Update modeling_gpt2.py

updating based on transformers==4.52.4

modeling_gpt2.py

@@ -15,20 +15,50 @@
 # limitations under the License.
 """PyTorch OpenAI GPT-2 model."""
 
+import math
+import os
+import warnings
 from dataclasses import dataclass
 from typing import Callable, Optional, Tuple, Union
 
 import torch
 from torch import nn
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
+from transformers.activations import ACT2FN, get_activation
 from transformers.cache_utils import Cache, DynamicCache, EncoderDecoderCache, StaticCache
-from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS
-from transformers.pytorch_utils import Conv1D, find_pruneable_heads_and_indices, prune_conv1d_layer
+from transformers.generation import GenerationMixin
+from transformers.modeling_attn_mask_utils import (
+    AttentionMaskConverter,
+    _prepare_4d_attention_mask_for_sdpa,
+)
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPastAndCrossAttentions,
+    CausalLMOutputWithCrossAttentions,
+    QuestionAnsweringModelOutput,
+    SequenceClassifierOutputWithPast,
+    TokenClassifierOutput,
+)
+from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
+from transformers.pytorch_utils import (
+    Conv1D,
+    find_pruneable_heads_and_indices,
+    prune_conv1d_layer,
+)
 from transformers.utils import (
+    ModelOutput,
+    add_start_docstrings,
+    auto_docstring,
     logging,
 )
 from transformers.utils.deprecation import deprecate_kwarg
-from transformers.models.gpt2.modeling_gpt2 import load_tf_weights_in_gpt2, eager_attention_forward, GPT2Block, GPT2MLP, GPT2SequenceSummary,GPT2PreTrainedModel,GPT2DoubleHeadsModelOuptut,GPT2DoubleHeadsModel, GPT2Model,GPT2LMHeadModel, GPT2DoubleHeadsModel, GPT2ForSequenceClassification,GPT2ForTokenClassification,GPT2ForQuestionAnswering
+from transformers.utils.model_parallel_utils import assert_device_map, get_device_map
+from .configuration_gpt2 import GPT2Config
+from transformers.models.gpt2.modeling_gpt2 import (
+    load_tf_weights_in_gpt2,
+    eager_attention_forward,
+)
+
 
 logger = logging.get_logger(__name__)
 
@@ -40,9 +70,9 @@ class GPT2Attention(nn.Module):
         max_positions = config.max_position_embeddings
         self.register_buffer(
             "bias",
-            torch.tril(torch.ones((max_positions, max_positions), dtype=torch.bool)).view(
-                1, 1, max_positions, max_positions
-            ),
+            torch.tril(
+                torch.ones((max_positions, max_positions), dtype=torch.bool)
+            ).view(1, 1, max_positions, max_positions),
             persistent=False,
         )
         self.register_buffer("masked_bias", torch.tensor(-1e4), persistent=False)
@@ -81,25 +111,39 @@ class GPT2Attention(nn.Module):
     def prune_heads(self, heads):
         if len(heads) == 0:
             return
-        heads, index = find_pruneable_heads_and_indices(heads, self.num_heads, self.head_dim, self.pruned_heads)
-        index_attn = torch.cat([index, index + self.split_size, index + (2 * self.split_size)])
+        heads, index = find_pruneable_heads_and_indices(
+            heads, self.num_heads, self.head_dim, self.pruned_heads
+        )
+        index_attn = torch.cat(
+            [index, index + self.split_size, index + (2 * self.split_size)]
+        )
 
         # Prune conv1d layers
         self.c_attn = prune_conv1d_layer(self.c_attn, index_attn, dim=1)
         self.c_proj = prune_conv1d_layer(self.c_proj, index, dim=0)
 
         # Update hyper params
-        self.split_size = (self.split_size // self.num_heads) * (self.num_heads - len(heads))
+        self.split_size = (self.split_size // self.num_heads) * (
+            self.num_heads - len(heads)
+        )
         self.num_heads = self.num_heads - len(heads)
         self.pruned_heads = self.pruned_heads.union(heads)
 
-    def _upcast_and_reordered_attn(self, query, key, value, attention_mask=None, head_mask=None):
+    def _upcast_and_reordered_attn(
+        self, query, key, value, attention_mask=None, head_mask=None
+    ):
         # Use `torch.baddbmm` (a bit more efficient w/ alpha param for scaling -- from Megatron-LM)
         bsz, num_heads, q_seq_len, dk = query.size()
         _, _, k_seq_len, _ = key.size()
 
         # Preallocate attn_weights for `baddbmm`
-        attn_weights = torch.empty(bsz * num_heads, q_seq_len, k_seq_len, dtype=torch.float32, device=query.device)
+        attn_weights = torch.empty(
+            bsz * num_heads,
+            q_seq_len,
+            k_seq_len,
+            dtype=torch.float32,
+            device=query.device,
+        )
 
         # Compute Scale Factor
         scale_factor = 1.0
@@ -111,18 +155,26 @@ class GPT2Attention(nn.Module):
 
         # Upcast (turn off autocast) and reorder (Scale K by 1 / root(dk))
         with torch.amp.autocast(query.device.type, enabled=False):
-            q, k = query.reshape(-1, q_seq_len, dk), key.transpose(-1, -2).reshape(-1, dk, k_seq_len)
-            attn_weights = torch.baddbmm(attn_weights, q.float(), k.float(), beta=0, alpha=scale_factor)
+            q, k = query.reshape(-1, q_seq_len, dk), key.transpose(-1, -2).reshape(
+                -1, dk, k_seq_len
+            )
+            attn_weights = torch.baddbmm(
+                attn_weights, q.float(), k.float(), beta=0, alpha=scale_factor
+            )
             attn_weights = attn_weights.reshape(bsz, num_heads, q_seq_len, k_seq_len)
 
         if not self.is_cross_attention:
             # if only "normal" attention layer implements causal mask
             query_length, key_length = query.size(-2), key.size(-2)
-            causal_mask = self.bias[:, :, key_length - query_length : key_length, :key_length]
+            causal_mask = self.bias[
+                :, :, key_length - query_length : key_length, :key_length
+            ]
             mask_value = torch.finfo(attn_weights.dtype).min
             # Need to be a tensor, otherwise we get error: `RuntimeError: expected scalar type float but found double`.
             # Need to be on the same device, otherwise `RuntimeError: ..., x and y to be on the same device`
-            mask_value = torch.tensor(mask_value, dtype=attn_weights.dtype, device=attn_weights.device)
+            mask_value = torch.tensor(
+                mask_value, dtype=attn_weights.dtype, device=attn_weights.device
+            )
             attn_weights = torch.where(causal_mask, attn_weights, mask_value)
 
         if attention_mask is not None:
@@ -133,7 +185,9 @@ class GPT2Attention(nn.Module):
 
