add weights and tokenizers

model/__init__.py

@@ -0,0 +1,111 @@
+# flake8: noqa
+# There's no way to ignore "F401 '...' imported but unused" warnings in this
+# module, but to preserve other warnings. So, don't check this module at all.
+
+# Copyright 2020 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import TYPE_CHECKING
+
+from ...file_utils import (
+    _BaseLazyModule,
+    is_flax_available,
+    is_tf_available,
+    is_tokenizers_available,
+    is_torch_available,
+)
+
+
+_import_structure = {
+    "configuration_gpt2": ["GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP", "GPT2Config"],
+    "tokenization_gpt2": ["GPT2Tokenizer"],
+}
+
+if is_tokenizers_available():
+    _import_structure["tokenization_gpt2_fast"] = ["GPT2TokenizerFast"]
+
+if is_torch_available():
+    _import_structure["modeling_gpt2"] = [
+        "GPT2_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "GPT2DoubleHeadsModel",
+        "GPT2ForSequenceClassification",
+        "GPT2LMHeadModel",
+        "GPT2Model",
+        "GPT2PreTrainedModel",
+        "load_tf_weights_in_gpt2",
+    ]
+
+if is_tf_available():
+    _import_structure["modeling_tf_gpt2"] = [
+        "TF_GPT2_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "TFGPT2DoubleHeadsModel",
+        "TFGPT2ForSequenceClassification",
+        "TFGPT2LMHeadModel",
+        "TFGPT2MainLayer",
+        "TFGPT2Model",
+        "TFGPT2PreTrainedModel",
+    ]
+
+if is_flax_available():
+    _import_structure["modeling_flax_gpt2"] = ["FlaxGPT2LMHeadModel", "FlaxGPT2Model", "FlaxGPT2PreTrainedModel"]
+
+if TYPE_CHECKING:
+    from .configuration_gpt2 import GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP, GPT2Config
+    from .tokenization_gpt2 import GPT2Tokenizer
+
+    if is_tokenizers_available():
+        from .tokenization_gpt2_fast import GPT2TokenizerFast
+
+    if is_torch_available():
+        from .modeling_gpt2 import (
+            GPT2_PRETRAINED_MODEL_ARCHIVE_LIST,
+            GPT2DoubleHeadsModel,
+            GPT2ForSequenceClassification,
+            GPT2LMHeadModel,
+            GPT2Model,
+            GPT2PreTrainedModel,
+            load_tf_weights_in_gpt2,
+        )
+
+    if is_tf_available():
+        from .modeling_tf_gpt2 import (
+            TF_GPT2_PRETRAINED_MODEL_ARCHIVE_LIST,
+            TFGPT2DoubleHeadsModel,
+            TFGPT2ForSequenceClassification,
+            TFGPT2LMHeadModel,
+            TFGPT2MainLayer,
+            TFGPT2Model,
+            TFGPT2PreTrainedModel,
+        )
+
+    if is_flax_available():
+        from .modeling_flax_gpt2 import FlaxGPT2LMHeadModel, FlaxGPT2Model, FlaxGPT2PreTrainedModel
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

model/configuration_gpt2.py

@@ -0,0 +1,197 @@
+# coding=utf-8
+# Copyright 2018 The OpenAI Team Authors and HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" OpenAI GPT-2 configuration """
+
+from ...configuration_utils import PretrainedConfig
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "gpt2": "https://huggingface.co/gpt2/resolve/main/config.json",
+    "gpt2-medium": "https://huggingface.co/gpt2-medium/resolve/main/config.json",
+    "gpt2-large": "https://huggingface.co/gpt2-large/resolve/main/config.json",
+    "gpt2-xl": "https://huggingface.co/gpt2-xl/resolve/main/config.json",
+    "distilgpt2": "https://huggingface.co/distilgpt2/resolve/main/config.json",
+}
+
+
+class GPT2Config(PretrainedConfig):
+    """
+    This is the configuration class to store the configuration of a :class:`~transformers.GPT2Model` or a
+    :class:`~transformers.TFGPT2Model`. It is used to instantiate a GPT-2 model according to the specified arguments,
+    defining the model architecture. Instantiating a configuration with the defaults will yield a similar configuration
+    to that of the GPT-2 `small <https://huggingface.co/gpt2>`__ architecture.
+
+    Configuration objects inherit from :class:`~transformers.PretrainedConfig` and can be used to control the model
+    outputs. Read the documentation from :class:`~transformers.PretrainedConfig` for more information.
+
+
+    Args:
+        vocab_size (:obj:`int`, `optional`, defaults to 50257):
+            Vocabulary size of the GPT-2 model. Defines the number of different tokens that can be represented by the
+            :obj:`inputs_ids` passed when calling :class:`~transformers.GPT2Model` or
+            :class:`~transformers.TFGPT2Model`.
+        n_positions (:obj:`int`, `optional`, defaults to 1024):
+            The maximum sequence length that this model might ever be used with. Typically set this to something large
+            just in case (e.g., 512 or 1024 or 2048).
+        n_ctx (:obj:`int`, `optional`, defaults to 1024):
+            Dimensionality of the causal mask (usually same as n_positions).
+        n_embd (:obj:`int`, `optional`, defaults to 768):
+            Dimensionality of the embeddings and hidden states.
+        n_layer (:obj:`int`, `optional`, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        n_head (:obj:`int`, `optional`, defaults to 12):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        n_inner (:obj:`int`, `optional`, defaults to None):
+            Dimensionality of the inner feed-forward layers. :obj:`None` will set it to 4 times n_embd
+        activation_function (:obj:`str`, `optional`, defaults to :obj:`"gelu"`):
+            Activation function, to be selected in the list :obj:`["relu", "silu", "gelu", "tanh", "gelu_new"]`.
+        resid_pdrop (:obj:`float`, `optional`, defaults to 0.1):
+            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
+        embd_pdrop (:obj:`int`, `optional`, defaults to 0.1):
+            The dropout ratio for the embeddings.
+        attn_pdrop (:obj:`float`, `optional`, defaults to 0.1):
+            The dropout ratio for the attention.
+        layer_norm_epsilon (:obj:`float`, `optional`, defaults to 1e-5):
+            The epsilon to use in the layer normalization layers
+        initializer_range (:obj:`float`, `optional`, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        summary_type (:obj:`string`, `optional`, defaults to :obj:`"cls_index"`):
+            Argument used when doing sequence summary, used in the models :class:`~transformers.GPT2DoubleHeadsModel`
+            and :class:`~transformers.TFGPT2DoubleHeadsModel`.
+
+            Has to be one of the following options:
+
+                - :obj:`"last"`: Take the last token hidden state (like XLNet).
+                - :obj:`"first"`: Take the first token hidden state (like BERT).
+                - :obj:`"mean"`: Take the mean of all tokens hidden states.
+                - :obj:`"cls_index"`: Supply a Tensor of classification token position (like GPT/GPT-2).
+                - :obj:`"attn"`: Not implemented now, use multi-head attention.
+        summary_use_proj (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Argument used when doing sequence summary, used in the models :class:`~transformers.GPT2DoubleHeadsModel`
+            and :class:`~transformers.TFGPT2DoubleHeadsModel`.
+
+            Whether or not to add a projection after the vector extraction.
+        summary_activation (:obj:`str`, `optional`):
+            Argument used when doing sequence summary. Used in for the multiple choice head in
+            :class:`~transformers.GPT2DoubleHeadsModel`.
+
+            Pass :obj:`"tanh"` for a tanh activation to the output, any other value will result in no activation.
+        summary_proj_to_labels (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Argument used when doing sequence summary, used in the models :class:`~transformers.GPT2DoubleHeadsModel`
+            and :class:`~transformers.TFGPT2DoubleHeadsModel`.
+
+            Whether the projection outputs should have :obj:`config.num_labels` or :obj:`config.hidden_size` classes.
+        summary_first_dropout (:obj:`float`, `optional`, defaults to 0.1):
+            Argument used when doing sequence summary, used in the models :class:`~transformers.GPT2DoubleHeadsModel`
+            and :class:`~transformers.TFGPT2DoubleHeadsModel`.
+
+            The dropout ratio to be used after the projection and activation.
+        scale_attn_weights (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Scale attention weights by dividing by sqrt(hidden_size).
+        gradient_checkpointing (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.
+        use_cache (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Whether or not the model should return the last key/values attentions (not used by all models).
+
+    Example::
+
+        >>> from transformers import GPT2Model, GPT2Config
+
+        >>> # Initializing a GPT2 configuration
+        >>> configuration = GPT2Config()
+
+        >>> # Initializing a model from the configuration
+        >>> model = GPT2Model(configuration)
+
+        >>> # Accessing the model configuration
+        >>> configuration = model.config
+    """
+
+    model_type = "gpt2"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        vocab_size=50257,
+        n_positions=1024,
+        n_ctx=1024,
+        n_embd=768,
+        n_layer=12,
+        n_head=12,
+        n_inner=None,
+        activation_function="gelu_new",
+        resid_pdrop=0.1,
+        embd_pdrop=0.1,
+        attn_pdrop=0.1,
+        layer_norm_epsilon=1e-5,
+        initializer_range=0.02,
+        summary_type="cls_index",
+        summary_use_proj=True,
+        summary_activation=None,
+        summary_proj_to_labels=True,
+        summary_first_dropout=0.1,
+        scale_attn_weights=True,
+        gradient_checkpointing=False,
+        use_cache=True,
+        bos_token_id=50256,
+        eos_token_id=50256,
+        **kwargs
+    ):
+        super().__init__(bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)
+
+        self.vocab_size = vocab_size
+        self.n_ctx = n_ctx
+        self.n_positions = n_positions
+        self.n_embd = n_embd
+        self.n_layer = n_layer
+        self.n_head = n_head
+        self.n_inner = n_inner
+        self.activation_function = activation_function
+        self.resid_pdrop = resid_pdrop
+        self.embd_pdrop = embd_pdrop
+        self.attn_pdrop = attn_pdrop
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.initializer_range = initializer_range
+        self.summary_type = summary_type
+        self.summary_use_proj = summary_use_proj
+        self.summary_activation = summary_activation
+        self.summary_first_dropout = summary_first_dropout
+        self.summary_proj_to_labels = summary_proj_to_labels
+        self.gradient_checkpointing = gradient_checkpointing
+        self.scale_attn_weights = scale_attn_weights
+        self.use_cache = use_cache
+
+        self.bos_token_id = bos_token_id
+        self.eos_token_id = eos_token_id
+
+    @property
+    def max_position_embeddings(self):
+        return self.n_positions
+
+    @property
+    def hidden_size(self):
+        return self.n_embd
+
+    @property
+    def num_attention_heads(self):
+        return self.n_head
+
+    @property
+    def num_hidden_layers(self):
+        return self.n_layer

model/convert_gpt2_original_tf_checkpoint_to_pytorch.py

@@ -0,0 +1,68 @@
+# coding=utf-8
+# Copyright 2018 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Convert OpenAI GPT checkpoint."""
+
+
+import argparse
+
+import torch
+
+from transformers import GPT2Config, GPT2Model, load_tf_weights_in_gpt2
+from transformers.file_utils import CONFIG_NAME, WEIGHTS_NAME
+from transformers.utils import logging
+
+
+logging.set_verbosity_info()
+
+
+def convert_gpt2_checkpoint_to_pytorch(gpt2_checkpoint_path, gpt2_config_file, pytorch_dump_folder_path):
+    # Construct model
+    if gpt2_config_file == "":
+        config = GPT2Config()
+    else:
+        config = GPT2Config.from_json_file(gpt2_config_file)
+    model = GPT2Model(config)
+
+    # Load weights from numpy
+    load_tf_weights_in_gpt2(model, config, gpt2_checkpoint_path)
+
+    # Save pytorch-model
+    pytorch_weights_dump_path = pytorch_dump_folder_path + "/" + WEIGHTS_NAME
+    pytorch_config_dump_path = pytorch_dump_folder_path + "/" + CONFIG_NAME
+    print(f"Save PyTorch model to {pytorch_weights_dump_path}")
+    torch.save(model.state_dict(), pytorch_weights_dump_path)
+    print(f"Save configuration file to {pytorch_config_dump_path}")
+    with open(pytorch_config_dump_path, "w", encoding="utf-8") as f:
+        f.write(config.to_json_string())
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    # Required parameters
+    parser.add_argument(
+        "--gpt2_checkpoint_path", default=None, type=str, required=True, help="Path to the TensorFlow checkpoint path."
+    )
+    parser.add_argument(
+        "--pytorch_dump_folder_path", default=None, type=str, required=True, help="Path to the output PyTorch model."
+    )
+    parser.add_argument(
+        "--gpt2_config_file",
+        default="",
+        type=str,
+        help="An optional config json file corresponding to the pre-trained OpenAI model. \n"
+        "This specifies the model architecture.",
+    )
+    args = parser.parse_args()
+    convert_gpt2_checkpoint_to_pytorch(args.gpt2_checkpoint_path, args.gpt2_config_file, args.pytorch_dump_folder_path)

