Aluode/PerceptionLabPortable
0
1# coding=utf-82# Copyright 2025 Westlake Representational Learning Lab (Fajie Yuan Lab) team and the HuggingFace Inc. team. All rights reserved.3#4# Licensed under the Apache License, Version 2.0 (the "License");5# you may not use this file except in compliance with the License.6# You may obtain a copy of the License at7#8# http://www.apache.org/licenses/LICENSE-2.09#10# Unless required by applicable law or agreed to in writing, software11# distributed under the License is distributed on an "AS IS" BASIS,12# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.13# See the License for the specific language governing permissions and14# limitations under the License.15 16import warnings17from dataclasses import dataclass18from typing import Optional, Union19 20import torch21from torch import Tensor, nn22 23from ...cache_utils import Cache, DynamicCache24from ...generation import GenerationMixin25from ...masking_utils import create_causal_mask26from ...modeling_outputs import (27 BaseModelOutputWithPast,28 BaseModelOutputWithPoolingAndCrossAttentions,29 CausalLMOutputWithPast,30 ModelOutput,31)32from ...modeling_utils import ModuleUtilsMixin, PreTrainedModel, get_parameter_dtype33from ...utils import (34 auto_docstring,35 can_return_tuple,36 logging,37)38from ...utils.deprecation import deprecate_kwarg39from ...utils.generic import OutputRecorder, check_model_inputs40from ..esm.modeling_esm import (41 EsmAttention,42 EsmEmbeddings,43 EsmEncoder,44 EsmIntermediate,45 EsmLayer,46 EsmOutput,47 EsmPooler,48 EsmSelfAttention,49 EsmSelfOutput,50)51from ..llama.modeling_llama import (52 LlamaAttention,53 LlamaDecoderLayer,54 LlamaMLP,55 LlamaPreTrainedModel,56 LlamaRMSNorm,57 LlamaRotaryEmbedding,58)59from .configuration_evolla import EvollaConfig, SaProtConfig60 61 62logger = logging.get_logger(__name__)63 64 65class EvollaSaProtEmbeddings(EsmEmbeddings):66 def __init__(self, config):67 super().__init__(config)68 # remove the position_ids in EsmEmbeddings69 self.position_ids = None70 71 72def rotate_half_esm(x):73 x1, x2 = x.chunk(2, dim=-1)74 return torch.cat((-x2, x1), dim=-1)75 76 77def apply_rotary_pos_emb_esm(x, cos, sin):78 cos = cos[:, :, : x.shape[-2], :]79 sin = sin[:, :, : x.shape[-2], :]80 81 return (x * cos) + (rotate_half_esm(x) * sin)82 83 84class EvollaSaProtRotaryEmbedding(nn.Module):85 """86 Rotary position embeddings based on those in87 [RoFormer](https://huggingface.co/docs/transformers/model_doc/roformer). Query and keys are transformed by rotation88 matrices which depend on their relative positions.89 """90 91 inv_freq: torch.Tensor # fix linting for `register_buffer`92 93 def __init__(self, dim: int):94 super().__init__()95 # Generate and save the inverse frequency buffer (non trainable)96 inv_freq = 1.0 / (10000 ** (torch.arange(0, dim, 2, dtype=torch.int64).float() / dim))97 self.register_buffer("inv_freq", inv_freq)98 99 self._seq_len_cached = None100 self._cos_cached = None101 self._sin_cached = None102 103 def _update_cos_sin_tables(self, x, seq_dimension=2):104 seq_len = x.shape[seq_dimension]105 106 # Reset the tables if the sequence length has changed,107 # or if we're on a new device (possibly due to tracing for instance)108 if seq_len != self._seq_len_cached or self._cos_cached.device != x.device:109 self._seq_len_cached = seq_len110 t = torch.arange(x.shape[seq_dimension], device=x.device).type_as(self.inv_freq)111 freqs = torch.outer(t, self.inv_freq)112 emb = torch.cat((freqs, freqs), dim=-1).to(x.device)113 114 self._cos_cached = emb.cos()[None, None, :, :]115 self._sin_cached = emb.sin()[None, None, :, :]116 117 return self._cos_cached, self._sin_cached118 119 def forward(self, q: torch.Tensor, k: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:120 self._cos_cached, self._sin_cached = self._update_cos_sin_tables(k, seq_dimension=-2)121 122 return (123 apply_rotary_pos_emb_esm(q, self._cos_cached, self._sin_cached).to(dtype=q.dtype),124 apply_rotary_pos_emb_esm(k, self._cos_cached, self._sin_cached).to(dtype=k.dtype),125 )126 127 128class EvollaSaProtSelfAttention(EsmSelfAttention):129 def __init__(self, config, position_embedding_type=None, layer_idx=None, is_cross_attention=False):130 nn.Module.