image-match/python/py/Lib/site-packages/transformers/models/esm/modeling_esm.py

# Copyright 2022 Meta and The HuggingFace Inc. team. All rights reserved.
# Copyright (c) 2025, 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 ESM model."""

import math
from collections.abc import Callable

import torch
from torch import nn
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss

from ... import initialization as init
from ...masking_utils import create_bidirectional_mask, create_causal_mask
from ...modeling_layers import GradientCheckpointingLayer
from ...modeling_outputs import (
    BaseModelOutputWithCrossAttentions,
    BaseModelOutputWithPoolingAndCrossAttentions,
    MaskedLMOutput,
    SequenceClassifierOutput,
    TokenClassifierOutput,
)
from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
from ...processing_utils import Unpack
from ...utils import TransformersKwargs, auto_docstring, can_return_tuple, logging
from ...utils.generic import merge_with_config_defaults
from ...utils.output_capturing import OutputRecorder, capture_outputs
from .configuration_esm import EsmConfig


logger = logging.get_logger(__name__)


def rotate_half(x):
    x1, x2 = x.chunk(2, dim=-1)
    return torch.cat((-x2, x1), 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):
    """
    This is the gelu implementation from the original ESM repo. Using F.gelu yields subtly wrong results.
    """
    return x * 0.5 * (1.0 + torch.erf(x / math.sqrt(2.0)))


def symmetrize(x):
    "Make layer symmetric in final two dimensions, used for contact prediction."
    return x + x.transpose(-1, -2)


def average_product_correct(x):
    "Perform average product correct, used for contact prediction."
    a1 = x.sum(-1, keepdims=True)
    a2 = x.sum(-2, keepdims=True)
    a12 = x.sum((-1, -2), keepdims=True)

    avg = a1 * a2
    avg.div_(a12)  # in-place to reduce memory
    normalized = x - avg
    return normalized


class RotaryEmbedding(torch.nn.Module):
    """
    Rotary position embeddings based on those in
    [RoFormer](https://huggingface.co/docs/transformers/model_doc/roformer). Query and keys are transformed by rotation
    matrices which depend on their relative positions.
    """

    inv_freq: torch.Tensor  # fix linting for `register_buffer`

    def __init__(self, dim: int):
        super().__init__()
        self.dim = dim
        # Generate and save the inverse frequency buffer (non trainable)
        inv_freq = 1.0 / (10000 ** (torch.arange(0, dim, 2, dtype=torch.int64).float() / dim))
        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).to(dtype=q.dtype),
            apply_rotary_pos_emb(k, self._cos_cached, self._sin_cached).to(dtype=k.dtype),
        )


class EsmContactPredictionHead(nn.Module):
    """Performs symmetrization, apc, and computes a logistic regression on the output features"""

    def __init__(
        self,
        in_features: int,
        bias=True,
        eos_idx: int = 2,
    ):
        super().__init__()
        self.in_features = in_features
        self.eos_idx = eos_idx
        self.regression = nn.Linear(in_features, 1, bias)
        self.activation = nn.Sigmoid()

    def forward(self, tokens, attentions):
        # remove eos token attentions
        eos_mask = tokens.ne(self.eos_idx).to(attentions)
        eos_mask = eos_mask.unsqueeze(1) * eos_mask.unsqueeze(2)
        attentions = attentions * eos_mask[:, None, None, :, :]
        attentions = attentions[..., :-1, :-1]
        # remove cls token attentions
        attentions = attentions[..., 1:, 1:]
        batch_size, layers, heads, seqlen, _ = attentions.size()
        attentions = attentions.view(batch_size, layers * heads, seqlen, seqlen)

        # features: batch x channels x tokens x tokens (symmetric)
        attentions = attentions.to(
            self.regression.weight.device
        )  # attentions always float32, may need to convert to float16
        attentions = average_product_correct(symmetrize(attentions))
        attentions = attentions.permute(0, 2, 3, 1)
        return self.activation(self.regression(attentions).squeeze(3))


class EsmEmbeddings(nn.Module):
    """
    Same as BertEmbeddings with a tiny tweak for positional embeddings indexing.
    """

    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)
        # position_ids (1, len position emb) is contiguous in memory and exported when serialized
        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,
    ):
        if position_ids is None:
            if input_ids is not None:
                # Create the position ids from the input token ids. Any padded tokens remain padded.
                position_ids = create_position_ids_from_input_ids(input_ids, self.padding_idx)
            else:
                position_ids = self.create_position_ids_from_inputs_embeds(inputs_embeds)