         # Downcast (if necessary) back to V's dtype (if in mixed-precision) -- No-Op if otherwise
         if attn_weights.dtype != torch.float32:
-            raise RuntimeError("Error with upcasting, attn_weights does not have dtype torch.float32")
+            raise RuntimeError(
+                "Error with upcasting, attn_weights does not have dtype torch.float32"
+            )
         attn_weights = attn_weights.type(value.dtype)
         attn_weights = self.attn_dropout(attn_weights)
 
@@ -146,7 +200,12 @@ class GPT2Attention(nn.Module):
 
         return attn_output, attn_weights
 
-    @deprecate_kwarg("layer_past", new_name="past_key_value", version="4.53.0", raise_if_both_names=True)
+    @deprecate_kwarg(
+        "layer_past",
+        new_name="past_key_value",
+        version="4.53.0",
+        raise_if_both_names=True,
+    )
     def forward(
         self,
         hidden_states: Optional[Tuple[torch.FloatTensor]],
@@ -168,13 +227,17 @@ class GPT2Attention(nn.Module):
                 )
 
             query_states = self.q_attn(hidden_states)
-            key_states, value_states = self.c_attn(encoder_hidden_states).split(self.split_size, dim=2)
+            key_states, value_states = self.c_attn(encoder_hidden_states).split(
+                self.split_size, dim=2
+            )
             attention_mask = encoder_attention_mask
         else:
-            query_states, key_states, value_states = self.c_attn(hidden_states).split(self.split_size, dim=2)
+            query_states, key_states, value_states = self.c_attn(hidden_states).split(
+                self.split_size, dim=2
+            )
 
         shape_q = (query_states.shape[0],query_states.shape[1], -1, self.head_dim)
-        shape_kv = (query_states.shape[0],query_states.shape[1], -1, self.head_dim)
+        shape_kv = (key_states.shape[0], key_states.shape[1],-1, self.head_dim)
 
         query_states = query_states.view(shape_q).transpose(1, 2)
         key_states = key_states.view(shape_kv).transpose(1, 2)
@@ -191,12 +254,18 @@ class GPT2Attention(nn.Module):
                 key_states, value_states, self.layer_idx, cache_kwargs=cache_kwargs
             )
 
-        is_causal = attention_mask is None and query_states.shape[-2] > 1 and not is_cross_attention
+        is_causal = (
+            attention_mask is None
+            and query_states.shape[-2] > 1
+            and not is_cross_attention
+        )
 
         using_eager = self.config._attn_implementation == "eager"
         attention_interface: Callable = eager_attention_forward
         if self.config._attn_implementation != "eager":
-            if self.config._attn_implementation == "sdpa" and (output_attentions or head_mask is not None):
+            if self.config._attn_implementation == "sdpa" and (
+                output_attentions or head_mask is not None
+            ):
                 using_eager = True
                 logger.warning_once(
                     "`torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to "
@@ -206,7 +275,9 @@ class GPT2Attention(nn.Module):
                 # Attention functions are consistent with previous equivalent attention classes, however they do not support some options
                 # (e.g. layer scaling, head mask) that eager supports. These implementations are thus equivalent to previous code, but
                 # not necessarily to eager (if mentioned options are provided).
-                attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+                attention_interface = ALL_ATTENTION_FUNCTIONS[
+                    self.config._attn_implementation
+                ]
 
         if using_eager and self.reorder_and_upcast_attn:
             attn_output, attn_weights = self._upcast_and_reordered_attn(
@@ -231,6 +302,1576 @@ class GPT2Attention(nn.Module):
 