__init__(self)131 self.config = config132 133 if config.hidden_size % config.num_attention_heads != 0 and not hasattr(config, "embedding_size"):134 raise ValueError(135 f"The hidden size ({config.hidden_size}) is not a multiple of the number of attention "136 f"heads ({config.num_attention_heads})"137 )138 139 self.num_attention_heads = config.num_attention_heads140 self.attention_head_size = int(config.hidden_size / config.num_attention_heads)141 self.all_head_size = self.num_attention_heads * self.attention_head_size142 143 self.query = nn.Linear(config.hidden_size, self.all_head_size)144 self.key = nn.Linear(config.hidden_size, self.all_head_size)145 self.value = nn.Linear(config.hidden_size, self.all_head_size)146 147 self.dropout = config.attention_probs_dropout_prob148 self.position_embedding_type = position_embedding_type or getattr(149 config, "position_embedding_type", "absolute"150 )151 self.rotary_embeddings = None152 if self.position_embedding_type == "relative_key" or self.position_embedding_type == "relative_key_query":153 self.max_position_embeddings = config.max_position_embeddings154 self.distance_embedding = nn.Embedding(2 * config.max_position_embeddings - 1, self.attention_head_size)155 elif self.position_embedding_type == "rotary":156 self.rotary_embeddings = EvollaSaProtRotaryEmbedding(dim=self.attention_head_size)157 158 self.is_decoder = config.is_decoder159 self.layer_idx = layer_idx160 self.scaling = 1.0161 self.is_causal = self.is_decoder and not is_cross_attention162 163 164class EvollaSaProtSelfOutput(EsmSelfOutput):165 pass166 167 168class EvollaSaProtAttention(EsmAttention):169 pass170 171 172class EvollaSaProtIntermediate(EsmIntermediate):173 pass174 175 176class EvollaSaProtOutput(EsmOutput):177 pass178 179 180class EvollaSaProtLayer(EsmLayer):181 pass182 183 184class EvollaSaProtEncoder(EsmEncoder):185 pass186 187 188class EvollaSaProtPooler(EsmPooler):189 pass190 191 192@auto_docstring193class EvollaSaProtPreTrainedModel(PreTrainedModel):194 config: SaProtConfig195 _no_split_modules = ["EvollaSaProtLayer"]196 _supports_flash_attn = True197 _supports_sdpa = True198 _supports_attention_backend = True199 200 _can_record_outputs = {201 "hidden_states": EvollaSaProtLayer,202 "attentions": [OutputRecorder(EvollaSaProtSelfAttention, index=1, layer_name="attention")],203 "cross_attentions": [204 OutputRecorder(EvollaSaProtSelfAttention, index=1, layer_name="crossattention"),205 ],206 }207 208 def _init_weights(self, module):209 """Initialize the weights"""210 std = self.config.initializer_range211 if isinstance(module, nn.Linear):212 module.weight.data.normal_(mean=0.0, std=std)213 if module.bias is not None:214 module.bias.data.zero_()215 elif isinstance(module, nn.Embedding):216 module.weight.data.normal_(mean=0.0, std=std)217 if module.padding_idx is not None:218 module.weight.data[module.padding_idx].zero_()219 elif isinstance(module, nn.LayerNorm):220 module.bias.data.zero_()221 module.weight.data.fill_(1.0)222 223 224class EvollaSaProtProteinEncoder(EvollaSaProtPreTrainedModel):225 def __init__(self, config: SaProtConfig):226 super().__init__(config)227 self.embeddings = EvollaSaProtEmbeddings(config)228 self.encoder = EvollaSaProtEncoder(config)229 230 def get_input_embeddings(self):231 return self.embeddings.word_embeddings232 233 def set_input_embeddings(self, value):234 self.embeddings.word_embeddings = value235 236 def _prune_heads(self, heads_to_prune):237 """238 Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base239 class PreTrainedModel240 """241 for layer, heads in heads_to_prune.items():242 self.encoder.layer[layer].attention.prune_heads(heads)243 244 @check_model_inputs()245 def forward(246 self,247 input_ids: Optional[torch.Tensor],248 attention_mask: Optional[torch.Tensor] = None,249 ) -> Union[tuple[torch.Tensor], BaseModelOutputWithPoolingAndCrossAttentions]:250 input_shape = input_ids.size()251 batch_size, seq_length = input_shape252 253 device = input_ids.device254 if attention_mask is None:255 attention_mask = torch.ones(((batch_size, seq_length)), device=device)256 257 inputs_embeds = self.embeddings(input_ids=input_ids, attention_mask=attention_mask)258 extended_attention_mask = self.get_extended_attention_mask(attention_mask, input_shape)259 encoder_outputs = self.encoder(inputs_embeds, attention_mask=extended_attention_mask)260 sequence_output = encoder_outputs[0]261 262 return BaseModelOutputWithPoolingAndCrossAttentions(263 last_hidden_state=sequence_output,264 hidden_states=encoder_outputs.hidden_states,265 attentions=encoder_outputs.attentions,266 cross_attentions=encoder_outputs.cross_attentions,267 )268 269 def get_extended_attention_mask(270 self,271 attention_mask: Tensor,272 input_shape: tuple[int],273 device: Optional[torch.device] = None,274 dtype: Optional[torch.dtype] = None,275 ) -> Tensor:276 """277 Makes broadcastable attention and causal masks so that future and masked tokens are ignored.278 279 Arguments:280 