modeling_thermoformer.py

import math
from typing import List, Optional, Tuple, Union
import torch.nn.functional as F
import torch
import torch.utils.checkpoint
from torch import nn
from torch.nn import CrossEntropyLoss
from dataclasses import dataclass
from transformers.modeling_outputs import (
    BaseModelOutputWithPastAndCrossAttentions,
    BaseModelOutputWithPoolingAndCrossAttentions,
    MaskedLMOutput,
    ModelOutput,
)
from transformers.modeling_utils import (
    PreTrainedModel,
    find_pruneable_heads_and_indices,
    prune_linear_layer,
)
from transformers.utils import logging
from .configuration_thermoformer import ThermoFormerConfig
from torch.nn.functional import scaled_dot_product_attention

logger = logging.get_logger(__name__)


def rotate_half(x):
    return torch.cat((-x[..., x.shape[-1] // 2 :], x[..., : x.shape[-1] // 2]), dim=-1)


def apply_rotary_pos_emb(x, cos, sin):
    cos = cos[:, :, : x.shape[-2], :]
    sin = sin[:, :, : x.shape[-2], :]
    return (x * cos) + (rotate_half(x) * sin)


def gelu(x):
    return x * 0.5 * (1.0 + torch.erf(x / math.sqrt(2.0)))


class RotaryEmbedding(torch.nn.Module):
    def __init__(self, dim: int):
        super().__init__()
        # Generate and save the inverse frequency buffer (non trainable)
        inv_freq = 1.0 / (
            10000 ** (torch.arange(0, dim, 2, dtype=torch.int64).float() / dim)
        )
        inv_freq = inv_freq
        self.register_buffer("inv_freq", inv_freq)

        self._seq_len_cached = None
        self._cos_cached = None
        self._sin_cached = None

    def _update_cos_sin_tables(self, x, seq_dimension=2):
        seq_len = x.shape[seq_dimension]

        # Reset the tables if the sequence length has changed,
        # or if we're on a new device (possibly due to tracing for instance)
        if seq_len != self._seq_len_cached or self._cos_cached.device != x.device:
            self._seq_len_cached = seq_len
            t = torch.arange(x.shape[seq_dimension], device=x.device).type_as(
                self.inv_freq
            )
            freqs = torch.outer(t, self.inv_freq)
            emb = torch.cat((freqs, freqs), dim=-1).to(x.device)

            self._cos_cached = emb.cos()[None, None, :, :]
            self._sin_cached = emb.sin()[None, None, :, :]

        return self._cos_cached, self._sin_cached

    def forward(
        self, q: torch.Tensor, k: torch.Tensor
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        self._cos_cached, self._sin_cached = self._update_cos_sin_tables(
            k, seq_dimension=-2
        )

        return (
            apply_rotary_pos_emb(q, self._cos_cached, self._sin_cached),
            apply_rotary_pos_emb(k, self._cos_cached, self._sin_cached),
        )


class ThermoFormerEmbeddings(nn.Module):

    def __init__(self, config):
        super().__init__()
        self.word_embeddings = nn.Embedding(
            config.vocab_size, config.hidden_size, padding_idx=config.pad_token_id
        )

        if config.emb_layer_norm_before:
            self.layer_norm = nn.LayerNorm(
                config.hidden_size, eps=config.layer_norm_eps
            )
        else:
            self.layer_norm = None
        self.dropout = nn.Dropout(config.hidden_dropout_prob)
        self.position_embedding_type = getattr(
            config, "position_embedding_type", "absolute"
        )
        self.register_buffer(
            "position_ids",
            torch.arange(config.max_position_embeddings).expand((1, -1)),
            persistent=False,
        )

        self.padding_idx = config.pad_token_id
        if self.position_embedding_type == "absolute":
            self.position_embeddings = nn.Embedding(
                config.max_position_embeddings,
                config.hidden_size,
                padding_idx=self.padding_idx,
            )
        self.token_dropout = config.token_dropout
        self.mask_token_id = config.mask_token_id

