Update modeling_gpt2.py

updating based on transformers==4.52.4

modeling_gpt2.py

@@ -19,19 +19,18 @@ import math
 import os
 import warnings
 from dataclasses import dataclass
-from typing import Optional, Tuple, Union
+from typing import Callable, Optional, Tuple, Union
 
 import torch
-import torch.utils.checkpoint
-from packaging import version
 from torch import nn
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
-from transformers.activations import ACT2FN
+from transformers.activations import ACT2FN, get_activation
+from transformers.cache_utils import Cache, DynamicCache, EncoderDecoderCache, StaticCache
 from transformers.generation import GenerationMixin
 from transformers.modeling_attn_mask_utils import (
+    AttentionMaskConverter,
     _prepare_4d_attention_mask_for_sdpa,
-    _prepare_4d_causal_attention_mask_for_sdpa,
 )
 from transformers.modeling_outputs import (
     BaseModelOutputWithPastAndCrossAttentions,
@@ -40,7 +39,7 @@ from transformers.modeling_outputs import (
     SequenceClassifierOutputWithPast,
     TokenClassifierOutput,
 )
-from transformers.modeling_utils import PreTrainedModel, SequenceSummary
+from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
 from transformers.pytorch_utils import (
     Conv1D,
     find_pruneable_heads_and_indices,
@@ -48,86 +47,21 @@ from transformers.pytorch_utils import (
 )
 from transformers.utils import (
     ModelOutput,
-    add_code_sample_docstrings,
     add_start_docstrings,
-    add_start_docstrings_to_model_forward,
-    get_torch_version,
-    is_flash_attn_2_available,
-    is_flash_attn_greater_or_equal_2_10,
+    auto_docstring,
     logging,
-    replace_return_docstrings,
 )
+from transformers.utils.deprecation import deprecate_kwarg
 from transformers.utils.model_parallel_utils import assert_device_map, get_device_map
 from .configuration_gpt2 import GPT2Config
-
-
-if is_flash_attn_2_available():
-    from transformers.modeling_flash_attention_utils import _flash_attention_forward
+from transformers.models.gpt2.modeling_gpt2 import (
+    load_tf_weights_in_gpt2,
+    eager_attention_forward,
+)
 
 
 logger = logging.get_logger(__name__)
 
-_CHECKPOINT_FOR_DOC = "openai-community/gpt2"
-_CONFIG_FOR_DOC = "GPT2Config"
-
-
-def load_tf_weights_in_gpt2(model, config, gpt2_checkpoint_path):
-    """Load tf checkpoints in a pytorch model"""
-    try:
-        import re
-
-        import tensorflow as tf
-    except ImportError:
-        logger.error(
-            "Loading a TensorFlow model in PyTorch, requires TensorFlow to be installed. Please see "
-            "https://www.tensorflow.org/install/ for installation instructions."
-        )
-        raise
-    tf_path = os.path.abspath(gpt2_checkpoint_path)
-    logger.info(f"Converting TensorFlow checkpoint from {tf_path}")
-    # Load weights from TF model
-    init_vars = tf.train.list_variables(tf_path)
-    names = []
-    arrays = []
-    for name, shape in init_vars:
-        logger.info(f"Loading TF weight {name} with shape {shape}")
-        array = tf.train.load_variable(tf_path, name)
-        names.append(name)
-        arrays.append(array.squeeze())
-
-    for name, array in zip(names, arrays):
-        name = name[6:]  # skip "model/"
-        name = name.split("/")
-        pointer = model
-        for m_name in name:
-            if re.fullmatch(r"[A-Za-z]+\d+", m_name):
-                scope_names = re.split(r"(\d+)", m_name)
-            else:
-                scope_names = [m_name]
-            if scope_names[0] == "w" or scope_names[0] == "g":
-                pointer = getattr(pointer, "weight")
-            elif scope_names[0] == "b":
-                pointer = getattr(pointer, "bias")
-            elif scope_names[0] == "wpe" or scope_names[0] == "wte":
-                pointer = getattr(pointer, scope_names[0])
-                pointer = getattr(pointer, "weight")
-            else:
-                pointer = getattr(pointer, scope_names[0])
-            if len(scope_names) >= 2:
-                num = int(scope_names[1])
-                pointer = pointer[num]
-        try:
-            if pointer.shape != array.shape:
-                raise ValueError(
-                    f"Pointer shape {pointer.shape} and array shape {array.shape} mismatched"
-                )
-        except ValueError as e:
-            e.args += (pointer.shape, array.shape)
-            raise
-        logger.info(f"Initialize PyTorch weight {name}")
-        pointer.data = torch.from_numpy(array)
-    return model
-
 
 class GPT2Attention(nn.Module):
     def __init__(self, config, is_cross_attention=False, layer_idx=None):
@@ -195,55 +129,6 @@ class GPT2Attention(nn.Module):
         self.num_heads = self.num_heads - len(heads)
         self.pruned_heads = self.pruned_heads.union(heads)
 