attention_mask (`torch.Tensor`):281 Mask with ones indicating tokens to attend to, zeros for tokens to ignore.282 input_shape (`Tuple[int]`):283 The shape of the input to the model.284 285 Returns:286 `torch.Tensor` The extended attention mask, with a the same dtype as `attention_mask.dtype`.287 """288 if dtype is None:289 dtype = get_parameter_dtype(self)290 291 if not (attention_mask.dim() == 2 and self.config.is_decoder):292 # show warning only if it won't be shown in `create_extended_attention_mask_for_decoder`293 if device is not None:294 warnings.warn(295 "The `device` argument is deprecated and will be removed in v5 of Transformers.", FutureWarning296 )297 # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]298 # ourselves in which case we just need to make it broadcastable to all heads.299 if attention_mask.dim() == 3:300 extended_attention_mask = attention_mask[:, None, :, :]301 elif attention_mask.dim() == 2:302 # Provided a padding mask of dimensions [batch_size, seq_length]303 # - if the model is a decoder, apply a causal mask in addition to the padding mask304 # - if the model is an encoder, make the mask broadcastable to [batch_size, num_heads, seq_length, seq_length]305 if self.config.is_decoder:306 extended_attention_mask = ModuleUtilsMixin.create_extended_attention_mask_for_decoder(307 input_shape, attention_mask, device308 )309 else:310 extended_attention_mask = attention_mask[:, None, None, :]311 else:312 raise ValueError(313 f"Wrong shape for input_ids (shape {input_shape}) or attention_mask (shape {attention_mask.shape})"314 )315 316 # Since attention_mask is 1.0 for positions we want to attend and 0.0 for317 # masked positions, this operation will create a tensor which is 0.0 for318 # positions we want to attend and the dtype's smallest value for masked positions.319 # Since we are adding it to the raw scores before the softmax, this is320 # effectively the same as removing these entirely.321 extended_attention_mask = extended_attention_mask.to(dtype=dtype) # fp16 compatibility322 extended_attention_mask = (1.0 - extended_attention_mask) * torch.finfo(dtype).min323 return extended_attention_mask324 325 326class EvollaSequenceCompressorAttention(nn.Module):327 def __init__(self, dim, dim_head=64, heads=8):328 super().__init__()329 self.scale = dim_head**-0.5330 self.heads = heads331 inner_dim = dim_head * heads332 333 self.norm_media = nn.LayerNorm(dim)334 self.norm_latents = nn.LayerNorm(dim)335 336 self.to_q = nn.Linear(dim, inner_dim, bias=False)337 self.to_kv = nn.Linear(dim, inner_dim * 2, bias=False)338 self.to_out = nn.Linear(inner_dim, dim, bias=False)339 340 def forward(self, x, latents, mask):341 """342 Args:343 x (torch.Tensor): image features344 shape (b, n1, D)345 latent (torch.Tensor): latent features346 shape (b, n2, D); n2: num of latent tokens347 """348 x = self.norm_media(x)349 latents = self.norm_latents(latents)350 351 h = self.heads352 353 q = self.to_q(latents)354 kv_input = torch.cat((x, latents), dim=-2)355 k, v = self.to_kv(kv_input).chunk(356 2, dim=-1357 ) # each: batch_size, max_protein_length+num_latents, dim_head*num_heads358 359 q = q.view(q.size(0), q.size(1), h, -1).permute(0, 2, 1, 3)360 k = k.view(k.size(0), k.size(1), h, -1).permute(0, 2, 1, 3)361 v = v.view(v.size(0), v.size(1), h, -1).permute(0, 2, 1, 3)362 q = q * self.scale # batch_size, num_heads, num_latents, dim_head363 364 # attention365 sim = torch.matmul(q, k.transpose(-1, -2))366 sim = sim - sim.amax(dim=-1, keepdim=True).detach()367 bs, nh, skd, okd = sim.shape368 ones = torch.ones(nh, skd).to(mask.device) # Create a tensor of ones with shape (nh, skd)369 mask_exp = mask[:, None, None, :]370 ones_exp = ones[None, :, :, None]371 mask = mask_exp * ones_exp372 373 sim = sim.masked_fill((1 - mask).bool(), -1e4)374 attn = sim.softmax(dim=-1)375 out = torch.matmul(attn, v)376 out = out.permute(0, 2, 1, 3)377 378 # [batch, seq, head, features] -> [batch, seq, head*features]379 out = out.reshape(out.size(0), out.size(1), -1)380 381 return self.to_out(out)382 383 384class EvollaFeedForward(nn.Module):385 def __init__(self, dim, mult=4):386 super().__init__()387 inner_dim = int(dim * mult)388 389 self.norm = nn.LayerNorm(dim)390 self.fc1 = nn.Linear(dim, inner_dim, bias=False)391 self.activation = nn.GELU()392 self.fc2 = nn.Linear(inner_dim, dim, bias=False)393 394 def forward(self, x):395 return self.fc2(self.activation(self.fc1(self.norm(x))))396 397 398class EvollaSequenceCompressorResampler(nn.Module):399 def __init__(self, config: EvollaConfig):400 super().