is bumped to 2.1. For details, please see the comment in LlamaFlashAttention2 __init__.
+            causal = self.is_causal and query_length != 1
+
+        # Contains at least one padding token in the sequence
+        if attention_mask is not None:
+            batch_size = query_states.shape[0]
+            (
+                query_states,
+                key_states,
+                value_states,
+                indices_q,
+                cu_seq_lens,
+                max_seq_lens,
+            ) = self._upad_input(
+                query_states, key_states, value_states, attention_mask, query_length
+            )
+
+            cu_seqlens_q, cu_seqlens_k = cu_seq_lens
+            max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens
+
+            attn_output_unpad = flash_attn_varlen_func(
+                query_states,
+                key_states,
+                value_states,
+                cu_seqlens_q=cu_seqlens_q,
+                cu_seqlens_k=cu_seqlens_k,
+                max_seqlen_q=max_seqlen_in_batch_q,
+                max_seqlen_k=max_seqlen_in_batch_k,
+                dropout_p=dropout,
+                softmax_scale=softmax_scale,
+                causal=causal,
+            )
+
+            attn_output = pad_input(
+                attn_output_unpad, indices_q, batch_size, query_length
+            )
+        else:
+            attn_output = flash_attn_func(
+                query_states,
+                key_states,
+                value_states,
+                dropout,
+                softmax_scale=softmax_scale,
+                causal=causal,
+            )
+
+        return attn_output
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2._upad_input
+    def _upad_input(
+        self, query_layer, key_layer, value_layer, attention_mask, query_length
+    ):
+        indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)
+        batch_size, kv_seq_len, num_key_value_heads, head_dim = key_layer.shape
+
+        key_layer = index_first_axis(
+            key_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim),
+            indices_k,
+        )
+        value_layer = index_first_axis(
+            value_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim),
+            indices_k,
+        )
+        if query_length == kv_seq_len:
+            query_layer = index_first_axis(
+                query_layer.reshape(batch_size * kv_seq_len, self.num_heads, head_dim),
+                indices_k,
+            )
+            cu_seqlens_q = cu_seqlens_k
+            max_seqlen_in_batch_q = max_seqlen_in_batch_k
+            indices_q = indices_k
+        elif query_length == 1:
+            max_seqlen_in_batch_q = 1
+            cu_seqlens_q = torch.arange(
+                batch_size + 1, dtype=torch.int32, device=query_layer.device
+            )  # There is a memcpy here, that is very bad.
+            indices_q = cu_seqlens_q[:-1]
+            query_layer = query_layer.squeeze(1)
+        else:
+            # The -q_len: slice assumes left padding.
+            attention_mask = attention_mask[:, -query_length:]
+            query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(
+                query_layer, attention_mask
+            )
+
+        return (
+            query_layer,
+            key_layer,
+            value_layer,
+            indices_q,
+            (cu_seqlens_q, cu_seqlens_k),
+            (max_seqlen_in_batch_q, max_seqlen_in_batch_k),
+        )
+
+
+OPT_ATTENTION_CLASSES = {
+    "eager": OPTAttention,
+    "flash_attention_2": OptFlashAttention2,
+}
+
+
+class OPTDecoderLayer(nn.Module):
+    def __init__(self, config: OPTConfig):
+        super().__init__()
+        self.embed_dim = config.hidden_size
+
+        self.self_attn = OPT_ATTENTION_CLASSES[config._attn_implementation](
+            config=config, is_decoder=True
+        )
+
+        self.do_layer_norm_before = config.do_layer_norm_before
+        self.dropout = config.dropout
+        self.activation_fn = ACT2FN[config.activation_function]
+
+        self.self_attn_layer_norm = nn.LayerNorm(
+            self.embed_dim, elementwise_affine=config.layer_norm_elementwise_affine
+        )
+        self.fc1 = nn.Linear(self.embed_dim, config.ffn_dim, bias=config.enable_bias)
+        self.fc2 = nn.Linear(config.ffn_dim, self.embed_dim, bias=config.enable_bias)
+        self.final_layer_norm = nn.LayerNorm(
+            self.embed_dim, elementwise_affine=config.layer_norm_elementwise_affine
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        layer_head_mask: Optional[torch.Tensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+    ) -> Tuple[
+        torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]
+    ]:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+            attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            layer_head_mask (`torch.FloatTensor`, *optional*): mask for attention heads in a given layer of size
+                `(encoder_attention_heads,)`.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            use_cache (`bool`, *optional*):
+                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
+                (see `past_key_values`).
