modeling_jbert.py

# Copyright 2022 MosaicML Examples authors
# SPDX-License-Identifier: Apache-2.0

# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
# Copyright (c) 2018-2021, NVIDIA CORPORATION.  All rights reserved.
# Copyright (c) 2022, Tri Dao.

import copy
import logging
import math
import warnings
from typing import List, Optional, Tuple, Union

import torch
import torch.nn as nn
from einops import rearrange
from transformers.activations import ACT2FN
from transformers.modeling_outputs import (
    MaskedLMOutput,
    SequenceClassifierOutput,
    BaseModelOutputWithPastAndCrossAttentions,
    BaseModelOutputWithPoolingAndCrossAttentions,
)
from transformers.models.bert.modeling_bert import BertPreTrainedModel

from .bert_padding import (index_first_axis, index_put_first_axis, pad_input,
                           unpad_input, unpad_input_only)
from .configuration_jbert import JBertConfig

logger = logging.getLogger(__name__)


class JBertEmbeddings(nn.Module):
    """Construct the embeddings for words, ignoring position.

    There are no positional embeddings since we use ALiBi and token_type
    embeddings.

    This module is modeled after the Hugging Face BERT's
    :class:`~transformers.model.bert.modeling_bert.BertEmbeddings`, but is
    modified to implement ALiBi. The key change is
    that position embeddings are removed. Position information instead comes
    from attention biases that scale linearly with the position distance
    between query and key tokens.

    This module ignores the `position_ids` input to the `forward` method.
    """

    def __init__(self, config):
        super().__init__()
        self.word_embeddings = nn.Embedding(
            config.vocab_size, config.hidden_size, padding_idx=config.pad_token_id
        )
        # ALiBi doesn't use position embeddings
        self.token_type_embeddings = nn.Embedding(
            config.type_vocab_size, config.hidden_size
        )

        # self.LayerNorm is not snake-cased to stick with TensorFlow model
        # variable name and be able to load any TensorFlow checkpoint file
        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)
        self.register_buffer(
            "token_type_ids", torch.zeros((1, config.model_max_length), dtype=torch.long), persistent=False
        )

    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        token_type_ids: Optional[torch.LongTensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        past_key_values_length: int = 0,
    ) -> torch.Tensor:
        if (input_ids is not None) == (inputs_embeds is not None):
            raise ValueError('Must specify either input_ids or input_embeds!')
        if input_ids is not None:
            input_shape = input_ids.size()
        else:
            assert inputs_embeds is not None  # just for type checking
            input_shape = inputs_embeds.size()[:-1]

        seq_length = input_shape[1]

        if position_ids is not None:
            warnings.warn('position_ids is not used in JBertEmbeddings as it does not have position embeddings.')

        # Setting the token_type_ids to the registered buffer in constructor
        # where it is all zeros, which usually occurs when it's auto-generated;
        # registered buffer helps users when tracing the model without passing
        # token_type_ids, solves issue #5664
        if token_type_ids is None:
            if hasattr(self, 'token_type_ids'):
                buffered_token_type_ids = self.token_type_ids[:, :seq_length]
                buffered_token_type_ids_expanded = buffered_token_type_ids.expand(
                    input_shape[0], seq_length
                )
                token_type_ids = buffered_token_type_ids_expanded  # type: ignore
            else:
                token_type_ids = torch.zeros(
                    input_shape,  # type: ignore
                    dtype=torch.long,
                    device=self.word_embeddings.device,
                )  # type: ignore  # yapf: disable

        if inputs_embeds is None:
            inputs_embeds = self.word_embeddings(input_ids)
        token_type_embeddings = self.token_type_embeddings(token_type_ids)

        embeddings = inputs_embeds + token_type_embeddings
        embeddings = self.LayerNorm(embeddings)
        embeddings = self.dropout(embeddings)
        return embeddings


class BertUnpadSelfAttention(nn.Module):
    """Performs multi-headed self attention on a batch of unpadded sequences.