__init__()401 protein_repr_dim = config.protein_encoder_config.hidden_size402 self.num_latents = config.resampler_num_latents403 self.latents = nn.Parameter(torch.randn(self.num_latents, protein_repr_dim), requires_grad=True)404 self.layers = nn.ModuleList([])405 for _ in range(config.resampler_depth):406 self.layers.append(407 nn.ModuleList(408 [409 EvollaSequenceCompressorAttention(410 dim=protein_repr_dim, dim_head=config.resampler_dim_head, heads=config.resampler_heads411 ),412 EvollaFeedForward(dim=protein_repr_dim, mult=config.resampler_ff_mult),413 ]414 )415 )416 417 self.norm = nn.LayerNorm(config.hidden_size)418 self.protein_projector = nn.Linear(protein_repr_dim, config.hidden_size)419 420 def forward(self, embeds, mask):421 b = embeds.shape[0]422 423 bs, _ = mask.shape # bs, max_protein_length424 latent_mask = torch.ones(bs, self.num_latents).to(mask.device)425 mask = torch.cat((mask, latent_mask), dim=1) # bs, max_protein_length + num_latents426 427 # blocks428 ones = torch.ones(b).to(self.latents.device)429 latents = self.latents[None] * ones.view(-1, 1, 1) # [b,n,d]430 latents = latents.to(embeds.dtype)431 for attn, ff in self.layers:432 latents = attn(embeds, latents, mask) + latents433 latents = ff(latents) + latents434 435 transformed_feature = self.protein_projector(latents)436 437 return self.norm(transformed_feature)438 439 440@dataclass441@auto_docstring442class EvollaProteinEncoderModelOutput(ModelOutput):443 sequence_compressor_output: Optional[torch.FloatTensor] = None444 last_hidden_state: Optional[torch.FloatTensor] = None445 hidden_states: Optional[tuple[torch.FloatTensor, ...]] = None446 attentions: Optional[tuple[torch.FloatTensor, ...]] = None447 448 449class EvollaProteinEncoder(nn.Module):450 def __init__(self, config: EvollaConfig):451 super().__init__()452 self.model = EvollaSaProtProteinEncoder(config=config.protein_encoder_config)453 self.sequence_compressor_resampler = EvollaSequenceCompressorResampler(config=config)454 455 @can_return_tuple456 def forward(self, input_ids: torch.LongTensor, attention_mask: torch.FloatTensor, **kwargs):457 protein_output = self.model(input_ids=input_ids, attention_mask=attention_mask)458 protein_embeds = protein_output.last_hidden_state459 sequence_repr = self.sequence_compressor_resampler(protein_embeds, attention_mask)460 461 return EvollaProteinEncoderModelOutput(462 sequence_compressor_output=sequence_repr,463 last_hidden_state=protein_output.last_hidden_state,464 )465 466 467class EvollaSequenceAlignerCrossAttention(nn.Module):468 def __init__(469 self,470 config,471 protein_encoder_dim: Optional[int] = None,472 structure_encoder_dim: Optional[int] = None,473 msa_encoder_dim: Optional[int] = None,474 ):475 super().__init__()476 477 self.hidden_size = config.hidden_size478 self.num_attention_heads = config.num_attention_heads479 self.scale = self.num_attention_heads**-0.5480 self.attention_head_size = int(self.hidden_size / self.num_attention_heads)481 self.all_head_size = self.num_attention_heads * self.attention_head_size482 483 attention_probs_dropout_prob = config.aligner_attention_probs_dropout_prob484 enable_bias = config.aligner_enable_bias485 ffn_mult = config.aligner_ffn_mult486 487 self.query = nn.Linear(self.hidden_size, self.all_head_size)488 if protein_encoder_dim is not None:489 self.key_protein = nn.Linear(protein_encoder_dim, self.all_head_size)490 self.value_protein = nn.Linear(protein_encoder_dim, self.all_head_size)491 else:492 self.key_protein = None493 self.value_protein = None494 495 if structure_encoder_dim is not None:496 self.key_structure = nn.Linear(structure_encoder_dim, self.all_head_size)497 self.value_structure = nn.Linear(structure_encoder_dim, self.all_head_size)498 else:499 self.key_structure = None500 self.value_structure = None501 502 if msa_encoder_dim is not None:503 self.key_msa = nn.Linear(msa_encoder_dim, self.all_head_size)504 self.value_msa = nn.Linear(msa_encoder_dim, self.all_head_size)505 else:506 self.key_msa = None507 self.value_msa = None508 509 self.attention_norm = EvollaRMSNorm(self.hidden_size)510 511 self.dropout = nn.Dropout(attention_probs_dropout_prob)512 513 self.out_proj = nn.Linear(self.hidden_size, self.hidden_size, bias=enable_bias)514 515 self.ff = EvollaFeedForward(self.hidden_size, ffn_mult)516 self.gate_attention = nn.Parameter(torch.tensor([0.0]))517 self.gate_ffw = nn.Parameter(torch.tensor([0.0]))518 519 def cross_attention(520 self,521 query_states,522 protein_key_value_states,523 structure_key_value_states,524 msa_key_value_states,525 query_attn_mask,526 protein_kv_attn_mask,527 structure_kv_attn_mask,528 msa_kv_attn_mask,529 ):530 """531 query_states: text532 key_value_states: protein533 query_states: [bs, query_seq_len, dim]534 key_value_states: [bs, kv_seq_len, dim]535 query_attn_mask: [bs, query_seq_len]536 kv_attn_mask: [bs, kv_seq_len]537 """538 539 # Concatenate protein and structure540 kv_attn_mask = [protein_kv_attn_mask, structure_kv_attn_mask, msa_kv_attn_mask]541 kv_attn_mask = [_ for _ in kv_attn_mask if _ is not None]542 if not kv_attn_mask:543 raise ValueError("At least one modality should be provided for cross attention.")544 kv_attn_mask = torch.cat(kv_attn_mask, dim=1)545 546 query_layer = self.attention_norm(query_states)547 548 # Warning: This place might cause issues, refers to549 # https://discuss.pytorch.org/t/cuda-error-cublas-status-not-supported-when-calling-cublasltmatmul-from-torch-nn-functional-linear/170214/13550 # Solution: add `DISABLE_ADDMM_CUDA_LT=1` as environment variable551 # Apply linear transformation to input_query, input_key, and input_value552 query_layer = self.query(query_layer) # [bs, querylength, dim]553 554 if self.key_protein is not None and self.value_protein is not None:555 protein_key_value_states = protein_key_value_states.to(query_states)556 key_layer_protein = self.key_protein(protein_key_value_states) # [bs, keylength, dim]557 value_layer_protein = self.value_protein(protein_key_value_states) # [bs, keylength, dim]558 else:559 key_layer_protein = None560 value_layer_protein = None561 562 if self.key_structure is not None and self.value_structure is not None:563 structure_key_value_states = structure_key_value_states.to(query_states)564 key_layer_structure = self.key_structure(structure_key_value_states) # [bs, keylength, dim]565 value_layer_structure = self.value_structure(structure_key_value_states) # [bs, keylength, dim]566 else:567 key_layer_structure = None568 value_layer_structure = None569 570 if self.key_msa is not None and self.value_msa is not None:571 msa_key_value_states = msa_key_value_states.to(query_states)572 key_layer_msa = self.key_msa(msa_key_value_states) # [bs, keylength, dim]573 value_layer_msa = self.value_msa(msa_key_value_states) # [bs, keylength, dim]574 else:575 key_layer_msa = None576 value_layer_msa = None577 578 key_layer = [key_layer_protein, key_layer_structure, key_layer_msa]579 key_layer = [_ for _ in key_layer if _ is not None]580 key_layer = torch.cat(key_layer, dim=1)581 582 value_layer = [value_layer_protein, value_layer_structure, value_layer_msa]583 value_layer = [_ for _ in value_layer if _ is not None]584 value_layer = torch.cat(value_layer, dim=1)585 586 new_query_layer_shape = query_layer.size()[:-1] + (587 self.num_attention_heads,588 self.attention_head_size,589 )590 query_layer = query_layer.view(*new_query_layer_shape).permute(0, 2, 1, 3)591 592 new_key_layer_shape = key_layer.size()[:-1] + (593 self.num_attention_heads,594 self.attention_head_size,595 )596 key_layer = key_layer.view(*new_key_layer_shape).permute(0, 2, 1, 3)597 598 new_value_layer_shape = value_layer.size()[:-1] + (599 self.num_attention_heads,600 self.attention_head_size,601 )602 value_layer = value_layer.view(*new_value_layer_shape).permute(0, 2, 1, 3)603 604 query_layer = query_layer * self.scale605 606 # attention_mask: [bs, 1, querylength, keylength]607 if query_attn_mask is None:608 query_attn_mask = torch.ones(query_states.size(0), query_states.size(1)).to(query_states.device)609 attention_mask = query_attn_mask[:, None, :, None] * kv_attn_mask[:, None, None, :]610 # Compute the scaled dot-product attention scores611 attn_weights = torch.matmul(query_layer, key_layer.transpose(-1, -2)) # [bs, numheads, querylength, keylength]612 attn_weights = attn_weights - attn_weights.amax(dim=-1, keepdim=True).detach() # To stabilize score613 attention_scores = attn_weights.masked_fill(614 (1 - attention_mask).bool(), torch.finfo(attn_weights.dtype).min615 ) # [bs, numheads, querylength, keylength]616 617 attention_probs = nn.Softmax(dim=-1)(attention_scores)618 619 # attention_probs_dropped = self.dropout(attention_probs)620 621 context_layer = torch.matmul(attention_probs, value_layer) # [bs, numheads, querylength, dim/numheads]622 623 context_layer = context_layer.permute(0, 2, 1, 3).contiguous()624 new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)625 context_layer = context_layer.view(*new_context_layer_shape)626 627 context_layer = self.out_proj(context_layer)628 629 return context_layer630 631 @deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")632 def forward(633 self,634 query_states,635 protein_kv_states,636 structure_kv_states,637 msa_kv_states,638 