model/modeling_flax_gpt2.py

@@ -0,0 +1,641 @@
+# coding=utf-8
+# Copyright 2021 The Google Flax Team Authors and The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Any, Optional, Tuple
+
+import flax.linen as nn
+import jax
+import jax.numpy as jnp
+from flax.core.frozen_dict import FrozenDict, unfreeze
+from flax.linen import combine_masks, make_causal_mask
+from flax.linen.attention import dot_product_attention_weights
+from jax import lax
+
+from ...file_utils import add_start_docstrings, add_start_docstrings_to_model_forward
+from ...modeling_flax_outputs import FlaxBaseModelOutput, FlaxBaseModelOutputWithPast, FlaxCausalLMOutput
+from ...modeling_flax_utils import ACT2FN, FlaxPreTrainedModel, append_call_sample_docstring
+from ...utils import logging
+from .configuration_gpt2 import GPT2Config
+
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "gpt2"
+_CONFIG_FOR_DOC = "GPT2Config"
+_TOKENIZER_FOR_DOC = "GPT2Tokenizer"
+
+
+GPT2_START_DOCSTRING = r"""
+
+    This model inherits from :class:`~transformers.FlaxPreTrainedModel`. 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 Flax Linen `flax.nn.Module
+    <https://flax.readthedocs.io/en/latest/_autosummary/flax.nn.module.html>`__ subclass. Use it as a regular Flax
+    Module and refer to the Flax documentation for all matter related to general usage and behavior.
+
+    Finally, this model supports inherent JAX features such as:
+
+    - `Just-In-Time (JIT) compilation <https://jax.readthedocs.io/en/latest/jax.html#just-in-time-compilation-jit>`__
+    - `Automatic Differentiation <https://jax.readthedocs.io/en/latest/jax.html#automatic-differentiation>`__
+    - `Vectorization <https://jax.readthedocs.io/en/latest/jax.html#vectorization-vmap>`__
+    - `Parallelization <https://jax.readthedocs.io/en/latest/jax.html#parallelization-pmap>`__
+
+    Parameters:
+        config (:class:`~transformers.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 :meth:`~transformers.FlaxPreTrainedModel.from_pretrained` method to load the
+            model weights.
+"""
+
+GPT2_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (:obj:`numpy.ndarray` of shape :obj:`(batch_size, input_ids_length)`):
+            :obj:`input_ids_length` = ``sequence_length``. Indices of input sequence tokens in the vocabulary.
+
+            Indices can be obtained using :class:`~transformers.GPT2Tokenizer`. See
+            :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__` for
+            details.
+
+            `What are input IDs? <../glossary.html#input-ids>`__
+        attention_mask (:obj:`numpy.ndarray` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Mask to avoid performing attention on padding token indices. Mask values selected in ``[0, 1]``:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            `What are attention masks? <../glossary.html#attention-mask>`__
+        position_ids (:obj:`numpy.ndarray` of shape :obj:`(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]``.
+        past_key_values (:obj:`Dict[str, np.ndarray]`, `optional`, returned by ``init_cache`` or when passing previous ``past_key_values``):
+            Dictionary of pre-computed hidden-states (key and values in the attention blocks) that can be used for fast
+            auto-regressive decoding. Pre-computed key and value hidden-states are of shape `[batch_size, max_length]`.
+        output_attentions (:obj:`bool`, `optional`):
+            Whether or not to return the attentions tensors of all attention layers. See ``attentions`` under returned
+            tensors for more detail.
+        output_hidden_states (:obj:`bool`, `optional`):
+            Whether or not to return the hidden states of all layers. See ``hidden_states`` under returned tensors for
+            more detail.
+        return_dict (:obj:`bool`, `optional`):
+            Whether or not to return a :class:`~transformers.file_utils.ModelOutput` instead of a plain tuple.
+"""
+
+
+class FlaxConv1D(nn.Module):
+    features: int
+    use_bias: bool = True
+    dtype: Any = jnp.float32
+    precision: Any = None
+
+    @nn.compact
+    def __call__(self, inputs):
+        inputs = jnp.asarray(inputs, self.dtype)
+        kernel = self.param("kernel", jax.nn.initializers.normal(stddev=0.02), (self.features, inputs.shape[-1]))
+        kernel = jnp.asarray(kernel.transpose(), self.dtype)
+        y = lax.dot_general(inputs, kernel, (((inputs.ndim - 1,), (0,)), ((), ())), precision=self.precision)
+        if self.use_bias:
+            bias = self.param("bias", jax.nn.initializers.zeros, (self.features,))
+            bias = jnp.asarray(bias, self.dtype)
+            y = y + bias
+        return y
+
+
+class FlaxGPT2Attention(nn.Module):
+    config: GPT2Config
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        config = self.config
+        self.embed_dim = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.embed_dim // self.num_heads
+
+        self.c_attn = FlaxConv1D(features=3 * self.embed_dim, dtype=self.dtype)
+        self.c_proj = FlaxConv1D(self.embed_dim, dtype=self.dtype)
+        self.resid_dropout = nn.Dropout(rate=config.resid_pdrop)
+        self.causal_mask = make_causal_mask(jnp.ones((1, config.max_position_embeddings), dtype="bool"), dtype="bool")
+
+    def _split_heads(self, hidden_states):
+        return hidden_states.reshape(hidden_states.shape[:2] + (self.num_heads, self.head_dim))
+
+    def _merge_heads(self, hidden_states):
+        return hidden_states.reshape(hidden_states.shape[:2] + (self.embed_dim,))
+
+    @nn.compact
+    def _concatenate_to_cache(self, key, value, query, attention_mask):
+        """
+        This function takes projected key, value states from a single input token and concatenates the states to cached
+        states from previous steps. This function is slighly adapted from the official Flax repository:
+        https://github.com/google/flax/blob/491ce18759622506588784b4fca0e4bf05f8c8cd/flax/linen/attention.py#L252
+        """
+        # detect if we're initializing by absence of existing cache data.
+        is_initialized = self.has_variable("cache", "cached_key")
+        cached_key = self.variable("cache", "cached_key", jnp.zeros, key.shape, key.dtype)
+        cached_value = self.variable("cache", "cached_value", jnp.zeros, value.shape, value.dtype)
+        cache_index = self.variable("cache", "cache_index", lambda: jnp.array(0, dtype=jnp.int32))
+
+        if is_initialized:
+            *batch_dims, max_length, num_heads, depth_per_head = cached_key.value.shape
+            # update key, value caches with our new 1d spatial slices
+            cur_index = cache_index.value
+            indices = (0,) * len(batch_dims) + (cur_index, 0, 0)
+            key = lax.dynamic_update_slice(cached_key.value, key, indices)
+            value = lax.dynamic_update_slice(cached_value.value, value, indices)
+            cached_key.value = key
+            cached_value.value = value
+            num_updated_cache_vectors = query.shape[1]
+            cache_index.value = cache_index.value + num_updated_cache_vectors
+            # causal mask for cached decoder self-attention: our single query position should only attend to those key positions that have already been generated and cached, not the remaining zero elements.
+            pad_mask = jnp.broadcast_to(
+                jnp.arange(max_length) < cur_index + num_updated_cache_vectors,
+                tuple(batch_dims) + (1, num_updated_cache_vectors, max_length),
+            )
+            attention_mask = combine_masks(pad_mask, attention_mask)
+        return key, value, attention_mask
+
+    def __call__(
+        self,
+        hidden_states,
+        attention_mask=None,
+        deterministic: bool = True,
+        init_cache: bool = False,
+        output_attentions: bool = False,
+    ):
+        qkv_out = self.c_attn(hidden_states)
+        query, key, value = jnp.split(qkv_out, 3, axis=2)
+
+        query = self._split_heads(query)
+        key = self._split_heads(key)
+        value = self._split_heads(value)
+
+        query_length, key_length = query.shape[1], key.shape[1]
+
+        if self.has_variable("cache", "cached_key"):
+            mask_shift = self.variables["cache"]["cache_index"]
+            max_decoder_length = self.variables["cache"]["cached_key"].shape[1]
+            causal_mask = lax.dynamic_slice(
+                self.causal_mask, (0, 0, mask_shift, 0), (1, 1, query_length, max_decoder_length)
+            )
+        else:
+            causal_mask = self.causal_mask[:, :, :query_length, :key_length]
+
+        batch_size = hidden_states.shape[0]
+        causal_mask = jnp.broadcast_to(causal_mask, (batch_size,) + causal_mask.shape[1:])
+
+        attention_mask = jnp.broadcast_to(jnp.expand_dims(attention_mask, axis=(-3, -2)), causal_mask.shape)
+        attention_mask = combine_masks(attention_mask, causal_mask)
+
+        dropout_rng = None
+        if not deterministic and self.config.attn_pdrop > 0.0:
+            dropout_rng = self.make_rng("dropout")
+
+        # During fast autoregressive decoding, we feed one position at a time,
+        # and cache the keys and values step by step.
+        if self.has_variable("cache", "cached_key") or init_cache:
+            key, value, attention_mask = self._concatenate_to_cache(key, value, query, attention_mask)
+
+        # transform boolean mask into float mask
+        attention_bias = lax.select(
+            attention_mask > 0,
+            jnp.full(attention_mask.shape, 0.0).astype(self.dtype),
+            jnp.full(attention_mask.shape, -1e4).astype(self.dtype),
+        )
+
+        # usual dot product attention
+        attn_weights = dot_product_attention_weights(
+            query,
+            key,
+            bias=attention_bias,
+            dropout_rng=dropout_rng,
+            dropout_rate=self.config.attn_pdrop,
+            deterministic=deterministic,
+            dtype=self.dtype,
+            precision=None,
+        )
+
+        attn_output = jnp.einsum("...hqk,...khd->...qhd", attn_weights, value)
+        attn_output = self._merge_heads(attn_output)
+        attn_output = self.c_proj(attn_output)
+        attn_output = self.resid_dropout(attn_output, deterministic=deterministic)
+
+        outputs = (attn_output, attn_weights) if output_attentions else (attn_output,)
+        return outputs
+
+
+class FlaxGPT2MLP(nn.Module):
+    config: GPT2Config
+    intermediate_size: int
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        embed_dim = self.config.hidden_size
+        self.c_fc = FlaxConv1D(self.intermediate_size, dtype=self.dtype)
+        self.c_proj = FlaxConv1D(embed_dim, dtype=self.dtype)
+        self.act = ACT2FN[self.config.activation_function]
+        self.dropout = nn.Dropout(rate=self.config.resid_pdrop)
+
+    def __call__(self, hidden_states, deterministic: bool = True):
+        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, deterministic=deterministic)
+        return hidden_states
+
+
+class FlaxGPT2Block(nn.Module):
+    config: GPT2Config
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        hidden_size = self.config.hidden_size
+        inner_dim = self.config.n_inner if self.config.n_inner is not None else 4 * hidden_size
+
+        self.ln_1 = nn.LayerNorm(epsilon=self.config.layer_norm_epsilon, dtype=self.dtype)
+        self.attn = FlaxGPT2Attention(self.config, dtype=self.dtype)
+        self.ln_2 = nn.LayerNorm(epsilon=self.config.layer_norm_epsilon, dtype=self.dtype)
+        self.mlp = FlaxGPT2MLP(self.config, inner_dim, dtype=self.dtype)
+
+    def __call__(
+        self,
+        hidden_states,
+        attention_mask=None,
+        deterministic: bool = True,
+        init_cache: bool = False,
+        output_attentions: bool = False,
+    ):
+        residual = hidden_states
+        hidden_states = self.ln_1(hidden_states)
+        outputs = self.attn(
+            hidden_states,
