File size: 24,652 Bytes
3b2d368 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 | import math
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.nn.attention.flex_attention import flex_attention
from transformers import PreTrainedModel, GenerationMixin
from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
from transformers.cache_utils import Cache, DynamicCache
from rotary_embedding_torch import RotaryEmbedding
from .config import FSTConfig
# Allows for easier hooking
class Residual(nn.Module):
def __init__(self):
super().__init__()
def forward(self, x, delta):
return x + delta
class PhiAttention(nn.Module):
def __init__(self, config, layer_idx=None):
super().__init__()
self.layer_idx = layer_idx
self.use_causal_attention = config.use_causal_attention
self.hidden_size = config.hidden_size_phi
self.embedding_size = config.embedding_size_phi
self.num_heads = config.num_attention_heads_phi
self.head_dim = self.hidden_size // self.num_heads
assert self.head_dim * self.num_heads == self.hidden_size
self.q_proj = nn.Linear(self.embedding_size, self.hidden_size, bias=False)
self.k_proj = nn.Linear(self.embedding_size, self.hidden_size, bias=False)
self.v_proj = nn.Linear(self.embedding_size, self.hidden_size, bias=True)
self.o_proj = nn.Linear(self.hidden_size, self.embedding_size, bias=True)
self.rotary_emb = RotaryEmbedding(dim=self.head_dim)
self.scale = self.head_dim ** -0.5
def forward(self, x, attention_mask=None, past_key_values=None):
B, T, _ = x.size()
q = self.q_proj(x)
k = self.k_proj(x)
v = self.v_proj(x)
q = q.view(B, T, self.num_heads, self.head_dim).transpose(1, 2)
k = k.view(B, T, self.num_heads, self.head_dim).transpose(1, 2)
v = v.view(B, T, self.num_heads, self.head_dim).transpose(1, 2)
if past_key_values is None:
q = self.rotary_emb.rotate_queries_or_keys(q)
k = self.rotary_emb.rotate_queries_or_keys(k)
else:
k_cache, v_cache = past_key_values[self.layer_idx] if self.layer_idx < len(past_key_values) else (None, None)
k_len = k_cache.shape[-2] if k_cache is not None else 0
q = self.rotary_emb.rotate_queries_or_keys(q, offset=k_len)
k = self.rotary_emb.rotate_queries_or_keys(k, offset=k_len)
past_key_values.update(k, v, self.layer_idx)
if k_cache is not None and v_cache is not None:
k = torch.cat([k_cache, k], dim=-2)
v = torch.cat([v_cache, v], dim=-2)
# Uses "is_causal" when possible for efficiency
attn_output = F.scaled_dot_product_attention(q, k, v, attn_mask=attention_mask, scale=self.scale, is_causal=(self.use_causal_attention and attention_mask is None))
attn_output = attn_output.transpose(1, 2).contiguous().view(B, T, self.hidden_size)
out = self.o_proj(attn_output)
return out
class PhiMLP(nn.Module):
def __init__(self, config):
super().__init__()
self.fc_up = nn.Linear(config.embedding_size_phi, config.intermediate_size_phi)
self.activation = nn.GELU()
self.fc_down = nn.Linear(config.intermediate_size_phi, config.embedding_size_phi)
def forward(self, x):
return self.fc_down(self.activation(self.fc_up(x)))
class PhiBlock(nn.Module):
def __init__(self, config, layer_idx=None):
super().__init__()
self.layer_idx = layer_idx
self.ln_attn = nn.LayerNorm(config.embedding_size_phi)
self.attn = PhiAttention(config, layer_idx=layer_idx)
self.resid_attn = Residual()
self.ln_mlp = nn.LayerNorm(config.embedding_size_phi)
self.mlp = PhiMLP(config)
self.resid_mlp = Residual()
def forward(self, x, attention_mask=None, past_key_values=None):
attn_out = self.attn(self.ln_attn(x), attention_mask=attention_mask, past_key_values=past_key_values)
x = self.resid_attn(x, attn_out)
mlp_out = self.mlp(self.ln_mlp(x))
x = self.resid_mlp(x, mlp_out)
return x
class ICLAttention(nn.Module):
def __init__(self, config, layer_idx=None):
super().__init__()
self.layer_idx = layer_idx
self.use_causal_attention = config.use_causal_attention
self.hidden_size = config.hidden_size_f
self.embedding_size_f = config.embedding_size_f
self.embedding_size_phi = config.embedding_size_phi
self.num_heads = config.num_attention_heads_f