+            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
+        """
+
+        residual = hidden_states
+
+        # 125m, 1.7B, ..., 175B applies layer norm BEFORE attention
+        if self.do_layer_norm_before:
+            hidden_states = self.self_attn_layer_norm(hidden_states)
+
+        # Self Attention
+        hidden_states, self_attn_weights, present_key_value = self.self_attn(
+            hidden_states=hidden_states,
+            past_key_value=past_key_value,
+            attention_mask=attention_mask,
+            layer_head_mask=layer_head_mask,
+            output_attentions=output_attentions,
+        )
+        hidden_states = nn.functional.dropout(
+            hidden_states, p=self.dropout, training=self.training
+        )
+        hidden_states = residual + hidden_states
+
+        # 350m applies layer norm AFTER attention
+        if not self.do_layer_norm_before:
+            hidden_states = self.self_attn_layer_norm(hidden_states)
+
+        # Fully Connected
+        hidden_states_shape = hidden_states.shape
+        hidden_states = hidden_states.reshape(-1, hidden_states.size(-1))
+        residual = hidden_states
+
+        # 125m, 1.7B, ..., 175B applies layer norm BEFORE attention
+        if self.do_layer_norm_before:
+            hidden_states = self.final_layer_norm(hidden_states)
+
+        hidden_states = self.fc1(hidden_states)
+        hidden_states = self.activation_fn(hidden_states)
+
+        hidden_states = self.fc2(hidden_states)
+        hidden_states = nn.functional.dropout(
+            hidden_states, p=self.dropout, training=self.training
+        )
+
+        hidden_states = (residual + hidden_states).view(hidden_states_shape)
+
+        # 350m applies layer norm AFTER attention
+        if not self.do_layer_norm_before:
+            hidden_states = self.final_layer_norm(hidden_states)
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        if use_cache:
+            outputs += (present_key_value,)
+
+        return outputs
+
+
+OPT_START_DOCSTRING = r"""
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+
+    Parameters:
+        config ([`OPTConfig`]):
+            Model configuration class with all the parameters of the model. Initializing with a config file does not
+            load the weights associated with the model, only the configuration. Check out the
+            [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+
+@add_start_docstrings(
+    "The bare OPT Model outputting raw hidden-states without any specific head on top.",
+    OPT_START_DOCSTRING,
+)
+class OPTPreTrainedModel(PreTrainedModel):
+    config_class = OPTConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["OPTDecoderLayer"]
+    _supports_flash_attn_2 = True
+
+    def _init_weights(self, module):
+        std = self.config.init_std
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+
+OPT_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
+            `past_key_values`).
+
+            If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
+            and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
+            information on the default strategy.
+        head_mask (`torch.Tensor` of shape `(encoder_layers, encoder_attention_heads)`, *optional*):
+            Mask to nullify selected heads of the attention modules in the encoder. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
+            Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
+            `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
+            `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
+
+            Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
+            blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
+
+            If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
+            don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
+            `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
+            model's internal embedding lookup matrix.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+            `past_key_values`).