+            attention_mask=attention_mask,
+            deterministic=deterministic,
+            init_cache=init_cache,
+            output_attentions=output_attentions,
+        )
+        # residual connection
+        attn_output = outputs[0]
+        hidden_states = attn_output + residual
+
+        residual = hidden_states
+        hidden_states = self.ln_2(hidden_states)
+        feed_forward_hidden_states = self.mlp(hidden_states, deterministic=deterministic)
+        # residual connection
+        hidden_states = residual + feed_forward_hidden_states
+
+        return (hidden_states,) + outputs[1:]
+
+
+class FlaxGPT2PreTrainedModel(FlaxPreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = GPT2Config
+    base_model_prefix = "transformer"
+    module_class: nn.Module = None
+
+    def __init__(
+        self,
+        config: GPT2Config,
+        input_shape: Tuple = (1, 1),
+        seed: int = 0,
+        dtype: jnp.dtype = jnp.float32,
+        **kwargs,
+    ):
+        module = self.module_class(config=config, dtype=dtype, **kwargs)
+        super().__init__(config, module, input_shape=input_shape, seed=seed, dtype=dtype)
+
+    def init_weights(self, rng: jax.random.PRNGKey, input_shape: Tuple) -> FrozenDict:
+        # init input tensors
+        input_ids = jnp.zeros(input_shape, dtype="i4")
+        attention_mask = jnp.ones_like(input_ids)
+        position_ids = jnp.broadcast_to(jnp.arange(jnp.atleast_2d(input_ids).shape[-1]), input_shape)
+        params_rng, dropout_rng = jax.random.split(rng)
+        rngs = {"params": params_rng, "dropout": dropout_rng}
+
+        return self.module.init(rngs, input_ids, attention_mask, position_ids, return_dict=False)["params"]
+
+    def init_cache(self, batch_size, max_length):
+        r"""
+        Args:
+            batch_size (:obj:`int`):
+                batch_size used for fast auto-regressive decoding. Defines the batch size of the initialized cache.
+            max_length (:obj:`int`):
+                maximum possible length for auto-regressive decoding. Defines the sequence length of the initialized
+                cache.
+        """
+        # init input variables to retrieve cache
+        input_ids = jnp.ones((batch_size, max_length))
+        attention_mask = jnp.ones_like(input_ids)
+        position_ids = jnp.broadcast_to(jnp.arange(jnp.atleast_2d(input_ids).shape[-1]), input_ids.shape)
+
+        init_variables = self.module.init(
+            jax.random.PRNGKey(0), input_ids, attention_mask, position_ids, return_dict=False, init_cache=True
+        )
+        return init_variables["cache"]
+
+    @add_start_docstrings_to_model_forward(GPT2_INPUTS_DOCSTRING)
+    def __call__(
+        self,
+        input_ids,
+        attention_mask=None,
+        position_ids=None,
+        params: dict = None,
+        past_key_values: dict = None,
+        dropout_rng: jax.random.PRNGKey = None,
+        train: bool = False,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ):
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.return_dict
+
+        batch_size, sequence_length = input_ids.shape
+
+        if position_ids is None:
+            if past_key_values is not None:
+                raise ValueError("Make sure to provide `position_ids` when passing `past_key_values`.")
+
+            position_ids = jnp.broadcast_to(jnp.arange(sequence_length)[None, :], (batch_size, sequence_length))
+
+        if attention_mask is None:
+            attention_mask = jnp.ones((batch_size, sequence_length))
+
+        # Handle any PRNG if needed
+        rngs = {}
+        if dropout_rng is not None:
+            rngs["dropout"] = dropout_rng
+
+        inputs = {"params": params or self.params}
+
+        # if past_key_values are passed then cache is already initialized a private flag init_cache has to be passed down to ensure cache is used. It has to be made sure that cache is marked as mutable so that it can be changed by FlaxGPT2Attention module
+        if past_key_values:
+            inputs["cache"] = past_key_values
+            mutable = ["cache"]
+        else:
+            mutable = False
+
+        outputs = self.module.apply(
+            inputs,
+            jnp.array(input_ids, dtype="i4"),
+            jnp.array(attention_mask, dtype="i4"),
+            jnp.array(position_ids, dtype="i4"),
+            not train,
+            False,
+            output_attentions,
+            output_hidden_states,
+            return_dict,
+            rngs=rngs,
+            mutable=mutable,
+        )
+
+        # add updated cache to model output
+        if past_key_values is not None and return_dict:
+            outputs, past_key_values = outputs
+            outputs["past_key_values"] = unfreeze(past_key_values["cache"])
+            return outputs
+        elif past_key_values is not None and not return_dict:
+            outputs, past_key_values = outputs
+            outputs = outputs[:1] + (unfreeze(past_key_values["cache"]),) + outputs[1:]
+
+        return outputs
+
+
+class FlaxGPT2BlockCollection(nn.Module):
+    config: GPT2Config
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        self.blocks = [
+            FlaxGPT2Block(self.config, name=str(i), dtype=self.dtype) for i in range(self.config.num_hidden_layers)
+        ]
+
+    def __call__(
+        self,
+        hidden_states,
+        attention_mask=None,
+        deterministic: bool = True,
+        init_cache: bool = False,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ):
+        all_attentions = () if output_attentions else None
+        all_hidden_states = () if output_hidden_states else None
+
+        for block in self.blocks:
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            layer_outputs = block(
+                hidden_states,
+                attention_mask,
+                deterministic=deterministic,
+                init_cache=init_cache,
+                output_attentions=output_attentions,
+            )
+            hidden_states = layer_outputs[0]
+
+            if output_attentions:
+                all_attentions += (layer_outputs[1],)
+
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        outputs = (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in outputs if v is not None)
+
+        return FlaxBaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=None,
+            hidden_states=all_hidden_states,
+            attentions=all_attentions,
+        )
+
+
+class FlaxGPT2Module(nn.Module):
+    config: GPT2Config
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        self.embed_dim = self.config.hidden_size
+
+        self.wte = nn.Embed(
+            self.config.vocab_size,
+            self.embed_dim,
+            embedding_init=jax.nn.initializers.normal(stddev=self.config.initializer_range),
+            dtype=self.dtype,
+        )
+        self.wpe = nn.Embed(
+            self.config.max_position_embeddings,
+            self.embed_dim,
+            embedding_init=jax.nn.initializers.normal(stddev=self.config.initializer_range),
+            dtype=self.dtype,
+        )
+        self.dropout = nn.Dropout(rate=self.config.embd_pdrop)
+        self.h = FlaxGPT2BlockCollection(self.config, dtype=self.dtype)
+        self.ln_f = nn.LayerNorm(epsilon=self.config.layer_norm_epsilon, dtype=self.dtype)
+
+    def __call__(
+        self,
+        input_ids,
+        attention_mask,
+        position_ids,
+        deterministic=True,
+        init_cache: bool = False,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ):
+        input_embeds = self.wte(input_ids.astype("i4"))
+        position_embeds = self.wpe(position_ids.astype("i4"))
+
+        hidden_states = input_embeds + position_embeds
+        hidden_states = self.dropout(hidden_states, deterministic=deterministic)
+
+        outputs = self.h(
+            hidden_states,
+            attention_mask,
+            deterministic=deterministic,
+            init_cache=init_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = outputs[0]
+        hidden_states = self.ln_f(hidden_states)
+
+        if not return_dict:
+            return (hidden_states,) + outputs[1:]
+
+        return FlaxBaseModelOutput(
+            last_hidden_state=hidden_states,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    "The bare GPT2 Model transformer outputting raw hidden-states without any specific head on top.",
+    GPT2_START_DOCSTRING,
+)
+class FlaxGPT2Model(FlaxGPT2PreTrainedModel):
+    module_class = FlaxGPT2Module
+
+
+append_call_sample_docstring(
+    FlaxGPT2Model, _TOKENIZER_FOR_DOC, _CHECKPOINT_FOR_DOC, FlaxBaseModelOutput, _CONFIG_FOR_DOC
+)
+
+
+class FlaxGPT2LMHeadModule(nn.Module):
+    config: GPT2Config
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        self.transformer = FlaxGPT2Module(self.config, dtype=self.dtype)
+        self.lm_head = nn.Dense(
+            self.config.vocab_size,
+            use_bias=False,
+            dtype=self.dtype,
+            kernel_init=jax.nn.initializers.normal(stddev=self.config.initializer_range, dtype=self.dtype),
+        )
+
+    def __call__(
+        self,
+        input_ids,
+        attention_mask,
+        position_ids,
+        deterministic: bool = True,
+        init_cache: bool = False,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ):
+        outputs = self.transformer(
+            input_ids,
+            attention_mask,
+            position_ids,
+            deterministic=deterministic,
+            init_cache=init_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = outputs[0]
+
+        if self.config.tie_word_embeddings:
+            shared_kernel = self.transformer.variables["params"]["wte"]["embedding"].T
+            lm_logits = self.lm_head.apply({"params": {"kernel": shared_kernel}}, hidden_states)
+        else:
+            lm_logits = self.lm_head(hidden_states)
+
+        if not return_dict:
+            return (lm_logits,) + outputs[1:]
+
+        return FlaxCausalLMOutput(logits=lm_logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions)
+
+
+@add_start_docstrings(
+    """
+    The GPT2 Model transformer with a language modeling head on top (linear layer with weights tied to the input
+    embeddings).
+    """,
+    GPT2_START_DOCSTRING,
+)
+class FlaxGPT2LMHeadModel(FlaxGPT2PreTrainedModel):
+    module_class = FlaxGPT2LMHeadModule
+
+    def prepare_inputs_for_generation(self, input_ids, max_length, attention_mask: Optional[jnp.DeviceArray] = None):
+        # initializing the cache
+        batch_size, seq_length = input_ids.shape
+
+        past_key_values = self.init_cache(batch_size, max_length)
+        # Note that usually one would have to put 0's in the attention_mask for x > input_ids.shape[-1] and x < cache_length.
+        # But since GPT2 uses a causal mask, those positions are masked anyways.
+        # Thus we can create a single static attention_mask here, which is more efficient for compilation
+        extended_attention_mask = jnp.ones((batch_size, max_length), dtype="i4")
+        if attention_mask is not None:
+            position_ids = attention_mask.cumsum(axis=-1) - 1
+            extended_attention_mask = lax.dynamic_update_slice(extended_attention_mask, attention_mask, (0, 0))
+        else:
+            position_ids = jnp.broadcast_to(jnp.arange(seq_length, dtype="i4")[None, :], (batch_size, seq_length))
+
+        return {
+            "past_key_values": past_key_values,
+            "attention_mask": extended_attention_mask,
+            "position_ids": position_ids,
+        }
+
+    def update_inputs_for_generation(self, model_outputs, model_kwargs):
+        model_kwargs["past_key_values"] = model_outputs.past_key_values
+        model_kwargs["position_ids"] = model_kwargs["position_ids"][:, -1:] + 1
+        return model_kwargs
+
+
+append_call_sample_docstring(
+    FlaxGPT2LMHeadModel, _TOKENIZER_FOR_DOC, _CHECKPOINT_FOR_DOC, FlaxCausalLMOutput, _CONFIG_FOR_DOC
+)