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

        # Note that if we want to support ESM-1 (not 1b!) in future then we need to support an
        # embedding_scale factor here.
        embeddings = inputs_embeds

        # Matt: ESM has the option to handle masking in MLM in a slightly unusual way. If the token_dropout
        # flag is False then it is handled in the same was as BERT/RoBERTa. If it is set to True, however,
        # masked tokens are treated as if they were selected for input dropout and zeroed out.
        # This "mask-dropout" is compensated for when masked tokens are not present, by scaling embeddings by
        # a factor of (fraction of unmasked tokens during training) / (fraction of unmasked tokens in sample).
        # This is analogous to the way that dropout layers scale down outputs during evaluation when not
        # actually dropping out values (or, equivalently, scale up their un-dropped outputs in training).
        if self.token_dropout and input_ids is not None:
            embeddings = embeddings.masked_fill((input_ids == self.mask_token_id).unsqueeze(-1), 0.0)
            mask_ratio_train = 0.15 * 0.8  # Hardcoded as the ratio used in all ESM model training runs
            src_lengths = attention_mask.sum(-1) if attention_mask is not None else input_ids.shape[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):
        """
        We are provided embeddings directly. We cannot infer which are padded so just generate sequential position ids.

        Args:
            inputs_embeds: torch.Tensor

        Returns: torch.Tensor
        """
        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)


# Copied from transformers.models.bert.modeling_bert.eager_attention_forward
def eager_attention_forward(
    module: nn.Module,
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    attention_mask: torch.Tensor | None,
    scaling: float | None = None,
    dropout: float = 0.0,
    **kwargs: Unpack[TransformersKwargs],
):
    if scaling is None:
        scaling = query.size(-1) ** -0.5

    # Take the dot product between "query" and "key" to get the raw attention scores.
    attn_weights = torch.matmul(query, key.transpose(2, 3)) * scaling

    if attention_mask is not None:
        attn_weights = attn_weights + attention_mask

    attn_weights = nn.functional.softmax(attn_weights, dim=-1)
    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)

    attn_output = torch.matmul(attn_weights, value)
    attn_output = attn_output.transpose(1, 2).contiguous()

    return attn_output, attn_weights


class EsmSelfAttention(nn.Module):
    def __init__(self, config, position_embedding_type=None, layer_idx=None, is_cross_attention=False):
        super().__init__()
        self.config = config

        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 = config.attention_probs_dropout_prob

        self.rotary_embeddings = None
        self.position_embedding_type = position_embedding_type or getattr(
            config, "position_embedding_type", "absolute"
        )
        if self.position_embedding_type == "rotary":
            self.rotary_embeddings = RotaryEmbedding(dim=self.attention_head_size)

        self.scaling = 1.0  # For BC we apply scaling before RoPE
        self.is_decoder = config.is_decoder
        self.layer_idx = layer_idx
        self.is_causal = self.is_decoder and not is_cross_attention

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.FloatTensor | None = None,
        encoder_hidden_states: torch.FloatTensor | None = None,
        encoder_attention_mask: torch.FloatTensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple[torch.Tensor]:
        input_shape = hidden_states.shape[:-1]
        hidden_shape = (*input_shape, -1, self.attention_head_size)

        query_layer = self.query(hidden_states).view(hidden_shape).transpose(1, 2)

        is_cross_attention = encoder_hidden_states is not None
        current_states = encoder_hidden_states if is_cross_attention else hidden_states
        attention_mask = encoder_attention_mask if is_cross_attention else attention_mask
        key_layer = self.key(current_states).view(hidden_shape).transpose(1, 2)
        value_layer = self.value(current_states).view(hidden_shape).transpose(1, 2)

        # Matt: Our BERT model (which this code was derived from) scales attention logits down by sqrt(head_dim).
        # ESM scales the query down by the same factor instead. Modulo numerical stability these are equivalent,
        # but not when rotary embeddings get involved. Therefore, we scale the query here to match the original
        # ESM code and fix rotary embeddings.
        query_layer = query_layer * self.attention_head_size**-0.5

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

        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
            self.config._attn_implementation, eager_attention_forward
        )

        attn_output, attn_weights = attention_interface(
            self,
            query_layer,
            key_layer,
            value_layer,
            attention_mask,
            dropout=0.0 if not self.training else self.dropout,
            scaling=self.scaling,
            **kwargs,
        )