    If Triton is installed, this module uses Flash Attention to greatly improve throughput.
    The Flash Attention implementation used in Mosaic BERT supports arbitrary attention biases (which
    we use to implement ALiBi), but does not support attention dropout. If either Triton is not installed
    or `config.attention_probs_dropout_prob > 0`, the implementation will default to a
    math-equivalent pytorch version, which is much slower.

    See `forward` method for additional detail.
    """

    def __init__(self, config):
        super().__init__()
        if config.hidden_size % config.num_attention_heads != 0 and not hasattr(
            config, 'embedding_size'
        ):
            raise ValueError(
                f'The hidden size ({config.hidden_size}) is not a multiple of the number of attention '
                f'heads ({config.num_attention_heads})'
            )

        self.num_attention_heads = config.num_attention_heads
        self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
        # TODO: self.all_head_size == config.hidden_size? Why not just use config.hidden_size?
        self.all_head_size = self.num_attention_heads * self.attention_head_size

        self.Wqkv = nn.Linear(self.all_head_size, 3 * config.hidden_size)

        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)

    def forward(
        self,
        hidden_states: torch.Tensor,
        cu_seqlens: torch.Tensor,
        max_seqlen_in_batch: int,
        indices: torch.Tensor,
        attn_mask: torch.Tensor,
        bias: torch.Tensor,
    ) -> torch.Tensor:
        """Perform self-attention.

        If dropout is zero, then we can use the Triton kernel, so we do that. However, if not, we send through a standard PyTorch
        implementation of self-attention.

        The arguments are unpadded, and our implementations of attention require padded arguments,
        so we first call `pad_input`. Once we compute attention, we re-unpad our outputs for the other layers.
        The pad/unpad operations add overhead, but not sending pad tokens through ffs saves compute.
        It is possible to write an unpadded implementation of attention (in Triton and PyTorch), which we will eventually do.

        Args:
            hidden_states: (total_nnz, dim)
            cu_seqlens: (batch + 1,)
            max_seqlen_in_batch: int
            indices: (total_nnz,)
            attn_mask: (batch, max_seqlen_in_batch)
            bias: (batch, heads, max_seqlen_in_batch, max_seqlen_in_batch)

        Returns:
            attention: (total_nnz, dim)
        """
        qkv = self.Wqkv(hidden_states)
        qkv = pad_input(
            qkv, indices, cu_seqlens.shape[0] - 1, max_seqlen_in_batch
        )  # batch, max_seqlen_in_batch, thd
        qkv = rearrange(
            qkv, 'b s (t h d) -> b s t h d', t=3, h=self.num_attention_heads
        )
        # if we have nonzero attention dropout (e.g. during fine-tuning) or no Triton, compute attention in PyTorch
        q = qkv[:, :, 0, :, :].permute(0, 2, 1, 3)  # b h s d
        k = qkv[:, :, 1, :, :].permute(0, 2, 3, 1)  # b h d s
        v = qkv[:, :, 2, :, :].permute(0, 2, 1, 3)  # b h s d
        attention_scores = torch.matmul(q, k) / math.sqrt(self.attention_head_size)
        attention_scores = attention_scores + bias
        attention_probs = nn.functional.softmax(attention_scores, dim=-1)
        attention_probs = self.dropout(attention_probs)
        attention_probs = attention_probs.to(dtype=v.dtype)
        attention = torch.matmul(attention_probs, v).permute(0, 2, 1, 3)  # b s h

        # attn_mask is 1 for attend and 0 for don't
        attention = unpad_input_only(attention, torch.squeeze(attn_mask) == 1)
        return rearrange(attention, 'nnz h d -> nnz (h d)')


# Copy of transformer's library BertSelfOutput that will not be caught by surgery methods looking for HF BERT modules.
class BertSelfOutput(nn.Module):
    """Computes the output of the attention layer.