         return attn_output, attn_weights
 
+
+class GPT2MLP(nn.Module):
+    def __init__(self, intermediate_size, config):
+        super().__init__()
+        embed_dim = config.hidden_size
+        self.c_fc = Conv1D(intermediate_size, embed_dim)
+        self.c_proj = Conv1D(embed_dim, intermediate_size)
+        self.act = ACT2FN[config.activation_function]
+        self.dropout = nn.Dropout(config.resid_pdrop)
+
+    def forward(
+        self, hidden_states: Optional[Tuple[torch.FloatTensor]]
+    ) -> torch.FloatTensor:
+        hidden_states = self.c_fc(hidden_states)
+        hidden_states = self.act(hidden_states)
+        hidden_states = self.c_proj(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        return hidden_states
+
+
+class GPT2Block(nn.Module):
+    def __init__(self, config, layer_idx=None):
+        super().__init__()
+        hidden_size = config.hidden_size
+        inner_dim = config.n_inner if config.n_inner is not None else 4 * hidden_size
+
+        self.ln_1 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
+        self.attn = GPT2Attention(config=config, layer_idx=layer_idx)
+        self.ln_2 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
+
+        if config.add_cross_attention:
+            self.crossattention = GPT2Attention(
+                config=config, is_cross_attention=True, layer_idx=layer_idx
+            )
+            self.ln_cross_attn = nn.LayerNorm(
+                hidden_size, eps=config.layer_norm_epsilon
+            )
+
+        self.mlp = GPT2MLP(inner_dim, config)
+
+    @deprecate_kwarg(
+        "layer_past",
+        new_name="past_key_value",
+        version="4.53.0",
+        raise_if_both_names=True,
+    )
+    def forward(
+        self,
+        hidden_states: Optional[Tuple[torch.FloatTensor]],
+        past_key_value: Optional[Cache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = False,
+        output_attentions: Optional[bool] = False,
+        **kwargs,
+    ) -> Union[
+        Tuple[torch.Tensor],
+        Optional[Tuple[torch.Tensor, Tuple[torch.FloatTensor, ...]]],
+    ]:
+        residual = hidden_states
+        hidden_states = self.ln_1(hidden_states)
+        attn_output, self_attn_weights = self.attn(
+            hidden_states,
+            past_key_value=past_key_value,
+            cache_position=cache_position,
+            attention_mask=attention_mask,
+            head_mask=head_mask,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            **kwargs,
+        )
+        # residual connection
+        hidden_states = attn_output + residual
+
+        if encoder_hidden_states is not None:
+            # add one self-attention block for cross-attention
+            if not hasattr(self, "crossattention"):
+                raise ValueError(
+                    f"If `encoder_hidden_states` are passed, {self} has to be instantiated with "
+                    "cross-attention layers by setting `config.add_cross_attention=True`"
+                )
+            residual = hidden_states
+            hidden_states = self.ln_cross_attn(hidden_states)
+            cross_attn_output, cross_attn_weights = self.crossattention(
+                hidden_states,
+                past_key_value=past_key_value,
+                attention_mask=attention_mask,
+                head_mask=head_mask,
+                encoder_hidden_states=encoder_hidden_states,
+                encoder_attention_mask=encoder_attention_mask,
+                output_attentions=output_attentions,
+            )
+            # residual connection
+            hidden_states = residual + cross_attn_output
+
+        residual = hidden_states
+        hidden_states = self.ln_2(hidden_states)
+        feed_forward_hidden_states = self.mlp(hidden_states)
+        # residual connection
+        hidden_states = residual + feed_forward_hidden_states
+
+        outputs = (hidden_states,)
+        if output_attentions:
+            outputs += (self_attn_weights,)
+            if encoder_hidden_states is not None:
+                outputs += (cross_attn_weights,)
+
+        return outputs
+
+
+# Copied from transformers.models.xlm.modeling_xlm.XLMSequenceSummary with XLM->GPT2
+class GPT2SequenceSummary(nn.Module):
+    r"""
+    Compute a single vector summary of a sequence hidden states.
+
+    Args:
+        config ([`GPT2Config`]):
+            The config used by the model. Relevant arguments in the config class of the model are (refer to the actual
+            config class of your model for the default values it uses):
+
+            - **summary_type** (`str`) -- The method to use to make this summary. Accepted values are:
+
+                - `"last"` -- Take the last token hidden state (like XLNet)
+                - `"first"` -- Take the first token hidden state (like Bert)
+                - `"mean"` -- Take the mean of all tokens hidden states
+                - `"cls_index"` -- Supply a Tensor of classification token position (GPT/GPT-2)
+                - `"attn"` -- Not implemented now, use multi-head attention
+
+            - **summary_use_proj** (`bool`) -- Add a projection after the vector extraction.
+            - **summary_proj_to_labels** (`bool`) -- If `True`, the projection outputs to `config.num_labels` classes
+              (otherwise to `config.hidden_size`).
+            - **summary_activation** (`Optional[str]`) -- Set to `"tanh"` to add a tanh activation to the output,
+              another string or `None` will add no activation.
+            - **summary_first_dropout** (`float`) -- Optional dropout probability before the projection and activation.
+            - **summary_last_dropout** (`float`)-- Optional dropout probability after the projection and activation.
+    """
+
+    def __init__(self, config: GPT2Config):
+        super().__init__()
+
+        self.summary_type = getattr(config, "summary_type", "last")
+        if self.summary_type == "attn":
+            # We should use a standard multi-head attention module with absolute positional embedding for that.
+            # Cf. https://github.com/zihangdai/xlnet/blob/master/modeling.py#L253-L276
+            # We can probably just use the multi-head attention module of PyTorch >=1.1.0
+            raise NotImplementedError
+
+        self.summary = nn.Identity()
+        if hasattr(config, "summary_use_proj") and config.summary_use_proj:
+            if (
+                hasattr(config, "summary_proj_to_labels")
+                and config.summary_proj_to_labels
+                and config.num_labels > 0
+            ):
+                num_classes = config.num_labels
+            else:
+                num_classes = config.hidden_size
+            self.summary = nn.Linear(config.hidden_size, num_classes)
+
+        activation_string = getattr(config, "summary_activation", None)
+        self.activation: Callable = (
+            get_activation(activation_string) if activation_string else nn.Identity()
+        )
+
+        self.first_dropout = nn.Identity()
+        if (
+            hasattr(config, "summary_first_dropout")
+            and config.summary_first_dropout > 0
+        ):
+            self.first_dropout = nn.Dropout(config.summary_first_dropout)
+
+        self.last_dropout = nn.Identity()
+        if hasattr(config, "summary_last_dropout") and config.summary_last_dropout > 0:
+            self.last_dropout = nn.Dropout(config.summary_last_dropout)
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        cls_index: Optional[torch.LongTensor] = None,
+    ) -> torch.FloatTensor:
+        """
+        Compute a single vector summary of a sequence hidden states.
+
+        Args:
+            hidden_states (`torch.FloatTensor` of shape `[batch_size, seq_len, hidden_size]`):
+                The hidden states of the last layer.
+            cls_index (`torch.LongTensor` of shape `[batch_size]` or `[batch_size, ...]` where ... are optional leading dimensions of `hidden_states`, *optional*):
+                Used if `summary_type == "cls_index"` and takes the last token of the sequence as classification token.
+
+        Returns:
+            `torch.FloatTensor`: The summary of the sequence hidden states.
+        """
+        if self.summary_type == "last":
+            output = hidden_states[:, -1]
+        elif self.summary_type == "first":
+            output = hidden_states[:, 0]
+        elif self.summary_type == "mean":
+            output = hidden_states.mean(dim=1)
+        elif self.summary_type == "cls_index":
+            if cls_index is None:
+                cls_index = torch.full_like(
+                    hidden_states[..., :1, :],
+                    hidden_states.shape[-2] - 1,
+                    dtype=torch.long,
+                )
+            else:
+                cls_index = cls_index.unsqueeze(-1).unsqueeze(-1)
+                cls_index = cls_index.expand(
+                    (-1,) * (cls_index.dim() - 1) + (hidden_states.size(-1),)