        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
        return attn_output, attn_weights


class EsmSelfOutput(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 EsmAttention(nn.Module):
    def __init__(self, config, layer_idx=None, is_cross_attention=False):
        super().__init__()
        self.self = EsmSelfAttention(config, layer_idx=layer_idx, is_cross_attention=is_cross_attention)
        self.output = EsmSelfOutput(config)

        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)

    def forward(
        self,
        hidden_states,
        attention_mask=None,
        encoder_hidden_states=None,
        encoder_attention_mask=None,
        **kwargs: Unpack[TransformersKwargs],
    ):
        hidden_states_ln = self.LayerNorm(hidden_states)
        attn_output, _ = self.self(
            hidden_states_ln,
            attention_mask=attention_mask,
            encoder_hidden_states=encoder_hidden_states,
            encoder_attention_mask=encoder_attention_mask,
            **kwargs,
        )
        attn_output = self.output(attn_output, hidden_states)
        return attn_output


class EsmIntermediate(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 EsmOutput(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 EsmLayer(GradientCheckpointingLayer):
    def __init__(self, config):
        super().__init__()
        self.chunk_size_feed_forward = config.chunk_size_feed_forward
        self.seq_len_dim = 1
        self.attention = EsmAttention(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 = EsmAttention(config, is_cross_attention=True)
        self.intermediate = EsmIntermediate(config)
        self.output = EsmOutput(config)
        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)

    def forward(
        self,
        hidden_states,
        attention_mask=None,
        encoder_hidden_states=None,
        encoder_attention_mask=None,
        **kwargs: Unpack[TransformersKwargs],
    ):
        attention_output = self.attention(
            hidden_states,
            attention_mask=attention_mask,
            **kwargs,
        )

        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`"
                )

            attention_output = self.crossattention(
                attention_output,
                attention_mask=attention_mask,
                encoder_hidden_states=encoder_hidden_states,
                encoder_attention_mask=encoder_attention_mask,
                **kwargs,
            )

        layer_output = self.feed_forward_chunk(attention_output)
        return layer_output

    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 EsmEncoder(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.layer = nn.ModuleList([EsmLayer(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 = False

    @can_return_tuple
    def forward(
        self,
        hidden_states,
        attention_mask=None,
        encoder_hidden_states=None,
        encoder_attention_mask=None,
        **kwargs: Unpack[TransformersKwargs],
    ):
        for i, layer_module in enumerate(self.layer):
            hidden_states = layer_module(
                hidden_states,
                attention_mask=attention_mask,
                encoder_hidden_states=encoder_hidden_states,
                encoder_attention_mask=encoder_attention_mask,
                **kwargs,
            )

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

        return BaseModelOutputWithCrossAttentions(last_hidden_state=hidden_states)


# Copied from transformers.models.bert.modeling_bert.BertPooler
class EsmPooler(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) -> torch.Tensor:
        # We "pool" the model by simply taking the hidden state corresponding
        # to the first token.
        first_token_tensor = hidden_states[:, 0]
        pooled_output = self.dense(first_token_tensor)
        pooled_output = self.activation(pooled_output)
        return pooled_output


@auto_docstring
class EsmPreTrainedModel(PreTrainedModel):
    config: EsmConfig
    base_model_prefix = "esm"
    supports_gradient_checkpointing = True
    accepts_loss_kwargs = False
    _no_split_modules = ["EsmLayer", "EsmFoldTriangularSelfAttentionBlock", "EsmEmbeddings"]
    _keys_to_ignore_on_load_unexpected = ["position_embeddings.weight"]
    _supports_flash_attn = True
    _supports_sdpa = True
    _supports_flex_attn = True
    _supports_attention_backend = True

    _can_record_outputs = {
        "hidden_states": EsmLayer,
        "attentions": [OutputRecorder(EsmSelfAttention, index=1, layer_name="attention")],
        "cross_attentions": [
            OutputRecorder(EsmSelfAttention, index=1, layer_name="crossattention"),
        ],
    }

    @torch.no_grad()
    def _init_weights(self, module):
        """Initialize the weights"""
        super()._init_weights(module)
        if isinstance(module, EsmLMHead):
            init.zeros_(module.bias)
        elif isinstance(module, EsmEmbeddings):
            init.copy_(module.position_ids, torch.arange(module.position_ids.shape[-1]).expand((1, -1)))
        elif isinstance(module, RotaryEmbedding):
            inv_freq = 1.0 / (10000 ** (torch.arange(0, module.dim, 2, dtype=torch.int64).float() / module.dim))
            init.copy_(module.inv_freq, inv_freq)

    def get_output_embeddings(self):
        # NOTE: get_output_embeddings() must return None to prevent accidental weight tying.
        # See e.g. https://github.com/huggingface/transformers/pull/39339#discussion_r2219126400
        return None


@auto_docstring
class EsmModel(EsmPreTrainedModel):
    """

    The model can behave as an encoder (with only self-attention) as well as a decoder, in which case a layer of
    cross-attention is added between the self-attention layers, following the architecture described in [Attention is
    all you need](https://huggingface.co/papers/1706.03762) by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit,
    Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin.