    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        position_ids=None,
        inputs_embeds=None,
        past_key_values_length=0,
    ):
        if position_ids is None:
            if input_ids is not None:
                position_ids = create_position_ids_from_input_ids(
                    input_ids, self.padding_idx, past_key_values_length
                )
            else:
                position_ids = self.create_position_ids_from_inputs_embeds(
                    inputs_embeds
                )

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

        embeddings = inputs_embeds

        if self.token_dropout:
            embeddings = embeddings.masked_fill(
                (input_ids == self.mask_token_id).unsqueeze(-1), 0.0
            )
            mask_ratio_train = 0.15 * 0.8
            src_lengths = attention_mask.sum(-1)
            mask_ratio_observed = (input_ids == self.mask_token_id).sum(
                -1
            ).float() / src_lengths
            embeddings = (
                embeddings
                * (1 - mask_ratio_train)
                / (1 - mask_ratio_observed)[:, None, None]
            ).to(embeddings.dtype)

        if self.position_embedding_type == "absolute":
            position_embeddings = self.position_embeddings(position_ids)
            embeddings = embeddings + position_embeddings

        if self.layer_norm is not None:
            embeddings = self.layer_norm(embeddings)
        if attention_mask is not None:
            embeddings = (embeddings * attention_mask.unsqueeze(-1)).to(
                embeddings.dtype
            )
        # Matt: I think this line was copied incorrectly from BERT, disabling it for now.
        # embeddings = self.dropout(embeddings)
        return embeddings

    def create_position_ids_from_inputs_embeds(self, inputs_embeds):
        input_shape = inputs_embeds.size()[:-1]
        sequence_length = input_shape[1]

        position_ids = torch.arange(
            self.padding_idx + 1,
            sequence_length + self.padding_idx + 1,
            dtype=torch.long,
            device=inputs_embeds.device,
        )
        return position_ids.unsqueeze(0).expand(input_shape)


class ThermoFormerSelfAttention(nn.Module):
    def __init__(self, config, position_embedding_type=None):
        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)
        self.all_head_size = self.num_attention_heads * self.attention_head_size

        self.query = nn.Linear(config.hidden_size, self.all_head_size)
        self.key = nn.Linear(config.hidden_size, self.all_head_size)
        self.value = nn.Linear(config.hidden_size, self.all_head_size)

        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
        self.position_embedding_type = position_embedding_type or getattr(
            config, "position_embedding_type", "absolute"
        )
        self.rotary_embeddings = None
        if (
            self.position_embedding_type == "relative_key"
            or self.position_embedding_type == "relative_key_query"
        ):
            self.max_position_embeddings = config.max_position_embeddings
            self.distance_embedding = nn.Embedding(
                2 * config.max_position_embeddings - 1, self.attention_head_size
            )
        elif self.position_embedding_type == "rotary":
            self.rotary_embeddings = RotaryEmbedding(dim=self.attention_head_size)
        self.flash_attention = config.flash_attention
        self.is_decoder = config.is_decoder
        self.config = config

    def transpose_for_scores(self, x: torch.Tensor) -> torch.Tensor:
        new_x_shape = x.size()[:-1] + (
            self.num_attention_heads,
            self.attention_head_size,
        )
        x = x.view(new_x_shape)
        return x.permute(0, 2, 1, 3)

    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_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
        output_attentions: Optional[bool] = False,
    ) -> Tuple[torch.Tensor]:
        mixed_query_layer = self.query(hidden_states)

        # If this is instantiated as a cross-attention module, the keys
        # and values come from an encoder; the attention mask needs to be
        # such that the encoder's padding tokens are not attended to.
        is_cross_attention = encoder_hidden_states is not None

        if is_cross_attention and past_key_value is not None:
            # reuse k,v, cross_attentions
            key_layer = past_key_value[0]
            value_layer = past_key_value[1]
            attention_mask = encoder_attention_mask
        elif is_cross_attention:
            key_layer = self.transpose_for_scores(self.key(encoder_hidden_states))
            value_layer = self.transpose_for_scores(self.value(encoder_hidden_states))
            attention_mask = encoder_attention_mask
        elif past_key_value is not None:
            key_layer = self.transpose_for_scores(self.key(hidden_states))
            value_layer = self.transpose_for_scores(self.value(hidden_states))
            key_layer = torch.cat([past_key_value[0], key_layer], dim=2)
            value_layer = torch.cat([past_key_value[1], value_layer], dim=2)
        else:
            key_layer = self.transpose_for_scores(self.key(hidden_states))
            value_layer = self.transpose_for_scores(self.value(hidden_states))