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+class OPTDecoder(OPTPreTrainedModel):
+    """
+    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`OPTDecoderLayer`]
+
+    Args:
+        config: OPTConfig
+    """
+
+    def __init__(self, config: OPTConfig):
+        super().__init__(config)
+        self.dropout = config.dropout
+        self.layerdrop = config.layerdrop
+        self.padding_idx = config.pad_token_id
+        self.max_target_positions = config.max_position_embeddings
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = nn.Embedding(
+            config.vocab_size, config.word_embed_proj_dim, self.padding_idx
+        )
+        self._embed_positions = OPTLearnedPositionalEmbedding(
+            config.max_position_embeddings, config.hidden_size
+        )
+        self.embed_positions = self._embed_positions.embeddings
+
+        if config.word_embed_proj_dim != config.hidden_size:
+            self.project_out = nn.Linear(
+                config.hidden_size, config.word_embed_proj_dim, bias=False
+            )
+        else:
+            self.project_out = None
+
+        if config.word_embed_proj_dim != config.hidden_size:
+            self.project_in = nn.Linear(
+                config.word_embed_proj_dim, config.hidden_size, bias=False
+            )
+        else:
+            self.project_in = None
+
+        # Note that the only purpose of `config._remove_final_layer_norm` is to keep backward compatibility
+        # with checkpoints that have been fine-tuned before transformers v4.20.1
+        # see https://github.com/facebookresearch/metaseq/pull/164
+        if config.do_layer_norm_before and not config._remove_final_layer_norm:
+            self.final_layer_norm = nn.LayerNorm(
+                config.hidden_size,
+                elementwise_affine=config.layer_norm_elementwise_affine,
+            )
+        else:
+            self.final_layer_norm = None
+
+        self.layers = nn.ModuleList(
+            [OPTDecoderLayer(config) for _ in range(config.num_hidden_layers)]
+        )
+        self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"
+
+        self.gradient_checkpointing = False
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        r"""
+        Args:
+            input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+                Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you
+                provide it.
+
+                Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+                [`PreTrainedTokenizer.__call__`] for details.
+
+                [What are input IDs?](../glossary#input-ids)
+            attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+                - 1 for tokens that are **not masked**,
+                - 0 for tokens that are **masked**.
+
+                [What are attention masks?](../glossary#attention-mask)
+            head_mask (`torch.Tensor` of shape `(num_hidden_layers, num_attention_heads)`, *optional*):
+                Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:
+
+                - 1 indicates the head is **not masked**,
+                - 0 indicates the head is **masked**.
+
+            past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
+                Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
+                shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of
+
+                Contains pre-computed hidden-states (key and values in the self-attention blocks and in the
+                cross-attention blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
+
+                If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those
+                that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of
+                all `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+
+            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+                Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
+                This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+                than the model's internal embedding lookup matrix.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            output_hidden_states (`bool`, *optional*):
+                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
+                for more detail.
+            return_dict (`bool`, *optional*):
+                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+        """
+        output_attentions = (
+            output_attentions
+            if output_attentions is not None
+            else self.config.output_attentions
+        )
+        output_hidden_states = (
+            output_hidden_states
+            if output_hidden_states is not None
+            else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        # retrieve input_ids and inputs_embeds
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError(
+                "You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time"
+            )
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+            input_ids = input_ids.view(-1, input_shape[-1])
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+        else:
+            raise ValueError(
+                "You have to specify either decoder_input_ids or decoder_inputs_embeds"
+            )
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        batch_size, seq_length = input_shape
+        past_key_values_length = (
+            past_key_values[0][0].shape[2] if past_key_values is not None else 0
+        )
+        # required mask seq length can be calculated via length of past
+        mask_seq_length = past_key_values_length + seq_length
+
+        # embed positions
+        if self._use_flash_attention_2:
+            # 2d mask is passed through the layers
+            causal_attention_mask = (
+                attention_mask
+                if (attention_mask is not None and 0 in attention_mask)
+                else None
+            )
+            attention_mask = (
+                torch.ones(batch_size, mask_seq_length, device=inputs_embeds.device)
+                if attention_mask is None
+                else attention_mask
+            )
+        else:
+            # 4d mask is passed through the layers
+            if attention_mask is None:
+                attention_mask = torch.ones(
+                    batch_size, mask_seq_length, device=inputs_embeds.device
+                )
+            elif attention_mask.shape[1] != mask_seq_length:
+                raise ValueError(
+                    f"The provided attention mask has length {attention_mask.shape[1]}, but its length should be "
+                    f"{mask_seq_length} (sum of the lengths of current and past inputs)"
+                )
+            causal_attention_mask = _prepare_4d_causal_attention_mask(
+                attention_mask, input_shape, inputs_embeds, past_key_values_length
+            )
+
+        pos_embeds = self._embed_positions(attention_mask, past_key_values_length)
+
+        if self.project_in is not None:
+            inputs_embeds = self.project_in(inputs_embeds)
+
+        hidden_states = inputs_embeds + pos_embeds
+
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                logger.warning_once(
+                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                )
+                use_cache = False
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        next_decoder_cache = () if use_cache else None
+
+        # check if head_mask has a correct number of layers specified if desired
+        for attn_mask, mask_name in zip([head_mask], ["head_mask"]):
+            if attn_mask is not None:
+                if attn_mask.size()[0] != (len(self.layers)):
+                    raise ValueError(
+                        f"The `{mask_name}` should be specified for {len(self.layers)} layers, but it is for"
+                        f" {head_mask.size()[0]}."
+                    )
+
+        for idx, decoder_layer in enumerate(self.layers):
+            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            if self.training:
+                dropout_probability = torch.rand([])
+                if dropout_probability < self.layerdrop:
+                    continue
+
+            past_key_value = (
+                past_key_values[idx] if past_key_values is not None else None
+            )
+
+            if self.gradient_checkpointing and self.training:
+                layer_outputs = self._gradient_checkpointing_func(
+                    decoder_layer.__call__,
+                    hidden_states,
+                    causal_attention_mask,
+                    head_mask[idx] if head_mask is not None else None,
+                    None,
+                    output_attentions,
+                    use_cache,
+                )
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask=causal_attention_mask,
+                    layer_head_mask=(head_mask[idx] if head_mask is not None else None),
+                    past_key_value=past_key_value,
+                    output_attentions=output_attentions,
+                    use_cache=use_cache,
+                )
+
+            hidden_states = layer_outputs[0]
+
+            if use_cache:
+                next_decoder_cache += (layer_outputs[2 if output_attentions else 1],)
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+        if self.final_layer_norm is not None:
+            hidden_states = self.final_layer_norm(hidden_states)
+
+        if self.project_out is not None:
+            hidden_states = self.project_out(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]
+                if v is not None
+            )
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+        )
+
+
+@add_start_docstrings(
+    "The bare OPT Model outputting raw hidden-states without any specific head on top.",
+    OPT_START_DOCSTRING,
+)
+class OPTModel(OPTPreTrainedModel):
+    def __init__(self, config: OPTConfig):
+        super().__init__(config)
+        self.decoder = OPTDecoder(config)
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.decoder.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.decoder.embed_tokens = value
+
+    def get_decoder(self):
+        return self.decoder
+
+    @add_start_docstrings_to_model_forward(OPT_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=BaseModelOutputWithPast,
+        config_class=_CONFIG_FOR_DOC,
+        expected_output=_EXPECTED_OUTPUT_SHAPE,
+    )
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        output_attentions = (
+            output_attentions
+            if output_attentions is not None
+            else self.config.output_attentions
+        )
+        output_hidden_states = (
+            output_hidden_states
+            if output_hidden_states is not None
+            else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        # decoder outputs consists of (dec_features, past_key_value, dec_hidden, dec_attn)
+        decoder_outputs = self.decoder(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            head_mask=head_mask,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        if not return_dict:
+            return decoder_outputs
+
+        return BaseModelOutputWithPast(
+            last_hidden_state=decoder_outputs.last_hidden_state,
+            past_key_values=decoder_outputs.past_key_values,
+            hidden_states=decoder_outputs.hidden_states,
+            attentions=decoder_outputs.attentions,
+        )
+
+
+class OPTForCausalLM(OPTPreTrainedModel):
+    _tied_weights_keys = ["lm_head.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = OPTModel(config)
+
+        # the lm_head weight is automatically tied to the embed tokens weight
+        self.lm_head = nn.Linear(
+            config.word_embed_proj_dim, config.vocab_size, bias=False
+        )
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.decoder.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.decoder.embed_tokens = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def set_decoder(self, decoder):
+        self.model.decoder = decoder
+
+    def get_decoder(self):
+        return self.model.decoder
+
+    @replace_return_docstrings(
+        output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
+    )
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        r"""
+        Args:
+            input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+                Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you
+                provide it.