model/modeling_gpt2.py

@@ -0,0 +1,1328 @@
+# coding=utf-8
+# Copyright 2018 The OpenAI Team Authors and HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""PyTorch OpenAI GPT-2 model."""
+
+import os
+from dataclasses import dataclass
+from typing import Optional, Tuple
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import CrossEntropyLoss, MSELoss
+
+from ...activations import ACT2FN
+from ...file_utils import (
+    ModelOutput,
+    add_code_sample_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    replace_return_docstrings,
+)
+from ...modeling_outputs import (
+    BaseModelOutputWithPastAndCrossAttentions,
+    CausalLMOutputWithCrossAttentions,
+    SequenceClassifierOutputWithPast,
+)
+from ...modeling_utils import (
+    Conv1D,
+    PreTrainedModel,
+    SequenceSummary,
+    find_pruneable_heads_and_indices,
+    prune_conv1d_layer,
+)
+from ...utils import logging
+from ...utils.model_parallel_utils import assert_device_map, get_device_map
+from .configuration_gpt2 import GPT2Config
+
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "gpt2"
+_CONFIG_FOR_DOC = "GPT2Config"
+_TOKENIZER_FOR_DOC = "GPT2Tokenizer"
+
+GPT2_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "gpt2",
+    "gpt2-medium",
+    "gpt2-large",
+    "gpt2-xl",
+    "distilgpt2",
+    # See all GPT-2 models at https://huggingface.co/models?filter=gpt2
+]
+
+
+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:
+            assert (
+                pointer.shape == array.shape
+            ), f"Pointer shape {pointer.shape} and array shape {array.shape} mismatched"
+        except AssertionError 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):
+        super().__init__()
+
+        max_positions = config.max_position_embeddings
+        self.register_buffer(
+            "bias",
+            torch.tril(torch.ones((max_positions, max_positions), dtype=torch.uint8)).view(
+                1, 1, max_positions, max_positions
+            ),
+        )
+        self.register_buffer("masked_bias", torch.tensor(-1e4))
+
+        self.embed_dim = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.embed_dim // self.num_heads
+        self.split_size = self.embed_dim
+        if self.head_dim * self.num_heads != self.embed_dim:
+            raise ValueError(
+                f"`embed_dim` must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`: {self.num_heads})."
+            )
+
+        self.scale_attn_weights = config.scale_attn_weights
+        self.is_cross_attention = is_cross_attention
+
+        if self.is_cross_attention:
+            self.c_attn = Conv1D(2 * self.embed_dim, self.embed_dim)
+            self.q_attn = Conv1D(self.embed_dim, self.embed_dim)
+        else:
+            self.c_attn = Conv1D(3 * self.embed_dim, self.embed_dim)
+        self.c_proj = Conv1D(self.embed_dim, self.embed_dim)
+
+        self.attn_dropout = nn.Dropout(config.attn_pdrop)
+        self.resid_dropout = nn.Dropout(config.resid_pdrop)
+
+        self.pruned_heads = set()
+
+    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)])
+
+        # 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.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 / (float(value.size(-1)) ** 0.5)
+
+        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].bool()
+            attn_weights = torch.where(causal_mask, attn_weights, self.masked_bias.to(attn_weights.dtype))
+
+        if attention_mask is not None:
+            # Apply the attention mask
+            attn_weights = attn_weights + attention_mask
+
+        attn_weights = nn.Softmax(dim=-1)(attn_weights)
+        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 _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)
+
+    def forward(
+        self,
+        hidden_states,
+        layer_past=None,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        use_cache=False,
+        output_attentions=False,
+    ):
+        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(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
+
+        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 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):
+        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):
+        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)
+        self.ln_2 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
+
+        if config.add_cross_attention:
+            self.crossattention = GPT2Attention(config, is_cross_attention=True)
+            self.ln_cross_attn = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
+
+        self.mlp = GPT2MLP(inner_dim, config)
+
+    def forward(
+        self,
+        hidden_states,
+        layer_past=None,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        use_cache=False,
+        output_attentions=False,
+    ):
+        residual = hidden_states
+        hidden_states = self.ln_1(hidden_states)
+        attn_outputs = self.attn(
+            hidden_states,
+            layer_past=layer_past,
+            attention_mask=attention_mask,
+            head_mask=head_mask,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+        )
+        attn_output = attn_outputs[0]  # output_attn: a, present, (attentions)
+        outputs = attn_outputs[1:]
+        # 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_outputs = self.crossattention(
+                hidden_states,
+                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
+
+        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
+
+        if use_cache:
+            outputs = (hidden_states,) + outputs
+        else:
+            outputs = (hidden_states,) + outputs[1:]
+
+        return outputs  # hidden_states, present, (attentions, cross_attentions)
+
+
+class GPT2PreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = GPT2Config
+    load_tf_weights = load_tf_weights_in_gpt2
+    base_model_prefix = "transformer"
+    is_parallelizable = 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)
+
+
+@dataclass
+class GPT2DoubleHeadsModelOutput(ModelOutput):
+    """
+    Base class for outputs of models predicting if two sentences are consecutive or not.
+
+    Args:
+        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when ``labels`` is provided):
+            Language modeling loss.
+        mc_loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`mc_labels` is provided):
+            Multiple choice classification loss.
+        logits (:obj:`torch.FloatTensor` of shape :obj:`(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 (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, num_choices)`):
+            Prediction scores of the multiple choice classification head (scores for each choice before SoftMax).
+        past_key_values (:obj:`Tuple[Tuple[torch.Tensor]]`, `optional`, returned when ``use_cache=True`` is passed or when ``config.use_cache=True``):
+            Tuple of length :obj:`config.n_layers`, containing tuples of tensors of shape :obj:`(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
+            :obj:`past_key_values` input) to speed up sequential decoding.
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            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: torch.FloatTensor = None
+    mc_logits: 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 :class:`~transformers.PreTrainedModel`. Check the superclass documentation for the generic
+    methods the library implements for all its model (such as downloading or saving, resizing the input embeddings,
+    pruning heads etc.)
+
+    This model is also a PyTorch `torch.nn.Module <https://pytorch.org/docs/stable/nn.html#torch.nn.Module>`__
+    subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to
+    general usage and behavior.
+
+    Parameters:
+        config (:class:`~transformers.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 :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model
+            weights.
+"""
+
+GPT2_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, input_ids_length)`):
+            :obj:`input_ids_length` = ``sequence_length`` if :obj:`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 :obj:`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 :class:`~transformers.GPT2Tokenizer`. See
+            :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__` for
+            details.
+
+            `What are input IDs? <../glossary.html#input-ids>`__
+        past_key_values (:obj:`Tuple[Tuple[torch.Tensor]]` of length :obj:`config.n_layers`):
+            Contains precomputed hidden-states (key and values in the attention blocks) as computed by the model (see
+            :obj:`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 (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Mask to avoid performing attention on padding token indices. Mask values selected in ``[0, 1]``:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            `What are attention masks? <../glossary.html#attention-mask>`__
+        token_type_ids (:obj:`torch.LongTensor` of shape :obj:`(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.html#token-type-ids>`_
+        position_ids (:obj:`torch.LongTensor` of shape :obj:`(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.html#position-ids>`_
+        head_mask (:obj:`torch.FloatTensor` of shape :obj:`(num_heads,)` or :obj:`(num_layers, num_heads)`, `optional`):
+            Mask to nullify selected heads of the self-attention modules. Mask values selected in ``[0, 1]``:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):
+            Optionally, instead of passing :obj:`input_ids` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert :obj:`input_ids` indices into associated
+            vectors than the model's internal embedding lookup matrix.
+
+            If :obj:`past_key_values` is used, optionally only the last :obj:`inputs_embeds` have to be input (see
+            :obj:`past_key_values`).
+        use_cache (:obj:`bool`, `optional`):
+            If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up
+            decoding (see :obj:`past_key_values`).
+        output_attentions (:obj:`bool`, `optional`):
+            Whether or not to return the attentions tensors of all attention layers. See ``attentions`` under returned
+            tensors for more detail.
+        output_hidden_states (:obj:`bool`, `optional`):
+            Whether or not to return the hidden states of all layers. See ``hidden_states`` under returned tensors for
+            more detail.
+        return_dict (:obj:`bool`, `optional`):
+            Whether or not to return a :class:`~transformers.file_utils.ModelOutput` instead of a plain tuple.
+"""
+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 (:obj:`Dict[int, list]`, optional, defaults to None):
+            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:
+
+                - gpt2: 12
+                - gpt2-medium: 24
+                - gpt2-large: 36
+                - gpt2-xl: 48
+
+    Example::
+
+            # 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('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::
+
+        # On a 4 GPU machine with gpt2-large:
+        model = GPT2LMHeadModel.from_pretrained('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()
+"""
+
+
+@add_start_docstrings(
+    "The bare GPT2 Model transformer outputting raw hidden-states without any specific head on top.",
+    GPT2_START_DOCSTRING,
+)
+class GPT2Model(GPT2PreTrainedModel):
+    _keys_to_ignore_on_load_missing = ["attn.masked_bias"]
+
+    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) for _ in range(config.num_hidden_layers)])
+        self.ln_f = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon)
+
+        self.init_weights()
+
+        # Model parallel
+        self.model_parallel = False
+        self.device_map = None
+
+    @add_start_docstrings(PARALLELIZE_DOCSTRING)
+    def parallelize(self, device_map=None):
+        # Check validity of device_map
+        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):
+        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)
+
+    @add_start_docstrings_to_model_forward(GPT2_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=BaseModelOutputWithPastAndCrossAttentions,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids=None,
+        past_key_values=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+            input_ids = input_ids.view(-1, input_shape[-1])
+            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 position_ids is not None:
+            position_ids = position_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).view(-1, input_shape[-1])
+
+        # GPT2Attention mask.
+        if attention_mask is not None:
+            assert batch_size > 0, "batch_size has to be defined and > 0"
+            attention_mask = attention_mask.view(batch_size, -1)
+            # 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 -10000.0 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) * -10000.0
+
+        # If a 2D ou 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()
+            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)
+            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 inputs_embeds is None:
+            inputs_embeds = self.wte(input_ids)
+        position_embeds = self.wpe(position_ids)
+        hidden_states = inputs_embeds + position_embeds
+
+        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 = input_shape + (hidden_states.size(-1),)
+
+        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, (block, layer_past) in enumerate(zip(self.h, past_key_values)):
+
+            # 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)
+                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 getattr(self.config, "gradient_checkpointing", False) and self.training:
+
+                if use_cache:
+                    logger.warning(
+                        "`use_cache=True` is incompatible with `config.gradient_checkpointing=True`. Setting "
+                        "`use_cache=False`..."
+                    )
+                    use_cache = False
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        # None for past_key_value
+                        return module(*inputs, use_cache, output_attentions)
+
+                    return custom_forward
+
+                outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    hidden_states,
+                    None,
+                    attention_mask,
+                    head_mask[i],
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                )
+            else:
+                outputs = block(
+                    hidden_states,
+                    layer_past=layer_past,
+                    attention_mask=attention_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,
+                )
+
+            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],)
+                if self.config.add_cross_attention:
+                    all_cross_attentions = all_cross_attentions + (outputs[3 if use_cache else 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,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, presents, all_hidden_states, all_self_attentions] if v is not None)
+
+        return BaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            past_key_values=presents,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attentions,
+            cross_attentions=all_cross_attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    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):
+    _keys_to_ignore_on_load_missing = [r"attn.masked_bias", r"attn.bias", r"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)
+
+        self.init_weights()
+
+        # Model parallel
+        self.model_parallel = False
+        self.device_map = None
+
+    @add_start_docstrings(PARALLELIZE_DOCSTRING)
+    def parallelize(self, device_map=None):
+        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):
+        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
+
+    def prepare_inputs_for_generation(self, input_ids, past=None, **kwargs):
+        token_type_ids = kwargs.get("token_type_ids", None)
+        # only last token for inputs_ids if past is defined in kwargs
+        if past:
+            input_ids = input_ids[:, -1].unsqueeze(-1)
+            if token_type_ids is not None:
+                token_type_ids = token_type_ids[:, -1].unsqueeze(-1)
+
+        attention_mask = kwargs.get("attention_mask", None)
+        position_ids = kwargs.get("position_ids", None)
+
+        if attention_mask is not None and position_ids is None:
+            # create position_ids on the fly for batch generation
+            position_ids = attention_mask.long().cumsum(-1) - 1
+            position_ids.masked_fill_(attention_mask == 0, 1)
+            if past:
+                position_ids = position_ids[:, -1].unsqueeze(-1)
+        else:
+            position_ids = None
+        return {
+            "input_ids": input_ids,
+            "past_key_values": past,
+            "use_cache": kwargs.get("use_cache"),
+            "position_ids": position_ids,
+            "attention_mask": attention_mask,
+            "token_type_ids": token_type_ids,
+        }
+
+    @add_start_docstrings_to_model_forward(GPT2_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=CausalLMOutputWithCrossAttentions,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids=None,
+        past_key_values=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        labels=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Labels for 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,
+            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:
+            # 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))
+
+        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,
+        )
+
+    @staticmethod
+    def _reorder_cache(past: Tuple[Tuple[torch.Tensor]], beam_idx: torch.Tensor) -> Tuple[Tuple[torch.Tensor]]:
+        """
+        This function is used to re-order the :obj:`past_key_values` cache if
+        :meth:`~transformers.PreTrainedModel.beam_search` or :meth:`~transformers.PreTrainedModel.beam_sample` is
+        called. This is required to match :obj:`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
+        )
+
+
+@add_start_docstrings(
+    """
+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):
+    _keys_to_ignore_on_load_missing = [r"attn.masked_bias", r"attn.bias", r"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 = SequenceSummary(config)
+
+        self.init_weights()
+
+        # Model parallel
+        self.model_parallel = False
+        self.device_map = None
+
+    @add_start_docstrings(PARALLELIZE_DOCSTRING)
+    def parallelize(self, device_map=None):
+        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):
+        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
+
+    def prepare_inputs_for_generation(self, input_ids, past=None, **kwargs):
+        token_type_ids = kwargs.get("token_type_ids", None)
+        # only last token for inputs_ids if past is defined in kwargs
+        if past:
+            input_ids = input_ids[:, -1].unsqueeze(-1)
+            if token_type_ids is not None:
+                token_type_ids = token_type_ids[:, -1].unsqueeze(-1)
+
+        attention_mask = kwargs.get("attention_mask", None)
+        position_ids = kwargs.get("position_ids", None)
+
+        if attention_mask is not None and position_ids is None:
+            # create position_ids on the fly for batch generation
+            position_ids = attention_mask.long().cumsum(-1) - 1
+            position_ids.masked_fill_(attention_mask == 0, 1)
+            if past:
+                position_ids = position_ids[:, -1].unsqueeze(-1)
+        else:
+            position_ids = None
+
+        return {
+            "input_ids": input_ids,
+            "past_key_values": past,
+            "use_cache": kwargs.get("use_cache"),
+            "position_ids": position_ids,
+            "attention_mask": attention_mask,
+            "token_type_ids": token_type_ids,
+        }
+
+    @add_start_docstrings_to_model_forward(GPT2_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=GPT2DoubleHeadsModelOutput, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_ids=None,
+        past_key_values=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        mc_token_ids=None,
+        labels=None,
+        mc_labels=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+        **kwargs,
+    ):
+        r"""
+        mc_token_ids (:obj:`torch.LongTensor` of shape :obj:`(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 (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_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 (:obj:`torch.LongTensor` of shape :obj:`(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)
+
+        Return:
+
+        Example::
+
+            >>> import torch
+            >>> from transformers import GPT2Tokenizer, GPT2DoubleHeadsModel
+
+            >>> tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
+            >>> model = GPT2DoubleHeadsModel.from_pretrained('gpt2')
+
+            >>> # Add a [CLS] to the vocabulary (we should train it also!)
+            >>> num_added_tokens = tokenizer.add_special_tokens({'cls_token': '[CLS]'})
+
+            >>> embedding_layer = model.resize_token_embeddings(len(tokenizer))  # Update the model embeddings with the new vocabulary size
+
+            >>> 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,
+            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:
+            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: Tuple[Tuple[torch.Tensor]], beam_idx: torch.Tensor) -> Tuple[Tuple[torch.Tensor]]:
+        """
+        This function is used to re-order the :obj:`past_key_values` cache if
+        :meth:`~transformers.PreTrainedModel.beam_search` or :meth:`~transformers.PreTrainedModel.beam_sample` is
+        called. This is required to match :obj:`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
+        )
+
+
+@add_start_docstrings(
+    """
+    The GPT2 Model transformer with a sequence classification head on top (linear layer).
+
+    :class:`~transformers.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
+    :obj:`pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each
+    row. If no :obj:`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 :obj:`inputs_embeds` are passed instead of :obj:`input_ids`, it does the same (take
+    the last value in each row of the batch).
+    """,
+    GPT2_START_DOCSTRING,
+)
+class GPT2ForSequenceClassification(GPT2PreTrainedModel):
+    _keys_to_ignore_on_load_missing = [r"h\.\d+\.attn\.masked_bias", r"lm_head\.weight"]
+
+    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)
+
+        self.init_weights()
+
+        # Model parallel
+        self.model_parallel = False
+        self.device_map = None
+
+    @add_start_docstrings_to_model_forward(GPT2_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="microsoft/DialogRPT-updown",
+        output_type=SequenceClassifierOutputWithPast,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids=None,
+        past_key_values=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        labels=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
+            Labels for computing the sequence classification/regression loss. Indices should be in :obj:`[0, ...,
+            config.num_labels - 1]`. If :obj:`config.num_labels == 1` a regression loss is computed (Mean-Square loss),
+            If :obj:`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        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]
+
+        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:
+            sequence_lengths = -1
+        else:
+            if input_ids is not None:
+                sequence_lengths = torch.ne(input_ids, self.config.pad_token_id).sum(-1) - 1
+            else:
+                sequence_lengths = -1
+                logger.warning(
+                    f"{self.__class__.__name__} will not detect padding tokens in `inputs_embeds`. Results may be "
+                    f"unexpected if using padding tokens in conjunction with `inputs_embeds.`"
+                )
+
+        pooled_logits = logits[range(batch_size), sequence_lengths]
+
+        loss = None
+        if labels is not None:
+            if self.num_labels == 1:
+                #  We are doing regression
+                loss_fct = MSELoss()
+                loss = loss_fct(pooled_logits.view(-1), labels.to(self.dtype).view(-1))
+            else:
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(pooled_logits.view(-1, self.num_labels), labels.view(-1))
+
+        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,
+        )