    This module is modeled after the Hugging Face BERT's
    :class:`~transformers.model.bert.modeling_bert.BertSelfOutput`.
    The implementation is identical. Rather than use the original module
    directly, we re-implement it here so that Mosaic BERT's modules will not
    be affected by any Composer surgery algorithm that modifies Hugging Face
    BERT modules.
    """

    def __init__(self, config):
        super().__init__()
        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)

    def forward(
        self, hidden_states: torch.Tensor, input_tensor: torch.Tensor
    ) -> torch.Tensor:
        hidden_states = self.dense(hidden_states)
        hidden_states = self.dropout(hidden_states)
        hidden_states = self.LayerNorm(hidden_states + input_tensor)
        return hidden_states


class BertUnpadAttention(nn.Module):
    """Chains attention, Dropout, and LayerNorm for Mosaic BERT."""

    def __init__(self, config):
        super().__init__()
        self.self = BertUnpadSelfAttention(config)
        self.output = BertSelfOutput(config)

    def forward(
        self,
        input_tensor: torch.Tensor,
        cu_seqlens: torch.Tensor,
        max_s: int,
        subset_idx: Optional[torch.Tensor] = None,
        indices: Optional[torch.Tensor] = None,
        attn_mask: Optional[torch.Tensor] = None,
        bias: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
        """Forward pass for scaled self-attention without padding.

        Arguments:
            input_tensor: (total_nnz, dim)
            cu_seqlens: (batch + 1,)
            max_s: int
            subset_idx: () set of indices whose values we care about at the end of the layer
                        (e.g., the masked tokens, if this is the final layer).
            indices: None or (total_nnz,)
            attn_mask: None or (batch, max_seqlen_in_batch)
            bias: None or (batch, heads, max_seqlen_in_batch, max_seqlen_in_batch)
        """
        self_output = self.self(
            input_tensor, cu_seqlens, max_s, indices, attn_mask, bias
        )
        if subset_idx is not None:
            return self.output(
                index_first_axis(self_output, subset_idx),
                index_first_axis(input_tensor, subset_idx),
            )
        else:
            return self.output(self_output, input_tensor)


class BertGatedLinearUnitMLP(nn.Module):
    """Applies the FFN at the end of each Mosaic BERT layer.

    Compared to the default BERT architecture, this block replaces :class:`~transformers.model.bert.modeling_bert.BertIntermediate`
    and :class:`~transformers.model.bert.modeling_bert.SelfOutput` with a single module that has similar functionality, but
    introduces Gated Linear Units.

    Note: Mosaic BERT adds parameters in order to implement Gated Linear Units. To keep parameter count consistent with that of a
    standard Hugging Face BERT, scale down `config.intermediate_size` by 2/3. For example, a Mosaic BERT constructed with
    `config.intermediate_size=2048` will have the same parameter footprint as its Hugging Face BERT counterpart constructed
    with the `config.intermediate_size=3072`.
    However, in most cases it will not be necessary to adjust `config.intermediate_size` since, despite the increased
    parameter size, Mosaic BERT typically offers a net higher throughput than a Hugging Face BERT built from the same `config`.
    """

    def __init__(self, config):
        super().__init__()
        self.config = config
        self.gated_layers = nn.Linear(
            config.hidden_size, config.intermediate_size * 2, bias=False
        )
        self.act = nn.GELU(approximate='none')
        self.wo = nn.Linear(config.intermediate_size, config.hidden_size)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)
        self.layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        """Compute new hidden states from current hidden states.

        Args:
            hidden_states (torch.Tensor): The (unpadded) hidden states from
                the attention layer [nnz, dim].
        """
        residual_connection = hidden_states
        # compute the activation
        hidden_states = self.gated_layers(hidden_states)
        gated = hidden_states[:, : self.config.intermediate_size]
        non_gated = hidden_states[:, self.config.intermediate_size :]
        hidden_states = self.act(gated) * non_gated
        hidden_states = self.dropout(hidden_states)
        # multiply by the second matrix
        hidden_states = self.wo(hidden_states)
        # add the residual connection and post-LN
        hidden_states = self.layernorm(hidden_states + residual_connection)
        return hidden_states


class BertLayer(nn.Module):
    """Composes the Mosaic BERT attention and FFN blocks into a single layer."""

    def __init__(self, config: JBertConfig):
        super().__init__()
        self.attention = BertUnpadAttention(config)
        self.mlp = BertGatedLinearUnitMLP(config)

    def forward(
        self,
        hidden_states: torch.Tensor,
        cu_seqlens: torch.Tensor,
        seqlen: int,
        subset_idx: Optional[torch.Tensor] = None,
        indices: Optional[torch.Tensor] = None,
        attn_mask: Optional[torch.Tensor] = None,
        bias: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
        """Forward pass for a BERT layer, including both attention and MLP.