+                )
+            # shape of cls_index: (bsz, XX, 1, hidden_size) where XX are optional leading dim of hidden_states
+            output = hidden_states.gather(-2, cls_index).squeeze(
+                -2
+            )  # shape (bsz, XX, hidden_size)
+        elif self.summary_type == "attn":
+            raise NotImplementedError
+
+        output = self.first_dropout(output)
+        output = self.summary(output)
+        output = self.activation(output)
+        output = self.last_dropout(output)
+
+        return output
+
+
+@auto_docstring
+class GPT2PreTrainedModel(PreTrainedModel):
+    config_class = GPT2Config
+    load_tf_weights = load_tf_weights_in_gpt2
+    base_model_prefix = "transformer"
+    is_parallelizable = True
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["GPT2Block"]
+    _skip_keys_device_placement = "past_key_values"
+    _supports_flash_attn_2 = True
+    _supports_sdpa = True
+    _supports_attention_backend = True
+    _supports_cache_class = True
+    _supports_static_cache = True
+
+    def __init__(self, *inputs, **kwargs):
+        super().__init__(*inputs, **kwargs)
+
+    def _init_weights(self, module):
+        """Initialize the weights."""
+        if isinstance(module, (nn.Linear, Conv1D)):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+
+        # Reinitialize selected weights subject to the OpenAI GPT-2 Paper Scheme:
+        #   > A modified initialization which accounts for the accumulation on the residual path with model depth. Scale
+        #   > the weights of residual layers at initialization by a factor of 1/√N where N is the # of residual layers.
+        #   >   -- GPT-2 :: https://openai.com/blog/better-language-models/
+        #
+        # Reference (Megatron-LM): https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/model/gpt_model.py
+        for name, p in module.named_parameters():
+            if name == "c_proj.weight":
+                # Special Scaled Initialization --> There are 2 Layer Norms per Transformer Block
+                p.data.normal_(
+                    mean=0.0,
+                    std=(
+                        self.config.initializer_range
+                        / math.sqrt(2 * self.config.n_layer)
+                    ),
+                )
+
+
+@dataclass
+class GPT2DoubleHeadsModelOutput(ModelOutput):
+    """
+    Base class for outputs of models predicting if two sentences are consecutive or not.
+
+    Args:
+        loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
+            Language modeling loss.
+        mc_loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `mc_labels` is provided):
+            Multiple choice classification loss.
+        logits (`torch.FloatTensor` of shape `(batch_size, num_choices, sequence_length, config.vocab_size)`):
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        mc_logits (`torch.FloatTensor` of shape `(batch_size, num_choices)`):
+            Prediction scores of the multiple choice classification head (scores for each choice before SoftMax).
+        past_key_values (`Tuple[Tuple[torch.Tensor]]`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
+            Tuple of length `config.n_layers`, containing tuples of tensors of shape `(batch_size, num_heads,
+            sequence_length, embed_size_per_head)`).
+
+            Contains pre-computed hidden-states (key and values in the attention blocks) that can be used (see
+            `past_key_values` input) to speed up sequential decoding.
+        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer) of
+            shape `(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
+            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
+            sequence_length)`.
+
+            GPT2Attentions weights after the attention softmax, used to compute the weighted average in the
+            self-attention heads.
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    mc_loss: Optional[torch.FloatTensor] = None
+    logits: Optional[torch.FloatTensor] = None
+    mc_logits: Optional[torch.FloatTensor] = None
+    past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+PARALLELIZE_DOCSTRING = r"""
+    This is an experimental feature and is a subject to change at a moment's notice.
+
+    Uses a device map to distribute attention modules of the model across several devices. If no device map is given,
+    it will evenly distribute blocks across all devices.
+
+    Args:
+        device_map (`Dict[int, list]`, *optional*):
+            A dictionary that maps attention modules to devices. Note that the embedding module and LMHead are always
+            automatically mapped to the first device (for esoteric reasons). That means that the first device should
+            have fewer attention modules mapped to it than other devices. For reference, the gpt2 models have the
+            following number of attention modules:
+
+                - openai-community/gpt2: 12
+                - openai-community/gpt2-medium: 24
+                - openai-community/gpt2-large: 36
+                - openai-community/gpt2-xl: 48
+
+    Example:
+
+    ```python
+    # Here is an example of a device map on a machine with 4 GPUs using gpt2-xl, which has a total of 48 attention modules:
+    model = GPT2LMHeadModel.from_pretrained("openai-community/gpt2-xl")
+    device_map = {
+        0: [0, 1, 2, 3, 4, 5, 6, 7, 8],
+        1: [9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21],
+        2: [22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34],
+        3: [35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47],
+    }
+    model.parallelize(device_map)
+    ```
+"""
+DEPARALLELIZE_DOCSTRING = r"""
+    Moves the model to cpu from a model parallel state.
+
+    Example:
+
+    ```python
+    # On a 4 GPU machine with openai-community/gpt2-large:
+    model = GPT2LMHeadModel.from_pretrained("openai-community/gpt2-large")
+    device_map = {
+        0: [0, 1, 2, 3, 4, 5, 6, 7],
+        1: [8, 9, 10, 11, 12, 13, 14, 15],
+        2: [16, 17, 18, 19, 20, 21, 22, 23],
+        3: [24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35],
+    }
+    model.parallelize(device_map)  # Splits the model across several devices
+    model.deparallelize()  # Put the model back on cpu and cleans memory by calling torch.cuda.empty_cache()
+    ```
+"""
+
+
+@auto_docstring
+class GPT2Model(GPT2PreTrainedModel):
+    _supports_param_buffer_assignment = False
+
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.embed_dim = config.hidden_size
+
+        self.wte = nn.Embedding(config.vocab_size, self.embed_dim)
+        self.wpe = nn.Embedding(config.max_position_embeddings, self.embed_dim)
+
+        self.drop = nn.Dropout(config.embd_pdrop)
+        self.h = nn.ModuleList(
+            [GPT2Block(config, layer_idx=i) for i in range(config.num_hidden_layers)]
+        )
+        self.ln_f = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon)
+
+        # Model parallel
+        self.model_parallel = False
+        self.device_map = None
+        self.gradient_checkpointing = False
+        self._attn_implementation = config._attn_implementation
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings(PARALLELIZE_DOCSTRING)
+    def parallelize(self, device_map=None):
+        # Check validity of device_map
+        warnings.warn(
+            "`GPT2Model.parallelize` is deprecated and will be removed in v5 of Transformers, you should load your"
+            " model with `device_map='balanced'` in the call to `from_pretrained`. You can also provide your own"
+            " `device_map` but it needs to be a dictionary module_name to device, so for instance {'h.0': 0, 'h.1': 1,"
+            " ...}",
+            FutureWarning,
+        )
+        self.device_map = (
+            get_device_map(len(self.h), range(torch.cuda.device_count()))
+            if device_map is None
+            else device_map
+        )
+        assert_device_map(self.device_map, len(self.h))
+        self.model_parallel = True
+        self.first_device = (
+            "cpu"
+            if "cpu" in self.device_map.keys()
+            else "cuda:" + str(min(self.device_map.keys()))
+        )
+        self.last_device = "cuda:" + str(max(self.device_map.keys()))
+        self.wte = self.wte.to(self.first_device)
+        self.wpe = self.wpe.to(self.first_device)
+        # Load onto devices
+        for k, v in self.device_map.items():
+            for block in v:
+                cuda_device = "cuda:" + str(k)
+                self.h[block] = self.h[block].to(cuda_device)
+        # ln_f to last
+        self.ln_f = self.ln_f.to(self.last_device)
+
+    @add_start_docstrings(DEPARALLELIZE_DOCSTRING)
+    def deparallelize(self):
+        warnings.warn(
+            "Like `parallelize`, `deparallelize` is deprecated and will be removed in v5 of Transformers.",
+            FutureWarning,
+        )