self.head_dim = self.hidden_size // self.num_heads
assert self.head_dim * self.num_heads == self.hidden_size
self.q_proj = nn.Linear(self.embedding_size_phi, self.hidden_size, bias=False)
self.k_proj = nn.Linear(self.embedding_size_phi, self.hidden_size, bias=False)
self.v_proj = nn.Linear(self.embedding_size_f, self.hidden_size, bias=True)
self.o_proj = nn.Linear(self.hidden_size, self.embedding_size_f, bias=True)
self.rotary_emb = RotaryEmbedding(dim=self.head_dim)
self.scale = self.head_dim ** -0.5
def forward(self, q, k, v, attention_mask=None, past_key_values=None):
B, T, _ = q.size()
q = self.q_proj(q)
k = self.k_proj(k)
v = self.v_proj(v)
q = q.view(B, T, self.num_heads, self.head_dim).transpose(1, 2)
k = k.view(B, T, self.num_heads, self.head_dim).transpose(1, 2)
v = v.view(B, T, self.num_heads, self.head_dim).transpose(1, 2)
if past_key_values is None:
q = self.rotary_emb.rotate_queries_or_keys(q)
k = self.rotary_emb.rotate_queries_or_keys(k)
else:
k_cache, v_cache = past_key_values[self.layer_idx] if self.layer_idx < len(past_key_values) else (None, None)
k_len = k_cache.shape[-2] if k_cache is not None else 0
q = self.rotary_emb.rotate_queries_or_keys(q, offset=k_len)
k = self.rotary_emb.rotate_queries_or_keys(k, offset=k_len)
past_key_values.update(k, v, self.layer_idx)
if k_cache is not None and v_cache is not None:
k = torch.cat([k_cache, k], dim=-2)
v = torch.cat([v_cache, v], dim=-2)
# Uses "is_causal" when possible for efficiency
attn_output = F.scaled_dot_product_attention(q, k, v, attn_mask=attention_mask, scale=self.scale, is_causal=(self.use_causal_attention and attention_mask is None))
attn_output = attn_output.transpose(1, 2).contiguous().view(B, T, self.hidden_size)
out = self.o_proj(attn_output)
return out
class ICLMLP(nn.Module):
def __init__(self, config):
super().__init__()
self.fc_up = nn.Linear(config.embedding_size_f, config.intermediate_size_f)
self.activation = nn.GELU()
self.fc_down = nn.Linear(config.intermediate_size_f, config.embedding_size_f)
def forward(self, x):
return self.fc_down(self.activation(self.fc_up(x)))
class ICLBlock(nn.Module):
def __init__(self, config, layer_idx=None):
super().__init__()
self.layer_idx = layer_idx
self.ln_attn_qk = nn.LayerNorm(config.embedding_size_phi)
self.ln_attn_v = nn.LayerNorm(config.embedding_size_f)
self.attn = ICLAttention(config, layer_idx=layer_idx)
self.resid_attn = Residual()
self.ln_mlp = nn.LayerNorm(config.embedding_size_f)
self.mlp = ICLMLP(config)
self.resid_mlp = Residual()
def forward(self, phi, e, f, attention_mask=None, past_key_values=None):
qk = self.ln_attn_qk(phi)
v = self.ln_attn_v(e)
attn_out = self.attn(
q = qk,
k = qk,
v = v,
attention_mask=attention_mask,
past_key_values=past_key_values
)
f = self.resid_attn(f, attn_out)
mlp_out = self.mlp(self.ln_mlp(f))
f = self.resid_mlp(f, mlp_out)
return f
class FSTPreTrainedModel(PreTrainedModel):
config_class = FSTConfig
base_model_prefix = "model"
_no_split_modules = ["PhiBlock", "ICLBlock"]
_skip_keys_device_placement = ["past_key_values"]
_supports_flash_attn_2 = True
_supports_cache_class = True
# Initialization borrowed from Deepseek and Falcon
def _init_weights(self, module):
std = self.config.initializer_range
if isinstance(module, nn.Linear):
module.weight.data.normal_(mean=0.0, std=std)
if module.bias is not None:
module.bias.data.zero_()
elif isinstance(module, nn.Embedding):
module.weight.data.normal_(mean=0.0, std=std)
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)
def calculate_loss(self, logits, target_tokens, l1_loss_lambda=None):
# print(logits)
loss = F.cross_entropy(
logits.reshape(-1, logits.size(-1)),
target_tokens.reshape(-1),
reduction='mean'
)
return loss
def count_parameters(self):
total_params = sum(p.numel() for p in self.parameters())
embed_params = sum(p.numel() for name, p in self.named_parameters() if "embed" in name.lower())
non_embed_params = total_params - embed_params
return total_params, embed_params, non_embed_params