+
+                Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+                [`PreTrainedTokenizer.__call__`] for details.
+
+                [What are input IDs?](../glossary#input-ids)
+            attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+                - 1 for tokens that are **not masked**,
+                - 0 for tokens that are **masked**.
+
+                [What are attention masks?](../glossary#attention-mask)
+            head_mask (`torch.Tensor` of shape `(num_hidden_layers, num_attention_heads)`, *optional*):
+                Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:
+
+                - 1 indicates the head is **not masked**,
+                - 0 indicates the head is **masked**.
+
+            past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
+                Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
+                shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of
+                shape `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`. The two additional
+                tensors are only required when the model is used as a decoder in a Sequence to Sequence model.
+
+                Contains pre-computed hidden-states (key and values in the self-attention blocks and in the
+                cross-attention blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
+
+                If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those
+                that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of
+                all `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+                Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
+                This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+                than the model's internal embedding lookup matrix.
+            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+                config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+                (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+            use_cache (`bool`, *optional*):
+                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
+                (see `past_key_values`).
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            output_hidden_states (`bool`, *optional*):
+                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
+                for more detail.
+            return_dict (`bool`, *optional*):
+                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer, OPTForCausalLM
+
+        >>> model = OPTForCausalLM.from_pretrained("facebook/opt-350m")
+        >>> tokenizer = AutoTokenizer.from_pretrained("facebook/opt-350m")
+
+        >>> prompt = "Hey, are you conscious? Can you talk to me?"
+        >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+        >>> # Generate
+        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+        "Hey, are you conscious? Can you talk to me?\nI'm not conscious. I'm just a little bit of a weirdo."
+        ```"""
+
+        output_attentions = (
+            output_attentions
+            if output_attentions is not None
+            else self.config.output_attentions
+        )
+        output_hidden_states = (
+            output_hidden_states
+            if output_hidden_states is not None
+            else self.config.output_hidden_states
+        )
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        outputs = self.model.decoder(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            head_mask=head_mask,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        logits = self.lm_head(outputs[0]).contiguous()
+
+        loss = None
+        if labels is not None:
+            # move labels to correct device to enable model parallelism
+            labels = labels.to(logits.device)
+            # Shift so that tokens < n predict n
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(
+                shift_logits.view(-1, self.config.vocab_size), shift_labels.view(-1)
+            )
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    def prepare_inputs_for_generation(
+        self,
+        input_ids,
+        past_key_values=None,
+        attention_mask=None,
+        inputs_embeds=None,
+        **kwargs,
+    ):
+        if past_key_values is not None:
+            past_length = past_key_values[0][0].shape[2]
+
+            # Some generation methods already pass only the last input ID
+            if input_ids.shape[1] > past_length:
+                remove_prefix_length = past_length
+            else:
+                # Default to old behavior: keep only final ID
+                remove_prefix_length = input_ids.shape[1] - 1
+
+            input_ids = input_ids[:, remove_prefix_length:]
+
+        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
+        if inputs_embeds is not None and past_key_values is None:
+            model_inputs = {"inputs_embeds": inputs_embeds}
+        else:
+            model_inputs = {"input_ids": input_ids}
+
+        model_inputs.update(
+            {
+                "past_key_values": past_key_values,
+                "use_cache": kwargs.get("use_cache"),
+                "attention_mask": attention_mask,
+            }
+        )
+        return model_inputs
+
+    @staticmethod
+    def _reorder_cache(past_key_values, beam_idx):
+        reordered_past = ()
+        for layer_past in past_key_values:
+            reordered_past += (
+                tuple(
+                    past_state.index_select(0, beam_idx.to(past_state.device))
+                    for past_state in layer_past
+                ),
+            )
+        return reordered_past
+
+
+@add_start_docstrings(
+    """
+    The OPT Model transformer with a sequence classification head on top (linear layer).