-    def _attn(self, query, key, value, attention_mask=None, head_mask=None):
-        attn_weights = torch.matmul(query, key.transpose(-1, -2))
-
-        if self.scale_attn_weights:
-            attn_weights = attn_weights / torch.full(
-                [],
-                value.size(-1) ** 0.5,
-                dtype=attn_weights.dtype,
-                device=attn_weights.device,
-            )
-
-        # Layer-wise attention scaling
-        if self.scale_attn_by_inverse_layer_idx:
-            attn_weights = attn_weights / float(self.layer_idx + 1)
-
-        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
-            ]
-            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.full(
-                [], mask_value, dtype=attn_weights.dtype, device=attn_weights.device
-            )
-            attn_weights = torch.where(
-                causal_mask, attn_weights.to(attn_weights.dtype), mask_value
-            )
-
-        if attention_mask is not None:
-            # Apply the attention mask
-            attn_weights = attn_weights + attention_mask
-
-        attn_weights = nn.functional.softmax(attn_weights, dim=-1)
-
-        # Downcast (if necessary) back to V's dtype (if in mixed-precision) -- No-Op otherwise
-        attn_weights = attn_weights.type(value.dtype)
-        attn_weights = self.attn_dropout(attn_weights)
-
-        # Mask heads if we want to
-        if head_mask is not None:
-            attn_weights = attn_weights * head_mask
-
-        attn_output = torch.matmul(attn_weights, value)
-
-        return attn_output, attn_weights
-
     def _upcast_and_reordered_attn(
         self, query, key, value, attention_mask=None, head_mask=None
     ):
@@ -287,8 +172,8 @@ class GPT2Attention(nn.Module):
             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).to(
-                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)
 
@@ -311,321 +196,111 @@ class GPT2Attention(nn.Module):
             attn_weights = attn_weights * head_mask
 
         attn_output = torch.matmul(attn_weights, value)
+        attn_output = attn_output.transpose(1, 2)
 
         return attn_output, attn_weights
 
-    def _split_heads(self, tensor, num_heads, attn_head_size):
-        """
-        Splits hidden_size dim into attn_head_size and num_heads
-        """
-        new_shape = tensor.size()[:-1] + (num_heads, attn_head_size)
-        tensor = tensor.view(new_shape)
-        return tensor.permute(0, 2, 1, 3)  # (batch, head, seq_length, head_features)
-
-    def _merge_heads(self, tensor, num_heads, attn_head_size):
-        """
-        Merges attn_head_size dim and num_attn_heads dim into hidden_size
-        """
-        tensor = tensor.permute(0, 2, 1, 3).contiguous()
-        new_shape = tensor.size()[:-2] + (num_heads * attn_head_size,)
-        return tensor.view(new_shape)
-
+    @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]],
-        layer_past: Optional[Tuple[torch.Tensor]] = None,
+        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,
     ) -> Tuple[Union[torch.Tensor, Tuple[torch.Tensor]], ...]:
-        if encoder_hidden_states is not None:
+        is_cross_attention = encoder_hidden_states is not None
+        if is_cross_attention:
             if not hasattr(self, "q_attn"):
                 raise ValueError(
                     "If class is used as cross attention, the weights `q_attn` have to be defined. "
                     "Please make sure to instantiate class with `GPT2Attention(..., is_cross_attention=True)`."
                 )
 