+        self.model_parallel = False
+        self.device_map = None
+        self.first_device = "cpu"
+        self.last_device = "cpu"
+        self.wte = self.wte.to("cpu")
+        self.wpe = self.wpe.to("cpu")
+        for index in range(len(self.h)):
+            self.h[index] = self.h[index].to("cpu")
+        self.ln_f = self.ln_f.to("cpu")
+        torch.cuda.empty_cache()
+
+    def get_input_embeddings(self):
+        return self.wte
+
+    def set_input_embeddings(self, new_embeddings):
+        self.wte = new_embeddings
+
+    def _prune_heads(self, heads_to_prune):
+        """
+        Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer}
+        """
+        for layer, heads in heads_to_prune.items():
+            self.h[layer].attn.prune_heads(heads)
+
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Union[Tuple[Tuple[torch.Tensor]], Cache]] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        encoder_attention_mask: 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,
+        **kwargs,
+    ) -> Union[Tuple, BaseModelOutputWithPastAndCrossAttentions]:
+        r"""
+        input_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`):
+            `input_ids_length` = `sequence_length` if `past_key_values` is `None` else
+            `past_key_values[0][0].shape[-2]` (`sequence_length` of input past key value states). Indices of input
+            sequence tokens in the vocabulary.
+
+            If `past_key_values` is used, only `input_ids` that do not have their past calculated should be passed as
+            `input_ids`.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        """
+        output_attentions = (
+            output_attentions
+            if output_attentions is not None
+            else self.config.output_attentions
+        )
+        output_hidden_states = (
+            output_hidden_states
+            if output_hidden_states is not None
+            else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError(
+                "You cannot specify both input_ids and inputs_embeds at the same time"
+            )
+        elif input_ids is not None:
+            self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask)
+            input_shape = input_ids.size()
+            input_ids = input_ids.view(-1, input_shape[-1])
+            batch_size = input_ids.shape[0]
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+            batch_size = inputs_embeds.shape[0]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+
+        if token_type_ids is not None:
+            token_type_ids = token_type_ids.view(-1, input_shape[-1])
+
+        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
+
+        # based on pattern from src/transformers/models/whisper/modeling_whisper.py::WhisperDecoder
+        return_legacy_cache = False
+        if use_cache:
+            if past_key_values is None:
+                return_legacy_cache = True
+                past_key_values = DynamicCache()
+            elif not isinstance(past_key_values, Cache):
+                return_legacy_cache = True
+                logger.warning_once(
+                    "Passing a tuple of `past_key_values` is deprecated and will be removed in Transformers v4.53.0. "
+                    "You should pass an instance of `Cache` instead, e.g. "
+                    "`past_key_values=DynamicCache.from_legacy_cache(past_key_values)`."
+                )
+                past_key_values = DynamicCache.from_legacy_cache(past_key_values)
+
+            if self.config.add_cross_attention and not isinstance(
+                past_key_values, EncoderDecoderCache
+            ):
+                past_key_values = EncoderDecoderCache(past_key_values, DynamicCache())
+
+        if inputs_embeds is None:
+            inputs_embeds = self.wte(input_ids)
+
+        if cache_position is None:
+            past_seen_tokens = (
+                past_key_values.get_seq_length() if past_key_values is not None else 0
+            )
+            cache_position = torch.arange(
+                past_seen_tokens,
+                past_seen_tokens + inputs_embeds.shape[1],
+                device=inputs_embeds.device,
+            )
+        if position_ids is None:
+            position_ids = cache_position.unsqueeze(0)
+
+        position_embeds = self.wpe(position_ids)
+        hidden_states = inputs_embeds + position_embeds.to(inputs_embeds.device)
+
+        # Attention mask.
+        # ._update_causal_mask() and ._prepare_4d_causal_attention_mask_with_cache_position() copied from LlamaModel
+        if attention_mask is not None and attention_mask.ndim < 4:
+            attention_mask = attention_mask.view(batch_size, -1)
+        causal_mask = self._update_causal_mask(
+            attention_mask,
+            inputs_embeds,
+            cache_position,
+            past_key_values,
+            output_attentions,
+        )
+
+        # If a 2D or 3D attention mask is provided for the cross-attention
+        # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
+        _use_sdpa = (
+            self._attn_implementation == "sdpa"
+            and output_attentions is False
+            and head_mask is None
+        )
+        if self.config.add_cross_attention and encoder_hidden_states is not None:
+            encoder_batch_size, encoder_sequence_length, _ = (
+                encoder_hidden_states.size()
+            )
+            encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
+            if encoder_attention_mask is None:
+                encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)
+            if _use_sdpa:
+                encoder_attention_mask = _prepare_4d_attention_mask_for_sdpa(
+                    mask=encoder_attention_mask,
+                    dtype=inputs_embeds.dtype,
+                    tgt_len=input_shape[-1],
+                )
+            elif not self._attn_implementation == "flash_attention_2":
+                encoder_attention_mask = self.invert_attention_mask(
+                    encoder_attention_mask
+                )
+        else:
+            encoder_attention_mask = None
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # head_mask has shape n_layer x batch x n_heads x N x N
+        head_mask = self.get_head_mask(head_mask, self.config.n_layer)
+
+        if token_type_ids is not None:
+            token_type_embeds = self.wte(token_type_ids)
+            hidden_states = hidden_states + token_type_embeds
+
+        hidden_states = self.drop(hidden_states)
+
+        output_shape = (-1,) + input_shape[1:] + (hidden_states.size(-1),)
+
+        all_self_attentions = () if output_attentions else None
+        all_cross_attentions = (
+            () if output_attentions and self.config.add_cross_attention else None
+        )
+        all_hidden_states = () if output_hidden_states else None
+        for i, block in enumerate(self.h):
+            # Model parallel
+            if self.model_parallel:
+                torch.cuda.set_device(hidden_states.device)
+                # Ensure that attention_mask is always on the same device as hidden_states
+                if attention_mask is not None:
+                    attention_mask = attention_mask.to(hidden_states.device)
+                if isinstance(head_mask, torch.Tensor):
+                    head_mask = head_mask.to(hidden_states.device)
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+                outputs = self._gradient_checkpointing_func(
+                    block.__call__,
+                    hidden_states,
+                    past_key_values,
+                    cache_position,
+                    causal_mask,
+                    head_mask[i],
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    use_cache,
+                    output_attentions,
+                )
+            else:
+                outputs = block(
+                    hidden_states,
+                    past_key_value=past_key_values,
+                    cache_position=cache_position,
+                    attention_mask=causal_mask,
+                    head_mask=head_mask[i],
+                    encoder_hidden_states=encoder_hidden_states,
+                    encoder_attention_mask=encoder_attention_mask,
+                    use_cache=use_cache,
+                    output_attentions=output_attentions,
+                    **kwargs,
+                )
+
+            hidden_states = outputs[0]
+
+            if output_attentions:
+                all_self_attentions = all_self_attentions + (outputs[1],)
+                if self.config.add_cross_attention:
+                    all_cross_attentions = all_cross_attentions + (outputs[2],)
+
+            # Model Parallel: If it's the last layer for that device, put things on the next device
+            if self.model_parallel:
+                for k, v in self.device_map.items():
+                    if i == v[-1] and "cuda:" + str(k) != self.last_device:
+                        hidden_states = hidden_states.to("cuda:" + str(k + 1))
+
+        hidden_states = self.ln_f(hidden_states)
+
+        hidden_states = hidden_states.view(output_shape)
+        # Add last hidden state
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        past_key_values = past_key_values if use_cache else None
+        if return_legacy_cache:
+            past_key_values = (
+                past_key_values.self_attention_cache.to_legacy_cache()
+                if self.config.add_cross_attention
+                else past_key_values.to_legacy_cache()
+            )
+        if not return_dict:
+            return tuple(
+                v
+                for v in [
+                    hidden_states,
+                    past_key_values,
+                    all_hidden_states,
+                    all_self_attentions,
+                    all_cross_attentions,
+                ]
+                if v is not None
+            )
+
+        return BaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            past_key_values=past_key_values,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attentions,
+            cross_attentions=all_cross_attentions,
+        )
+
+    def _update_causal_mask(
+        self,
+        attention_mask: torch.Tensor,
+        input_tensor: torch.Tensor,
+        cache_position: torch.Tensor,
+        past_key_values: Cache,
+        output_attentions: bool,
+    ):
+        if self.config._attn_implementation == "flash_attention_2":
+            if attention_mask is not None and 0.0 in attention_mask:
+                return attention_mask
+            return None
+
+        # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in
+        # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail
+        # to infer the attention mask.
+        past_seen_tokens = (
+            past_key_values.get_seq_length() if past_key_values is not None else 0
+        )
+        using_static_cache = isinstance(past_key_values, StaticCache)
+
+        # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward
+        if (
+            self.config._attn_implementation == "sdpa"
+            and not using_static_cache
+            and not output_attentions
+        ):
+            if AttentionMaskConverter._ignore_causal_mask_sdpa(
+                attention_mask,
+                inputs_embeds=input_tensor,
+                past_key_values_length=past_seen_tokens,
+                is_training=self.training,
+            ):
+                return None
+
+        dtype = input_tensor.dtype
+        sequence_length = input_tensor.shape[1]
+        if using_static_cache:
+            target_length = past_key_values.get_max_cache_shape()
+        else:
+            target_length = (
+                attention_mask.shape[-1]
+                if isinstance(attention_mask, torch.Tensor)
+                else past_seen_tokens + sequence_length + 1
+            )
+
+        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
+        causal_mask = self._prepare_4d_causal_attention_mask_with_cache_position(
+            attention_mask,
+            sequence_length=sequence_length,
+            target_length=target_length,
+            dtype=dtype,
+            cache_position=cache_position,
+            batch_size=input_tensor.shape[0],
+        )
+
+        if (
+            self.config._attn_implementation == "sdpa"
+            and attention_mask is not None
+            and attention_mask.device.type == "cuda"
+            and not output_attentions
+        ):
+            # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
+            # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
+            # Details: https://github.com/pytorch/pytorch/issues/110213
+            min_dtype = torch.finfo(dtype).min
+            causal_mask = AttentionMaskConverter._unmask_unattended(
+                causal_mask, min_dtype
+            )
+
+        return causal_mask
+
+    @staticmethod
+    def _prepare_4d_causal_attention_mask_with_cache_position(
+        attention_mask: torch.Tensor,
+        sequence_length: int,
+        target_length: int,
+        dtype: torch.dtype,
+        cache_position: torch.Tensor,
+        batch_size: int,
+        **kwargs,
+    ):
+        """
+        Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
+        `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.
+
+        Args:
+            attention_mask (`torch.Tensor`):
+                A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape
+                `(batch_size, 1, query_length, key_value_length)`.
+            sequence_length (`int`):
+                The sequence length being processed.
+            target_length (`int`):
+                The target length: when generating with static cache, the mask should be as long as the static cache,
+                to account for the 0 padding, the part of the cache that is not filled yet.
+            dtype (`torch.dtype`):
+                The dtype to use for the 4D attention mask.
+            cache_position (`torch.Tensor`):
+                Indices depicting the position of the input sequence tokens in the sequence.
+            batch_size (`torch.Tensor`):
+                Batch size.
+        """
+        if attention_mask is not None and attention_mask.dim() == 4:
+            # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
+            causal_mask = attention_mask
+        else:
+            min_dtype = torch.finfo(dtype).min
+            causal_mask = torch.full(
+                (sequence_length, target_length),
+                fill_value=min_dtype,
+                dtype=dtype,
+                device=cache_position.device,
+            )
+            if sequence_length != 1:
+                causal_mask = torch.triu(causal_mask, diagonal=1)
+            causal_mask *= torch.arange(
+                target_length, device=cache_position.device
+            ) > cache_position.reshape(-1, 1)
+            causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
+            if attention_mask is not None:
+                causal_mask = (
+                    causal_mask.clone()
+                )  # copy to contiguous memory for in-place edit
+                mask_length = attention_mask.shape[-1]
+                padding_mask = (
+                    causal_mask[:, :, :, :mask_length]
+                    + attention_mask[:, None, None, :]
+                )
+                padding_mask = padding_mask == 0
+                causal_mask[:, :, :, :mask_length] = causal_mask[
+                    :, :, :, :mask_length
+                ].masked_fill(padding_mask, min_dtype)
+
+        return causal_mask
+
+
+@auto_docstring(
+    custom_intro="""
+    The GPT2 Model transformer with a language modeling head on top (linear layer with weights tied to the input
+    embeddings).
+    """
+)
+class GPT2LMHeadModel(GPT2PreTrainedModel, GenerationMixin):
+    _tied_weights_keys = ["lm_head.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.transformer = GPT2Model(config)
+        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
+
+        # Model parallel
+        self.model_parallel = False
+        self.device_map = None
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings(PARALLELIZE_DOCSTRING)
+    def parallelize(self, device_map=None):
+        warnings.warn(
+            "`GPT2LMHeadModel.parallelize` is deprecated and will be removed in v5 of Transformers, you should load"
+            " your model with `device_map='balanced'` in the call to `from_pretrained`. You can also provide your own"
+            " `device_map` but it needs to be a dictionary module_name to device, so for instance {'transformer.h.0':"
+            " 0, 'transformer.h.1': 1, ...}",
+            FutureWarning,
+        )
+        self.device_map = (
+            get_device_map(len(self.transformer.h), range(torch.cuda.device_count()))
+            if device_map is None
+            else device_map
+        )
+        assert_device_map(self.device_map, len(self.transformer.h))
+        self.transformer.parallelize(self.device_map)
+        self.lm_head = self.lm_head.to(self.transformer.first_device)
+        self.model_parallel = True
+
+    @add_start_docstrings(DEPARALLELIZE_DOCSTRING)
+    def deparallelize(self):
+        warnings.warn(
+            "Like `parallelize`, `deparallelize` is deprecated and will be removed in v5 of Transformers.",
+            FutureWarning,
+        )
+        self.transformer.deparallelize()
+        self.transformer = self.transformer.to("cpu")
+        self.lm_head = self.lm_head.to("cpu")
+        self.model_parallel = False
+        torch.cuda.empty_cache()
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        encoder_attention_mask: 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,
+        **kwargs,
+    ) -> Union[Tuple, CausalLMOutputWithCrossAttentions]:
+        r"""
+        input_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`):
+            `input_ids_length` = `sequence_length` if `past_key_values` is `None` else
+            `past_key_values[0][0].shape[-2]` (`sequence_length` of input past key value states). Indices of input
+            sequence tokens in the vocabulary.
+
+            If `past_key_values` is used, only `input_ids` that do not have their past calculated should be passed as
+            `input_ids`.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        labels (`torch.LongTensor` of shape `(batch_size, input_ids_length)`, *optional*):
+            Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
+            `labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100`
+            are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]`
+        """
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        transformer_outputs = self.transformer(
+            input_ids,
+            past_key_values=past_key_values,
+            attention_mask=attention_mask,
+            cache_position=cache_position,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        hidden_states = transformer_outputs[0]
+
+        # Set device for model parallelism
+        if self.model_parallel:
+            torch.cuda.set_device(self.transformer.first_device)
+            hidden_states = hidden_states.to(self.lm_head.weight.device)
+
+        lm_logits = self.lm_head(hidden_states)
+
+        loss = None
+        if labels is not None:
+            # Flatten the tokens
+            loss = self.loss_function(
+                lm_logits,
+                labels,
+                vocab_size=self.config.vocab_size,
+                **kwargs,
+            )
+
+        if not return_dict:
+            output = (lm_logits,) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return CausalLMOutputWithCrossAttentions(
+            loss=loss,
+            logits=lm_logits,
+            past_key_values=transformer_outputs.past_key_values,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+            cross_attentions=transformer_outputs.cross_attentions,
+        )
+
+
+@auto_docstring(
+    custom_intro="""
+        The GPT2 Model transformer with a language modeling and a multiple-choice classification head on top e.g. for
+    RocStories/SWAG tasks. The two heads are two linear layers. The language modeling head has its weights tied to the
+    input embeddings, the classification head takes as input the input of a specified classification token index in the
+    input sequence).
+    """
+)
+class GPT2DoubleHeadsModel(GPT2PreTrainedModel, GenerationMixin):
+    _tied_weights_keys = ["lm_head.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        config.num_labels = 1
+        self.transformer = GPT2Model(config)
+        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
+        self.multiple_choice_head = GPT2SequenceSummary(config)
+
+        # Model parallel
+        self.model_parallel = False
+        self.device_map = None
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings(PARALLELIZE_DOCSTRING)
+    def parallelize(self, device_map=None):
+        warnings.warn(
+            "`GPT2DoubleHeadsModel.parallelize` is deprecated and will be removed in v5 of Transformers, you should"
+            " load your model with `device_map='balanced'` in the call to `from_pretrained`. You can also provide your"
+            " own `device_map` but it needs to be a dictionary module_name to device, so for instance"
+            " {'transformer.h.0': 0, 'transformer.h.1': 1, ...}",
+            FutureWarning,
+        )
+        self.device_map = (
+            get_device_map(len(self.transformer.h), range(torch.cuda.device_count()))
+            if device_map is None
+            else device_map
+        )
+        assert_device_map(self.device_map, len(self.transformer.h))
+        self.transformer.parallelize(self.device_map)
+        self.lm_head = self.lm_head.to(self.transformer.first_device)
+        self.multiple_choice_head = self.multiple_choice_head.to(
+            self.transformer.first_device
+        )
+        self.model_parallel = True
+
+    @add_start_docstrings(DEPARALLELIZE_DOCSTRING)
+    def deparallelize(self):
+        warnings.warn(
+            "Like `parallelize`, `deparallelize` is deprecated and will be removed in v5 of Transformers.",
+            FutureWarning,
+        )
+        self.transformer.deparallelize()
+        self.transformer = self.transformer.to("cpu")
+        self.lm_head = self.lm_head.to("cpu")
+        self.multiple_choice_head = self.multiple_choice_head.to("cpu")
+        self.model_parallel = False
+        torch.cuda.empty_cache()
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        mc_token_ids: Optional[torch.LongTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        mc_labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        **kwargs,
+    ) -> Union[Tuple, GPT2DoubleHeadsModelOutput]:
+        r"""
+        input_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`):
+            `input_ids_length` = `sequence_length` if `past_key_values` is `None` else
+            `past_key_values[0][0].shape[-2]` (`sequence_length` of input past key value states). Indices of input
+            sequence tokens in the vocabulary.
+
+            If `past_key_values` is used, only `input_ids` that do not have their past calculated should be passed as
+            `input_ids`.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        mc_token_ids (`torch.LongTensor` of shape `(batch_size, num_choices)`, *optional*, default to index of the last token of the input):
+            Index of the classification token in each input sequence. Selected in the range `[0, input_ids.size(-1) -
+            1]`.
+        labels (`torch.LongTensor` of shape `(batch_size, input_ids_length)`, *optional*):
+            Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
+            `labels = input_ids`. Indices are selected in `[-100, 0, ..., config.vocab_size - 1]`. All labels set to
+            `-100` are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size - 1]`
+        mc_labels (`torch.LongTensor` of shape `(batch_size)`, *optional*):
+            Labels for computing the multiple choice classification loss. Indices should be in `[0, ..., num_choices]`
+            where *num_choices* is the size of the second dimension of the input tensors. (see *input_ids* above)
+
+        Example:
+
+        ```python
+        >>> import torch
+        >>> from transformers import AutoTokenizer, GPT2DoubleHeadsModel
+
+        >>> tokenizer = AutoTokenizer.from_pretrained("openai-community/gpt2")
+        >>> model = GPT2DoubleHeadsModel.from_pretrained("openai-community/gpt2")
+
+        >>> # Add a [CLS] to the vocabulary (we should train it also!)
+        >>> num_added_tokens = tokenizer.add_special_tokens({"cls_token": "[CLS]"})
+        >>> # Update the model embeddings with the new vocabulary size
+        >>> embedding_layer = model.resize_token_embeddings(len(tokenizer))
+
+        >>> choices = ["Hello, my dog is cute [CLS]", "Hello, my cat is cute [CLS]"]
+        >>> encoded_choices = [tokenizer.encode(s) for s in choices]
+        >>> cls_token_location = [tokens.index(tokenizer.cls_token_id) for tokens in encoded_choices]
+
+        >>> input_ids = torch.tensor(encoded_choices).unsqueeze(0)  # Batch size: 1, number of choices: 2
+        >>> mc_token_ids = torch.tensor([cls_token_location])  # Batch size: 1
+
+        >>> outputs = model(input_ids, mc_token_ids=mc_token_ids)
+        >>> lm_logits = outputs.logits
+        >>> mc_logits = outputs.mc_logits
+        ```"""
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        transformer_outputs = self.transformer(
+            input_ids,
+            past_key_values=past_key_values,
+            cache_position=cache_position,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = transformer_outputs[0]
+
+        # Set device for model parallelism
+        if self.model_parallel:
+            torch.cuda.set_device(self.transformer.first_device)
+            hidden_states = hidden_states.to(self.lm_head.weight.device)
+
+        lm_logits = self.lm_head(hidden_states)
+        mc_logits = self.multiple_choice_head(hidden_states, mc_token_ids).squeeze(-1)
+
+        mc_loss = None
+        if mc_labels is not None:
+            loss_fct = CrossEntropyLoss()
+            mc_loss = loss_fct(
+                mc_logits.view(-1, mc_logits.size(-1)), mc_labels.view(-1)
+            )
+        lm_loss = None
+        if labels is not None:
+            labels = labels.to(lm_logits.device)
+            shift_logits = lm_logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            loss_fct = CrossEntropyLoss()
+            lm_loss = loss_fct(
+                shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1)