+
+    [`OPTForSequenceClassification`] uses the last token in order to do the classification, as other causal models
+    (e.g. GPT-2) do.
+
+    Since it does classification on the last token, it requires to know the position of the last token. If a
+    `pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
+    no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
+    padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in
+    each row of the batch).
+    """,
+    OPT_START_DOCSTRING,
+)
+class OPTForSequenceClassification(OPTPreTrainedModel):
+    def __init__(self, config: OPTConfig):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.model = OPTModel(config)
+        self.score = nn.Linear(config.word_embed_proj_dim, self.num_labels, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(OPT_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_SEQUENCE_CLASSIFICATION,
+        output_type=SequenceClassifierOutputWithPast,
+        config_class=_CONFIG_FOR_DOC,
+        expected_output=_SEQ_CLASS_EXPECTED_OUTPUT,
+        expected_loss=_SEQ_CLASS_EXPECTED_LOSS,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, SequenceClassifierOutputWithPast]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        transformer_outputs = self.model(
+            input_ids,
+            past_key_values=past_key_values,
+            attention_mask=attention_mask,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        hidden_states = transformer_outputs[0]
+        logits = self.score(hidden_states)
+
+        if input_ids is not None:
+            batch_size, sequence_length = input_ids.shape[:2]
+        else:
+            batch_size, sequence_length = inputs_embeds.shape[:2]
+
+        if self.config.pad_token_id is None:
+            sequence_lengths = -1
+        else:
+            if input_ids is not None:
+                # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
+                sequence_lengths = (
+                    torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1
+                )
+                sequence_lengths = sequence_lengths % input_ids.shape[-1]
+                sequence_lengths = sequence_lengths.to(logits.device)
+            else:
+                sequence_lengths = -1
+                logger.warning(
+                    f"{self.__class__.__name__} will not detect padding tokens in `inputs_embeds`. Results may be "
+                    "unexpected if using padding tokens in conjunction with `inputs_embeds.`"
+                )
+
+        pooled_logits = logits[
+            torch.arange(batch_size, device=logits.device), sequence_lengths
+        ]
+
+        loss = None
+        if labels is not None:
+            if self.config.problem_type is None:
+                if self.num_labels == 1:
+                    self.config.problem_type = "regression"
+                elif self.num_labels > 1 and (
+                    labels.dtype == torch.long or labels.dtype == torch.int
+                ):
+                    self.config.problem_type = "single_label_classification"
+                else:
+                    self.config.problem_type = "multi_label_classification"
+
+            if self.config.problem_type == "regression":
+                loss_fct = MSELoss()
+                if self.num_labels == 1:
+                    loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(pooled_logits, labels)
+            elif self.config.problem_type == "single_label_classification":
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(
+                    pooled_logits.view(-1, self.num_labels), labels.view(-1)
+                )
+            elif self.config.problem_type == "multi_label_classification":
+                loss_fct = BCEWithLogitsLoss()
+                loss = loss_fct(pooled_logits, labels)
+        if not return_dict:
+            output = (pooled_logits,) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return SequenceClassifierOutputWithPast(
+            loss=loss,
+            logits=pooled_logits,
+            past_key_values=transformer_outputs.past_key_values,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
+
+    def get_input_embeddings(self):
+        return self.model.decoder.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.decoder.embed_tokens = value
+
+
+@add_start_docstrings(
+    """
+    The OPT Model transformer with a span classification head on top for extractive question-answering tasks like SQuAD
+    (a linear layers on top of the hidden-states output to compute `span start logits` and `span end logits`).