-            query = self.q_attn(hidden_states)
-            key, value = self.c_attn(encoder_hidden_states).split(
+            query_states = self.q_attn(hidden_states)
+            key_states, value_states = self.c_attn(encoder_hidden_states).split(
                 self.split_size, dim=2
             )
             attention_mask = encoder_attention_mask
         else:
-            query, key, value = self.c_attn(hidden_states).split(self.split_size, dim=2)
-
-        query = self._split_heads(query, self.num_heads, self.head_dim)
-        key = self._split_heads(key, self.num_heads, self.head_dim)
-        value = self._split_heads(value, self.num_heads, self.head_dim)
-
-        if layer_past is not None:
-            past_key, past_value = layer_past
-            key = torch.cat((past_key, key), dim=-2)
-            value = torch.cat((past_value, value), dim=-2)
-
-        if use_cache is True:
-            present = (key, value)
-        else:
-            present = None
-
-        if self.reorder_and_upcast_attn:
-            attn_output, attn_weights = self._upcast_and_reordered_attn(
-                query, key, value, attention_mask, head_mask
-            )
-        else:
-            attn_output, attn_weights = self._attn(
-                query, key, value, attention_mask, head_mask
-            )
-
-        attn_output = self._merge_heads(attn_output, self.num_heads, self.head_dim)
-        attn_output = self.c_proj(attn_output)
-        attn_output = self.resid_dropout(attn_output)
-
-        outputs = (attn_output, present)
-        if output_attentions:
-            outputs += (attn_weights,)
-
-        return outputs  # a, present, (attentions)
-
-
-class GPT2FlashAttention2(GPT2Attention):
-    """
-    GPT2 flash attention module. This module inherits from `GPT2Attention` 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: Optional[Tuple[torch.FloatTensor]],
-        layer_past: Optional[Tuple[torch.Tensor]] = 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,
-    ) -> Tuple[Union[torch.Tensor, Tuple[torch.Tensor]], ...]:
-        bsz, _, _ = hidden_states.size()
-        if encoder_hidden_states is not None:
-            if not hasattr(self, "q_attn"):
-                raise ValueError(
-                    "If class is used as cross attention, the weights `q_attn` have to be defined. "
-                    "Please make sure to instantiate class with `GPT2Attention(..., is_cross_attention=True)`."
-                )
-
-            query = self.q_attn(hidden_states)
-            key, value = self.c_attn(encoder_hidden_states).split(
+            query_states, key_states, value_states = self.c_attn(hidden_states).split(
                 self.split_size, dim=2
             )
-            attention_mask = encoder_attention_mask
-        else:
-            query, key, value = self.c_attn(hidden_states).split(self.split_size, dim=2)
-
-        query = self._split_heads(query, self.num_heads, self.head_dim)
-        key = self._split_heads(key, self.num_heads, self.head_dim)
-        value = self._split_heads(value, self.num_heads, self.head_dim)
-
-        if layer_past is not None:
-            past_key = layer_past[0]
-            past_value = layer_past[1]
-            key = torch.cat((past_key, key), dim=-2)
-            value = torch.cat((past_value, value), dim=-2)
 
-        present = None
-        if use_cache is True:
-            present = (key, value)
+        shape_q = (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_length = query.shape[2]
-        tgt_len = key.shape[2]
+        query_states = query_states.view(shape_q).transpose(1, 2)
+        key_states = key_states.view(shape_kv).transpose(1, 2)
+        value_states = value_states.view(shape_kv).transpose(1, 2)
 
-        # Flash attention requires the input to have the shape
-        # batch_size x seq_length x head_dim x hidden_dim
-        query = query.transpose(1, 2).view(
-            bsz, query_length, self.num_heads, self.head_dim
-        )
-        key = key.transpose(1, 2).view(bsz, tgt_len, self.num_heads, self.head_dim)
-        value = value.transpose(1, 2).view(bsz, tgt_len, self.num_heads, self.head_dim)
-
-        attn_dropout = self.attn_dropout.p 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 the correct dtype just to be sure everything works as expected.
-        # This might slowdown training & inference so it is recommended to not cast the LayerNorms
-        # in fp32. (LlamaRMSNorm handles it correctly)
-
-        if query.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.c_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}."
+        if past_key_value is not None:
+            if isinstance(past_key_value, EncoderDecoderCache):
+                if is_cross_attention:
+                    past_key_value = past_key_value.cross_attention_cache
+                else:
+                    past_key_value = past_key_value.self_attention_cache
+            cache_kwargs = {"cache_position": cache_position}
+            key_states, value_states = past_key_value.update(
+                key_states, value_states, self.layer_idx, cache_kwargs=cache_kwargs
             )
 
-            query = query.to(target_dtype)
-            key = key.to(target_dtype)
-            value = value.to(target_dtype)
-
-        attn_output = _flash_attention_forward(
-            query,
-            key,
-            value,
-            attention_mask,
-            query_length,
-            dropout=attn_dropout,
-            is_causal=self.is_causal,
-            use_top_left_mask=self._flash_attn_uses_top_left_mask,
-        )
-
-        attn_weights_reshaped = attn_output.reshape(
-            bsz, query_length, self.num_heads * self.head_dim
+        is_causal = (
+            attention_mask is None
+            and query_states.shape[-2] > 1
+            and not is_cross_attention
         )
-        attn_output = self.c_proj(attn_weights_reshaped)
-        attn_output = self.resid_dropout(attn_output)
-
-        outputs = (attn_output, present)
-        if output_attentions:
-            outputs += (attn_weights_reshaped,)
-
-        return outputs
-
-
-class GPT2SdpaAttention(GPT2Attention):
-    """
-    GPT2 attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
-    `GPT2Attention` as the weights of the module stays untouched. The only changes are on the forward pass
-    to adapt to the SDPA API.
-    """
-
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-
-        # Idea adapted from transformers.models.bert.modeling_bert.BertSdpaSelfAttention.__init__
-        # SDPA with memory-efficient backend is broken in torch==2.1.2 when using non-contiguous inputs and a custom
-        # attn_mask, so we need to call `.contiguous()`. This was fixed in torch==2.2.0.
-        # Reference: https://github.com/pytorch/pytorch/issues/112577
-        self.require_contiguous_qkv = version.parse(
-            get_torch_version()
-        ) < version.parse("2.2.0")
-
-    def forward(
-        self,
-        hidden_states: Optional[Tuple[torch.FloatTensor]],
-        layer_past: Optional[Tuple[torch.Tensor]] = 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,
-    ) -> Tuple[Union[torch.Tensor, Tuple[torch.Tensor]], ...]:
-        if output_attentions or head_mask is not None:
-            logger.warning_once(
-                "`GPT2SdpaAttention` is used but `torch.nn.functional.scaled_dot_product_attention` does not support "
-                "`output_attentions=True` or `head_mask`. Falling back to the manual attention implementation, but "
-                "specifying the manual implementation will be required from Transformers version v5.0.0 onwards. "
-                'This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
-            )
-            return super().forward(
-                hidden_states=hidden_states,
-                layer_past=layer_past,
-                attention_mask=attention_mask,
-                head_mask=head_mask,
-                encoder_hidden_states=encoder_hidden_states,
-                encoder_attention_mask=encoder_attention_mask,
-                use_cache=use_cache,
-                output_attentions=output_attentions,
-            )
 