        query_layer = self.transpose_for_scores(mixed_query_layer)

        query_layer = query_layer * self.attention_head_size**-0.5

        if self.is_decoder:
            past_key_value = (key_layer, value_layer)

        if self.position_embedding_type == "rotary":
            query_layer, key_layer = self.rotary_embeddings(query_layer, key_layer)

        if not self.flash_attention:
            # Take the dot product between "query" and "key" to get the raw attention scores.
            attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))

            if (
                self.position_embedding_type == "relative_key"
                or self.position_embedding_type == "relative_key_query"
            ):
                seq_length = hidden_states.size()[1]
                position_ids_l = torch.arange(
                    seq_length, dtype=torch.long, device=hidden_states.device
                ).view(-1, 1)
                position_ids_r = torch.arange(
                    seq_length, dtype=torch.long, device=hidden_states.device
                ).view(1, -1)
                distance = position_ids_l - position_ids_r
                positional_embedding = self.distance_embedding(
                    distance + self.max_position_embeddings - 1
                )
                positional_embedding = positional_embedding.to(
                    dtype=query_layer.dtype
                )  # fp16 compatibility

                if self.position_embedding_type == "relative_key":
                    relative_position_scores = torch.einsum(
                        "bhld,lrd->bhlr", query_layer, positional_embedding
                    )
                    attention_scores = attention_scores + relative_position_scores
                elif self.position_embedding_type == "relative_key_query":
                    relative_position_scores_query = torch.einsum(
                        "bhld,lrd->bhlr", query_layer, positional_embedding
                    )
                    relative_position_scores_key = torch.einsum(
                        "bhrd,lrd->bhlr", key_layer, positional_embedding
                    )
                    attention_scores = (
                        attention_scores
                        + relative_position_scores_query
                        + relative_position_scores_key
                    )

            if attention_mask is not None:
                attention_scores = attention_scores + attention_mask

            # Normalize the attention scores to probabilities.
            attention_probs = nn.functional.softmax(attention_scores, dim=-1)

            # This is actually dropping out entire tokens to attend to, which might
            # seem a bit unusual, but is taken from the original Transformer paper.
            attention_probs = self.dropout(attention_probs)

            # Mask heads if we want to
            if head_mask is not None:
                attention_probs = attention_probs * head_mask

            context_layer = torch.matmul(attention_probs, value_layer)
        else:
            if self.training:
                context_layer = scaled_dot_product_attention(
                    query_layer,
                    key_layer,
                    value_layer,
                    attn_mask=attention_mask,
                    dropout_p=self.config.attention_probs_dropout_prob,
                    scale=1,  # we have query_layer = query_layer * self.attention_head_size**-0.5
                )
            else:
                context_layer = scaled_dot_product_attention(
                    query_layer,
                    key_layer,
                    value_layer,
                    attn_mask=attention_mask,
                    scale=1,  # we have query_layer = query_layer * self.attention_head_size**-0.5
                )

        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
        context_layer = context_layer.view(new_context_layer_shape)

        outputs = (
            (context_layer, attention_probs) if output_attentions else (context_layer,)
        )

        if self.is_decoder:
            outputs = outputs + (past_key_value,)
        return outputs


class ThermoFormerSelfOutput(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)

    def forward(self, hidden_states, input_tensor):
        hidden_states = self.dense(hidden_states)
        hidden_states = self.dropout(hidden_states)
        hidden_states = hidden_states + input_tensor
        return hidden_states


class ThermoFormerAttention(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.self = ThermoFormerSelfAttention(config)
        self.output = ThermoFormerSelfOutput(config)
        self.pruned_heads = set()
        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)

    def prune_heads(self, heads):
        if len(heads) == 0:
            return
        heads, index = find_pruneable_heads_and_indices(
            heads,
            self.self.num_attention_heads,
            self.self.attention_head_size,
            self.pruned_heads,
        )