+            )
+
+        if not return_dict:
+            output = (lm_logits, mc_logits) + transformer_outputs[1:]
+            if mc_loss is not None:
+                output = (mc_loss,) + output
+            return ((lm_loss,) + output) if lm_loss is not None else output
+
+        return GPT2DoubleHeadsModelOutput(
+            loss=lm_loss,
+            mc_loss=mc_loss,
+            logits=lm_logits,
+            mc_logits=mc_logits,
+            past_key_values=transformer_outputs.past_key_values,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
+
+    @staticmethod
+    def _reorder_cache(
+        past_key_values: Tuple[Tuple[torch.Tensor]], beam_idx: torch.Tensor
+    ) -> Tuple[Tuple[torch.Tensor]]:
+        """
+        This function is used to re-order the `past_key_values` cache if [`~PreTrainedModel.beam_search`] or
+        [`~PreTrainedModel.beam_sample`] is called. This is required to match `past_key_values` with the correct
+        beam_idx at every generation step.
+        """
+        return tuple(
+            tuple(
+                past_state.index_select(0, beam_idx.to(past_state.device))
+                for past_state in layer_past
+            )
+            for layer_past in past_key_values
+        )
+
+
+@auto_docstring(
+    custom_intro="""
+    The GPT2 Model transformer with a sequence classification head on top (linear layer).
+
+    [`GPT2ForSequenceClassification`] uses the last token in order to do the classification, as other causal models
+    (e.g. GPT-1) 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).
+    """
+)
+class GPT2ForSequenceClassification(GPT2PreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.transformer = GPT2Model(config)
+        self.score = nn.Linear(config.n_embd, self.num_labels, bias=False)
+
+        # Model parallel
+        self.model_parallel = False
+        self.device_map = None
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, SequenceClassifierOutputWithPast]:
+        r"""
+        input_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`):
+            `input_ids_length` = `sequence_length` if `past_key_values` is `None` else
+            `past_key_values[0][0].shape[-2]` (`sequence_length` of input past key value states). Indices of input
+            sequence tokens in the vocabulary.
+
+            If `past_key_values` is used, only `input_ids` that do not have their past calculated should be passed as
+            `input_ids`.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        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.transformer(
+            input_ids,
+            past_key_values=past_key_values,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        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 and batch_size != 1:
+            raise ValueError(
+                "Cannot handle batch sizes > 1 if no padding token is defined."
+            )
+        if self.config.pad_token_id is None:
+            last_non_pad_token = -1
+        elif input_ids is not None:
+            # To handle both left- and right- padding, we take the rightmost token that is not equal to pad_token_id
+            non_pad_mask = (input_ids != self.config.pad_token_id).to(
+                logits.device, torch.int32
+            )
+            token_indices = torch.arange(
+                input_ids.shape[-1], device=logits.device, dtype=torch.int32
+            )
+            last_non_pad_token = (token_indices * non_pad_mask).argmax(-1)
+        else:
+            last_non_pad_token = -1
+            logger.warning_once(
+                f"{self.__class__.__name__} will not detect padding tokens in `inputs_embeds`. Results may be "
+                "unexpected if using padding tokens in conjunction with `inputs_embeds.`"
+            )
+
+        pooled_logits = logits[
+            torch.arange(batch_size, device=logits.device), last_non_pad_token
+        ]
+
+        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,
+        )
+
+
+@auto_docstring
+class GPT2ForTokenClassification(GPT2PreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.transformer = GPT2Model(config)
+        if (
+            hasattr(config, "classifier_dropout")
+            and config.classifier_dropout is not None
+        ):
+            classifier_dropout = config.classifier_dropout
+        elif hasattr(config, "hidden_dropout") and config.hidden_dropout is not None:
+            classifier_dropout = config.hidden_dropout
+        else:
+            classifier_dropout = 0.1
+        self.dropout = nn.Dropout(classifier_dropout)
+        self.classifier = nn.Linear(config.hidden_size, config.num_labels)
+
+        # Model parallel
+        self.model_parallel = False
+        self.device_map = None
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, TokenClassifierOutput]:
+        r"""
+        input_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`):
+            `input_ids_length` = `sequence_length` if `past_key_values` is `None` else
+            `past_key_values[0][0].shape[-2]` (`sequence_length` of input past key value states). Indices of input
+            sequence tokens in the vocabulary.
+
+            If `past_key_values` is used, only `input_ids` that do not have their past calculated should be passed as
+            `input_ids`.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *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.transformer(
+            input_ids,
+            past_key_values=past_key_values,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = transformer_outputs[0]
+        hidden_states = self.dropout(hidden_states)
+        logits = self.classifier(hidden_states)
+
+        loss = None
+        if labels is not None:
+            labels = labels.to(logits.device)
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+
+        if not return_dict:
+            output = (logits,) + transformer_outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TokenClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
+
+
+@auto_docstring
+class GPT2ForQuestionAnswering(GPT2PreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.transformer = GPT2Model(config)
+        self.qa_outputs = nn.Linear(config.hidden_size, 2)
+
+        # Model parallel
+        self.model_parallel = False
+        self.device_map = None
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        start_positions: Optional[torch.LongTensor] = None,
+        end_positions: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, QuestionAnsweringModelOutput]:
+        r"""
+        input_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`):
+            `input_ids_length` = `sequence_length` if `past_key_values` is `None` else
+            `past_key_values[0][0].shape[-2]` (`sequence_length` of input past key value states). Indices of input
+            sequence tokens in the vocabulary.
+
+            If `past_key_values` is used, only `input_ids` that do not have their past calculated should be passed as
+            `input_ids`.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        """
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        outputs = self.transformer(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        sequence_output = outputs[0]
+
+        logits = self.qa_outputs(sequence_output)
+        start_logits, end_logits = logits.split(1, dim=-1)
+        start_logits = start_logits.squeeze(-1).contiguous()
+        end_logits = end_logits.squeeze(-1).contiguous()
+
+        total_loss = None
+        if start_positions is not None and end_positions is not None:
+            # If we are on multi-GPU, split add a dimension
+            if len(start_positions.size()) > 1:
+                start_positions = start_positions.squeeze(-1).to(start_logits.device)
+            if len(end_positions.size()) > 1:
+                end_positions = end_positions.squeeze(-1).to(end_logits.device)
+            # sometimes the start/end positions are outside our model inputs, we ignore these terms
+            ignored_index = start_logits.size(1)
+            start_positions = start_positions.clamp(0, ignored_index)
+            end_positions = end_positions.clamp(0, ignored_index)
+
+            loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
+            start_loss = loss_fct(start_logits, start_positions)
+            end_loss = loss_fct(end_logits, end_positions)
+            total_loss = (start_loss + end_loss) / 2
+
+        if not return_dict:
+            output = (start_logits, end_logits) + outputs[2:]
+            return ((total_loss,) + output) if total_loss is not None else output
+
+        return QuestionAnsweringModelOutput(
+            loss=total_loss,
+            start_logits=start_logits,
+            end_logits=end_logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
 __all__ = [
     "GPT2DoubleHeadsModel",
     "GPT2ForQuestionAnswering",