class FSTModel(FSTPreTrainedModel):
def __init__(self, config):
super().__init__(config)
self.config = config
self.embedding_f = nn.Embedding(config.vocab_size, config.embedding_size_f)
if not config.share_f_and_phi_embedding:
self.embedding_phi = nn.Embedding(config.vocab_size, config.embedding_size_phi)
self.phi_blocks = nn.ModuleList([PhiBlock(config, layer_idx) for layer_idx in range(0, config.num_hidden_layers, 2)])
self.icl_blocks = nn.ModuleList([ICLBlock(config, layer_idx) for layer_idx in range(1, config.num_hidden_layers, 2)])
self.ln_out = nn.LayerNorm(config.embedding_size)
self.post_init()
def _to_dynamic_cache(self, past_key_values):
cache = DynamicCache()
for i, (k, v) in enumerate(past_key_values):
cache.update({"prev_key": k, "prev_value": v}, layer_idx=i)
return cache
def _prepare_causal_attention_mask(self, x, attention_mask=None, past_key_values=None, strictly_causal=False):
device = x.device
B = x.shape[0]
T = x.shape[1]
T_past = past_key_values.get_seq_length() if past_key_values is not None else 0
T_total = T + T_past
# Diagonal shift effectively does nothing during training or inference, but here for compatibility
causal_mask = torch.triu(torch.ones((T, T_total), dtype=torch.bool, device=device), diagonal=(T_past + (0 if strictly_causal else 1))).unsqueeze(0).unsqueeze(0) # [1, 1, T, S]
# Combine with existing attention mask
if attention_mask is not None:
attn_len = attention_mask.shape[-1]
# If passed attention mask is too small (ex. excludes past tokens), pad left
if attn_len < T_total:
pad = torch.zeros(B, T_past, device=device, dtype=attention_mask.dtype)
attention_mask = torch.cat([pad, attention_mask], dim=-1)
# If passed attention mask is too big, clip to match sequence length
elif attn_len > T_total:
attention_mask = attention_mask[:, -T_total:]
expanded_mask = (attention_mask == 0).view(B, 1, 1, T_total)
causal_mask = causal_mask | expanded_mask
return causal_mask
def forward(
self,
input_ids=None,
attention_mask=None,
inputs_embeds=None,
past_key_values=None,
use_cache=None,
output_hidden_states=None,
return_dict=None,
**kwargs
):
use_cache = use_cache if use_cache is not None else self.config.use_cache
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 inputs_embeds is None:
e = self.embedding_f(input_ids)
else:
assert input_ids is None, "You cannot specify both input_ids and inputs_embeds"
assert self.config.share_f_and_phi_embedding, "You cannot specify inputs_embeds unless share_f_and_phi_embedding=True"
e = inputs_embeds
B, T, _ = e.shape
device = e.device
if self.config.share_f_and_phi_embedding:
phi = e
else:
phi = self.embedding_phi(input_ids)
f = torch.zeros(B, T, self.config.embedding_size_f, device=device) # Initialize f as zero
if not use_cache:
past_key_values=None
elif past_key_values is None:
past_key_values = DynamicCache()
elif isinstance(past_key_values, (tuple, list)):
past_key_values = self._to_dynamic_cache(past_key_values)
if self.config.use_causal_attention:
attention_mask_phi = None if attention_mask is None else self._prepare_causal_attention_mask(e, attention_mask=attention_mask, past_key_values=past_key_values) # Will automatically apply is_causal
attention_mask_f = None if attention_mask is None else self._prepare_causal_attention_mask(e, attention_mask=attention_mask, past_key_values=past_key_values)
else:
attention_mask_phi = None
attention_mask_f = None
hidden_states = [] if output_hidden_states else None
for phi_block, icl_block in zip(self.phi_blocks, self.icl_blocks):
phi = phi_block(phi, attention_mask=attention_mask_phi, past_key_values=past_key_values)
f = icl_block(phi, e, f, attention_mask=attention_mask_f, past_key_values=past_key_values)
if output_hidden_states:
hidden_states.append(phi)
hidden_states.append(f)
f = self.ln_out(f)
if return_dict:
return BaseModelOutputWithPast(
last_hidden_state=f,
past_key_values=past_key_values,
hidden_states=hidden_states
)
return f, past_key_values, hidden_states