model/modeling_tf_gpt2.py

@@ -0,0 +1,1081 @@
+# coding=utf-8
+# Copyright 2018 The OpenAI Team Authors and HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" TF 2.0 OpenAI GPT-2 model. """
+
+from dataclasses import dataclass
+from typing import List, Optional, Tuple
+
+import tensorflow as tf
+
+from ...activations_tf import get_tf_activation
+from ...file_utils import (
+    ModelOutput,
+    add_code_sample_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    replace_return_docstrings,
+)
+from ...modeling_tf_outputs import (
+    TFBaseModelOutputWithPast,
+    TFCausalLMOutputWithPast,
+    TFSequenceClassifierOutputWithPast,
+)
+from ...modeling_tf_utils import (
+    TFCausalLanguageModelingLoss,
+    TFConv1D,
+    TFPreTrainedModel,
+    TFSequenceClassificationLoss,
+    TFSequenceSummary,
+    TFSharedEmbeddings,
+    get_initializer,
+    input_processing,
+    keras_serializable,
+    shape_list,
+)
+from ...utils import logging
+from .configuration_gpt2 import GPT2Config
+
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "gpt2"
+_CONFIG_FOR_DOC = "GPT2Config"
+_TOKENIZER_FOR_DOC = "GPT2Tokenizer"
+
+TF_GPT2_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "gpt2",
+    "gpt2-medium",
+    "gpt2-large",
+    "gpt2-xl",
+    "distilgpt2",
+    # See all GPT-2 models at https://huggingface.co/models?filter=gpt2
+]
+
+
+class TFAttention(tf.keras.layers.Layer):
+    def __init__(self, nx, n_ctx, config, scale=False, **kwargs):
+        super().__init__(**kwargs)
+
+        n_state = nx  # in Attention: n_state=768 (nx=n_embd)
+        # [switch nx => n_state from Block to Attention to keep identical to TF implementation]
+        assert n_state % config.n_head == 0
+        self.n_ctx = n_ctx
+        self.n_head = config.n_head
+        self.split_size = n_state
+        self.scale = scale
+        self.output_attentions = config.output_attentions
+
+        self.c_attn = TFConv1D(n_state * 3, nx, initializer_range=config.initializer_range, name="c_attn")
+        self.c_proj = TFConv1D(n_state, nx, initializer_range=config.initializer_range, name="c_proj")
+        self.attn_dropout = tf.keras.layers.Dropout(config.attn_pdrop)
+        self.resid_dropout = tf.keras.layers.Dropout(config.resid_pdrop)
+        self.pruned_heads = set()
+
+    def prune_heads(self, heads):
+        pass
+
+    @staticmethod
+    def causal_attention_mask(nd, ns, dtype):
+        """
+        1's in the lower triangle, counting from the lower right corner. Same as tf.matrix_band_part(tf.ones([nd, ns]),
+        -1, ns-nd), but doesn't produce garbage on TPUs.
+        """
+        i = tf.range(nd)[:, None]
+        j = tf.range(ns)
+        m = i >= j - ns + nd
+        return tf.cast(m, dtype)
+
+    def _attn(self, q, k, v, attention_mask, head_mask, output_attentions, training=False):
+        # q, k, v have shape [batch, heads, sequence, features]
+        w = tf.matmul(q, k, transpose_b=True)
+        if self.scale:
+            dk = tf.cast(shape_list(k)[-1], dtype=w.dtype)  # scale attention_scores
+            w = w / tf.math.sqrt(dk)
+
+        # w has shape [batch, heads, dst_sequence, src_sequence], where information flows from src to dst.
+        _, _, nd, ns = shape_list(w)
+        b = self.causal_attention_mask(nd, ns, dtype=w.dtype)
+        b = tf.reshape(b, [1, 1, nd, ns])
+        w = w * b - 1e4 * (1 - b)
+
+        if attention_mask is not None:
+            # Apply the attention mask
+            attention_mask = tf.cast(attention_mask, dtype=w.dtype)
+            w = w + attention_mask
+
+        w = tf.nn.softmax(w, axis=-1)
+        w = self.attn_dropout(w, training=training)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            w = w * head_mask
+
+        outputs = [tf.matmul(w, v)]
+        if output_attentions:
+            outputs.append(w)
+        return outputs
+
+    def merge_heads(self, x):
+        x = tf.transpose(x, [0, 2, 1, 3])
+        x_shape = shape_list(x)
+        new_x_shape = x_shape[:-2] + [x_shape[-2] * x_shape[-1]]
+        return tf.reshape(x, new_x_shape)
+
+    def split_heads(self, x):
+        x_shape = shape_list(x)
+        new_x_shape = x_shape[:-1] + [self.n_head, x_shape[-1] // self.n_head]
+        x = tf.reshape(x, new_x_shape)
+        return tf.transpose(x, (0, 2, 1, 3))  # (batch, head, seq_length, head_features)
+
+    def call(self, x, layer_past, attention_mask, head_mask, use_cache, output_attentions, training=False):
+        x = self.c_attn(x)
+        query, key, value = tf.split(x, 3, axis=2)
+        query = self.split_heads(query)
+        key = self.split_heads(key)
+        value = self.split_heads(value)
+        if layer_past is not None:
+            past_key, past_value = tf.unstack(layer_past, axis=0)
+            key = tf.concat([past_key, key], axis=-2)
+            value = tf.concat([past_value, value], axis=-2)
+
+        # to cope with keras serialization
+        if use_cache:
+            present = tf.stack([key, value], axis=0)
+        else:
+            present = (None,)
+
+        attn_outputs = self._attn(query, key, value, attention_mask, head_mask, output_attentions, training=training)
+        a = attn_outputs[0]
+
+        a = self.merge_heads(a)
+        a = self.c_proj(a)
+        a = self.resid_dropout(a, training=training)
+
+        outputs = [a, present] + attn_outputs[1:]
+        return outputs  # a, present, (attentions)
+
+
+class TFMLP(tf.keras.layers.Layer):
+    def __init__(self, n_state, config, **kwargs):
+        super().__init__(**kwargs)
+        nx = config.n_embd
+        self.c_fc = TFConv1D(n_state, nx, initializer_range=config.initializer_range, name="c_fc")
+        self.c_proj = TFConv1D(nx, n_state, initializer_range=config.initializer_range, name="c_proj")
+        self.act = get_tf_activation("gelu")
+        self.dropout = tf.keras.layers.Dropout(config.resid_pdrop)
+
+    def call(self, x, training=False):
+        h = self.act(self.c_fc(x))
+        h2 = self.c_proj(h)
+        h2 = self.dropout(h2, training=training)
+        return h2
+
+
+class TFBlock(tf.keras.layers.Layer):
+    def __init__(self, n_ctx, config, scale=False, **kwargs):
+        super().__init__(**kwargs)
+        nx = config.n_embd
+        inner_dim = config.n_inner if config.n_inner is not None else 4 * nx
+        self.ln_1 = tf.keras.layers.LayerNormalization(epsilon=config.layer_norm_epsilon, name="ln_1")
+        self.attn = TFAttention(nx, n_ctx, config, scale, name="attn")
+        self.ln_2 = tf.keras.layers.LayerNormalization(epsilon=config.layer_norm_epsilon, name="ln_2")
+        self.mlp = TFMLP(inner_dim, config, name="mlp")
+
+    def call(self, x, layer_past, attention_mask, head_mask, use_cache, output_attentions, training=False):
+        a = self.ln_1(x)
+        output_attn = self.attn(
+            a, layer_past, attention_mask, head_mask, use_cache, output_attentions, training=training
+        )
+        a = output_attn[0]  # output_attn: a, present, (attentions)
+        x = x + a
+
+        m = self.ln_2(x)
+        m = self.mlp(m, training=training)
+        x = x + m
+
+        outputs = [x] + output_attn[1:]
+        return outputs  # x, present, (attentions)
+
+
+@keras_serializable
+class TFGPT2MainLayer(tf.keras.layers.Layer):
+    config_class = GPT2Config
+
+    def __init__(self, config, *inputs, **kwargs):
+        super().__init__(*inputs, **kwargs)
+
+        self.config = config
+        self.output_attentions = config.output_attentions
+        self.output_hidden_states = config.output_hidden_states
+        self.use_cache = config.use_cache
+        self.return_dict = config.use_return_dict
+
+        self.num_hidden_layers = config.n_layer
+        self.vocab_size = config.vocab_size
+        self.n_embd = config.n_embd
+        self.n_positions = config.n_positions
+        self.initializer_range = config.initializer_range
+
+        self.wte = TFSharedEmbeddings(
+            config.vocab_size, config.hidden_size, initializer_range=config.initializer_range, name="wte"
+        )
+        self.drop = tf.keras.layers.Dropout(config.embd_pdrop)
+        self.h = [TFBlock(config.n_ctx, config, scale=True, name=f"h_._{i}") for i in range(config.n_layer)]
+        self.ln_f = tf.keras.layers.LayerNormalization(epsilon=config.layer_norm_epsilon, name="ln_f")
+
+    def build(self, input_shape):
+        with tf.name_scope("wpe"):
+            self.wpe = self.add_weight(
+                name="embeddings",
+                shape=[self.n_positions, self.n_embd],
+                initializer=get_initializer(self.initializer_range),
+            )
+
+        super().build(input_shape)
+
+    def get_input_embeddings(self):
+        return self.wte
+
+    def set_input_embeddings(self, value):
+        self.wte.weight = value
+        self.wte.vocab_size = shape_list(value)[0]
+
+    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}
+        """
+        raise NotImplementedError
+
+    def call(
+        self,
+        input_ids=None,
+        past=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+        training=False,
+        **kwargs,
+    ):
+        inputs = input_processing(
+            func=self.call,
+            config=self.config,
+            input_ids=input_ids,
+            past=past,
+            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,
+            training=training,
+            kwargs_call=kwargs,
+        )
+
+        if inputs["input_ids"] is not None and inputs["inputs_embeds"] is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif inputs["input_ids"] is not None:
+            input_shape = shape_list(inputs["input_ids"])
+            inputs["input_ids"] = tf.reshape(inputs["input_ids"], [-1, input_shape[-1]])
+        elif inputs["inputs_embeds"] is not None:
+            input_shape = shape_list(inputs["inputs_embeds"])[:-1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        if inputs["past"] is None:
+            past_length = 0
+            inputs["past"] = [None] * len(self.h)
+        else:
+            past_length = shape_list(inputs["past"][0][0])[-2]
+
+        if inputs["position_ids"] is None:
+            inputs["position_ids"] = tf.expand_dims(tf.range(past_length, input_shape[-1] + past_length), axis=0)
+
+        if inputs["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_shape = shape_list(inputs["attention_mask"])
+            inputs["attention_mask"] = tf.reshape(
+                inputs["attention_mask"], (attention_mask_shape[0], 1, 1, attention_mask_shape[1])
+            )
+
+            # 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 -10000.0 for masked positions.
+            # Since we are adding it to the raw scores before the softmax, this is
+            # effectively the same as removing these entirely.
+            one_cst = tf.constant(1.0)
+            inputs["attention_mask"] = tf.cast(inputs["attention_mask"], dtype=one_cst.dtype)
+            inputs["attention_mask"] = tf.multiply(
+                tf.subtract(one_cst, inputs["attention_mask"]), tf.constant(-10000.0)
+            )
+
+        # 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
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+        if inputs["head_mask"] is not None:
+            raise NotImplementedError
+        else:
+            inputs["head_mask"] = [None] * self.num_hidden_layers
+            # head_mask = tf.constant([0] * self.num_hidden_layers)
+
+        inputs["position_ids"] = tf.reshape(inputs["position_ids"], [-1, shape_list(inputs["position_ids"])[-1]])
+
+        if inputs["inputs_embeds"] is None:
+            inputs["inputs_embeds"] = self.wte(inputs["input_ids"], mode="embedding")
+
+        position_embeds = tf.gather(self.wpe, inputs["position_ids"])
+
+        if inputs["token_type_ids"] is not None:
+            inputs["token_type_ids"] = tf.reshape(
+                inputs["token_type_ids"], [-1, shape_list(inputs["token_type_ids"])[-1]]
+            )
+            token_type_embeds = self.wte(inputs["token_type_ids"], mode="embedding")
+        else:
+            token_type_embeds = tf.constant(0.0)
+
+        position_embeds = tf.cast(position_embeds, dtype=inputs["inputs_embeds"].dtype)
+        token_type_embeds = tf.cast(token_type_embeds, dtype=inputs["inputs_embeds"].dtype)
+        hidden_states = inputs["inputs_embeds"] + position_embeds + token_type_embeds
+        hidden_states = self.drop(hidden_states, training=inputs["training"])
+
+        output_shape = input_shape + [shape_list(hidden_states)[-1]]
+
+        presents = () if inputs["use_cache"] else None
+        all_attentions = () if inputs["output_attentions"] else None
+        all_hidden_states = () if inputs["output_hidden_states"] else None
+        for i, (block, layer_past) in enumerate(zip(self.h, inputs["past"])):
+            if inputs["output_hidden_states"]:
+                all_hidden_states = all_hidden_states + (tf.reshape(hidden_states, output_shape),)
+
+            outputs = block(
+                hidden_states,
+                layer_past,
+                inputs["attention_mask"],
+                inputs["head_mask"][i],
+                inputs["use_cache"],
+                inputs["output_attentions"],
+                training=inputs["training"],
+            )
+
+            hidden_states, present = outputs[:2]
+            if inputs["use_cache"]:
+                presents = presents + (present,)
+
+            if inputs["output_attentions"]:
+                all_attentions = all_attentions + (outputs[2],)
+
+        hidden_states = self.ln_f(hidden_states)
+
+        hidden_states = tf.reshape(hidden_states, output_shape)
+        # Add last hidden state
+        if inputs["output_hidden_states"]:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if inputs["output_attentions"]:
+            # let the number of heads free (-1) so we can extract attention even after head pruning
+            attention_output_shape = input_shape[:-1] + [-1] + shape_list(all_attentions[0])[-2:]
+            all_attentions = tuple(tf.reshape(t, attention_output_shape) for t in all_attentions)
+
+        if not inputs["return_dict"]:
+            return tuple(v for v in [hidden_states, presents, all_hidden_states, all_attentions] if v is not None)
+
+        return TFBaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=presents,
+            hidden_states=all_hidden_states,
+            attentions=all_attentions,
+        )
+
+
+class TFGPT2PreTrainedModel(TFPreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = GPT2Config
+    base_model_prefix = "transformer"
+    # names with a '.' represents the authorized unexpected/missing layers when a TF model is loaded from a PT model
+    _keys_to_ignore_on_load_unexpected = [r"h.\d+.attn.bias"]
+
+    @tf.function(
+        input_signature=[
+            {
+                "input_ids": tf.TensorSpec((None, None), tf.int32, name="input_ids"),
+                "attention_mask": tf.TensorSpec((None, None), tf.int32, name="attention_mask"),
+            }
+        ]
+    )
+    def serving(self, inputs):
+        output = self.call(inputs)
+
+        return self.serving_output(output)
+
+
+@dataclass
+class TFGPT2DoubleHeadsModelOutput(ModelOutput):
+    """
+    Base class for outputs of models predicting if two sentences are consecutive or not.
+
+    Args:
+        logits (:obj:`tf.Tensor` of shape :obj:`(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 (:obj:`tf.Tensor` of shape :obj:`(batch_size, num_choices)`):
+            Prediction scores of the multiple choice classification head (scores for each choice before SoftMax).
+        past_key_values (:obj:`List[tf.Tensor]`, `optional`, returned when ``use_cache=True`` is passed or when ``config.use_cache=True``):
+            List of :obj:`tf.Tensor` of length :obj:`config.n_layers`, with each tensor of shape :obj:`(2, 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
+            :obj:`past_key_values` input) to speed up sequential decoding.
+        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer) of
+            shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`tf.Tensor` (one for each layer) of shape :obj:`(batch_size, num_heads, sequence_length,
+            sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    logits: tf.Tensor = None
+    mc_logits: tf.Tensor = None
+    past_key_values: Optional[List[tf.Tensor]] = None
+    hidden_states: Optional[Tuple[tf.Tensor]] = None
+    attentions: Optional[Tuple[tf.Tensor]] = None
+
+
+GPT2_START_DOCSTRING = r"""
+
+    This model inherits from :class:`~transformers.TFPreTrainedModel`. 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 `tf.keras.Model <https://www.tensorflow.org/api_docs/python/tf/keras/Model>`__ subclass. Use
+    it as a regular TF 2.0 Keras Model and refer to the TF 2.0 documentation for all matter related to general usage
+    and behavior.
+
+    .. note::
+
+        TF 2.0 models accepts two formats as inputs:
+
+        - having all inputs as keyword arguments (like PyTorch models), or
+        - having all inputs as a list, tuple or dict in the first positional arguments.
+
+        This second option is useful when using :meth:`tf.keras.Model.fit` method which currently requires having all
+        the tensors in the first argument of the model call function: :obj:`model(inputs)`.
+
+        If you choose this second option, there are three possibilities you can use to gather all the input Tensors in
+        the first positional argument :
+
+        - a single Tensor with :obj:`input_ids` only and nothing else: :obj:`model(inputs_ids)`
+        - a list of varying length with one or several input Tensors IN THE ORDER given in the docstring:
+          :obj:`model([input_ids, attention_mask])` or :obj:`model([input_ids, attention_mask, token_type_ids])`
+        - a dictionary with one or several input Tensors associated to the input names given in the docstring:
+          :obj:`model({"input_ids": input_ids, "token_type_ids": token_type_ids})`
+
+    Parameters:
+        config (:class:`~transformers.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 :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model
+            weights.
+"""
+
+GPT2_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, input_ids_length)`):
+            :obj:`input_ids_length` = ``sequence_length`` if ``past`` is ``None`` else ``past[0].shape[-2]``
+            (``sequence_length`` of input past key value states). Indices of input sequence tokens in the vocabulary.
+
+            If :obj:`past` is used, only input IDs that do not have their past calculated should be passed as
+            ``input_ids``.
+
+            Indices can be obtained using :class:`~transformers.GPT2Tokenizer`. See
+            :func:`transformers.PreTrainedTokenizer.__call__` and :func:`transformers.PreTrainedTokenizer.encode` for
+            details.
+
+            `What are input IDs? <../glossary.html#input-ids>`__
+        past (:obj:`List[tf.Tensor]` of length :obj:`config.n_layers`):