+    """,
+    OPT_START_DOCSTRING,
+)
+class OPTForQuestionAnswering(OPTPreTrainedModel):
+    def __init__(self, config: OPTConfig):
+        super().__init__(config)
+        self.model = OPTModel(config)
+        self.qa_outputs = nn.Linear(config.word_embed_proj_dim, 2)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(OPT_INPUTS_DOCSTRING)
+    @replace_return_docstrings(
+        output_type=QuestionAnsweringModelOutput, config_class=_CONFIG_FOR_DOC
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        start_positions: Optional[torch.LongTensor] = None,
+        end_positions: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, QuestionAnsweringModelOutput]:
+        r"""
+        start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for position (index) of the start of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
+            are not taken into account for computing the loss.
+        end_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for position (index) of the end of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
+            are not taken into account for computing the loss.
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer, OPTForQuestionAnswering
+        >>> import torch
+
+        >>> torch.manual_seed(4)  # doctest: +IGNORE_RESULT
+        >>> tokenizer = AutoTokenizer.from_pretrained("facebook/opt-350m")
+
+        >>> # note: we are loading a OPTForQuestionAnswering from the hub here,
+        >>> # so the head will be randomly initialized, hence the predictions will be random
+        >>> model = OPTForQuestionAnswering.from_pretrained("facebook/opt-350m")
+
+        >>> question, text = "Who was Jim Henson?", "Jim Henson was a nice puppet"
+
+        >>> inputs = tokenizer(question, text, return_tensors="pt")
+        >>> with torch.no_grad():
+        ...     outputs = model(**inputs)
+
+        >>> answer_start_index = outputs.start_logits.argmax()
+        >>> answer_end_index = outputs.end_logits.argmax()
+
+        >>> answer_offset = len(tokenizer(question)[0])
+
+        >>> predict_answer_tokens = inputs.input_ids[
+        ...     0, answer_offset + answer_start_index : answer_offset + answer_end_index + 1
+        ... ]
+        >>> predicted = tokenizer.decode(predict_answer_tokens)
+        >>> predicted
+        ' a nice puppet'
+        ```"""
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        transformer_outputs = self.model(
+            input_ids,
+            past_key_values=past_key_values,
+            attention_mask=attention_mask,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        hidden_states = transformer_outputs[0]
+
+        logits = self.qa_outputs(hidden_states)
+        start_logits, end_logits = logits.split(1, dim=-1)
+        start_logits = start_logits.squeeze(-1).contiguous()
+        end_logits = end_logits.squeeze(-1).contiguous()
+
+        total_loss = None
+        if start_positions is not None and end_positions is not None:
+            # If we are on multi-GPU, split add a dimension
+            if len(start_positions.size()) > 1:
+                start_positions = start_positions.squeeze(-1)
+            if len(end_positions.size()) > 1:
+                end_positions = end_positions.squeeze(-1)
+            # sometimes the start/end positions are outside our model inputs, we ignore these terms
+            ignored_index = start_logits.size(1)
+            start_positions = start_positions.clamp(0, ignored_index)
+            end_positions = end_positions.clamp(0, ignored_index)
+
+            loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
+            start_loss = loss_fct(start_logits, start_positions)
+            end_loss = loss_fct(end_logits, end_positions)
+            total_loss = (start_loss + end_loss) / 2
+
+        if not return_dict:
+            output = (start_logits, end_logits) + transformer_outputs[2:]
+            return ((total_loss,) + output) if total_loss is not None else output
+
+        return QuestionAnsweringModelOutput(
+            loss=total_loss,
+            start_logits=start_logits,
+            end_logits=end_logits,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
+
+    def get_input_embeddings(self):
+        return self.model.decoder.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.decoder.embed_tokens = value