diff --git "a/modeling_isoformer.py" "b/modeling_isoformer.py"
new file mode 100644--- /dev/null
+++ "b/modeling_isoformer.py"
@@ -0,0 +1,3518 @@
+from isoformer_config import IsoformerConfig
+from modeling_esm import NTForMaskedLM, MultiHeadAttention
+from modeling_esm_original import EsmForMaskedLM
+from enformer_pytorch import Enformer, str_to_one_hot, EnformerConfig
+from typing import Dict
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import SiLU
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "facebook/esm2_t6_8M_UR50D"
+_CONFIG_FOR_DOC = "NTConfig"
+
+ESM_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "facebook/esm2_t6_8M_UR50D",
+    "facebook/esm2_t12_35M_UR50D",
+    # This is not a complete list of all ESM models!
+    # See all ESM models at https://huggingface.co/models?filter=esm
+]
+
+
+import math
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+from transformers.file_utils import (
+    add_code_sample_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+)
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPastAndCrossAttentions,
+    BaseModelOutputWithPoolingAndCrossAttentions,
+    MaskedLMOutput,
+    SequenceClassifierOutput,
+    TokenClassifierOutput,
+)
+from transformers.modeling_utils import (
+    PreTrainedModel,
+    find_pruneable_heads_and_indices,
+    prune_linear_layer,
+)
+from transformers.models.esm.configuration_esm import EsmConfig
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "facebook/esm2_t6_8M_UR50D"
+_CONFIG_FOR_DOC = "EsmConfig"
+
+
+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.
+    """
+
+    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 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
+        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:
+                # 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, 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)
+
+        # 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:
+            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)
+            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
+            )
+        # FIRST DIFF BETWEEN JAX AND TORCH
+        # 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)
+
+
+class EsmSelfAttention(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.is_decoder = config.is_decoder
+
+    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)
+
+        # 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.is_decoder:
+            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
+            # Further calls to cross_attention layer can then reuse all cross-attention
+            # key/value_states (first "if" case)
+            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
+            # all previous decoder key/value_states. Further calls to uni-directional self-attention
+            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
+            # if encoder bi-directional self-attention `past_key_value` is always `None`
+            past_key_value = (key_layer, value_layer)
+
+        if self.position_embedding_type == "rotary":
+            query_layer, key_layer = self.rotary_embeddings(query_layer, key_layer)
+
+        # 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:
+            # Apply the attention mask is (precomputed for all layers in EsmModel forward() function)
+            attention_scores = attention_scores + attention_mask
+
+        # Normalize the attention scores to probabilities.
+        attention_probs = nn.functional.softmax(attention_scores, dim=-1)
+
+        attention_mask_widened = (
+            attention_mask.repeat(
+                attention_probs.shape[0],
+                attention_probs.shape[1],
+                attention_probs.shape[2],
+                1,
+            ).permute(0, 1, 3, 2)
+            == 0
+        )
+        attention_probs = torch.where(
+            attention_mask_widened, attention_probs, 0.00097656
+        )
+
+        # SECOND DIFF BETWEEN JAX AND TORCH
+        # 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.to(value_layer.dtype), value_layer)
+
+        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 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):
+        super().__init__()
+        self.self = EsmSelfAttention(config)
+        self.output = EsmSelfOutput(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 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(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 = 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)
+        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,
+        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 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
+
+    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,
+        )
+
+
+# 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
+
+
+class EsmPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = EsmConfig
+    base_model_prefix = "esm"
+    supports_gradient_checkpointing = True
+    _no_split_modules = [
+        "EsmLayer",
+        "EsmFoldTriangularSelfAttentionBlock",
+        "EsmEmbeddings",
+    ]
+
+    # 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)
+
+
+ESM_START_DOCSTRING = r"""
+
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+
+    Parameters:
+        config ([`EsmConfig`]): Model configuration class with all the parameters of the
+            model. Initializing with a config file does not load the weights associated with the model, only the
+            configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+ESM_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `({0})`):
+            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)
+        attention_mask (`torch.FloatTensor` of shape `({0})`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+        position_ids (`torch.LongTensor` of shape `({0})`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.max_position_embeddings - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
+            Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        inputs_embeds (`torch.FloatTensor` of shape `({0}, 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.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~file_utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+@add_start_docstrings(
+    "The bare ESM Model transformer outputting raw hidden-states without any specific head on top.",
+    ESM_START_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://arxiv.org/abs/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):
+        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
+
+    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)
+
+    @add_start_docstrings_to_model_forward(
+        ESM_INPUTS_DOCSTRING.format("(batch_size, sequence_length)")
+    )
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=BaseModelOutputWithPoolingAndCrossAttentions,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    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]:
+        r"""
+        encoder_hidden_states  (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if
+            the model is configured as a decoder.