        Args:
            hidden_states: (total_nnz, dim)
            cu_seqlens: (batch + 1,)
            seqlen: int
            subset_idx: () set of indices whose values we care about at the end of the layer
                        (e.g., the masked tokens, if this is the final layer).
            indices: None or (total_nnz,)
            attn_mask: None or (batch, max_seqlen_in_batch)
            bias: None or (batch, heads, max_seqlen_in_batch, max_seqlen_in_batch)
        """
        attention_output = self.attention(
            hidden_states, cu_seqlens, seqlen, subset_idx, indices, attn_mask, bias
        )
        layer_output = self.mlp(attention_output)
        return layer_output


class JBertEncoder(nn.Module):
    """A stack of BERT layers providing the backbone.

    This module is modeled after the Hugging Face BERT's :class:`~transformers.model.bert.modeling_bert.BertEncoder`,
    but with substantial modifications to implement unpadding and ALiBi.

    Compared to the analogous Hugging Face BERT module, this module handles unpadding to reduce unnecessary computation
    at padded tokens, and pre-computes attention biases to implement ALiBi.
    """

    def __init__(self, config: JBertConfig):
        super().__init__()
        self.layer = nn.ModuleList(
            [BertLayer(config) for _ in range(config.num_hidden_layers)]
        )

        self.num_attention_heads = config.num_attention_heads

        # The alibi mask will be dynamically expanded if it is too small for
        # the input the model receives. But it generally helps to initialize it
        # to a reasonably large size to help pre-allocate CUDA memory.
        # The default `model_max_length` is 8192.
        self._current_alibi_size = int(config.model_max_length)
        self.alibi = torch.zeros(
            (
                1,
                self.num_attention_heads,
                self._current_alibi_size,
                self._current_alibi_size,
            )
        )
        self.rebuild_alibi_tensor(size=config.model_max_length)

    def rebuild_alibi_tensor(
        self, size: int, device: Optional[Union[torch.device, str]] = None
    ):
        # Alibi
        # Following https://github.com/ofirpress/attention_with_linear_biases/issues/5 (Implementation 1)
        # In the causal case, you can exploit the fact that softmax is invariant to a uniform translation
        # of the logits, which makes the math work out *after* applying causal masking. If no causal masking
        # will be applied, it is necessary to construct the diagonal mask.
        n_heads = self.num_attention_heads

        def _get_alibi_head_slopes(n_heads: int) -> List[float]:
            def get_slopes_power_of_2(n_heads: int) -> List[float]:
                start = 2 ** (-(2 ** -(math.log2(n_heads) - 3)))
                ratio = start
                return [start * ratio**i for i in range(n_heads)]

            # In the paper, they only train models that have 2^a heads for some a. This function
            # has some good properties that only occur when the input is a power of 2. To
            # maintain that even when the number of heads is not a power of 2, we use a
            # workaround.
            if math.log2(n_heads).is_integer():
                return get_slopes_power_of_2(n_heads)

            closest_power_of_2 = 2 ** math.floor(math.log2(n_heads))
            slopes_a = get_slopes_power_of_2(closest_power_of_2)
            slopes_b = _get_alibi_head_slopes(2 * closest_power_of_2)
            slopes_b = slopes_b[0::2][: n_heads - closest_power_of_2]
            return slopes_a + slopes_b