+            Contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model (see
+            :obj:`past` output below). Can be used to speed up sequential decoding. The token ids which have their past
+            given to this model should not be passed as input ids as they have already been computed.
+        attention_mask (:obj:`tf.Tensor` or :obj:`Numpy array` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Mask to avoid performing attention on padding token indices. Mask values selected in ``[0, 1]``:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            `What are attention masks? <../glossary.html#attention-mask>`__
+        token_type_ids (:obj:`tf.Tensor` or :obj:`Numpy array` of shape :obj:`(batch_size, sequence_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.html#token-type-ids>`__
+        position_ids (:obj:`tf.Tensor` or :obj:`Numpy array` of shape :obj:`(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.html#position-ids>`__
+        head_mask (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(num_heads,)` or :obj:`(num_layers, num_heads)`, `optional`):
+            Mask to nullify selected heads of the self-attention modules. Mask values selected in ``[0, 1]``:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        inputs_embeds (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):
+            Optionally, instead of passing :obj:`input_ids` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert :obj:`input_ids` indices into associated
+            vectors than the model's internal embedding lookup matrix.
+        output_attentions (:obj:`bool`, `optional`):
+            Whether or not to return the attentions tensors of all attention layers. See ``attentions`` under returned
+            tensors for more detail. This argument can be used only in eager mode, in graph mode the value in the
+            config will be used instead.
+        output_hidden_states (:obj:`bool`, `optional`):
+            Whether or not to return the hidden states of all layers. See ``hidden_states`` under returned tensors for
+            more detail. This argument can be used only in eager mode, in graph mode the value in the config will be
+            used instead.
+        return_dict (:obj:`bool`, `optional`):
+            Whether or not to return a :class:`~transformers.file_utils.ModelOutput` instead of a plain tuple. This
+            argument can be used in eager mode, in graph mode the value will always be set to True.
+        training (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Whether or not to use the model in training mode (some modules like dropout modules have different
+            behaviors between training and evaluation).
+"""
+
+
+@add_start_docstrings(
+    "The bare GPT2 Model transformer outputting raw hidden-states without any specific head on top.",
+    GPT2_START_DOCSTRING,
+)
+class TFGPT2Model(TFGPT2PreTrainedModel):
+    def __init__(self, config, *inputs, **kwargs):
+        super().__init__(config, *inputs, **kwargs)
+        self.transformer = TFGPT2MainLayer(config, name="transformer")
+
+    @add_start_docstrings_to_model_forward(GPT2_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=TFBaseModelOutputWithPast,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def call(
+        self,
+        input_ids=None,
+        past=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+        training=False,
+        **kwargs,
+    ):
+        inputs = input_processing(
+            func=self.call,
+            config=self.config,
+            input_ids=input_ids,
+            past=past,
+            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,
+            training=training,
+            kwargs_call=kwargs,
+        )
+        outputs = self.transformer(
+            input_ids=inputs["input_ids"],
+            past=inputs["past"],
+            attention_mask=inputs["attention_mask"],
+            token_type_ids=inputs["token_type_ids"],
+            position_ids=inputs["position_ids"],
+            head_mask=inputs["head_mask"],
+            inputs_embeds=inputs["inputs_embeds"],
+            use_cache=inputs["use_cache"],
+            output_attentions=inputs["output_attentions"],
+            output_hidden_states=inputs["output_hidden_states"],
+            return_dict=inputs["return_dict"],
+            training=inputs["training"],
+        )
+
+        return outputs
+
+    def serving_output(self, output):
+        pkv = tf.convert_to_tensor(output.past_key_values) if self.config.use_cache else None
+        hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
+        attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
+
+        return TFBaseModelOutputWithPast(
+            last_hidden_state=output.last_hidden_state, past_key_values=pkv, hidden_states=hs, attentions=attns
+        )
+
+
+@add_start_docstrings(
+    """
+    The GPT2 Model transformer with a language modeling head on top (linear layer with weights tied to the input
+    embeddings).
+    """,
+    GPT2_START_DOCSTRING,
+)
+class TFGPT2LMHeadModel(TFGPT2PreTrainedModel, TFCausalLanguageModelingLoss):
+    def __init__(self, config, *inputs, **kwargs):
+        super().__init__(config, *inputs, **kwargs)
+        self.transformer = TFGPT2MainLayer(config, name="transformer")
+
+    def get_output_embeddings(self):
+        return self.get_input_embeddings()
+
+    def set_output_embeddings(self, value):
+        self.set_input_embeddings(value)
+
+    def prepare_inputs_for_generation(self, inputs, past, **kwargs):
+        # only last token for inputs_ids if past is defined in kwargs
+        if past:
+            inputs = tf.expand_dims(inputs[:, -1], -1)
+
+        return {"input_ids": inputs, "past": past, "use_cache": kwargs["use_cache"]}
+
+    @add_start_docstrings_to_model_forward(GPT2_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=TFCausalLMOutputWithPast,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def call(
+        self,
+        input_ids=None,
+        past=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+        labels=None,
+        training=False,
+        **kwargs,
+    ):
+        r"""
+        labels (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Labels for computing the cross entropy classification loss. Indices should be in ``[0, ...,
+            config.vocab_size - 1]``.
+        """
+        inputs = input_processing(
+            func=self.call,
+            config=self.config,
+            input_ids=input_ids,
+            past=past,
+            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,
+            labels=labels,
+            training=training,
+            kwargs_call=kwargs,
+        )
+        transformer_outputs = self.transformer(
+            input_ids=inputs["input_ids"],
+            past=inputs["past"],
+            attention_mask=inputs["attention_mask"],
+            token_type_ids=inputs["token_type_ids"],
+            position_ids=inputs["position_ids"],
+            head_mask=inputs["head_mask"],
+            inputs_embeds=inputs["inputs_embeds"],
+            use_cache=inputs["use_cache"],
+            output_attentions=inputs["output_attentions"],
+            output_hidden_states=inputs["output_hidden_states"],
+            return_dict=inputs["return_dict"],
+            training=inputs["training"],
+        )
+        hidden_states = transformer_outputs[0]
+        logits = self.transformer.wte(hidden_states, mode="linear")
+
+        loss = None
+        if inputs["labels"] is not None:
+            # shift labels to the left and cut last logit token
+            logits = logits[:, :-1]
+            labels = inputs["labels"][:, 1:]
+            loss = self.compute_loss(labels, logits)
+
+        if not inputs["return_dict"]:
+            output = (logits,) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TFCausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=transformer_outputs.past_key_values,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
+
+    def serving_output(self, output):
+        pkv = tf.convert_to_tensor(output.past_key_values) if self.config.use_cache else None
+        hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
+        attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
+
+        return TFCausalLMOutputWithPast(logits=output.logits, past_key_values=pkv, hidden_states=hs, attentions=attns)
+
+
+@add_start_docstrings(
+    """
+    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 TFGPT2DoubleHeadsModel(TFGPT2PreTrainedModel):
+    def __init__(self, config, *inputs, **kwargs):
+        super().__init__(config, *inputs, **kwargs)
+        config.num_labels = 1
+        self.transformer = TFGPT2MainLayer(config, name="transformer")
+        self.multiple_choice_head = TFSequenceSummary(
+            config, initializer_range=config.initializer_range, name="multiple_choice_head"
+        )
+
+    @add_start_docstrings_to_model_forward(GPT2_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=TFGPT2DoubleHeadsModelOutput, config_class=_CONFIG_FOR_DOC)
+    def call(
+        self,
+        input_ids=None,
+        past=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        mc_token_ids=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+        training=False,
+        **kwargs,
+    ):
+        r"""
+        mc_token_ids (:obj:`tf.Tensor` or :obj:`Numpy array` of shape :obj:`(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[``.
+
+        Return:
+
+        Examples::
+
+            >>> import tensorflow as tf
+            >>> from transformers import GPT2Tokenizer, TFGPT2DoubleHeadsModel
+
+            >>> tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
+            >>> model = TFGPT2DoubleHeadsModel.from_pretrained('gpt2')
+
+            >>> # Add a [CLS] to the vocabulary (we should train it also!)
+            >>> num_added_tokens = tokenizer.add_special_tokens({'cls_token': '[CLS]'})
+
+            >>> embedding_layer = model.resize_token_embeddings(len(tokenizer))  # Update the model embeddings with the new vocabulary size
+
+            >>> 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 = tf.constant(encoded_choices)[None, :]  # Batch size: 1, number of choices: 2
+            >>> mc_token_ids = tf.constant([cls_token_location])  # Batch size: 1
+
+            >>> outputs = model(input_ids, mc_token_ids=mc_token_ids)
+            >>> lm_prediction_scores, mc_prediction_scores = outputs[:2]
+
+        """
+        inputs = input_processing(
+            func=self.call,
+            config=self.config,
+            input_ids=input_ids,
+            past=past,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            mc_token_ids=mc_token_ids,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            training=training,
+            kwargs_call=kwargs,
+        )
+
+        if inputs["input_ids"] is not None:
+            input_shapes = shape_list(inputs["input_ids"])
+        else:
+            input_shapes = shape_list(inputs["inputs_embeds"])[:-1]
+
+        seq_length = input_shapes[-1]
+        flat_input_ids = tf.reshape(inputs["input_ids"], (-1, seq_length)) if inputs["input_ids"] is not None else None
+        flat_attention_mask = (
+            tf.reshape(inputs["attention_mask"], (-1, seq_length)) if inputs["attention_mask"] is not None else None
+        )
+        flat_token_type_ids = (
+            tf.reshape(inputs["token_type_ids"], (-1, seq_length)) if inputs["token_type_ids"] is not None else None
+        )
+        flat_position_ids = (
+            tf.reshape(inputs["position_ids"], (-1, seq_length)) if inputs["position_ids"] is not None else None
+        )
+        transformer_outputs = self.transformer(
+            flat_input_ids,
+            inputs["past"],
+            flat_attention_mask,
+            flat_token_type_ids,
+            flat_position_ids,
+            inputs["head_mask"],
+            inputs["inputs_embeds"],
+            inputs["use_cache"],
+            inputs["output_attentions"],
+            inputs["output_hidden_states"],
+            return_dict=inputs["return_dict"],
+            training=inputs["training"],
+        )
+        hidden_states = transformer_outputs[0]
+        hidden_states = tf.reshape(hidden_states, input_shapes + shape_list(hidden_states)[-1:])
+        lm_logits = self.transformer.wte(hidden_states, mode="linear")
+        mc_logits = self.multiple_choice_head(hidden_states, inputs["mc_token_ids"], training=inputs["training"])
+        mc_logits = tf.squeeze(mc_logits, axis=-1)
+
+        if not inputs["return_dict"]:
+            return (lm_logits, mc_logits) + transformer_outputs[1:]
+
+        return TFGPT2DoubleHeadsModelOutput(
+            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,
+        )
+
+    @tf.function(
+        input_signature=[
+            {
+                "input_ids": tf.TensorSpec((None, None, None), tf.int32, name="input_ids"),
+                "attention_mask": tf.TensorSpec((None, None, None), tf.int32, name="attention_mask"),
+                "mc_token_ids": tf.TensorSpec((None, None), tf.int32, name="mc_token_ids"),
+            }
+        ]
+    )
+    def serving(self, inputs):
+        output = self.call(inputs)
+
+        return self.serving_output(output)
+
+    def serving_output(self, output):
+        pkv = tf.convert_to_tensor(output.past_key_values) if self.config.use_cache else None
+        hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
+        attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
+
+        return TFGPT2DoubleHeadsModelOutput(
+            logits=output.logits,
+            mc_logits=output.mc_logits,
+            past_key_values=pkv,
+            hidden_states=hs,
+            attentions=attns,
+        )
+
+
+@add_start_docstrings(
+    """
+    The GPT2 Model transformer with a sequence classification head on top (linear layer).
+
+    :class:`~transformers.TFGPT2ForSequenceClassification` 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
+    :obj:`pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each
+    row. If no :obj:`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 :obj:`inputs_embeds` are passed instead of :obj:`input_ids`, it does the same (take
+    the last value in each row of the batch).
+    """,
+    GPT2_START_DOCSTRING,
+)
+class TFGPT2ForSequenceClassification(TFGPT2PreTrainedModel, TFSequenceClassificationLoss):
+    def __init__(self, config, *inputs, **kwargs):
+        super().__init__(config, *inputs, **kwargs)
+        self.num_labels = config.num_labels
+        self.score = tf.keras.layers.Dense(
+            config.num_labels,
+            kernel_initializer=get_initializer(config.initializer_range),
+            name="score",
+            use_bias=False,
+        )
+        self.transformer = TFGPT2MainLayer(config, name="transformer")
+
+    @add_start_docstrings_to_model_forward(GPT2_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="microsoft/DialogRPT-updown",
+        output_type=TFSequenceClassifierOutputWithPast,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def call(
+        self,
+        input_ids=None,
+        past=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+        labels=None,
+        training=False,
+        **kwargs,
+    ):
+        r"""
+        labels (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Labels for computing the cross entropy classification loss. Indices should be in ``[0, ...,
+            config.vocab_size - 1]``.
+        """
+        inputs = input_processing(
+            func=self.call,
+            config=self.config,
+            input_ids=input_ids,
+            past=past,
+            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,
+            labels=labels,
+            training=training,
+            kwargs_call=kwargs,
+        )
+
+        transformer_outputs = self.transformer(
+            input_ids=inputs["input_ids"],
+            past=inputs["past"],
+            attention_mask=inputs["attention_mask"],
+            token_type_ids=inputs["token_type_ids"],
+            position_ids=inputs["position_ids"],
+            head_mask=inputs["head_mask"],
+            inputs_embeds=inputs["inputs_embeds"],
+            use_cache=inputs["use_cache"],
+            output_attentions=inputs["output_attentions"],
+            output_hidden_states=inputs["output_hidden_states"],
+            return_dict=inputs["return_dict"],
+            training=inputs["training"],
+        )
+
+        hidden_states = transformer_outputs[0]
+        logits = self.score(hidden_states)
+        logits_shape = shape_list(logits)
+        in_logits = None
+        if self.config.pad_token_id is None:
+            sequence_lengths = -1
+        else:
+            if inputs["input_ids"] is not None:
+                sequence_lengths = (
+                    tf.reduce_sum(
+                        tf.cast(
+                            tf.math.not_equal(inputs["input_ids"], self.config.pad_token_id),
+                            dtype=inputs["input_ids"].dtype,
+                        ),
+                        -1,
+                        keepdims=False,
+                    )
+                    - 1
+                )
+                in_logits = tf.gather(logits, sequence_lengths, batch_dims=1, axis=1)
+            else:
+                sequence_lengths = -1
+                logger.warning(
+                    f"{self.__class__.__name__} will not detect padding tokens in `inputs_embeds`. Results may be "
+                    f"unexpected if using padding tokens in conjunction with `inputs_embeds.`"
+                )
+        loss = None
+
+        if inputs["labels"] is not None:
+            assert (
+                self.config.pad_token_id is not None or logits_shape[0] == 1
+            ), "Cannot handle batch sizes > 1 if no padding token is defined."
+
+            if not tf.is_tensor(sequence_lengths):
+                in_logits = logits[0 : logits_shape[0], sequence_lengths]
+
+            loss = self.compute_loss(tf.reshape(inputs["labels"], [-1]), tf.reshape(in_logits, [-1, self.num_labels]))
+        pooled_logits = in_logits if in_logits is not None else logits
+
+        if not inputs["return_dict"]:
+            output = (pooled_logits,) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TFSequenceClassifierOutputWithPast(
+            loss=loss,
+            logits=pooled_logits,
+            past_key_values=transformer_outputs.past_key_values,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
+
+    def serving_output(self, output):
+        pkv = tf.convert_to_tensor(output.past_key_values) if self.config.use_cache else None
+        hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
+        attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
+
+        return TFSequenceClassifierOutputWithPast(
+            logits=output.logits, past_key_values=pkv, hidden_states=hs, attentions=attns
+        )