+        encoder_attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in
+            the cross-attention if the model is configured as a decoder. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+        past_key_values (`tuple(tuple(torch.FloatTensor))` of length `config.n_layers` with each tuple having 4 tensors of shape `(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
+            Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
+
+            If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
+            don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
+            `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+            `past_key_values`).
+        """
+        output_attentions = (
+            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
+            )
+
+        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+        # ourselves in which case we just need to make it broadcastable to all heads.
+        extended_attention_mask: torch.Tensor = self.get_extended_attention_mask(
+            attention_mask, input_shape
+        )
+
+        # If a 2D or 3D attention mask is provided for the cross-attention
+        # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
+        if self.config.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
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+        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]
+        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 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,
+        )
+
+    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)
+
+
+@add_start_docstrings(
+    """ESM Model with a `language modeling` head on top.""", ESM_START_DOCSTRING
+)
+class EsmForMaskedLM(EsmPreTrainedModel):
+    _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 `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.init_weights()
+
+    def get_output_embeddings(self):
+        return self.lm_head.decoder
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head.decoder = new_embeddings
+
+    @add_start_docstrings_to_model_forward(
+        ESM_INPUTS_DOCSTRING.format("batch_size, sequence_length")
+    )
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=MaskedLMOutput,
+        config_class=_CONFIG_FOR_DOC,
+        mask="<mask>",
+    )
+    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]:
+        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]`
+        kwargs (`Dict[str, any]`, optional, defaults to *{}*):
+            Used to hide legacy arguments that have been deprecated.
+        """
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        outputs = self.esm(
+            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,
+        )
+
+    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
+
+
+@add_start_docstrings(
+    """
+    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.
+    """,
+    ESM_START_DOCSTRING,
+)
+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.init_weights()
+
+    @add_start_docstrings_to_model_forward(
+        ESM_INPUTS_DOCSTRING.format("batch_size, sequence_length")
+    )
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=SequenceClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    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,
+        labels: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[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).
+        """
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        outputs = self.esm(
+            input_ids,
+            attention_mask=attention_mask,
+            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,
+        )
+        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)
+
+        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=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    ESM Model with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for
+    Named-Entity-Recognition (NER) tasks.
+    """,
+    ESM_START_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.init_weights()
+
+    @add_start_docstrings_to_model_forward(
+        ESM_INPUTS_DOCSTRING.format("batch_size, sequence_length")
+    )
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=TokenClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    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,
+        labels: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[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]`.
+        """
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        outputs = self.esm(
+            input_ids,
+            attention_mask=attention_mask,
+            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,
+        )
+
+        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))
+
+        if not return_dict:
+            output = (logits,) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        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, 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
+
+
+from dataclasses import asdict, dataclass
+from typing import Optional
+
+from transformers import PretrainedConfig, logging
+
+logger = logging.get_logger(__name__)
+
+# TODO Update this
+ESM_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "facebook/esm-1b": "https://huggingface.co/facebook/esm-1b/resolve/main/config.json",
+    # See all ESM models at https://huggingface.co/models?filter=esm
+}
+
+
+class NTConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`ESMModel`]. It is used to instantiate a ESM model
+    according to the specified arguments, defining the model architecture. Instantiating a configuration with the
+    defaults will yield a similar configuration to that of the ESM
+    [facebook/esm-1b](https://huggingface.co/facebook/esm-1b) architecture.
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+    Args:
+        vocab_size (`int`, *optional*):
+            Vocabulary size of the ESM model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`ESMModel`].
+        mask_token_id (`int`, *optional*):
+            The index of the mask token in the vocabulary. This must be included in the config because of the
+            "mask-dropout" scaling trick, which will scale the inputs depending on the number of masked tokens.
+        pad_token_id (`int`, *optional*):
+            The index of the padding token in the vocabulary. This must be included in the config because certain parts
+            of the ESM code use this instead of the attention mask.
+        hidden_size (`int`, *optional*, defaults to 768):
+            Dimensionality of the encoder layers and the pooler layer.
+        num_hidden_layers (`int`, *optional*, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 12):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        intermediate_size (`int`, *optional*, defaults to 3072):
+            Dimensionality of the "intermediate" (often named feed-forward) layer in the Transformer encoder.