        context_position = torch.arange(size, device=device)[:, None]
        memory_position = torch.arange(size, device=device)[None, :]
        relative_position = torch.abs(memory_position - context_position)
        # [n_heads, max_token_length, max_token_length]
        relative_position = relative_position.unsqueeze(0).expand(n_heads, -1, -1)
        slopes = torch.Tensor(_get_alibi_head_slopes(n_heads)).to(device)
        alibi = slopes.unsqueeze(1).unsqueeze(1) * -relative_position
        # [1, n_heads, max_token_length, max_token_length]
        alibi = alibi.unsqueeze(0)
        assert alibi.shape == torch.Size([1, n_heads, size, size])

        self._current_alibi_size = size
        self.alibi = alibi

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.FloatTensor] = None,
        head_mask: Optional[torch.FloatTensor] = None,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
        past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = False,
        output_hidden_states: Optional[bool] = False,
        return_dict: Optional[bool] = True,
    ) -> List[torch.Tensor]:
        all_hidden_states = [] if output_hidden_states else None

        extended_attention_mask = attention_mask.unsqueeze(1).unsqueeze(2)
        extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0

        attention_mask_bool = attention_mask.bool()
        batch, seqlen = hidden_states.shape[:2]
        # Unpad inputs and mask. It will remove tokens that are padded.
        # Assume ntokens is total number of tokens (padded and non-padded)
        # and ntokens_unpad is total number of non-padded tokens.
        # Then unpadding performs the following compression of the inputs:
        # hidden_states[ntokens,hidden] -> hidden_states[ntokens_unpad,hidden]
        hidden_states, indices, cu_seqlens, _ = unpad_input(
            hidden_states, attention_mask_bool
        )

        # Add alibi matrix to extended_attention_mask
        if self._current_alibi_size < seqlen:
            # Rebuild the alibi tensor when needed
            warnings.warn(
                f'Increasing alibi size from {self._current_alibi_size} to {seqlen}'
            )
            self.rebuild_alibi_tensor(size=seqlen, device=hidden_states.device)
        elif self.alibi.device != hidden_states.device:
            # Device catch-up
            self.alibi = self.alibi.to(hidden_states.device)
        alibi_bias = self.alibi[:, :, :seqlen, :seqlen]
        attn_bias = extended_attention_mask[:, :, :seqlen, :seqlen]
        alibi_attn_mask = attn_bias + alibi_bias

        for layer_module in self.layer:
            if output_hidden_states:
                all_hidden_states.append(rearrange(hidden_states, '(b n) d -> b n d', b=batch))
            hidden_states = layer_module(
                hidden_states,
                cu_seqlens,
                seqlen,
                None,
                indices,
                attn_mask=attention_mask,
                bias=alibi_attn_mask,
            )
        # Pad inputs and mask. It will insert back zero-padded tokens.
        # Assume ntokens is total number of tokens (padded and non-padded)
        # and ntokens_unpad is total number of non-padded tokens.
        # Then padding performs the following de-compression:
        #     hidden_states[ntokens_unpad,hidden] -> hidden_states[ntokens,hidden]
        hidden_states = pad_input(hidden_states, indices, batch, seqlen)

        if output_hidden_states:
            all_hidden_states.append(hidden_states)
        
        if not return_dict:
            return tuple(
                v for v in [hidden_states, all_hidden_states] if v is not None
            )
        return BaseModelOutputWithPastAndCrossAttentions(
            last_hidden_state=hidden_states,
            past_key_values=None,
            hidden_states=all_hidden_states,
            attentions=None,
            cross_attentions=None,
        )


class JBertPooler(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
        self.activation = nn.Tanh()

    def forward(
        self, hidden_states: torch.Tensor, pool: Optional[bool] = True
    ) -> torch.Tensor:
        # We "pool" the model by simply taking the hidden state corresponding
        # to the first token.
        first_token_tensor = hidden_states[:, 0] if pool else hidden_states
        pooled_output = self.dense(first_token_tensor)
        pooled_output = self.activation(pooled_output)
        return pooled_output


class BertPredictionHeadTransform(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
        if isinstance(config.hidden_act, str):
            self.transform_act_fn = ACT2FN[config.hidden_act]
        else:
            self.transform_act_fn = config.hidden_act
        self.LayerNorm = torch.nn.LayerNorm(config.hidden_size, eps=1e-12)

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        hidden_states = self.dense(hidden_states)
        hidden_states = self.transform_act_fn(hidden_states)
        hidden_states = self.LayerNorm(hidden_states)
        return hidden_states


class JBertModel(BertPreTrainedModel):
    """Overall BERT model.