-        bsz, q_len, _ = hidden_states.size()
-
-        # Initial attention projections
-        is_cross_attention = encoder_hidden_states is not None
-        if is_cross_attention:
-            if not hasattr(self, "q_attn"):
-                raise ValueError(
-                    "If class is used as cross attention, the weights `q_attn` have to be defined. "
-                    "Please make sure to instantiate class with `GPT2SdpaAttention(..., is_cross_attention=True)`."
+        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
+            ):
+                using_eager = True
+                logger.warning_once(
+                    "`torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to "
+                    'eager attention. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
                 )
+            else:
+                # 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
+                ]
 
-            query = self.q_attn(hidden_states)
-            key, value = self.c_attn(encoder_hidden_states).split(
-                self.split_size, dim=2
+        if using_eager and self.reorder_and_upcast_attn:
+            attn_output, attn_weights = self._upcast_and_reordered_attn(
+                query_states, key_states, value_states, attention_mask, head_mask
             )
-            attention_mask = encoder_attention_mask
         else:
-            query, key, value = self.c_attn(hidden_states).split(self.split_size, dim=2)
-
-        query = self._split_heads(query, self.num_heads, self.head_dim)
-        key = self._split_heads(key, self.num_heads, self.head_dim)
-        value = self._split_heads(value, self.num_heads, self.head_dim)
-
-        # Optional kv caching
-        if layer_past is not None:
-            past_key = layer_past[0]
-            past_value = layer_past[1]
-            key = torch.cat((past_key, key), dim=-2)
-            value = torch.cat((past_value, value), dim=-2)
-
-        present = None
-        if use_cache is True:
-            present = (key, value)
-
-        # Avoid torch==2.1.2 specific bug for the memory-efficient backend in SDPA
-        if (
-            self.require_contiguous_qkv
-            and query.device.type == "cuda"
-            and attention_mask is not None
-        ):
-            query = query.contiguous()
-            key = key.contiguous()
-            value = value.contiguous()
-
-        # We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment
-        # in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling.
-        is_causal = (
-            True
-            if attention_mask is None and q_len > 1 and not is_cross_attention
-            else False
-        )
-
-        attn_output = torch.nn.functional.scaled_dot_product_attention(
-            query,
-            key,
-            value,
-            attn_mask=attention_mask,
-            dropout_p=self.attn_dropout.p if self.training else 0.0,
-            is_causal=is_causal,
-        )
-
-        # Reshape outputs
-        attn_output = attn_output.transpose(1, 2).contiguous()
-        attn_output = attn_output.view(bsz, q_len, self.embed_dim)
+            attn_output, attn_weights = attention_interface(
+                self,
+                query_states,
+                key_states,
+                value_states,
+                attention_mask,
+                head_mask=head_mask,
+                dropout=self.attn_dropout.p if self.training else 0.0,
+                is_causal=is_causal,
+                **kwargs,
+            )
 
-        # Final projection
+        attn_output = attn_output.reshape(attn_output.shape[0],attn_output.shape[1], -1).contiguous()
         attn_output = self.c_proj(attn_output)
         attn_output = self.resid_dropout(attn_output)
 
-        return attn_output, present, None
+        return attn_output, attn_weights
 
 
 class GPT2MLP(nn.Module):
@@ -647,26 +322,18 @@ class GPT2MLP(nn.Module):
         return hidden_states
 