+        hidden_dropout_prob (`float`, *optional*, defaults to 0.1):
+            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_probs_dropout_prob (`float`, *optional*, defaults to 0.1):
+            The dropout ratio for the attention probabilities.
+        max_position_embeddings (`int`, *optional*, defaults to 1026):
+            The maximum sequence length that this model might ever be used with. Typically set this to something large
+            just in case (e.g., 512 or 1024 or 2048).
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        layer_norm_eps (`float`, *optional*, defaults to 1e-12):
+            The epsilon used by the layer normalization layers.
+        position_embedding_type (`str`, *optional*, defaults to `"absolute"`):
+            Type of position embedding. Choose one of `"absolute"`, `"relative_key"`, `"relative_key_query", "rotary"`.
+            For positional embeddings use `"absolute"`. For more information on `"relative_key"`, please refer to
+            [Self-Attention with Relative Position Representations (Shaw et al.)](https://arxiv.org/abs/1803.02155).
+            For more information on `"relative_key_query"`, please refer to *Method 4* in [Improve Transformer Models
+            with Better Relative Position Embeddings (Huang et al.)](https://arxiv.org/abs/2009.13658).
+        is_decoder (`bool`, *optional*, defaults to `False`):
+            Whether the model is used as a decoder or not. If `False`, the model is used as an encoder.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models). Only
+            relevant if `config.is_decoder=True`.
+        emb_layer_norm_before (`bool`, *optional*):
+            Whether to apply layer normalization after embeddings but before the main stem of the network.
+        token_dropout (`bool`, defaults to `False`):
+            When this is enabled, masked tokens are treated as if they had been dropped out by input dropout.
+    Examples:
+    ```python
+    >>> from transformers import EsmModel, EsmConfig
+    >>> # Initializing a ESM facebook/esm-1b style configuration >>> configuration = EsmConfig()
+    >>> # Initializing a model from the configuration >>> model = ESMModel(configuration)
+    >>> # Accessing the model configuration >>> configuration = model.config
+    ```"""
+    model_type = "esm"
+
+    def __init__(
+        self,
+        vocab_size=None,
+        mask_token_id=None,
+        pad_token_id=None,
+        hidden_size=768,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        intermediate_size=3072,
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        max_position_embeddings=1026,
+        initializer_range=0.02,
+        layer_norm_eps=1e-12,
+        position_embedding_type="absolute",
+        use_cache=True,
+        emb_layer_norm_before=None,
+        token_dropout=False,
+        is_folding_model=False,
+        esmfold_config=None,
+        vocab_list=None,
+        add_bias_fnn=True,
+        **kwargs,
+    ):
+        super().__init__(
+            pad_token_id=pad_token_id, mask_token_id=mask_token_id, **kwargs
+        )
+
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.intermediate_size = intermediate_size
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.max_position_embeddings = max_position_embeddings
+        self.initializer_range = initializer_range
+        self.layer_norm_eps = layer_norm_eps
+        self.position_embedding_type = position_embedding_type
+        self.use_cache = use_cache
+        self.emb_layer_norm_before = emb_layer_norm_before
+        self.token_dropout = token_dropout
+        self.is_folding_model = is_folding_model
+
+        # Arguments needed for Dalmatian
+        self.add_bias_fnn = add_bias_fnn
+        if is_folding_model:
+            if esmfold_config is None:
+                logger.info(
+                    "No esmfold_config supplied for folding model, using default values."
+                )
+                esmfold_config = EsmFoldConfig()
+            elif isinstance(esmfold_config, dict):
+                esmfold_config = EsmFoldConfig(**esmfold_config)
+            self.esmfold_config = esmfold_config
+            if vocab_list is None:
+                logger.warning(
+                    "No vocab_list supplied for folding model, assuming the ESM-2 vocabulary!"
+                )
+                self.vocab_list = get_default_vocab_list()
+            else:
+                self.vocab_list = vocab_list
+        else:
+            self.esmfold_config = None
+            self.vocab_list = None
+        if self.esmfold_config is not None and getattr(
+            self.esmfold_config, "use_esm_attn_map", False
+        ):
+            raise ValueError(
+                "The HuggingFace port of ESMFold does not support use_esm_attn_map at this time!"
+            )
+
+    def to_dict(self):
+        """
+        Serializes this instance to a Python dictionary. Override the default [`~PretrainedConfig.to_dict`].