class FSTForCausalLM(GenerationMixin, FSTPreTrainedModel):
accepts_loss_kwargs = False
def __init__(self, config):
super().__init__(config)
self.model = FSTModel(config)
self.lm_head = nn.Linear(config.embedding_size_f, config.vocab_size, bias=False)
print(config.tie_word_embeddings)
print('--------------------------------------------------------------------------------------------------------')
if config.tie_word_embeddings:
self.tie_weights()
self._dynamic_tied_weights_keys = {"lm_head.weight": "model.embedding_f.weight"} # Avoids safetensor naming issues
self.post_init()
def get_input_embeddings(self):
return self.model.embedding_f
def set_input_embeddings(self, new_embeddings):
self.model.embedding_f = new_embeddings
def get_output_embeddings(self):
return self.lm_head
def set_output_embeddings(self, new_embeddings):
self.lm_head = new_embeddings
def tie_weights(self):
self._tie_or_clone_weights(self.lm_head, self.get_input_embeddings())
def forward(
self,
input_ids=None,
attention_mask=None,
past_key_values=None,
inputs_embeds=None,
labels=None,
use_cache=None,
output_hidden_states=None,
return_dict=None,
**kwargs
):
if labels is not None:
return_dict = True
else:
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
model_output = self.model(
input_ids=input_ids,
attention_mask=attention_mask,
inputs_embeds=inputs_embeds,
past_key_values=past_key_values,
use_cache=use_cache,
output_hidden_states=output_hidden_states
)
logits = self.lm_head(model_output[0])
loss = None
if labels is not None:
shift_logits = logits[:, :-1, :].contiguous()
shift_labels = labels[:, 1:].contiguous()
loss = F.cross_entropy(
shift_logits.view(-1, shift_logits.size(-1)),
shift_labels.view(-1),
ignore_index=self.config.pad_token_id
)
if not return_dict:
output = (logits,) + model_output[1:]
return ((loss,) + output) if loss is not None else output
return logits
def _prepare_inputs_for_generation(self, input_ids, past_key_values=None, attention_mask=None, **kwargs):
if past_key_values is not None:
input_ids = input_ids[:, -1:]
model_inputs = {"input_ids": input_ids, "past_key_values": past_key_values, "use_cache": True}
if attention_mask is not None:
model_inputs["attention_mask"] = attention_mask
for key, value in kwargs.items():
model_inputs[key] = value
return model_inputs
def _reorder_cache(self, past_key_values, beam_idx):
reordered_past = []
for layer_past in past_key_values:
reordered_past.append(tuple(past_state.index_select(0, beam_idx) for past_state in layer_past))
return tuple(reordered_past)
@torch.no_grad()
def generate(
self,
input_ids,
max_generation_length,
tokenizer,
temperature=1.0,
top_p=0.9,
return_generation_only=False
):
self.eval()
batch_size = input_ids.size(0)
device = input_ids.device
generated = input_ids.clone()
finished = torch.zeros(batch_size, dtype=torch.bool, device=device)
for _ in range(max_generation_length):
logits = self(generated)[:, -1, :] / temperature
probs = F.softmax(logits, dim=-1)
sorted_probs, sorted_indices = torch.sort(probs, dim=-1, descending=True)
cumulative_probs = torch.cumsum(sorted_probs, dim=-1)
cutoff_mask = cumulative_probs > top_p
cutoff_mask[:, 1:] = cutoff_mask[:, :-1].clone()
cutoff_mask[:, 0] = False
sorted_probs = sorted_probs.masked_fill(cutoff_mask, 0.0)
normalized_probs = sorted_probs / sorted_probs.sum(dim=-1, keepdim=True)
probs = torch.zeros_like(normalized_probs).scatter(-1, sorted_indices, normalized_probs)
next_token = torch.multinomial(probs, num_samples=1).squeeze(-1)
next_token = torch.where(finished, torch.full_like(next_token, tokenizer.pad_token_id), next_token)
generated = torch.cat([generated, next_token.unsqueeze(1)], dim=1)
finished |= next_token == tokenizer.eos_token_id
if finished.all():
break
if return_generation_only:
return generated[:, input_ids.size(1):]
else:
return generated
from transformers.modeling_outputs import MaskedLMOutput
class FSTForMaskedLM(FSTPreTrainedModel):
"""
Masked Language Modeling head for Phi / ICL FST.