 
-GPT2_ATTENTION_CLASSES = {
-    "eager": GPT2Attention,
-    "flash_attention_2": GPT2FlashAttention2,
-    "sdpa": GPT2SdpaAttention,
-}
-
-
 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
-        attention_class = GPT2_ATTENTION_CLASSES[config._attn_implementation]
 
         self.ln_1 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
-        self.attn = attention_class(config=config, layer_idx=layer_idx)
+        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 = attention_class(
+            self.crossattention = GPT2Attention(
                 config=config, is_cross_attention=True, layer_idx=layer_idx
             )
             self.ln_cross_attn = nn.LayerNorm(
@@ -675,32 +342,40 @@ class GPT2Block(nn.Module):
 
         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]],
-        layer_past: Optional[Tuple[torch.Tensor]] = None,
+        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_outputs = self.attn(
+        attn_output, self_attn_weights = self.attn(
             hidden_states,
-            layer_past=layer_past,
+            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,
         )
-        attn_output = attn_outputs[0]  # output_attn: a, present, (attentions)
-        outputs = attn_outputs[1:]
         # residual connection
         hidden_states = attn_output + residual
 
@@ -713,20 +388,17 @@ class GPT2Block(nn.Module):
                 )
             residual = hidden_states
             hidden_states = self.ln_cross_attn(hidden_states)
-            cross_attn_outputs = self.crossattention(
+            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,
             )
-            attn_output = cross_attn_outputs[0]
             # residual connection
-            hidden_states = residual + attn_output
-            outputs = (
-                outputs + cross_attn_outputs[2:]
-            )  # add cross attentions if we output attention weights
+            hidden_states = residual + cross_attn_output
 
         residual = hidden_states
         hidden_states = self.ln_2(hidden_states)
@@ -734,20 +406,132 @@ class GPT2Block(nn.Module):
         # residual connection
         hidden_states = residual + feed_forward_hidden_states
 
-        if use_cache:
-            outputs = (hidden_states,) + outputs
-        else:
-            outputs = (hidden_states,) + outputs[1:]
+        outputs = (hidden_states,)
+        if output_attentions:
+            outputs += (self_attn_weights,)
+            if encoder_hidden_states is not None:
+                outputs += (cross_attn_weights,)
 
-        return outputs  # hidden_states, present, (attentions, cross_attentions)
+        return outputs
 
 
-class GPT2PreTrainedModel(PreTrainedModel):
-    """
-    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
-    models.
+# 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"
@@ -757,6 +541,9 @@ class GPT2PreTrainedModel(PreTrainedModel):
     _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)
@@ -830,96 +617,13 @@ class GPT2DoubleHeadsModelOutput(ModelOutput):
 
     loss: Optional[torch.FloatTensor] = None
     mc_loss: Optional[torch.FloatTensor] = None
-    logits: torch.FloatTensor = None
-    mc_logits: 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
 
 
-GPT2_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 ([`GPT2Config`]): 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.
-"""
-
-GPT2_INPUTS_DOCSTRING = r"""
-    Args:
-        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)
-        past_key_values (`Tuple[Tuple[torch.Tensor]]` of length `config.n_layers`):
-            Contains precomputed hidden-states (key and values in the attention blocks) as computed by the model (see
-            `past_key_values` output below). Can be used to speed up sequential decoding. The `input_ids` which have
-            their past given to this model should not be passed as `input_ids` as they have already been computed.
-        attention_mask (`torch.FloatTensor` 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**.
-
-            If `past_key_values` is used, `attention_mask` needs to contain the masking strategy that was used for
-            `past_key_values`. In other words, the `attention_mask` always has to have the length:
-            `len(past_key_values) + len(input_ids)`
-
-            [What are attention masks?](../glossary#attention-mask)
-        token_type_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`, *optional*):
-            Segment token indices to indicate first and second portions of the inputs. Indices are selected in `[0,
-            1]`:
-
-            - 0 corresponds to a *sentence A* token,
-            - 1 corresponds to a *sentence B* token.
-
-            [What are token type IDs?](../glossary#token-type-ids)
-        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
-            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
-            config.max_position_embeddings - 1]`.
-
-            [What are position IDs?](../glossary#position-ids)
-        head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
-            Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:
-
-            - 1 indicates the head is **not masked**,
-            - 0 indicates the head is **masked**.
-
-        inputs_embeds (`torch.FloatTensor` of shape `(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.
-
-            If `past_key_values` is used, optionally only the last `inputs_embeds` have to be input (see
-            `past_key_values`).
-        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.
-"""
 PARALLELIZE_DOCSTRING = r"""
     This is an experimental feature and is a subject to change at a moment's notice.
 