+        Returns:
+            `Dict[str, any]`: Dictionary of all the attributes that make up this configuration instance,
+        """
+        output = super().to_dict()
+        if isinstance(self.esmfold_config, EsmFoldConfig):
+            output["esmfold_config"] = self.esmfold_config.to_dict()
+        return output
+
+
+@dataclass
+class EsmFoldConfig:
+    esm_type: str = None
+    fp16_esm: bool = True
+    use_esm_attn_map: bool = False
+    esm_ablate_pairwise: bool = False
+    esm_ablate_sequence: bool = False
+    esm_input_dropout: float = 0
+
+    embed_aa: bool = True
+    bypass_lm: bool = False
+
+    lddt_head_hid_dim: int = 128
+    trunk: "TrunkConfig" = None
+
+    def __post_init__(self):
+        if self.trunk is None:
+            self.trunk = TrunkConfig()
+        elif isinstance(self.trunk, dict):
+            self.trunk = TrunkConfig(**self.trunk)
+
+    def to_dict(self):
+        """
+        Serializes this instance to a Python dictionary. Override the default [`~PretrainedConfig.to_dict`].
+        Returns:
+            `Dict[str, any]`: Dictionary of all the attributes that make up this configuration instance,
+        """
+        output = asdict(self)
+        output["trunk"] = self.trunk.to_dict()
+        return output
+
+
+@dataclass
+class TrunkConfig:
+    num_blocks: int = 48
+    sequence_state_dim: int = 1024
+    pairwise_state_dim: int = 128
+    sequence_head_width: int = 32
+    pairwise_head_width: int = 32
+    position_bins: int = 32
+    dropout: float = 0
+    layer_drop: float = 0
+    cpu_grad_checkpoint: bool = False
+    max_recycles: int = 4
+    chunk_size: Optional[int] = 128
+    structure_module: "StructureModuleConfig" = None
+
+    def __post_init__(self):
+        if self.structure_module is None:
+            self.structure_module = StructureModuleConfig()
+        elif isinstance(self.structure_module, dict):
+            self.structure_module = StructureModuleConfig(**self.structure_module)
+
+        if self.max_recycles <= 0:
+            raise ValueError(
+                f"`max_recycles` should be positive, got {self.max_recycles}."
+            )
+        if self.sequence_state_dim % self.sequence_state_dim != 0:
+            raise ValueError(
+                "`sequence_state_dim` should be a round multiple of `sequence_state_dim`, got"
+                f" {self.sequence_state_dim} and {self.sequence_state_dim}."
+            )
+        if self.pairwise_state_dim % self.pairwise_state_dim != 0:
+            raise ValueError(
+                "`pairwise_state_dim` should be a round multiple of `pairwise_state_dim`, got"
+                f" {self.pairwise_state_dim} and {self.pairwise_state_dim}."
+            )
+
+        sequence_num_heads = self.sequence_state_dim // self.sequence_head_width
+        pairwise_num_heads = self.pairwise_state_dim // self.pairwise_head_width
+
+        if self.sequence_state_dim != sequence_num_heads * self.sequence_head_width:
+            raise ValueError(
+                "`sequence_state_dim` should be equal to `sequence_num_heads * sequence_head_width, got"
+                f" {self.sequence_state_dim} != {sequence_num_heads} * {self.sequence_head_width}."
+            )
+        if self.pairwise_state_dim != pairwise_num_heads * self.pairwise_head_width:
+            raise ValueError(
+                "`pairwise_state_dim` should be equal to `pairwise_num_heads * pairwise_head_width, got"
+                f" {self.pairwise_state_dim} != {pairwise_num_heads} * {self.pairwise_head_width}."
+            )
+        if self.pairwise_state_dim % 2 != 0:
+            raise ValueError(
+                f"`pairwise_state_dim` should be even, got {self.pairwise_state_dim}."
+            )
+
+        if self.dropout >= 0.4:
+            raise ValueError(
+                f"`dropout` should not be greater than 0.4, got {self.dropout}."
+            )
+
+    def to_dict(self):
+        """
+        Serializes this instance to a Python dictionary. Override the default [`~PretrainedConfig.to_dict`].