model/tokenization_gpt2.py

@@ -0,0 +1,309 @@
+# coding=utf-8
+# Copyright 2018 The Open AI Team Authors and The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Tokenization classes for OpenAI GPT."""
+
+
+import json
+import os
+from functools import lru_cache
+from typing import TYPE_CHECKING, List, Optional, Tuple
+
+import regex as re
+
+from ...tokenization_utils import AddedToken, PreTrainedTokenizer
+from ...utils import logging
+
+
+if TYPE_CHECKING:
+    from transformers.pipelines.conversational import Conversation
+
+logger = logging.get_logger(__name__)
+
+VOCAB_FILES_NAMES = {
+    "vocab_file": "vocab.json",
+    "merges_file": "merges.txt",
+}
+
+PRETRAINED_VOCAB_FILES_MAP = {
+    "vocab_file": {
+        "gpt2": "https://huggingface.co/gpt2/resolve/main/vocab.json",
+        "gpt2-medium": "https://huggingface.co/gpt2-medium/resolve/main/vocab.json",
+        "gpt2-large": "https://huggingface.co/gpt2-large/resolve/main/vocab.json",
+        "gpt2-xl": "https://huggingface.co/gpt2-xl/resolve/main/vocab.json",
+        "distilgpt2": "https://huggingface.co/distilgpt2/resolve/main/vocab.json",
+    },
+    "merges_file": {
+        "gpt2": "https://huggingface.co/gpt2/resolve/main/merges.txt",
+        "gpt2-medium": "https://huggingface.co/gpt2-medium/resolve/main/merges.txt",
+        "gpt2-large": "https://huggingface.co/gpt2-large/resolve/main/merges.txt",
+        "gpt2-xl": "https://huggingface.co/gpt2-xl/resolve/main/merges.txt",
+        "distilgpt2": "https://huggingface.co/distilgpt2/resolve/main/merges.txt",
+    },
+}
+
+PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {
+    "gpt2": 1024,
+    "gpt2-medium": 1024,
+    "gpt2-large": 1024,
+    "gpt2-xl": 1024,
+    "distilgpt2": 1024,
+}
+
+
+@lru_cache()
+def bytes_to_unicode():
+    """
+    Returns list of utf-8 byte and a mapping to unicode strings. We specifically avoids mapping to whitespace/control
+    characters the bpe code barfs on.
+
+    The reversible bpe codes work on unicode strings. This means you need a large # of unicode characters in your vocab
+    if you want to avoid UNKs. When you're at something like a 10B token dataset you end up needing around 5K for
+    decent coverage. This is a significant percentage of your normal, say, 32K bpe vocab. To avoid that, we want lookup
+    tables between utf-8 bytes and unicode strings.
+    """
+    bs = (
+        list(range(ord("!"), ord("~") + 1)) + list(range(ord("¡"), ord("¬") + 1)) + list(range(ord("®"), ord("ÿ") + 1))
+    )
+    cs = bs[:]
+    n = 0
+    for b in range(2 ** 8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2 ** 8 + n)
+            n += 1
+    cs = [chr(n) for n in cs]
+    return dict(zip(bs, cs))
+
+
+def get_pairs(word):
+    """
+    Return set of symbol pairs in a word.
+
+    Word is represented as tuple of symbols (symbols being variable-length strings).
+    """
+    pairs = set()
+    prev_char = word[0]
+    for char in word[1:]:
+        pairs.add((prev_char, char))
+        prev_char = char
+    return pairs
+
+
+class GPT2Tokenizer(PreTrainedTokenizer):
+    """
+    Construct a GPT-2 tokenizer. Based on byte-level Byte-Pair-Encoding.
+
+    This tokenizer has been trained to treat spaces like parts of the tokens (a bit like sentencepiece) so a word will
+    be encoded differently whether it is at the beginning of the sentence (without space) or not:
+
+    ::
+
+        >>> from transformers import GPT2Tokenizer
+        >>> tokenizer = GPT2Tokenizer.from_pretrained("gpt2")
+        >>> tokenizer("Hello world")['input_ids']
+        [15496, 995]
+        >>> tokenizer(" Hello world")['input_ids']
+        [18435, 995]
+
+    You can get around that behavior by passing ``add_prefix_space=True`` when instantiating this tokenizer or when you
+    call it on some text, but since the model was not pretrained this way, it might yield a decrease in performance.
+
+    .. note::
+
+        When used with ``is_split_into_words=True``, this tokenizer will add a space before each word (even the first
+        one).
+
+    This tokenizer inherits from :class:`~transformers.PreTrainedTokenizer` which contains most of the main methods.
+    Users should refer to this superclass for more information regarding those methods.
+
+    Args:
+        vocab_file (:obj:`str`):
+            Path to the vocabulary file.
+        merges_file (:obj:`str`):
+            Path to the merges file.
+        errors (:obj:`str`, `optional`, defaults to :obj:`"replace"`):
+            Paradigm to follow when decoding bytes to UTF-8. See `bytes.decode
+            <https://docs.python.org/3/library/stdtypes.html#bytes.decode>`__ for more information.
+        unk_token (:obj:`str`, `optional`, defaults to :obj:`<|endoftext|>`):
+            The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this
+            token instead.
+        bos_token (:obj:`str`, `optional`, defaults to :obj:`<|endoftext|>`):
+            The beginning of sequence token.
+        eos_token (:obj:`str`, `optional`, defaults to :obj:`<|endoftext|>`):
+            The end of sequence token.
+        add_prefix_space (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Whether or not to add an initial space to the input. This allows to treat the leading word just as any
+            other word. (GPT2 tokenizer detect beginning of words by the preceding space).
+    """
+
+    vocab_files_names = VOCAB_FILES_NAMES
+    pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
+    max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
+    model_input_names = ["input_ids", "attention_mask"]
+
+    def __init__(
+        self,
+        vocab_file,
+        merges_file,
+        errors="replace",
+        unk_token="<|endoftext|>",
+        bos_token="<|endoftext|>",
+        eos_token="<|endoftext|>",
+        add_prefix_space=False,
+        **kwargs
+    ):
+        bos_token = AddedToken(bos_token, lstrip=False, rstrip=False) if isinstance(bos_token, str) else bos_token
+        eos_token = AddedToken(eos_token, lstrip=False, rstrip=False) if isinstance(eos_token, str) else eos_token
+        unk_token = AddedToken(unk_token, lstrip=False, rstrip=False) if isinstance(unk_token, str) else unk_token
+        super().__init__(
+            errors=errors,
+            unk_token=unk_token,
+            bos_token=bos_token,
+            eos_token=eos_token,
+            add_prefix_space=add_prefix_space,
+            **kwargs,
+        )
+
+        with open(vocab_file, encoding="utf-8") as vocab_handle:
+            self.encoder = json.load(vocab_handle)
+        self.decoder = {v: k for k, v in self.encoder.items()}
+        self.errors = errors  # how to handle errors in decoding
+        self.byte_encoder = bytes_to_unicode()
+        self.byte_decoder = {v: k for k, v in self.byte_encoder.items()}
+        with open(merges_file, encoding="utf-8") as merges_handle:
+            bpe_merges = merges_handle.read().split("\n")[1:-1]
+        bpe_merges = [tuple(merge.split()) for merge in bpe_merges]
+        self.bpe_ranks = dict(zip(bpe_merges, range(len(bpe_merges))))
+        self.cache = {}
+        self.add_prefix_space = add_prefix_space
+
+        # Should have added re.IGNORECASE so BPE merges can happen for capitalized versions of contractions
+        self.pat = re.compile(r"""'s|'t|'re|'ve|'m|'ll|'d| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+""")
+
+    @property
+    def vocab_size(self):
+        return len(self.encoder)
+
+    def get_vocab(self):
+        return dict(self.encoder, **self.added_tokens_encoder)
+
+    def bpe(self, token):
+        if token in self.cache:
+            return self.cache[token]
+        word = tuple(token)
+        pairs = get_pairs(word)
+
+        if not pairs:
+            return token
+
+        while True:
+            bigram = min(pairs, key=lambda pair: self.bpe_ranks.get(pair, float("inf")))
+            if bigram not in self.bpe_ranks:
+                break
+            first, second = bigram
+            new_word = []
+            i = 0
+            while i < len(word):
+                try:
+                    j = word.index(first, i)
+                except ValueError:
+                    new_word.extend(word[i:])
+                    break
+                else:
+                    new_word.extend(word[i:j])
+                    i = j
+
+                if word[i] == first and i < len(word) - 1 and word[i + 1] == second:
+                    new_word.append(first + second)
+                    i += 2
+                else:
+                    new_word.append(word[i])
+                    i += 1
+            new_word = tuple(new_word)
+            word = new_word
+            if len(word) == 1:
+                break
+            else:
+                pairs = get_pairs(word)
+        word = " ".join(word)
+        self.cache[token] = word
+        return word
+
+    def _tokenize(self, text):
+        """Tokenize a string."""
+        bpe_tokens = []
+        for token in re.findall(self.pat, text):
+            token = "".join(
+                self.byte_encoder[b] for b in token.encode("utf-8")
+            )  # Maps all our bytes to unicode strings, avoiding control tokens of the BPE (spaces in our case)
+            bpe_tokens.extend(bpe_token for bpe_token in self.bpe(token).split(" "))
+        return bpe_tokens
+
+    def _convert_token_to_id(self, token):
+        """Converts a token (str) in an id using the vocab."""
+        return self.encoder.get(token, self.encoder.get(self.unk_token))
+
+    def _convert_id_to_token(self, index):
+        """Converts an index (integer) in a token (str) using the vocab."""
+        return self.decoder.get(index)
+
+    def convert_tokens_to_string(self, tokens):
+        """Converts a sequence of tokens (string) in a single string."""
+        text = "".join(tokens)
+        text = bytearray([self.byte_decoder[c] for c in text]).decode("utf-8", errors=self.errors)
+        return text
+
+    def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
+        if not os.path.isdir(save_directory):
+            logger.error(f"Vocabulary path ({save_directory}) should be a directory")
+            return
+        vocab_file = os.path.join(
+            save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]
+        )
+        merge_file = os.path.join(
+            save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["merges_file"]
+        )
+
+        with open(vocab_file, "w", encoding="utf-8") as f:
+            f.write(json.dumps(self.encoder, ensure_ascii=False))
+
+        index = 0
+        with open(merge_file, "w", encoding="utf-8") as writer:
+            writer.write("#version: 0.2\n")
+            for bpe_tokens, token_index in sorted(self.bpe_ranks.items(), key=lambda kv: kv[1]):
+                if index != token_index:
+                    logger.warning(
+                        f"Saving vocabulary to {merge_file}: BPE merge indices are not consecutive."
+                        " Please check that the tokenizer is not corrupted!"
+                    )
+                    index = token_index
+                writer.write(" ".join(bpe_tokens) + "\n")
+                index += 1
+
+        return vocab_file, merge_file
+
+    def prepare_for_tokenization(self, text, is_split_into_words=False, **kwargs):
+        add_prefix_space = kwargs.pop("add_prefix_space", self.add_prefix_space)
+        if is_split_into_words or add_prefix_space:
+            text = " " + text
+        return (text, kwargs)
+
+    def _build_conversation_input_ids(self, conversation: "Conversation") -> List[int]:
+        input_ids = []
+        for is_user, text in conversation.iter_texts():
+            input_ids.extend(self.encode(text, add_special_tokens=False) + [self.eos_token_id])
+        if len(input_ids) > self.model_max_length:
+            input_ids = input_ids[-self.model_max_length :]
+        return input_ids