@@ -972,10 +676,7 @@ DEPARALLELIZE_DOCSTRING = r"""
 """
 
 
-@add_start_docstrings(
-    "The bare GPT2 Model transformer outputting raw hidden-states without any specific head on top.",
-    GPT2_START_DOCSTRING,
-)
+@auto_docstring
 class GPT2Model(GPT2PreTrainedModel):
     _supports_param_buffer_assignment = False
 
@@ -1065,16 +766,12 @@ class GPT2Model(GPT2PreTrainedModel):
         for layer, heads in heads_to_prune.items():
             self.h[layer].attn.prune_heads(heads)
 
-    @add_start_docstrings_to_model_forward(GPT2_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(
-        checkpoint=_CHECKPOINT_FOR_DOC,
-        output_type=BaseModelOutputWithPastAndCrossAttentions,
-        config_class=_CONFIG_FOR_DOC,
-    )
+    @auto_docstring
     def forward(
         self,
         input_ids: Optional[torch.LongTensor] = None,
-        past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = 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,
@@ -1086,7 +783,22 @@ class GPT2Model(GPT2PreTrainedModel):
         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
@@ -1122,68 +834,70 @@ class GPT2Model(GPT2PreTrainedModel):
         if token_type_ids is not None:
             token_type_ids = token_type_ids.view(-1, input_shape[-1])
 
-        if past_key_values is None:
-            past_length = 0
-            past_key_values = tuple([None] * len(self.h))
-        else:
-            past_length = past_key_values[0][0].size(-2)
-        if position_ids is None:
-            position_ids = torch.arange(
-                past_length,
-                input_shape[-1] + past_length,
-                dtype=torch.long,
-                device=device,
-            )
-            position_ids = position_ids.unsqueeze(0)
+        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
+        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
         )
-        attention_mask = (
-            attention_mask.view(batch_size, -1) if attention_mask is not None else None
-        )
-        if self._attn_implementation == "flash_attention_2":
-            attention_mask = (
-                attention_mask
-                if (attention_mask is not None and 0 in attention_mask)
-                else None
-            )
-        elif _use_sdpa:
-            attention_mask = _prepare_4d_causal_attention_mask_for_sdpa(
-                attention_mask=attention_mask,
-                input_shape=(batch_size, input_shape[-1]),
-                inputs_embeds=inputs_embeds,
-                past_key_values_length=past_length,
-            )
-        else:
-            if attention_mask is not None:
-                # We create a 3D attention mask from a 2D tensor mask.
-                # Sizes are [batch_size, 1, 1, to_seq_length]
-                # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
-                # this attention mask is more simple than the triangular masking of causal attention
-                # used in OpenAI GPT, we just need to prepare the broadcast dimension here.
-                attention_mask = attention_mask[:, None, None, :]
-
-                # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
-                # masked positions, this operation will create a tensor which is 0.0 for
-                # positions we want to attend and the dtype's smallest value for masked positions.
-                # Since we are adding it to the raw scores before the softmax, this is
-                # effectively the same as removing these entirely.
-                attention_mask = attention_mask.to(
-                    dtype=self.dtype
-                )  # fp16 compatibility
-                attention_mask = (1.0 - attention_mask) * torch.finfo(self.dtype).min
-
-        # 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]
         if self.config.add_cross_attention and encoder_hidden_states is not None:
             encoder_batch_size, encoder_sequence_length, _ = (
                 encoder_hidden_states.size()
@@ -1218,29 +932,15 @@ class GPT2Model(GPT2PreTrainedModel):
 
         output_shape = (-1,) + input_shape[1:] + (hidden_states.size(-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
-
-        presents = () if use_cache else None
         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 in range(len(self.h)):
-            block, layer_past = self.h[i], past_key_values[i]
+        for i, block in enumerate(self.h):
             # Model parallel
             if self.model_parallel:
                 torch.cuda.set_device(hidden_states.device)
-                # Ensure layer_past is on same device as hidden_states (might not be correct)
-                if layer_past is not None:
-                    layer_past = tuple(
-                        past_state.to(hidden_states.device) for past_state in layer_past
-                    )
                 # 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)
@@ -1253,8 +953,9 @@ class GPT2Model(GPT2PreTrainedModel):
                 outputs = self._gradient_checkpointing_func(
                     block.__call__,
                     hidden_states,
-                    None,
-                    attention_mask,
+                    past_key_values,
+                    cache_position,
+                    causal_mask,
                     head_mask[i],
                     encoder_hidden_states,
                     encoder_attention_mask,
@@ -1264,27 +965,23 @@ class GPT2Model(GPT2PreTrainedModel):
             else:
                 outputs = block(
                     hidden_states,
-                    layer_past=layer_past,
-                    attention_mask=attention_mask,
+                    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 use_cache is True:
-                presents = presents + (outputs[1],)
 
             if output_attentions:
-                all_self_attentions = all_self_attentions + (
-                    outputs[2 if use_cache else 1],
-                )
+                all_self_attentions = all_self_attentions + (outputs[1],)
                 if self.config.add_cross_attention:
-                    all_cross_attentions = all_cross_attentions + (
-                        outputs[3 if use_cache else 2],
-                    )
+                    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:
@@ -1299,12 +996,19 @@ class GPT2Model(GPT2PreTrainedModel):
         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,
-                    presents,
+                    past_key_values,
                     all_hidden_states,
                     all_self_attentions,
                     all_cross_attentions,
@@ -1314,19 +1018,153 @@ class GPT2Model(GPT2PreTrainedModel):
 
         return BaseModelOutputWithPastAndCrossAttentions(
             last_hidden_state=hidden_states,
-            past_key_values=presents,
+            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)
 