+        Returns:
+            `Dict[str, any]`: Dictionary of all the attributes that make up this configuration instance,
+        """
+        output = asdict(self)
+        output["structure_module"] = self.structure_module.to_dict()
+        return output
+
+
+@dataclass
+class StructureModuleConfig:
+    """
+    Args:
+        sequence_dim:
+            Single representation channel dimension
+        pairwise_dim:
+            Pair representation channel dimension
+        ipa_dim:
+            IPA hidden channel dimension
+        resnet_dim:
+            Angle resnet (Alg. 23 lines 11-14) hidden channel dimension
+        num_heads_ipa:
+            Number of IPA heads
+        num_qk_points:
+            Number of query/key points to generate during IPA
+        num_v_points:
+            Number of value points to generate during IPA
+        dropout_rate:
+            Dropout rate used throughout the layer
+        num_blocks:
+            Number of structure module blocks
+        num_transition_layers:
+            Number of layers in the single representation transition (Alg. 23 lines 8-9)
+        num_resnet_blocks:
+            Number of blocks in the angle resnet
+        num_angles:
+            Number of angles to generate in the angle resnet
+        trans_scale_factor:
+            Scale of single representation transition hidden dimension
+        epsilon:
+            Small number used in angle resnet normalization
+        inf:
+            Large number used for attention masking
+    """
+
+    sequence_dim: int = 384
+    pairwise_dim: int = 128
+    ipa_dim: int = 16
+    resnet_dim: int = 128
+    num_heads_ipa: int = 12
+    num_qk_points: int = 4
+    num_v_points: int = 8
+    dropout_rate: float = 0.1
+    num_blocks: int = 8
+    num_transition_layers: int = 1
+    num_resnet_blocks: int = 2
+    num_angles: int = 7
+    trans_scale_factor: int = 10
+    epsilon: float = 1e-8
+    inf: float = 1e5
+
+    def to_dict(self):
+        return asdict(self)
+
+
+def get_default_vocab_list():
+    return (
+        "<cls>",
+        "<pad>",
+        "<eos>",
+        "<unk>",
+        "L",
+        "A",
+        "G",
+        "V",
+        "S",
+        "E",
+        "R",
+        "T",
+        "I",
+        "D",
+        "P",
+        "K",
+        "Q",
+        "N",
+        "F",
+        "Y",
+        "M",
+        "H",
+        "W",
+        "C",
+        "X",
+        "B",
+        "U",
+        "Z",
+        "O",
+        ".",
+        "-",
+        "<null_1>",
+        "<mask>",
+    )
+
+
+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.
+    """
+
+    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).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 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
+        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:
+                # 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, 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)
+
+        # 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:
+            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)
+            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 += 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
+        #     )
+        # FIRST DIFF BETWEEN JAX AND TORCH
+        # 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)
+
+
+class EsmSelfAttention(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.is_decoder = config.is_decoder
+
+    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)
+
+        # 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.is_decoder:
+            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
+            # Further calls to cross_attention layer can then reuse all cross-attention
+            # key/value_states (first "if" case)
+            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
+            # all previous decoder key/value_states. Further calls to uni-directional self-attention
+            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
+            # if encoder bi-directional self-attention `past_key_value` is always `None`
+            past_key_value = (key_layer, value_layer)
+
+        if self.position_embedding_type == "rotary":
+            query_layer, key_layer = self.rotary_embeddings(query_layer, key_layer)
+
+        # 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)
+        attention_mask_widened = attention_mask.repeat(
+            attention_probs.shape[0],
+            attention_probs.shape[1],
+            attention_probs.shape[2],
+            1
+        ).permute(0,1,3,2) == 0
+        attention_probs = torch.where(attention_mask_widened, attention_probs, 0.00097656)
+        # SECOND DIFF BETWEEN JAX AND TORCH
+
+        # 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)
+
+        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 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 += input_tensor
+        return hidden_states
+
+
+class EsmAttention(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.self = EsmSelfAttention(config)
+        self.output = EsmSelfOutput(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 MultiHeadAttention(nn.Module):
+    def __init__(self, config, omics_of_interest_size: int, other_omic_size: int, 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(omics_of_interest_size, omics_of_interest_size) # 3072, 3072
+
+        self.key = nn.Linear(other_omic_size, omics_of_interest_size) # 768, 3072
+
+        self.value = nn.Linear(other_omic_size, omics_of_interest_size) # 768, 3072
+
+        self.dense = nn.Linear(omics_of_interest_size, omics_of_interest_size) # 3072, 3072
+
+
+        #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.is_decoder = config.is_decoder
+
+    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,
+    ) -> Dict[str, 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)
+
+        # 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.is_decoder:
+            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
+            # Further calls to cross_attention layer can then reuse all cross-attention
+            # key/value_states (first "if" case)
+            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
+            # all previous decoder key/value_states. Further calls to uni-directional self-attention
+            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
+            # if encoder bi-directional self-attention `past_key_value` is always `None`
+            past_key_value = (key_layer, value_layer)
+
+        if self.position_embedding_type == "rotary":
+            query_layer, key_layer = self.rotary_embeddings(query_layer, key_layer)
+
+        # 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:
+            # Apply the attention mask is (precomputed for all layers in NTModel forward() function)
+            #attention_scores = attention_scores + attention_mask
+            attention_scores = torch.where(attention_mask, attention_scores, -1e30)