        # Prune linear layers
        self.self.query = prune_linear_layer(self.self.query, index)
        self.self.key = prune_linear_layer(self.self.key, index)
        self.self.value = prune_linear_layer(self.self.value, index)
        self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)

        # Update hyper params and store pruned heads
        self.self.num_attention_heads = self.self.num_attention_heads - len(heads)
        self.self.all_head_size = (
            self.self.attention_head_size * self.self.num_attention_heads
        )
        self.pruned_heads = self.pruned_heads.union(heads)

    def forward(
        self,
        hidden_states,
        attention_mask=None,
        head_mask=None,
        encoder_hidden_states=None,
        encoder_attention_mask=None,
        past_key_value=None,
        output_attentions=False,
    ):
        hidden_states_ln = self.LayerNorm(hidden_states)
        self_outputs = self.self(
            hidden_states_ln,
            attention_mask,
            head_mask,
            encoder_hidden_states,
            encoder_attention_mask,
            past_key_value,
            output_attentions,
        )
        attention_output = self.output(self_outputs[0], hidden_states)
        outputs = (attention_output,) + self_outputs[
            1:
        ]  # add attentions if we output them
        return outputs


class ThermoFormerIntermediate(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.dense = nn.Linear(config.hidden_size, config.intermediate_size)

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


class ThermoFormerOutput(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)

    def forward(self, hidden_states, input_tensor):
        hidden_states = self.dense(hidden_states)
        hidden_states = self.dropout(hidden_states)
        hidden_states = hidden_states + input_tensor
        return hidden_states


class ThermoFormerLayer(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.chunk_size_feed_forward = config.chunk_size_feed_forward
        self.seq_len_dim = 1
        self.attention = ThermoFormerAttention(config)
        self.is_decoder = config.is_decoder
        self.add_cross_attention = config.add_cross_attention
        if self.add_cross_attention:
            if not self.is_decoder:
                raise RuntimeError(
                    f"{self} should be used as a decoder model if cross attention is added"
                )
            self.crossattention = ThermoFormerAttention(config)
        self.intermediate = ThermoFormerIntermediate(config)
        self.output = ThermoFormerOutput(config)
        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)

    def forward(
        self,
        hidden_states,
        attention_mask=None,
        head_mask=None,
        encoder_hidden_states=None,
        encoder_attention_mask=None,
        past_key_value=None,
        output_attentions=False,
    ):
        # decoder uni-directional self-attention cached key/values tuple is at positions 1,2
        self_attn_past_key_value = (
            past_key_value[:2] if past_key_value is not None else None
        )
        self_attention_outputs = self.attention(
            hidden_states,
            attention_mask,
            head_mask,
            output_attentions=output_attentions,
            past_key_value=self_attn_past_key_value,
        )
        attention_output = self_attention_outputs[0]

        # if decoder, the last output is tuple of self-attn cache
        if self.is_decoder:
            outputs = self_attention_outputs[1:-1]
            present_key_value = self_attention_outputs[-1]
        else:
            outputs = self_attention_outputs[
                1:
            ]  # add self attentions if we output attention weights

        cross_attn_present_key_value = None
        if self.is_decoder and encoder_hidden_states is not None:
            if not hasattr(self, "crossattention"):
                raise AttributeError(
                    f"If `encoder_hidden_states` are passed, {self} has to be instantiated"
                    " with cross-attention layers by setting `config.add_cross_attention=True`"
                )

            # cross_attn cached key/values tuple is at positions 3,4 of past_key_value tuple
            cross_attn_past_key_value = (
                past_key_value[-2:] if past_key_value is not None else None
            )
            cross_attention_outputs = self.crossattention(
                attention_output,
                attention_mask,
                head_mask,
                encoder_hidden_states,
                encoder_attention_mask,
                cross_attn_past_key_value,
                output_attentions,
            )
            attention_output = cross_attention_outputs[0]
            outputs = (
                outputs + cross_attention_outputs[1:-1]
            )  # add cross attentions if we output attention weights