    To behave as an decoder the model needs to be initialized with the `is_decoder` argument of the configuration set
    to `True`. To be used in a Seq2Seq model, the model needs to initialized with both `is_decoder` argument and
    `add_cross_attention` set to `True`; an `encoder_hidden_states` is then expected as an input to the forward pass.
    """

    def __init__(self, config, add_pooling_layer=True):
        r"""
        add_pooling_layer (bool, *optional*, defaults to `True`):
            Whether to add a pooling layer
        """
        super().__init__(config)
        self.config = config

        self.embeddings = EsmEmbeddings(config)
        self.encoder = EsmEncoder(config)

        self.pooler = EsmPooler(config) if add_pooling_layer else None

        self.contact_head = EsmContactPredictionHead(
            in_features=config.num_hidden_layers * config.num_attention_heads, bias=True
        )

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

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

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

    @merge_with_config_defaults
    @capture_outputs
    @auto_docstring
    def forward(
        self,
        input_ids: torch.Tensor | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.Tensor | None = None,
        inputs_embeds: torch.Tensor | None = None,
        encoder_hidden_states: torch.Tensor | None = None,
        encoder_attention_mask: torch.Tensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple[torch.Tensor] | BaseModelOutputWithPoolingAndCrossAttentions:
        r"""
        input_ids (`torch.LongTensor` of shape `((batch_size, sequence_length))`):
            Indices of input sequence tokens in the vocabulary.

            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
            [`PreTrainedTokenizer.__call__`] for details.

            [What are input IDs?](../glossary#input-ids)
        position_ids (`torch.LongTensor` of shape `((batch_size, sequence_length))`, *optional*):
            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
            config.max_position_embeddings - 1]`.

            [What are position IDs?](../glossary#position-ids)
        inputs_embeds (`torch.FloatTensor` of shape `((batch_size, sequence_length), hidden_size)`, *optional*):
            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
            model's internal embedding lookup matrix.
        """
        if (input_ids is None) ^ (inputs_embeds is not None):
            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")

        if inputs_embeds is None:
            # Important, attention_mask must be passed to the embedding class
            # This effects how the token_dropout is calculated
            inputs_embeds = self.embeddings(
                input_ids=input_ids,
                attention_mask=attention_mask,
                position_ids=position_ids,
            )

        attention_mask, encoder_attention_mask = self._create_attention_masks(
            attention_mask=attention_mask,
            encoder_attention_mask=encoder_attention_mask,
            embedding_output=inputs_embeds,
            encoder_hidden_states=encoder_hidden_states,
            past_key_values=None,
        )

        encoder_outputs = self.encoder(
            inputs_embeds,
            attention_mask=attention_mask,
            encoder_hidden_states=encoder_hidden_states,
            encoder_attention_mask=encoder_attention_mask,
            **kwargs,
        )
        sequence_output = encoder_outputs[0]
        pooled_output = self.pooler(sequence_output) if self.pooler is not None else None

        return BaseModelOutputWithPoolingAndCrossAttentions(
            last_hidden_state=sequence_output,
            pooler_output=pooled_output,
        )

    # Copied from transformers.models.bert.modeling_bert.BertModel._create_attention_masks
    def _create_attention_masks(
        self,
        attention_mask,
        encoder_attention_mask,
        embedding_output,
        encoder_hidden_states,
        past_key_values,
    ):
        if self.config.is_decoder:
            attention_mask = create_causal_mask(
                config=self.config,
                inputs_embeds=embedding_output,
                attention_mask=attention_mask,
                past_key_values=past_key_values,
            )
        else:
            attention_mask = create_bidirectional_mask(
                config=self.config,
                inputs_embeds=embedding_output,
                attention_mask=attention_mask,
            )

        if encoder_attention_mask is not None:
            encoder_attention_mask = create_bidirectional_mask(
                config=self.config,
                inputs_embeds=embedding_output,
                attention_mask=encoder_attention_mask,
                encoder_hidden_states=encoder_hidden_states,
            )