+            #attention_logits = jnp.where(attention_mask, attention_logits, -1e30)
+
+        # 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)
+
+        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 {
+            "embeddings": self.dense(context_layer) + hidden_states,
+            "query_heads": self.transpose_for_scores(mixed_query_layer),
+            "value_heads": self.transpose_for_scores(self.value(encoder_hidden_states)),
+            "key_heads": self.transpose_for_scores(self.key(encoder_hidden_states)),
+            "attention_probs": attention_probs,
+            "attention_scores": attention_scores,
+            "context_layer": context_layer,
+        }
+
+class EsmIntermediate(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(
+            config.hidden_size,
+            int(config.intermediate_size * 2),
+            bias=config.add_bias_fnn,
+        )
+        self.activation_fn = SiLU()
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+
+        # GLU
+        x1, x2 = hidden_states.split(int(hidden_states.size(-1) / 2), -1)
+        hidden_states = self.activation_fn(x1) * x2
+
+        return hidden_states
+
+
+class EsmOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(
+            config.intermediate_size, config.hidden_size, bias=config.add_bias_fnn
+        )
+        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 += input_tensor
+        return hidden_states
+
+
+class EsmLayer(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 = 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)
+        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,
+        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 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
+
+    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:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs, past_key_value, output_attentions)
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(layer_module),
+                    hidden_states,
+                    attention_mask,
+                    layer_head_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                )
+            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 += (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,
+        )
+
+
+# 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
+
+
+class EsmPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = NTConfig
+    base_model_prefix = "esm"
+    _no_split_modules = ["EsmLayer", "EsmFoldTriangularSelfAttentionBlock"]
+
+    # 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)
+
+
+ESM_START_DOCSTRING = r"""
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+    Parameters:
+        config ([`NTConfig`]): Model configuration class with all the parameters of the
+            model. Initializing with a config file does not load the weights associated with the model, only the
+            configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+ESM_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `({0})`):
+            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)
+        attention_mask (`torch.FloatTensor` of shape `({0})`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+            [What are attention masks?](../glossary#attention-mask)
+        position_ids (`torch.LongTensor` of shape `({0})`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.max_position_embeddings - 1]`.
+            [What are position IDs?](../glossary#position-ids)
+        head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
+            Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+        inputs_embeds (`torch.FloatTensor` of shape `({0}, 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.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~file_utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+@add_start_docstrings(
+    "The bare ESM Model transformer outputting raw hidden-states without any specific head on top.",
+    ESM_START_DOCSTRING,
+)
+class NTModel(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://arxiv.org/abs/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.
+    """
+
+    supports_gradient_checkpointing = False
+
+    def __init__(self, config, add_pooling_layer=True):
+        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 _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, EsmEncoder):
+            module.gradient_checkpointing = value
+
+    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)
+
+    @add_start_docstrings_to_model_forward(
+        ESM_INPUTS_DOCSTRING.format("(batch_size, sequence_length)")
+    )
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=BaseModelOutputWithPoolingAndCrossAttentions,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    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]:
+        r"""
+        encoder_hidden_states  (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if
+            the model is configured as a decoder.
+        encoder_attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in
+            the cross-attention if the model is configured as a decoder. Mask values selected in `[0, 1]`:
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+        past_key_values (`tuple(tuple(torch.FloatTensor))` of length `config.n_layers` with each tuple having 4 tensors of shape `(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
+            Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
+            If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
+            don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
+            `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+            `past_key_values`).
+        """
+        output_attentions = (
+            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:
+            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
+            )
+
+        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+        # ourselves in which case we just need to make it broadcastable to all heads.
+        extended_attention_mask: torch.Tensor = self.get_extended_attention_mask(
+            attention_mask, input_shape
+        )
+
+        # If a 2D or 3D attention mask is provided for the cross-attention
+        # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
+        if self.config.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
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+        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]
+        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 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,
+        )
+
+    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)
+
+
+@add_start_docstrings(
+    """ESM Model with a `language modeling` head on top.""", ESM_START_DOCSTRING
+)
+class NTForMaskedLM(EsmPreTrainedModel):
+    _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 `EsmForMaskedLM` make sure `config.is_decoder=False` for "
+                "bi-directional self-attention."