            # add cross-attn cache to positions 3,4 of present_key_value tuple
            cross_attn_present_key_value = cross_attention_outputs[-1]
            present_key_value = present_key_value + cross_attn_present_key_value

        layer_output = self.feed_forward_chunk(attention_output)

        outputs = (layer_output,) + outputs

        # if decoder, return the attn key/values as the last output
        if self.is_decoder:
            outputs = outputs + (present_key_value,)
        return outputs

    def feed_forward_chunk(self, attention_output):
        attention_output_ln = self.LayerNorm(attention_output)
        intermediate_output = self.intermediate(attention_output_ln)
        layer_output = self.output(intermediate_output, attention_output)
        return layer_output


class ThermoFormerEncoder(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.layer = nn.ModuleList(
            [ThermoFormerLayer(config) for _ in range(config.num_hidden_layers)]
        )
        self.emb_layer_norm_after = nn.LayerNorm(
            config.hidden_size, eps=config.layer_norm_eps
        )
        self.gradient_checkpointing = True

    def forward(
        self,
        hidden_states,
        attention_mask=None,
        head_mask=None,
        encoder_hidden_states=None,
        encoder_attention_mask=None,
        past_key_values=None,
        use_cache=None,
        output_attentions=False,
        output_hidden_states=False,
        return_dict=True,
    ):
        if self.gradient_checkpointing and self.training:
            if use_cache:
                logger.warning_once(
                    "`use_cache=True` is incompatible with `config.gradient_checkpointing=True`. Setting "
                    "`use_cache=False`..."
                )
                use_cache = False
        all_hidden_states = () if output_hidden_states else None
        all_self_attentions = () if output_attentions else None
        all_cross_attentions = (
            () if output_attentions and self.config.add_cross_attention else None
        )

        next_decoder_cache = () if use_cache else None
        for i, layer_module in enumerate(self.layer):
            if output_hidden_states:
                all_hidden_states = all_hidden_states + (hidden_states,)

            layer_head_mask = head_mask[i] if head_mask is not None else None
            past_key_value = past_key_values[i] if past_key_values is not None else None

            if self.gradient_checkpointing and self.training:
                layer_outputs = self._gradient_checkpointing_func(
                    layer_module.__call__,
                    hidden_states,
                    attention_mask,
                    layer_head_mask,
                    encoder_hidden_states,
                    encoder_attention_mask,
                    past_key_value,
                    output_attentions,
                )
            else:
                layer_outputs = layer_module(
                    hidden_states,
                    attention_mask,
                    layer_head_mask,
                    encoder_hidden_states,
                    encoder_attention_mask,
                    past_key_value,
                    output_attentions,
                )

            hidden_states = layer_outputs[0]
            if use_cache:
                next_decoder_cache = next_decoder_cache + (layer_outputs[-1],)
            if output_attentions:
                all_self_attentions = all_self_attentions + (layer_outputs[1],)
                if self.config.add_cross_attention:
                    all_cross_attentions = all_cross_attentions + (layer_outputs[2],)

        if self.emb_layer_norm_after:
            hidden_states = self.emb_layer_norm_after(hidden_states)

        if output_hidden_states:
            all_hidden_states = all_hidden_states + (hidden_states,)

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


class ThermoFormerPreTrainedModel(PreTrainedModel):
    """
    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
    models.
    """

    config_class = ThermoFormerConfig
    base_model_prefix = "ThermoFormer"
    supports_gradient_checkpointing = True
    _no_split_modules = [
        "ThermoFormerLayer",
        "ThermoFormerEmbeddings",
    ]