-@add_start_docstrings(
-    """
+        # 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).
-    """,
-    GPT2_START_DOCSTRING,
+    """
 )
 class GPT2LMHeadModel(GPT2PreTrainedModel, GenerationMixin):
     _tied_weights_keys = ["lm_head.weight"]
@@ -1380,16 +1218,12 @@ class GPT2LMHeadModel(GPT2PreTrainedModel, GenerationMixin):
     def set_output_embeddings(self, new_embeddings):
         self.lm_head = new_embeddings
 
-    @add_start_docstrings_to_model_forward(GPT2_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(
-        checkpoint=_CHECKPOINT_FOR_DOC,
-        output_type=CausalLMOutputWithCrossAttentions,
-        config_class=_CONFIG_FOR_DOC,
-    )
+    @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,
@@ -1402,9 +1236,22 @@ class GPT2LMHeadModel(GPT2PreTrainedModel, GenerationMixin):
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
+        **kwargs,
     ) -> Union[Tuple, CausalLMOutputWithCrossAttentions]:
         r"""
-        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+        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]`
@@ -1417,6 +1264,7 @@ class GPT2LMHeadModel(GPT2PreTrainedModel, GenerationMixin):
             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,
@@ -1439,15 +1287,12 @@ class GPT2LMHeadModel(GPT2PreTrainedModel, GenerationMixin):
 
         loss = None
         if labels is not None:
-            # move labels to correct device to enable model parallelism
-            labels = labels.to(lm_logits.device)
-            # Shift so that tokens < n predict n
-            shift_logits = lm_logits[..., :-1, :].contiguous()
-            shift_labels = labels[..., 1:].contiguous()
             # Flatten the tokens
-            loss_fct = CrossEntropyLoss()
-            loss = loss_fct(
-                shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1)
+            loss = self.loss_function(
+                lm_logits,
+                labels,
+                vocab_size=self.config.vocab_size,
+                **kwargs,
             )
 
         if not return_dict:
@@ -1463,32 +1308,14 @@ class GPT2LMHeadModel(GPT2PreTrainedModel, GenerationMixin):
             cross_attentions=transformer_outputs.cross_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
-        )
-
 
-@add_start_docstrings(
+@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).
     """
-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).
-""",
-    GPT2_START_DOCSTRING,
 )
 class GPT2DoubleHeadsModel(GPT2PreTrainedModel, GenerationMixin):
     _tied_weights_keys = ["lm_head.weight"]
@@ -1498,7 +1325,7 @@ class GPT2DoubleHeadsModel(GPT2PreTrainedModel, GenerationMixin):
         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 = SequenceSummary(config)
+        self.multiple_choice_head = GPT2SequenceSummary(config)
 
         # Model parallel
         self.model_parallel = False
@@ -1548,14 +1375,12 @@ class GPT2DoubleHeadsModel(GPT2PreTrainedModel, GenerationMixin):
     def set_output_embeddings(self, new_embeddings):
         self.lm_head = new_embeddings
 
-    @add_start_docstrings_to_model_forward(GPT2_INPUTS_DOCSTRING)
-    @replace_return_docstrings(
-        output_type=GPT2DoubleHeadsModelOutput, config_class=_CONFIG_FOR_DOC
-    )
+    @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,
@@ -1571,10 +1396,22 @@ class GPT2DoubleHeadsModel(GPT2PreTrainedModel, GenerationMixin):
         **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, sequence_length)`, *optional*):
+        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]`
@@ -1582,8 +1419,6 @@ class GPT2DoubleHeadsModel(GPT2PreTrainedModel, GenerationMixin):
             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)
 
-        Return:
-
         Example:
 
         ```python
@@ -1616,6 +1451,7 @@ class GPT2DoubleHeadsModel(GPT2PreTrainedModel, GenerationMixin):
         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,
@@ -1687,8 +1523,8 @@ class GPT2DoubleHeadsModel(GPT2PreTrainedModel, GenerationMixin):
         )
 
 
-@add_start_docstrings(
-    """
+@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
@@ -1699,8 +1535,7 @@ class GPT2DoubleHeadsModel(GPT2PreTrainedModel, GenerationMixin):
     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).
-    """,
-    GPT2_START_DOCSTRING,
+    """
 )
 class GPT2ForSequenceClassification(GPT2PreTrainedModel):
     def __init__(self, config):
@@ -1716,12 +1551,7 @@ class GPT2ForSequenceClassification(GPT2PreTrainedModel):
         # Initialize weights and apply final processing
         self.post_init()
 