model/tokenization_gpt2_fast.py

@@ -0,0 +1,187 @@
+# coding=utf-8
+# Copyright 2018 The Open AI Team Authors and The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Tokenization classes for OpenAI GPT."""
+
+
+import json
+from typing import TYPE_CHECKING, List, Optional, Tuple
+
+from tokenizers import pre_tokenizers
+
+from ...tokenization_utils_base import BatchEncoding
+from ...tokenization_utils_fast import PreTrainedTokenizerFast
+from ...utils import logging
+from .tokenization_gpt2 import GPT2Tokenizer
+
+
+if TYPE_CHECKING:
+    from transformers.pipelines.conversational import Conversation
+
+
+logger = logging.get_logger(__name__)
+
+VOCAB_FILES_NAMES = {"vocab_file": "vocab.json", "merges_file": "merges.txt", "tokenizer_file": "tokenizer.json"}
+
+PRETRAINED_VOCAB_FILES_MAP = {
+    "vocab_file": {
+        "gpt2": "https://huggingface.co/gpt2/resolve/main/vocab.json",
+        "gpt2-medium": "https://huggingface.co/gpt2-medium/resolve/main/vocab.json",
+        "gpt2-large": "https://huggingface.co/gpt2-large/resolve/main/vocab.json",
+        "gpt2-xl": "https://huggingface.co/gpt2-xl/resolve/main/vocab.json",
+        "distilgpt2": "https://huggingface.co/distilgpt2/resolve/main/vocab.json",
+    },
+    "merges_file": {
+        "gpt2": "https://huggingface.co/gpt2/resolve/main/merges.txt",
+        "gpt2-medium": "https://huggingface.co/gpt2-medium/resolve/main/merges.txt",
+        "gpt2-large": "https://huggingface.co/gpt2-large/resolve/main/merges.txt",
+        "gpt2-xl": "https://huggingface.co/gpt2-xl/resolve/main/merges.txt",
+        "distilgpt2": "https://huggingface.co/distilgpt2/resolve/main/merges.txt",
+    },
+    "tokenizer_file": {
+        "gpt2": "https://huggingface.co/gpt2/resolve/main/tokenizer.json",
+        "gpt2-medium": "https://huggingface.co/gpt2-medium/resolve/main/tokenizer.json",
+        "gpt2-large": "https://huggingface.co/gpt2-large/resolve/main/tokenizer.json",
+        "gpt2-xl": "https://huggingface.co/gpt2-xl/resolve/main/tokenizer.json",
+        "distilgpt2": "https://huggingface.co/distilgpt2/resolve/main/tokenizer.json",
+    },
+}
+
+PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {
+    "gpt2": 1024,
+    "gpt2-medium": 1024,
+    "gpt2-large": 1024,
+    "gpt2-xl": 1024,
+    "distilgpt2": 1024,
+}
+
+
+class GPT2TokenizerFast(PreTrainedTokenizerFast):
+    """
+    Construct a "fast" GPT-2 tokenizer (backed by HuggingFace's `tokenizers` library). Based on byte-level
+    Byte-Pair-Encoding.
+
+    This tokenizer has been trained to treat spaces like parts of the tokens (a bit like sentencepiece) so a word will
+    be encoded differently whether it is at the beginning of the sentence (without space) or not:
+
+    ::
+
+        >>> from transformers import GPT2TokenizerFast
+        >>> tokenizer = GPT2TokenizerFast.from_pretrained("gpt2")
+        >>> tokenizer("Hello world")['input_ids']
+        [15496, 995]
+        >>> tokenizer(" Hello world")['input_ids']
+        [18435, 995]
+
+    You can get around that behavior by passing ``add_prefix_space=True`` when instantiating this tokenizer or when you
+    call it on some text, but since the model was not pretrained this way, it might yield a decrease in performance.
+
+    .. note::
+
+        When used with ``is_split_into_words=True``, this tokenizer needs to be instantiated with
+        ``add_prefix_space=True``.
+
+    This tokenizer inherits from :class:`~transformers.PreTrainedTokenizerFast` which contains most of the main
+    methods. Users should refer to this superclass for more information regarding those methods.
+
+    Args:
+        vocab_file (:obj:`str`):
+            Path to the vocabulary file.
+        merges_file (:obj:`str`):
+            Path to the merges file.
+        errors (:obj:`str`, `optional`, defaults to :obj:`"replace"`):
+            Paradigm to follow when decoding bytes to UTF-8. See `bytes.decode
+            <https://docs.python.org/3/library/stdtypes.html#bytes.decode>`__ for more information.
+        unk_token (:obj:`str`, `optional`, defaults to :obj:`<|endoftext|>`):
+            The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this
+            token instead.
+        bos_token (:obj:`str`, `optional`, defaults to :obj:`<|endoftext|>`):
+            The beginning of sequence token.
+        eos_token (:obj:`str`, `optional`, defaults to :obj:`<|endoftext|>`):
+            The end of sequence token.
+        add_prefix_space (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Whether or not to add an initial space to the input. This allows to treat the leading word just as any
+            other word. (GPT2 tokenizer detect beginning of words by the preceding space).
+        trim_offsets (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Whether or not the post-processing step should trim offsets to avoid including whitespaces.
+    """
+
+    vocab_files_names = VOCAB_FILES_NAMES
+    pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
+    max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
+    model_input_names = ["input_ids", "attention_mask"]
+    slow_tokenizer_class = GPT2Tokenizer
+
+    def __init__(
+        self,
+        vocab_file=None,
+        merges_file=None,
+        tokenizer_file=None,
+        unk_token="<|endoftext|>",
+        bos_token="<|endoftext|>",
+        eos_token="<|endoftext|>",
+        add_prefix_space=False,
+        **kwargs
+    ):
+        super().__init__(
+            vocab_file,
+            merges_file,
+            tokenizer_file=tokenizer_file,
+            unk_token=unk_token,
+            bos_token=bos_token,
+            eos_token=eos_token,
+            add_prefix_space=add_prefix_space,
+            **kwargs,
+        )
+
+        pre_tok_state = json.loads(self.backend_tokenizer.pre_tokenizer.__getstate__())
+        if pre_tok_state.get("add_prefix_space", add_prefix_space) != add_prefix_space:
+            pre_tok_class = getattr(pre_tokenizers, pre_tok_state.pop("type"))
+            pre_tok_state["add_prefix_space"] = add_prefix_space
+            self.backend_tokenizer.pre_tokenizer = pre_tok_class(**pre_tok_state)
+
+        self.add_prefix_space = add_prefix_space
+
+    def _batch_encode_plus(self, *args, **kwargs) -> BatchEncoding:
+        is_split_into_words = kwargs.get("is_split_into_words", False)
+        assert self.add_prefix_space or not is_split_into_words, (
+            f"You need to instantiate {self.__class__.__name__} with add_prefix_space=True "
+            "to use it with pretokenized inputs."
+        )
+
+        return super()._batch_encode_plus(*args, **kwargs)
+
+    def _encode_plus(self, *args, **kwargs) -> BatchEncoding:
+        is_split_into_words = kwargs.get("is_split_into_words", False)
+
+        assert self.add_prefix_space or not is_split_into_words, (
+            f"You need to instantiate {self.__class__.__name__} with add_prefix_space=True "
+            "to use it with pretokenized inputs."
+        )
+
+        return super()._encode_plus(*args, **kwargs)
+
+    def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
+        files = self._tokenizer.model.save(save_directory, name=filename_prefix)
+        return tuple(files)
+
+    def _build_conversation_input_ids(self, conversation: "Conversation") -> List[int]:
+        """This corresponds to DialoGPT variants of models."""
+        input_ids = []
+        for is_user, text in conversation.iter_texts():
+            input_ids.extend(self.encode(text, add_special_tokens=False) + [self.eos_token_id])
+
+        if len(input_ids) > self.model_max_length:
+            input_ids = input_ids[-self.model_max_length :]
+        return input_ids