    # Copied from transformers.models.bert.modeling_bert.BertPreTrainedModel._init_weights
    def _init_weights(self, module):
        """Initialize the weights"""
        if isinstance(module, nn.Linear):
            # 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)


class ThermoFormerModel(ThermoFormerPreTrainedModel):
    base_model_prefix = "ThermoFormer"

    def __init__(self, config, add_pooling_layer=True):
        super().__init__(config)
        self.config = config
        self.embeddings = ThermoFormerEmbeddings(config)
        self.encoder = ThermoFormerEncoder(config)
        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 _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} See base
        class PreTrainedModel
        """
        for layer, heads in heads_to_prune.items():
            self.encoder.layer[layer].attention.prune_heads(heads)

    def forward(
        self,
        input_ids: Optional[torch.Tensor] = None,
        attention_mask: 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,
        past_key_values: Optional[List[torch.FloatTensor]] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple[torch.Tensor], BaseModelOutputWithPoolingAndCrossAttentions]:
        output_attentions = (
            output_attentions
            if output_attentions is not None
            else self.config.output_attentions
        )
        output_hidden_states = (
            output_hidden_states
            if output_hidden_states is not None
            else self.config.output_hidden_states
        )
        return_dict = (
            return_dict if return_dict is not None else self.config.use_return_dict
        )

        if self.config.is_decoder:
            use_cache = use_cache if use_cache is not None else self.config.use_cache
        else:
            use_cache = False

        if input_ids is not None and inputs_embeds is not None:
            raise ValueError(
                "You cannot specify both input_ids and inputs_embeds at the same time"
            )
        elif input_ids is not None:
            self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask)
            input_shape = input_ids.size()
        elif inputs_embeds is not None:
            input_shape = inputs_embeds.size()[:-1]
        else:
            raise ValueError("You have to specify either input_ids or inputs_embeds")

        batch_size, seq_length = input_shape
        device = input_ids.device if input_ids is not None else inputs_embeds.device

        # past_key_values_length
        past_key_values_length = (
            past_key_values[0][0].shape[2] if past_key_values is not None else 0
        )

        if attention_mask is None:
            attention_mask = torch.ones(
                ((batch_size, seq_length + past_key_values_length)), device=device
            )

        extended_attention_mask: torch.Tensor = self.get_extended_attention_mask(
            attention_mask, input_shape
        )

        if self.config.is_decoder 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_extended_attention_mask = self.invert_attention_mask(
                encoder_attention_mask
            )
        else:
            encoder_extended_attention_mask = None

        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)

        embedding_output = self.embeddings(
            input_ids=input_ids,
            position_ids=position_ids,
            attention_mask=attention_mask,
            inputs_embeds=inputs_embeds,
            past_key_values_length=past_key_values_length,
        )
        encoder_outputs = self.encoder(
            embedding_output,
            attention_mask=extended_attention_mask,
            head_mask=head_mask,
            encoder_hidden_states=encoder_hidden_states,
            encoder_attention_mask=encoder_extended_attention_mask,
            past_key_values=past_key_values,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )
        sequence_output = encoder_outputs[0]

        return BaseModelOutputWithPoolingAndCrossAttentions(
            last_hidden_state=sequence_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,
        )


class ThermoFormerForMaskedLM(ThermoFormerPreTrainedModel):
    _tied_weights_keys = ["lm_head.decoder.weight"]

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

        if config.is_decoder:
            logger.warning(
                "If you want to use `ThermoFormerForMaskedLM` make sure `config.is_decoder=False` for "
                "bi-directional self-attention."
            )

        self.model = ThermoFormerModel(config, add_pooling_layer=False)
        self.lm_head = ThermoFormerLMHead(config)
        self.init_weights()

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

    def get_output_embeddings(self):
        return self.lm_head.decoder

    def set_output_embeddings(self, new_embeddings):
        self.lm_head.decoder = new_embeddings

    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        head_mask: Optional[torch.Tensor] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        encoder_attention_mask: Optional[torch.Tensor] = None,
        labels: Optional[torch.LongTensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple, MaskedLMOutput]:
        return_dict = (
            return_dict if return_dict is not None else self.config.use_return_dict
        )

        outputs = self.model(
            input_ids,
            attention_mask=attention_mask,
            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.lm_head(sequence_output)

        masked_lm_loss = None
        if labels is not None:
            loss_fct = CrossEntropyLoss()

            labels = labels.to(prediction_scores.device)
            masked_lm_loss = loss_fct(
                prediction_scores.view(-1, self.config.vocab_size), labels.view(-1)
            )