    Args:
        config: a JBertConfig class instance with the configuration to build a new model

    Inputs:
        `input_ids`: a torch.LongTensor of shape [batch_size, sequence_length]
            with the word token indices in the vocabulary(see the tokens preprocessing logic in the scripts
            `extract_features.py`, `run_classifier.py` and `run_squad.py`)
        `token_type_ids`: an optional torch.LongTensor of shape [batch_size, sequence_length] with the token
            types indices selected in [0, 1]. Type 0 corresponds to a `sentence A` and type 1 corresponds to
            a `sentence B` token (see BERT paper for more details).
        `attention_mask`: an optional torch.LongTensor of shape [batch_size, sequence_length] with indices
            selected in [0, 1]. It's a mask to be used if the input sequence length is smaller than the max
            input sequence length in the current batch. It's the mask that we typically use for attention when
            a batch has varying length sentences.
        `output_all_encoded_layers`: boolean which controls the content of the `encoded_layers` output as described below. Default: `True`.

    Outputs: Tuple of (encoded_layers, pooled_output)
        `encoded_layers`: controlled by `output_all_encoded_layers` argument:
            - `output_all_encoded_layers=True`: outputs a list of the full sequences of encoded-hidden-states at the end
                of each attention block (i.e. 12 full sequences for BERT-base, 24 for BERT-large), each
                encoded-hidden-state is a torch.FloatTensor of size [batch_size, sequence_length, hidden_size],
            - `output_all_encoded_layers=False`: outputs only the full sequence of hidden-states corresponding
                to the last attention block of shape [batch_size, sequence_length, hidden_size],
        `pooled_output`: a torch.FloatTensor of size [batch_size, hidden_size] which is the output of a
            classifier pretrained on top of the hidden state associated to the first character of the
            input (`CLS`) to train on the Next-Sentence task (see BERT's paper).

    Example usage:
    ```python
    # Already been converted into WordPiece token ids
    input_ids = torch.LongTensor([[31, 51, 99], [15, 5, 0]])
    input_mask = torch.LongTensor([[1, 1, 1], [1, 1, 0]])
    token_type_ids = torch.LongTensor([[0, 0, 1], [0, 1, 0]])
    config = modeling.JBertConfig(vocab_size_or_config_json_file=32000, hidden_size=768,
        num_hidden_layers=12, num_attention_heads=12, intermediate_size=3072)
    model = JBertModel(config=config)
    all_encoder_layers, pooled_output = model(input_ids, token_type_ids, input_mask)
    ```
    """

    config_class = JBertConfig

    def __init__(self, config, add_pooling_layer=True):
        super().__init__(config)
        self.embeddings = JBertEmbeddings(config)
        self.encoder = JBertEncoder(config)
        self.pooler = JBertPooler(config) if add_pooling_layer else None
        self.post_init()

    def get_input_embeddings(self):
        return self.embeddings.word_embeddings

    def set_input_embeddings(self, value):
        self.embeddings.word_embeddings = value

    def forward(
        self,
        input_ids: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        token_type_ids: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.Tensor] = None,
        head_mask: Optional[torch.Tensor] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        encoder_attention_mask: Optional[torch.Tensor] = None,
        output_attentions: Optional[bool] = False,
        output_hidden_states: Optional[bool] = False,
        return_dict: Optional[bool] = True,
    ) -> Union[Tuple[torch.Tensor], BaseModelOutputWithPoolingAndCrossAttentions]:
        if attention_mask is None:
            attention_mask = torch.ones_like(input_ids)
        if token_type_ids is None:
            token_type_ids = torch.zeros_like(input_ids)

        embedding_output = self.embeddings(input_ids, token_type_ids, position_ids)

        encoder_outputs: BaseModelOutputWithPastAndCrossAttentions = self.encoder(
            hidden_states=embedding_output,
            attention_mask=attention_mask,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        sequence_output = encoder_outputs[0]
        pooled_output = (
            self.pooler(sequence_output) if self.pooler is not None else None
        )

        if not return_dict:
            return (sequence_output, pooled_output) + encoder_outputs[1:]