+            )
+
+        self.esm = NTModel(config, add_pooling_layer=False)
+        self.lm_head = EsmLMHead(config)
+
+        self.init_weights()
+
+    def get_output_embeddings(self):
+        return self.lm_head.decoder
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head.decoder = new_embeddings
+
+    @add_start_docstrings_to_model_forward(
+        ESM_INPUTS_DOCSTRING.format("batch_size, sequence_length")
+    )
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=MaskedLMOutput,
+        config_class=_CONFIG_FOR_DOC,
+        mask="<mask>",
+    )
+    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]:
+        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]`
+        kwargs (`Dict[str, any]`, optional, defaults to *{}*):
+            Used to hide legacy arguments that have been deprecated.
+        """
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        outputs = self.esm(
+            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,
+        )
+
+    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
+
+
+@add_start_docstrings(
+    """
+    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.
+    """,
+    ESM_START_DOCSTRING,
+)
+class EsmForSequenceClassification(EsmPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.config = config
+
+        self.esm = NTModel(config, add_pooling_layer=False)
+        self.classifier = EsmClassificationHead(config)
+
+        self.init_weights()
+
+    @add_start_docstrings_to_model_forward(
+        ESM_INPUTS_DOCSTRING.format("batch_size, sequence_length")
+    )
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=SequenceClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    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,
+        labels: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[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).
+        """
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        outputs = self.esm(
+            input_ids,
+            attention_mask=attention_mask,
+            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,
+        )
+        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)
+
+        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=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    ESM Model with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for
+    Named-Entity-Recognition (NER) tasks.
+    """,
+    ESM_START_DOCSTRING,
+)
+class EsmForTokenClassification(EsmPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.esm = NTModel(config, add_pooling_layer=False)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.classifier = nn.Linear(config.hidden_size, config.num_labels)
+
+        self.init_weights()
+
+    @add_start_docstrings_to_model_forward(
+        ESM_INPUTS_DOCSTRING.format("batch_size, sequence_length")
+    )
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=TokenClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    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,
+        labels: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[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]`.
+        """
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        outputs = self.esm(
+            input_ids,
+            attention_mask=attention_mask,
+            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,
+        )
+
+        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))
+
+        if not return_dict:
+            output = (logits,) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        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, 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
+
+
+class Isoformer(PreTrainedModel):
+    config_class = IsoformerConfig
+
+    def __init__(self, config):
+        super().__init__(config)
+
+
+        self.esm_config = EsmConfig(
+            vocab_size=config.esm_vocab_size,
+            mask_token_id=config.esm_mask_token_id,
+            pad_token_id=config.esm_pad_token_id,
+            hidden_size=config.esm_hidden_size,
+            num_hidden_layers=config.esm_num_hidden_layers,
+            num_attention_heads=config.esm_num_attention_heads,
+            intermediate_size=config.esm_intermediate_size,
+            max_position_embeddings=config.esm_max_position_embeddings,
+            token_dropout=config.esm_token_dropout,
+            emb_layer_norm_before=config.esm_emb_layer_norm_before,
+            attention_probs_dropout_prob=0.0,
+            hidden_dropout_prob=0.0,
+            use_cache=False,
+            add_bias_fnn=config.esm_add_bias_fnn,
+            position_embedding_type="rotary",
+            tie_word_embeddings=False,
+        )
+
+        self.nt_config = NTConfig(
+            vocab_size=config.nt_vocab_size,
+            mask_token_id=config.nt_mask_token_id,
+            pad_token_id=config.nt_pad_token_id,
+            hidden_size=config.nt_hidden_size,
+            num_hidden_layers=config.nt_num_hidden_layers,