query_attn_mask,639 protein_kv_attn_mask=None,640 structure_kv_attn_mask=None,641 msa_kv_attn_mask=None,642 protein_batch_mask=None,643 structure_batch_mask=None,644 msa_batch_mask=None,645 past_key_values=None,646 ):647 if protein_kv_states is not None:648 bs, protein_kv_seq_len, dim = protein_kv_states.shape649 if protein_kv_attn_mask is None:650 protein_kv_attn_mask = (651 torch.ones(bs, protein_kv_seq_len).to(protein_batch_mask.device)652 * protein_batch_mask.expand(size=(protein_kv_seq_len, bs)).T653 ).to(protein_kv_states.device)654 else:655 protein_kv_attn_mask = None656 657 if structure_kv_states is not None:658 bs, structure_kv_seq_len, dim = structure_kv_states.shape659 if structure_kv_attn_mask is None:660 structure_kv_attn_mask = (661 torch.ones(bs, structure_kv_seq_len).to(protein_batch_mask.device)662 * structure_batch_mask.expand(size=(structure_kv_seq_len, bs)).T663 ).to(structure_kv_states.device)664 else:665 structure_kv_attn_mask = None666 667 if msa_kv_states is not None:668 bs, msa_kv_seq_len, dim = msa_kv_states.shape669 if msa_kv_attn_mask is None:670 msa_kv_attn_mask = (671 torch.ones(bs, msa_kv_seq_len).to(protein_batch_mask.device)672 * msa_batch_mask.expand(size=(msa_kv_seq_len, bs)).T673 ).to(msa_kv_states.device)674 else:675 msa_kv_attn_mask = None676 hidden_states = query_states677 # only when there's at least one valid modality, crossattention will be performed678 if (679 (protein_kv_states is not None and protein_kv_attn_mask.any())680 or (structure_kv_states is not None and structure_kv_attn_mask.any())681 or (msa_kv_states is not None and msa_kv_attn_mask.any())682 ):683 residual = hidden_states684 hidden_states = self.cross_attention(685 query_states=hidden_states,686 protein_key_value_states=protein_kv_states,687 structure_key_value_states=structure_kv_states,688 msa_key_value_states=msa_kv_states,689 query_attn_mask=query_attn_mask,690 protein_kv_attn_mask=protein_kv_attn_mask,691 structure_kv_attn_mask=structure_kv_attn_mask,692 msa_kv_attn_mask=msa_kv_attn_mask,693 ) # [bs, query_seq_len, dim]694 # tanh gate695 hidden_states = torch.tanh(self.gate_attention) * hidden_states696 697 hidden_states = residual + hidden_states # input_query698 699 residual = hidden_states700 hidden_states = self.ff(hidden_states) * torch.tanh(self.gate_ffw)701 hidden_states = residual + hidden_states702 703 return hidden_states704 705 706class EvollaRMSNorm(LlamaRMSNorm):707 pass708 709 710class EvollaRotaryEmbedding(LlamaRotaryEmbedding):711 pass712 713 714class EvollaMLP(LlamaMLP):715 pass716 717 718class EvollaAttention(LlamaAttention):719 pass720 721 722class EvollaDecoderLayer(LlamaDecoderLayer):723 def __init__(self, config: EvollaConfig, layer_idx: int):724 super().__init__(config, layer_idx)725 if (layer_idx + 1) % max(config.num_hidden_layers // config.aligner_num_add_layers, 1) == 0:726 self.adapter = EvollaSequenceAlignerCrossAttention(727 config,728 protein_encoder_dim=config.hidden_size,729 )730 731 @deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")732 def forward(733 self,734 hidden_states: torch.Tensor,735 position_embeddings: tuple[torch.Tensor, torch.Tensor],736 attention_mask: Optional[torch.Tensor] = None,737 position_ids: Optional[torch.LongTensor] = None,738 past_key_values: Optional[Cache] = None,739 use_cache: Optional[bool] = False,740 cache_position: Optional[torch.LongTensor] = None,741 protein_kv_states: Optional[torch.Tensor] = None,742 structure_kv_states: Optional[torch.Tensor] = None,743 msa_kv_states: Optional[torch.Tensor] = None,744 protein_batch_mask: Optional[torch.Tensor] = None,745 structure_batch_mask: Optional[torch.Tensor] = None,746 msa_batch_mask: Optional[torch.Tensor] = None,747 query_attn_mask: Optional[torch.Tensor] = None,748 **kwargs,749 ):750 residual = hidden_states751 752 hidden_states = self.input_layernorm(hidden_states)753 754 # Self Attention755 hidden_states, _ = self.self_attn(756 hidden_states=hidden_states,757 attention_mask=attention_mask,758 position_ids=position_ids,759 past_key_values=past_key_values,760 use_cache=use_cache,761 cache_position=cache_position,762 position_embeddings=position_embeddings,763 **kwargs,764 )765 hidden_states = residual + hidden_states766 767 # Fully Connected768 residual = hidden_states769 hidden_states = self.post_attention_layernorm(hidden_states)770 hidden_states = self.mlp(hidden_states)771 hidden_states = residual + hidden_states772 773 if hasattr(self, "adapter"):774 hidden_states = self.adapter(775 query_states=hidden_states,776 protein_kv_states=protein_kv_states,777 structure_kv_states=structure_kv_states,778 