-    @add_start_docstrings_to_model_forward(GPT2_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(
-        checkpoint="microsoft/DialogRPT-updown",
-        output_type=SequenceClassifierOutputWithPast,
-        config_class=_CONFIG_FOR_DOC,
-    )
+    @auto_docstring
     def forward(
         self,
         input_ids: Optional[torch.LongTensor] = None,
@@ -1738,6 +1568,18 @@ class GPT2ForSequenceClassification(GPT2PreTrainedModel):
         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
@@ -1768,28 +1610,30 @@ class GPT2ForSequenceClassification(GPT2PreTrainedModel):
         else:
             batch_size, sequence_length = inputs_embeds.shape[:2]
 
-        assert (
-            self.config.pad_token_id is not None or batch_size == 1
-        ), "Cannot handle batch sizes > 1 if no padding token is defined."
+        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:
-            sequence_lengths = -1
+            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:
-            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_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.`"
-                )
+            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), sequence_lengths
+            torch.arange(batch_size, device=logits.device), last_non_pad_token
         ]
 
         loss = None
@@ -1831,13 +1675,7 @@ class GPT2ForSequenceClassification(GPT2PreTrainedModel):
         )
 
 
-@add_start_docstrings(
-    """
-    GPT2 Model with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for
-    Named-Entity-Recognition (NER) tasks.
-    """,
-    GPT2_START_DOCSTRING,
-)
+@auto_docstring
 class GPT2ForTokenClassification(GPT2PreTrainedModel):
     def __init__(self, config):
         super().__init__(config)
@@ -1863,29 +1701,7 @@ class GPT2ForTokenClassification(GPT2PreTrainedModel):
         # Initialize weights and apply final processing
         self.post_init()
 
-    @add_start_docstrings_to_model_forward(GPT2_INPUTS_DOCSTRING)
-    # fmt: off
-    @add_code_sample_docstrings(
-        checkpoint="brad1141/gpt2-finetuned-comp2",
-        output_type=TokenClassifierOutput,
-        config_class=_CONFIG_FOR_DOC,
-        expected_loss=0.25,
-        expected_output=[
-            "Lead",
-            "Lead",
-            "Lead",
-            "Position",
-            "Lead",
-            "Lead",
-            "Lead",
-            "Lead",
-            "Lead",
-            "Lead",
-            "Lead",
-            "Lead",
-        ],
-    )
-    # fmt: on
+    @auto_docstring
     def forward(
         self,
         input_ids: Optional[torch.LongTensor] = None,
@@ -1902,6 +1718,18 @@ class GPT2ForTokenClassification(GPT2PreTrainedModel):
         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
@@ -1947,13 +1775,7 @@ class GPT2ForTokenClassification(GPT2PreTrainedModel):
         )
 
 
-@add_start_docstrings(
-    """
-    The GPT-2 Model transformer with a span classification head on top for extractive question-answering tasks like
-    SQuAD (a linear layer on top of the hidden-states output to compute `span start logits` and `span end logits`).
-    """,
-    GPT2_START_DOCSTRING,
-)
+@auto_docstring
 class GPT2ForQuestionAnswering(GPT2PreTrainedModel):
     def __init__(self, config):
         super().__init__(config)
@@ -1968,15 +1790,7 @@ class GPT2ForQuestionAnswering(GPT2PreTrainedModel):
         # Initialize weights and apply final processing
         self.post_init()
 
-    @add_start_docstrings_to_model_forward(
-        GPT2_INPUTS_DOCSTRING.format("batch_size, sequence_length")
-    )
-    @add_code_sample_docstrings(
-        checkpoint=_CHECKPOINT_FOR_DOC,
-        output_type=QuestionAnsweringModelOutput,
-        config_class=_CONFIG_FOR_DOC,
-        real_checkpoint=_CHECKPOINT_FOR_DOC,
-    )
+    @auto_docstring
     def forward(
         self,
         input_ids: Optional[torch.LongTensor] = None,
@@ -1992,14 +1806,18 @@ class GPT2ForQuestionAnswering(GPT2PreTrainedModel):
         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.
+        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
@@ -2052,3 +1870,15 @@ class GPT2ForQuestionAnswering(GPT2PreTrainedModel):
             hidden_states=outputs.hidden_states,
             attentions=outputs.attentions,
         )
+
+
+__all__ = [
+    "GPT2DoubleHeadsModel",
+    "GPT2ForQuestionAnswering",
+    "GPT2ForSequenceClassification",
+    "GPT2ForTokenClassification",
+    "GPT2LMHeadModel",
+    "GPT2Model",
+    "GPT2PreTrainedModel",
+    "load_tf_weights_in_gpt2",
+]