        return attention_mask, encoder_attention_mask

    def predict_contacts(self, tokens, attention_mask):
        attns = self(tokens, attention_mask=attention_mask, return_dict=True, output_attentions=True).attentions
        attns = torch.stack(attns, dim=1)  # Matches the original model layout
        # In the original model, attentions for padding tokens are completely zeroed out.
        # This makes no difference most of the time because the other tokens won't attend to them,
        # but it does for the contact prediction task, which takes attentions as input,
        # so we have to mimic that here.
        attns *= attention_mask.unsqueeze(1).unsqueeze(2).unsqueeze(3)
        attns *= attention_mask.unsqueeze(1).unsqueeze(2).unsqueeze(4)
        return self.contact_head(tokens, attns)


@auto_docstring
class EsmForMaskedLM(EsmPreTrainedModel):
    _tied_weights_keys = {"lm_head.decoder.weight": "esm.embeddings.word_embeddings.weight"}

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

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

        self.esm = EsmModel(config, add_pooling_layer=False)
        self.lm_head = EsmLMHead(config)

        self.post_init()

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

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

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        encoder_hidden_states: torch.FloatTensor | None = None,
        encoder_attention_mask: torch.Tensor | None = None,
        labels: torch.LongTensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | MaskedLMOutput:
        r"""
        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Labels 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]`
        """

        outputs = self.esm(
            input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            inputs_embeds=inputs_embeds,
            encoder_hidden_states=encoder_hidden_states,
            encoder_attention_mask=encoder_attention_mask,
            **kwargs,
        )
        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))

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

    def predict_contacts(self, tokens, attention_mask):
        return self.esm.predict_contacts(tokens, attention_mask=attention_mask)


class EsmLMHead(nn.Module):
    """ESM Head for masked language modeling."""

    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


@auto_docstring(
    custom_intro="""
    ESM Model transformer with a sequence classification/regression head on top (a linear layer on top of the pooled
    output) e.g. for GLUE tasks.
    """
)
class EsmForSequenceClassification(EsmPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)
        self.num_labels = config.num_labels
        self.config = config

        self.esm = EsmModel(config, add_pooling_layer=False)
        self.classifier = EsmClassificationHead(config)

        self.post_init()

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        labels: torch.LongTensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | SequenceClassifierOutput:
        r"""
        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).
        """

        outputs = self.esm(
            input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            inputs_embeds=inputs_embeds,
            **kwargs,
        )
        sequence_output = outputs[0]
        logits = self.classifier(sequence_output)

        loss = None
        if labels is not None:
            labels = labels.to(logits.device)

            if self.config.problem_type is None:
                if self.num_labels == 1:
                    self.config.problem_type = "regression"
                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
                    self.config.problem_type = "single_label_classification"
                else:
                    self.config.problem_type = "multi_label_classification"

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

        return SequenceClassifierOutput(
            loss=loss,
            logits=logits,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )


@auto_docstring
class EsmForTokenClassification(EsmPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)
        self.num_labels = config.num_labels

        self.esm = EsmModel(config, add_pooling_layer=False)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)
        self.classifier = nn.Linear(config.hidden_size, config.num_labels)

        self.post_init()

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        labels: torch.LongTensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | TokenClassifierOutput:
        r"""
        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Labels for computing the token classification loss. Indices should be in `[0, ..., config.num_labels - 1]`.
        """

        outputs = self.esm(
            input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            inputs_embeds=inputs_embeds,
            **kwargs,
        )

        sequence_output = outputs[0]

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

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

            labels = labels.to(logits.device)
            loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))

        return TokenClassifierOutput(
            loss=loss,
            logits=logits,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )


class EsmClassificationHead(nn.Module):
    """Head for sentence-level classification tasks."""

    def __init__(self, config):
        super().__init__()
        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)
        self.out_proj = nn.Linear(config.hidden_size, config.num_labels)

    def forward(self, features, **kwargs):
        x = features[:, 0, :]  # take <s> token (equiv. to [CLS])
        x = self.dropout(x)
        x = self.dense(x)
        x = torch.tanh(x)
        x = self.dropout(x)
        x = self.out_proj(x)
        return x


def create_position_ids_from_input_ids(input_ids, padding_idx):
    """
    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) * mask
    return incremental_indices.long() + padding_idx


__all__ = [
    "EsmForMaskedLM",
    "EsmForSequenceClassification",
    "EsmForTokenClassification",
    "EsmModel",
    "EsmPreTrainedModel",
]