msa_kv_states=msa_kv_states,779 query_attn_mask=query_attn_mask,780 protein_batch_mask=protein_batch_mask,781 structure_batch_mask=structure_batch_mask,782 msa_batch_mask=msa_batch_mask,783 )784 785 return hidden_states786 787 788class EvollaPreTrainedModel(LlamaPreTrainedModel):789 _supports_flash_attn = False # see dependency on `EvollaSaProtProteinEncoder`790 _supports_flex_attn = False # see dependency on `EvollaSaProtProteinEncoder`791 _supports_attention_backend = False792 _no_split_modules = [793 "EvollaDecoderLayer",794 "EvollaSequenceCompressorResampler",795 "EvollaSequenceAlignerCrossAttention",796 ]797 798 def _init_weights(self, module):799 std = self.config.initializer_range800 PreTrainedModel._init_weights(self, module)801 if isinstance(module, EvollaSequenceAlignerCrossAttention):802 module.gate_attention.zero_()803 module.gate_ffw.zero_()804 module.attention_norm.weight.data.fill_(1.0)805 elif isinstance(module, EvollaSequenceCompressorResampler):806 module.latents.data.normal_(mean=0.0, std=std)807 808 809class EvollaModel(EvollaPreTrainedModel):810 def __init__(self, config: EvollaConfig):811 super().__init__(config)812 self.padding_idx = config.pad_token_id813 self.vocab_size = config.vocab_size814 self.embed_tokens = nn.Embedding(self.vocab_size, config.hidden_size, self.padding_idx)815 self.protein_encoder = EvollaProteinEncoder(config=config)816 self.layers = nn.ModuleList(817 [818 EvollaDecoderLayer(819 config=config,820 layer_idx=layer_idx,821 )822 for layer_idx in range(config.num_hidden_layers)823 ]824 )825 826 self.norm = EvollaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)827 self.rotary_emb = EvollaRotaryEmbedding(config=config)828 self.gradient_checkpointing = getattr(config, "gradient_checkpointing", False)829 self.post_init()830 831 def get_input_embeddings(self):832 return self.embed_tokens833 834 def set_input_embeddings(self, value):835 self.embed_tokens = value836 837 @auto_docstring838 @check_model_inputs()839 def forward(840 self,841 input_ids: Optional[torch.LongTensor] = None,842 attention_mask: Optional[torch.Tensor] = None,843 position_ids: Optional[torch.LongTensor] = None,844 past_key_values: Optional[Cache] = None,845 inputs_embeds: Optional[torch.FloatTensor] = None,846 use_cache: Optional[bool] = None,847 cache_position: Optional[torch.LongTensor] = None,848 protein_input_ids: Optional[torch.LongTensor] = None,849 protein_attention_mask: Optional[torch.Tensor] = None,850 structure_feats: Optional[torch.FloatTensor] = None,851 msa_feats: Optional[torch.FloatTensor] = None,852 structure_batch_mask: Optional[torch.Tensor] = None,853 msa_batch_mask: Optional[torch.Tensor] = None,854 **kwargs,855 ) -> Union[tuple, BaseModelOutputWithPast]:856 r"""857 protein_input_ids (torch.LongTensor):858 The input IDs for the protein sequence in structure-aware tokens. Should be of shape `(batch_size, protein_seq_length)` and type `torch.LongTensor`.859 protein_attention_mask (torch.Tensor):860 The attention mask for the protein sequence. Should be of shape `(batch_size, protein_seq_length)` and type `torch.Tensor`.861 structure_feats (torch.FloatTensor):862 The input IDs for purely structure-based features. Should be of shape `(batch_size, structure_seq_length, structure_feat_dim)` and type `torch.FloatTensor`. Dummy input for now.863 msa_feats (torch.FloatTensor):864 The input IDs for purely MSA-based features. Should be of shape `(batch_size, msa_seq_length, msa_feat_dim)` and type `torch.FloatTensor`. Dummy input for now.865 structure_batch_mask (torch.Tensor):866 The batch mask to decide which protein sequences are purely structure-based. Should be of shape `(batch_size)` and type `torch.Tensor`. Should be paired with `structure_feats`. Dummpy input for now.867 msa_batch_mask (torch.Tensor):868 The batch mask to decide which protein sequences are purely MSA-based. Should be of shape `(batch_size)` and type `torch.Tensor`. Should be paired with `msa_feats`. Dummpy input for now.869 """870 if (input_ids is None) ^ (inputs_embeds is not None):871 raise ValueError("You must specify exactly one of input_ids or inputs_embeds")872 873 if inputs_embeds is None:874 inputs_embeds = self.embed_tokens(input_ids)875 876 if use_cache and past_key_values is None:877 past_key_values = DynamicCache(config=self.config)878 879 if cache_position is None:880 past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0881 cache_position = torch.arange(882 past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device883 )884 885 if position_ids is None:886 position_ids = cache_position.unsqueeze(0)887 