        #return encoder_outputs, None
        return BaseModelOutputWithPoolingAndCrossAttentions(
            last_hidden_state=sequence_output,
            pooler_output=pooled_output,
            past_key_values=encoder_outputs.past_key_values,
            hidden_states=encoder_outputs.hidden_states,
            attentions=encoder_outputs.attentions,
            cross_attentions=encoder_outputs.cross_attentions,
        )


###################
# Bert Heads
###################
class BertLMPredictionHead(nn.Module):
    def __init__(self, config, bert_model_embedding_weights):
        super().__init__()
        self.transform = BertPredictionHeadTransform(config)
        # The output weights are the same as the input embeddings, but there is
        # an output-only bias for each token.
        self.decoder = nn.Linear(
            bert_model_embedding_weights.size(1), bert_model_embedding_weights.size(0)
        )
        self.decoder.weight = bert_model_embedding_weights

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        hidden_states = self.transform(hidden_states)
        hidden_states = self.decoder(hidden_states)
        return hidden_states


class BertOnlyMLMHead(nn.Module):
    def __init__(self, config, bert_model_embedding_weights):
        super().__init__()
        self.predictions = BertLMPredictionHead(config, bert_model_embedding_weights)

    def forward(self, sequence_output: torch.Tensor) -> torch.Tensor:
        prediction_scores = self.predictions(sequence_output)
        return prediction_scores


class BertOnlyNSPHead(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.seq_relationship = nn.Linear(config.hidden_size, 2)

    def forward(self, pooled_output: torch.Tensor) -> torch.Tensor:
        seq_relationship_score = self.seq_relationship(pooled_output)
        return seq_relationship_score


#####################
# Various Bert models
#####################
class JBertForMaskedLM(BertPreTrainedModel):
    config_class = JBertConfig

    def __init__(self, config):
        super().__init__(config)

        if config.is_decoder:
            warnings.warn(
                'If you want to use `JBertForMaskedLM` make sure `config.is_decoder=False` for '
                'bi-directional self-attention.'
            )

        self.bert = JBertModel(config, add_pooling_layer=False)
        self.cls = BertOnlyMLMHead(config, self.bert.embeddings.word_embeddings.weight)

        # Initialize weights and apply final processing
        self.post_init()

    def get_output_embeddings(self):
        return self.cls.predictions.decoder

    def set_output_embeddings(self, new_embeddings):
        self.cls.predictions.decoder = new_embeddings

    def forward(
        self,
        input_ids: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        token_type_ids: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.Tensor] = None,
        head_mask: Optional[torch.Tensor] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        encoder_attention_mask: Optional[torch.Tensor] = None,
        labels: Optional[torch.Tensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple[torch.Tensor], MaskedLMOutput]:
        # labels should be a `torch.LongTensor` of shape
        # `(batch_size, sequence_length)`. These are used for computing the
        #  masked language modeling loss.
        #
        # Indices should be in `[-100, 0, ..., config.vocab_size]` (see
        # `input_ids` docstring) Tokens with indices set to `-100` are ignored
        # (masked), the loss is only computed for the tokens with labels in `[0,
        # ..., config.vocab_size]`
        #
        # Prediction scores are only computed for masked tokens and the (bs,
        # seqlen) dimensions are flattened
        if (input_ids is not None) == (inputs_embeds is not None):
            raise ValueError('Must specify either input_ids or input_embeds!')