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

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


class ThermoFormerLMHead(nn.Module):

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

        self.decoder = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
        self.bias = nn.Parameter(torch.zeros(config.vocab_size))

    def forward(self, features, **kwargs):
        x = self.dense(features)
        x = gelu(x)
        x = self.layer_norm(x)

        # project back to size of vocabulary with bias
        x = self.decoder(x) + self.bias
        return x


class ThermoFormerStructureHead(nn.Module):

    def __init__(self, config):
        super().__init__()
        self.dense = nn.Linear(config.hidden_size, config.hidden_size, bias=False)
        self.layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.decoder = nn.Linear(
            config.hidden_size, config.structure_vocab_size, bias=False
        )
        self.bias = nn.Parameter(torch.zeros(config.structure_vocab_size))

    def forward(self, features, **kwargs):
        x = self.dense(features)
        x = gelu(x)
        x = self.layer_norm(x)

        # project back to size of vocabulary with bias
        x = self.decoder(x) + self.bias
        return x


def create_position_ids_from_input_ids(
    input_ids, padding_idx, past_key_values_length=0
):
    """
    Replace non-padding symbols with their position numbers. Position numbers begin at padding_idx+1. Padding symbols
    are ignored. This is modified from fairseq's `utils.make_positions`.

    Args:
        x: torch.Tensor x:

    Returns: torch.Tensor
    """
    # The series of casts and type-conversions here are carefully balanced to both work with ONNX export and XLA.
    mask = input_ids.ne(padding_idx).int()
    incremental_indices = (
        torch.cumsum(mask, dim=1).type_as(mask) + past_key_values_length
    ) * mask
    return incremental_indices.long() + padding_idx


# POOLING_HEAD
class MaskedConv1d(nn.Conv1d):
    """A masked 1-dimensional convolution layer.

    Takes the same arguments as torch.nn.Conv1D, except that the padding is set automatically.

         Shape:
            Input: (N, L, in_channels)
            input_mask: (N, L, 1), optional
            Output: (N, L, out_channels)
    """

    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        kernel_size: int,
        stride: int = 1,
        dilation: int = 1,
        groups: int = 1,
        bias: bool = True,
    ):
        """
        :param in_channels: input channels
        :param out_channels: output channels
        :param kernel_size: the kernel width
        :param stride: filter shift
        :param dilation: dilation factor
        :param groups: perform depth-wise convolutions
        :param bias: adds learnable bias to output
        """
        padding = dilation * (kernel_size - 1) // 2
        super().__init__(
            in_channels,
            out_channels,
            kernel_size,
            stride=stride,
            dilation=dilation,
            groups=groups,
            bias=bias,
            padding=padding,
        )

    def forward(self, x, input_mask=None):
        if input_mask is not None:
            x = x * input_mask
        return super().forward(x.transpose(1, 2)).transpose(1, 2)


class Attention1d(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.layer = MaskedConv1d(config.hidden_size, 1, 1)
        self.out = nn.Linear(config.hidden_size, config.hidden_size)

    def forward(self, x, input_mask=None, return_weights=False):
        batch_szie = x.shape[0]
        attn = self.layer(x)
        attn = attn.view(batch_szie, -1)
        if input_mask is not None:
            attn = attn.masked_fill_(
                ~input_mask.view(batch_szie, -1).bool(), float("-inf")
            )
        attn = F.softmax(attn, dim=-1).view(batch_szie, -1, 1)
        out = (attn * x).sum(dim=1)
        out = self.out(out)
        if return_weights:
            return out, attn
        return out


class FFN1d(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
        self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)
        self.act = nn.GELU()

    def forward(self, x):
        x = self.fc1(x)
        x = self.act(x)
        x = self.fc2(x)
        return x


class Attention1dPooling(nn.Module):
    """Outputs of the model with the attention1d"""

    def __init__(self, config):
        super(Attention1dPooling, self).__init__()
        self.attention1d = Attention1d(config)
        self.ffn = FFN1d(config)
        self.dropout1 = nn.Dropout(config.hidden_dropout_prob)
        self.dropout2 = nn.Dropout(config.hidden_dropout_prob)

    def forward(self, x, input_mask, return_weights=False):
        if return_weights:
            attn_out, weights = self.attention1d(
                x, input_mask=input_mask.unsqueeze(-1), return_weights=return_weights
            )
        else:
            attn_out = self.attention1d(x, input_mask=input_mask.unsqueeze(-1))
        x = self.dropout1(attn_out)
        ffn_out = self.ffn(x)
        x = x + self.dropout2(ffn_out)
        if return_weights:
            return x, weights
        return x