888 protein_feats = None889 protein_batch_mask = None890 # If provided, actually compute them891 if protein_input_ids is not None and protein_attention_mask is not None:892 protein_outputs = self.protein_encoder(893 input_ids=protein_input_ids,894 attention_mask=protein_attention_mask,895 )896 protein_feats = protein_outputs.sequence_compressor_output897 protein_batch_mask = torch.tensor([True] * protein_input_ids.shape[0], device=protein_input_ids.device)898 899 causal_mask = create_causal_mask(900 config=self.config,901 input_embeds=inputs_embeds,902 attention_mask=attention_mask,903 cache_position=cache_position,904 past_key_values=past_key_values,905 )906 907 hidden_states = inputs_embeds908 909 # create position embeddings to be shared across the decoder layers910 position_embeddings = self.rotary_emb(hidden_states, position_ids)911 912 for decoder_layer in self.layers:913 hidden_states = decoder_layer(914 hidden_states,915 attention_mask=causal_mask,916 position_ids=position_ids,917 past_key_values=past_key_values,918 use_cache=use_cache,919 cache_position=cache_position,920 position_embeddings=position_embeddings,921 protein_kv_states=protein_feats,922 structure_kv_states=structure_feats,923 msa_kv_states=msa_feats,924 protein_batch_mask=protein_batch_mask,925 structure_batch_mask=structure_batch_mask,926 msa_batch_mask=msa_batch_mask,927 query_attn_mask=attention_mask,928 **kwargs,929 )930 931 hidden_states = self.norm(hidden_states)932 933 output = BaseModelOutputWithPast(934 last_hidden_state=hidden_states,935 past_key_values=past_key_values,936 )937 return output938 939 940class EvollaForProteinText2Text(EvollaPreTrainedModel, GenerationMixin):941 def __init__(self, config):942 super().__init__(config)943 self.model = EvollaModel(config)944 self.vocab_size = config.vocab_size945 self.lm_head = nn.Linear(config.hidden_size, self.vocab_size, bias=False)946 947 self.post_init()948 949 def get_input_embeddings(self):950 return self.model.get_input_embeddings()951 952 def set_input_embeddings(self, value):953 return self.model.set_input_embeddings(value)954 955 @can_return_tuple956 @auto_docstring957 def forward(958 self,959 input_ids: Optional[torch.LongTensor] = None, # text input ids960 attention_mask: Optional[torch.Tensor] = None, # text attention mask961 inputs_embeds: Optional[torch.FloatTensor] = None, # text input embeddings962 labels: Optional[torch.LongTensor] = None,963 protein_input_ids: Optional[torch.LongTensor] = None,964 protein_attention_mask: Optional[torch.Tensor] = None,965 use_cache: Optional[bool] = None,966 **kwargs,967 ):968 r"""969 protein_input_ids (torch.LongTensor):970 The input IDs for the protein sequence. Should be of shape `(batch_size, protein_seq_length)` and type `torch.LongTensor`.971 protein_attention_mask (torch.Tensor):972 The attention mask for the protein sequence. Should be of shape `(batch_size, protein_seq_length)` and type `torch.Tensor`.973 974 Example:975 976 ```python977 >>> from transformers import EvollaProcessor, EvollaForProteinText2Text978 >>> model = EvollaForProteinText2Text.from_pretrained("westlake/Evolla-10B-hf")979 >>> processor = EvollaProcessor.from_pretrained("westlake/Evolla-10B-hf")980 981 >>> protein_information = {982 "aa_seq": "your amino acid sequence",983 "foldseek": "your foldseek sequence",984 }985 >>> question = "What is the function of this protein?"986 >>> message = [987 {"role": "system", "content": "You are an AI expert that can answer any questions about protein."},988 {"role": "user", "content": question},989 ]990 991 >>> inputs = processor(proteins=[protein_information], messages_list=[message], return_tensors="pt", padding="longest")992 >>> outputs = model.generate(**inputs)993 994 >>> print(processor.batch_decode(outputs, skip_special_tokens=True))995 ```"""996 997 outputs = self.model(998 input_ids=input_ids,999 attention_mask=attention_mask,1000 inputs_embeds=inputs_embeds,1001 protein_input_ids=protein_input_ids,1002 protein_attention_mask=protein_attention_mask,1003 use_cache=use_cache,1004 **kwargs,1005 )1006 hidden_states = outputs[0]1007 logits = self.lm_head(hidden_states)1008 1009 loss = None1010 if labels is not None:1011 loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.vocab_size, **kwargs)1012 1013 lm_outputs = CausalLMOutputWithPast(1014 loss=loss,1015 logits=logits,1016 past_key_values=outputs.past_key_values,1017 hidden_states=outputs.hidden_states,1018 attentions=outputs.attentions,1019 )1020 return lm_outputs1021 1022 1023__all__ = ["EvollaForProteinText2Text", "EvollaModel", "EvollaPreTrainedModel"]1024