        return_dict = (
            return_dict if return_dict is not None else self.config.use_return_dict
        )

        outputs = self.bert(
            input_ids,
            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,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        sequence_output = outputs[0]
        prediction_scores = self.cls(sequence_output)

        loss = None
        if labels is not None:
            # Compute loss
            loss_fct = nn.CrossEntropyLoss()
            loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), labels.view(-1))

        if not return_dict:
            output = (prediction_scores,) + outputs[2:]
            return ((loss,) + output) if loss is not None else output

        return MaskedLMOutput(
            loss=loss,
            logits=prediction_scores,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )

    def prepare_inputs_for_generation(
        self, input_ids: torch.Tensor, attention_mask: torch.Tensor, **model_kwargs
    ):
        input_shape = input_ids.shape
        effective_batch_size = input_shape[0]

        #  add a dummy token
        if self.config.pad_token_id is None:
            raise ValueError('The PAD token should be defined for generation')

        attention_mask = torch.cat(
            [attention_mask, attention_mask.new_zeros((attention_mask.shape[0], 1))],
            dim=-1,
        )
        dummy_token = torch.full(
            (effective_batch_size, 1),
            self.config.pad_token_id,
            dtype=torch.long,
            device=input_ids.device,
        )
        input_ids = torch.cat([input_ids, dummy_token], dim=1)

        return {'input_ids': input_ids, 'attention_mask': attention_mask}



class JBertForSequenceClassification(BertPreTrainedModel):
    """Bert Model transformer with a sequence classification/regression head.

    This head is just a linear layer on top of the pooled output. Used for,
    e.g., GLUE tasks.
    """

    config_class = JBertConfig

    def __init__(self, config):
        super().__init__(config)
        self.num_labels = config.num_labels
        self.config = config

        self.bert = JBertModel(config)
        classifier_dropout = (
            config.classifier_dropout
            if config.classifier_dropout is not None
            else config.hidden_dropout_prob
        )
        self.dropout = nn.Dropout(classifier_dropout)
        self.classifier = nn.Linear(config.hidden_size, config.num_labels)

        # Initialize weights and apply final processing
        self.post_init()

    def forward(
        self,
        input_ids: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        token_type_ids: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.Tensor] = None,
        head_mask: Optional[torch.Tensor] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
        labels: Optional[torch.Tensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple[torch.Tensor], SequenceClassifierOutput]:
        # labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
        # Labels for computing the sequence classification/regression loss.
        # Indices should be in `[0, ..., config.num_labels - 1]`.
        # If `config.num_labels == 1` a regression loss is computed
        # (mean-square loss). If `config.num_labels > 1` a classification loss
        # is computed (cross-entropy).

        return_dict = (
            return_dict if return_dict is not None else self.config.use_return_dict
        )

        outputs = self.bert(
            input_ids,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        pooled_output = outputs[1]

        pooled_output = self.dropout(pooled_output)
        logits = self.classifier(pooled_output)

        loss = None
        if labels is not None:
            # Compute loss
            if self.config.problem_type is None:
                if self.num_labels == 1:
                    self.config.problem_type = 'regression'
                elif self.num_labels > 1 and (
                    labels.dtype == torch.long or labels.dtype == torch.int
                ):
                    self.config.problem_type = 'single_label_classification'
                else:
                    self.config.problem_type = 'multi_label_classification'

            if self.config.problem_type == 'regression':
                loss_fct = nn.MSELoss()
                if self.num_labels == 1:
                    loss = loss_fct(logits.squeeze(), labels.squeeze())
                else:
                    loss = loss_fct(logits, labels)
            elif self.config.problem_type == 'single_label_classification':
                loss_fct = nn.CrossEntropyLoss()
                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
            elif self.config.problem_type == 'multi_label_classification':
                loss_fct = nn.BCEWithLogitsLoss()
                loss = loss_fct(logits, labels)

        if not return_dict:
            output = (logits,) + outputs[2:]
            return ((loss,) + output) if loss is not None else output

        return SequenceClassifierOutput(
            loss=loss,
            logits=logits,
            hidden_states=None,
            attentions=None,
        )