+            num_attention_heads=config.nt_num_attention_heads,
+            intermediate_size=config.nt_intermediate_size,
+            max_position_embeddings=config.nt_max_position_embeddings,
+            token_dropout=config.nt_token_dropout,
+            emb_layer_norm_before=config.nt_emb_layer_norm_before,
+            attention_probs_dropout_prob=0.0,
+            hidden_dropout_prob=0.0,
+            use_cache=False,
+            add_bias_fnn=config.nt_add_bias_fnn,
+            position_embedding_type="rotary",
+            tie_word_embeddings=False,
+        )
+        self.config = config
+
+        # self.enformer_config = EnformerConfig(
+        #     dim=config.enformer_dim,
+        #     depth=config.enformer_depth,
+        #     heads=config.enformer_heads,
+        #     output_heads=dict(
+        #         human=1,
+        #         mouse=1 # TODO CHANGE
+        #     ),
+        #     target_length=config.enformer_target_length,  # 896,
+        #     attn_dim_key=config.enformer_attn_dim_key,
+        #     dropout_rate=0.4,
+        #     attn_dropout=0.05,
+        #     pos_dropout=0.01,
+        #     use_checkpointing=config.enformer_use_checkpointing,
+        #     use_convnext=config.enformer_use_convnext,
+        #     num_downsamples=config.enformer_num_downsamples,
+        #     # genetic sequence is downsampled 2 ** 7 == 128x in default Enformer - can be changed for higher resolution
+        #     dim_divisible_by=config.enformer_dim_divisible_by,
+        #     use_tf_gamma=False,
+        # )
+
+        self.esm_model = EsmForMaskedLM(self.esm_config) # protein encoder
+        self.nt_model = NTForMaskedLM(self.nt_config) # rna encoder
+        #self.enformer_model = Enformer(self.enformer_config) # dna encoder
+        self.enformer_model = Enformer.from_pretrained("EleutherAI/enformer-official-rough")
+
+        self.cross_attention_layer_rna = MultiHeadAttention(
+            config=EsmConfig(
+                num_attention_heads=config.num_heads_omics_cross_attention,
+                attention_head_size=3072 // config.num_heads_omics_cross_attention,
+                hidden_size=3072,
+                attention_probs_dropout_prob=0,
+                max_position_embeddings=0
+            ),
+            omics_of_interest_size=3072,
+            other_omic_size=768
+        )
+        self.cross_attention_layer_protein = MultiHeadAttention(
+            config=EsmConfig(
+                num_attention_heads=config.num_heads_omics_cross_attention,
+                attention_head_size=3072 // config.num_heads_omics_cross_attention,
+                hidden_size=3072,
+                attention_probs_dropout_prob=0,
+                max_position_embeddings=0
+            ),
+            omics_of_interest_size=3072,
+            other_omic_size=640
+        )
+
+        self.head_layer_1 = nn.Linear(3072, 2 * 3072)
+        self.head_layer_2 = nn.Linear(2 * 3072, 30)
+
+    def forward(
+            self,
+            tensor_dna,
+            tensor_rna,
+            tensor_protein,
+            attention_mask_rna,
+            attention_mask_protein
+    ):
+        tensor_dna = tensor_dna[:, 1:] # remove CLS
+        dna_embedding = self.enformer_model(
+            tensor_dna,
+            return_only_embeddings=True
+            # attention_mask=attention_mask_dna,
+            # encoder_attention_mask=attention_mask_dna,
+            # output_hidden_states=True
+        )
+        protein_embedding = self.esm_model(
+            tensor_protein,
+            attention_mask=attention_mask_protein,
+            encoder_attention_mask=attention_mask_protein,
+            output_hidden_states=True
+        )
+        rna_embedding = self.nt_model(
+            tensor_rna,
+            attention_mask=attention_mask_rna,
+            encoder_attention_mask=attention_mask_rna,
+            output_hidden_states=True
+        )
+
+        encoder_attention_mask = torch.unsqueeze(torch.unsqueeze(tensor_rna != 1, 0),0).repeat(1,1,dna_embedding.shape[1],1)
+        rna_to_dna = self.cross_attention_layer_rna.forward(
+            hidden_states=dna_embedding,
+            encoder_hidden_states=rna_embedding["hidden_states"][-1],
+            encoder_attention_mask=encoder_attention_mask
+        )
+
+        final_dna_embeddings = self.cross_attention_layer_protein.forward(
+            hidden_states=rna_to_dna["embeddings"],
+            encoder_hidden_states=protein_embedding["hidden_states"][-1],
+        )["embeddings"]
+
+        sequence_mask = torch.zeros(final_dna_embeddings.shape[1])
+        sequence_mask[self.config.pool_window_start:self.config.pool_window_end] = 1
+        x = torch.sum(torch.einsum('ijk,j->ijk', final_dna_embeddings, sequence_mask),axis=1)/torch.sum(sequence_mask)
+        x = self.head_layer_1(x)
+        x = torch.nn.functional.softplus(x)
+        x = self.head_layer_2(x)
+
+
+        return {
+            "gene_expression_predictions": x,
+            "final_dna_embeddings": final_dna_embeddings,
+        }
\ No newline at end of file