"""
accepts_loss_kwargs = False
def __init__(self, config: FSTConfig):
super().__init__(config)
assert not config.use_causal_attention, \
"FSTForMaskedLM requires use_causal_attention=False"
assert not config.use_cache, \
"FSTForMaskedLM requires use_cache=False"
self.model = FSTModel(config)
self.lm_head = nn.Linear(
config.embedding_size_f,
config.vocab_size,
bias=False,
)
if config.tie_word_embeddings:
self.tie_weights()
self._dynamic_tied_weights_keys = {
"lm_head.weight": "model.embedding_f.weight"
}
self.post_init()
def get_input_embeddings(self):
return self.model.embedding_f
def get_output_embeddings(self):
return self.lm_head
def set_output_embeddings(self, new_embeddings):
self.lm_head = new_embeddings
def tie_weights(self):
self._tie_or_clone_weights(
self.lm_head,
self.get_input_embeddings(),
)
def forward(
self,
input_ids=None,
attention_mask=None,
inputs_embeds=None,
labels=None,
output_hidden_states=None,
return_dict=None,
**kwargs,
):
return_dict = (
return_dict
if return_dict is not None
else self.config.use_return_dict
)
outputs = self.model(
input_ids=input_ids,
attention_mask=attention_mask,
inputs_embeds=inputs_embeds,
past_key_values=None,
use_cache=False,
output_hidden_states=output_hidden_states,
)
hidden = (
outputs.last_hidden_state
if return_dict
else outputs[0]
)
logits = self.lm_head(hidden)
loss = None
if labels is not None:
loss = F.cross_entropy(
logits.view(-1, logits.size(-1)),
labels.view(-1),
ignore_index=self.config.pad_token_id,
)
if not return_dict:
output = (logits,) + outputs[1:]
return ((loss,) + output) if loss is not None else output
return MaskedLMOutput(
loss=loss,
logits=logits,
hidden_states=outputs.hidden_states,
)
from transformers.modeling_outputs import SequenceClassifierOutput
class FSTForSequenceClassification(FSTPreTrainedModel):
"""
Sequence classification head for Phi / ICL FST.
"""
accepts_loss_kwargs = False
def __init__(self, config: FSTConfig):
super().__init__(config)
self.num_labels = getattr(config, "num_labels", 2)
self.model = FSTModel(config)
self.classifier = nn.Linear(
config.embedding_size_f,
self.num_labels,
)
self.post_init()
def forward(
self,
input_ids=None,
attention_mask=None,
inputs_embeds=None,
labels=None,
output_hidden_states=None,
return_dict=None,
**kwargs,
):
return_dict = (
return_dict
if return_dict is not None
else self.config.use_return_dict
)
outputs = self.model(
input_ids=input_ids,
attention_mask=attention_mask,
inputs_embeds=inputs_embeds,
past_key_values=None,
use_cache=False,
output_hidden_states=output_hidden_states,
)
hidden = (
outputs.last_hidden_state
if return_dict
else outputs[0]
)
# === CLS pooling ===
pooled = hidden[:, 0, :]
logits = self.classifier(pooled)
loss = None
if labels is not None:
if self.num_labels == 1:
# Regression
loss = F.mse_loss(
logits.view(-1),
labels.view(-1),
)
else:
# Classification
loss = F.cross_entropy(
logits.view(-1, self.num_labels),
labels.view(-1).long(),
)
if not return_dict:
output = (logits,) + outputs[1:]
return ((loss,) + output) if loss is not None else output
return SequenceClassifierOutput(
loss=loss,
logits=logits,
hidden_states=outputs.hidden_states,
attentions=None,
)
|