@dataclass
class MaskedLMOutput(ModelOutput):
    loss: Optional[torch.FloatTensor] = None
    mlm_loss: Optional[torch.FloatTensor] = None
    value_loss: Optional[torch.FloatTensor] = None
    predicted_values: Optional[torch.FloatTensor] = None
    logits: torch.FloatTensor = None
    sequence_hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
    hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
    attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
    pooling_attentions: Optional[Tuple[torch.FloatTensor, ...]] = None


class ThermoFormer(ThermoFormerPreTrainedModel):
    _tied_weights_keys = ["lm_head.decoder.weight"]

    def __init__(self, config):
        super().__init__(config)
        self.model = ThermoFormerModel(config, add_pooling_layer=False)
        self.mlm = config.mlm
        if self.mlm:
            self.lm_head = ThermoFormerLMHead(config)
        else:
            self.lm_head = None
        self.sequence_pooling = Attention1dPooling(config)
        self.value_projection = nn.Sequential(
            nn.Linear(config.hidden_size, config.hidden_size),
            nn.Tanh(),
            nn.Linear(config.hidden_size, 1),
        )
        self.init_weights()

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

    def get_output_embeddings(self):
        return self.lm_head.decoder

    def set_output_embeddings(self, new_embeddings):
        self.lm_head.decoder = new_embeddings

    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        head_mask: Optional[torch.Tensor] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        encoder_attention_mask: Optional[torch.Tensor] = None,
        labels: Optional[torch.LongTensor] = None,
        values: Optional[torch.FloatTensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple, MaskedLMOutput]:
        return_dict = (
            return_dict if return_dict is not None else self.config.use_return_dict
        )

        outputs = self.model(
            input_ids,
            attention_mask=attention_mask,
            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]

        # Masked LM
        if labels is not None:
            assert self.lm_head is not None
            lm_prediction_scores = self.lm_head(sequence_output)
            loss_fct = CrossEntropyLoss()
            labels = labels.to(lm_prediction_scores.device)
            masked_lm_loss = loss_fct(
                lm_prediction_scores.view(-1, self.config.vocab_size), labels.view(-1)
            )
        else:
            lm_prediction_scores = None
            masked_lm_loss = None

        # Value prediction
        if values is not None:
            sequence_states, weights = self.sequence_pooling(
                sequence_output, attention_mask, return_weights=True
            )
            predicted_values = self.value_projection(sequence_states)
            values = values.to(predicted_values.dtype)
            values = values.reshape(-1, 1)
            value_loss = nn.MSELoss()(predicted_values, values)
        else:
            sequence_states, weights = self.sequence_pooling(
                sequence_output, attention_mask, return_weights=True
            )
            predicted_values = self.value_projection(sequence_states)
            value_loss = None

        if masked_lm_loss is not None and value_loss is not None:
            loss = masked_lm_loss + 0.01 * value_loss
        elif masked_lm_loss is not None and value_loss is None:
            loss = masked_lm_loss
        elif masked_lm_loss is None and value_loss is not None:
            loss = 0.01 * value_loss
        else:
            loss = None

        return MaskedLMOutput(
            loss=loss,
            mlm_loss=masked_lm_loss,
            value_loss=value_loss,
            logits=lm_prediction_scores,
            predicted_values=predicted_values.reshape(-1),
            hidden_states=outputs.hidden_states,
            sequence_hidden_states=sequence_states,
            attentions=outputs.attentions,
            pooling_attentions=weights,
        )


ThermoFormer.register_for_auto_class("AutoModel")