File size: 30,036 Bytes
85ba398 |
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 |
# Copyright (c) Facebook, Inc. and its affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
from typing import Dict, List, Tuple, Optional
import torch
import torch.nn as nn
from fairseq import utils
from fairseq.modules import LayerNorm, MultiheadAttention
from fairseq.modules.fairseq_dropout import FairseqDropout
from fairseq.modules.quant_noise import quant_noise
from torch import Tensor
class CrossChannelTransformerDecoderLayer(nn.Module):
"""Cross-Attention Transformer Decoder Layer block as described
in the paper: https://arxiv.org/pdf/2203.16502.pdf
Composed of a Multi-head Self Attention block followed by a
Multi-head Cross-Attention block which attends to the self-attention
outputs of the other channels. The weights of the attention blocks
in all channels are shared.
Args:
args (argparse.Namespace): parsed command-line arguments
no_encoder_attn (bool, optional): whether to attend to encoder outputs
(default: False).
"""
def __init__(
self, args, no_encoder_attn=False, add_bias_kv=False, add_zero_attn=False
):
super().__init__()
self.embed_dim = args.decoder_embed_dim
self.dropout_module = FairseqDropout(
args.dropout, module_name=self.__class__.__name__
)
self.quant_noise = getattr(args, "quant_noise_pq", 0)
self.quant_noise_block_size = getattr(args, "quant_noise_pq_block_size", 8)
# This cross_self_attention is used for encoder-decoder systems,
# It's not the cross-channel attention (defined below as cross_channel_attn)
self.cross_self_attention = getattr(args, "cross_self_attention", False)
self.self_attn = self.build_self_attention(
self.embed_dim,
args,
add_bias_kv=add_bias_kv,
add_zero_attn=add_zero_attn,
)
self.cross_channel_attn = self.build_cross_channel_attention(
self.embed_dim,
args,
add_bias_kv=add_bias_kv,
add_zero_attn=add_zero_attn,
)
self.activation_fn = utils.get_activation_fn(
activation=str(args.activation_fn)
if getattr(args, "activation_fn", None) is not None
else "relu"
)
activation_dropout_p = getattr(args, "activation_dropout", 0) or 0
if activation_dropout_p == 0:
# for backwards compatibility with models that use args.relu_dropout
activation_dropout_p = getattr(args, "relu_dropout", 0) or 0
self.activation_dropout_module = FairseqDropout(
float(activation_dropout_p), module_name=self.__class__.__name__
)
self.normalize_before = args.decoder_normalize_before
# use layerNorm rather than FusedLayerNorm for exporting.
# char_inputs can be used to determint this.
# TODO remove this once we update apex with the fix
export = getattr(args, "char_inputs", False)
self.self_attn_layer_norm = LayerNorm(self.embed_dim, export=export)
self.cross_channel_attn_layer_norm = LayerNorm(self.embed_dim, export=export)
if no_encoder_attn:
self.encoder_attn = None
self.encoder_attn_layer_norm = None
else:
self.encoder_attn = self.build_encoder_attention(self.embed_dim, args)
self.encoder_attn_layer_norm = LayerNorm(self.embed_dim, export=export)
self.fc1 = self.build_fc1(
self.embed_dim,
args.decoder_ffn_embed_dim,
self.quant_noise,
self.quant_noise_block_size,
)
self.fc2 = self.build_fc2(
args.decoder_ffn_embed_dim,
self.embed_dim,
self.quant_noise,
self.quant_noise_block_size,
)
self.final_layer_norm = LayerNorm(self.embed_dim, export=export)
self.need_attn = True
self.onnx_trace = False
def build_fc1(self, input_dim, output_dim, q_noise, qn_block_size):
return quant_noise(nn.Linear(input_dim, output_dim), q_noise, qn_block_size)
def build_fc2(self, input_dim, output_dim, q_noise, qn_block_size):
return quant_noise(nn.Linear(input_dim, output_dim), q_noise, qn_block_size)
def build_self_attention(
self, embed_dim, args, add_bias_kv=False, add_zero_attn=False
):
return MultiheadAttention(
embed_dim,
args.decoder_attention_heads,
dropout=args.attention_dropout,
add_bias_kv=add_bias_kv,
add_zero_attn=add_zero_attn,
self_attention=not getattr(args, "cross_self_attention", False),
q_noise=self.quant_noise,
qn_block_size=self.quant_noise_block_size,
)
def build_cross_channel_attention(
self, embed_dim, args, add_bias_kv=False, add_zero_attn=False
):
return MultiheadAttention(
embed_dim,
args.decoder_attention_heads,
dropout=args.attention_dropout,
add_bias_kv=add_bias_kv,
add_zero_attn=add_zero_attn,
self_attention=False,
q_noise=self.quant_noise,
qn_block_size=self.quant_noise_block_size,
)
def build_encoder_attention(self, embed_dim, args):
return MultiheadAttention(
embed_dim,
args.decoder_attention_heads,
kdim=getattr(args, "encoder_embed_dim", None),
vdim=getattr(args, "encoder_embed_dim", None),
dropout=args.attention_dropout,
encoder_decoder_attention=True,
q_noise=self.quant_noise,
qn_block_size=self.quant_noise_block_size,
)
def prepare_for_onnx_export_(self):
self.onnx_trace = True
def residual_connection(self, x, residual):
return residual + x
def forward(
self,
x_list_tensor: List[torch.Tensor],
encoder_out: Optional[torch.Tensor] = None,
encoder_padding_mask: Optional[torch.Tensor] = None,
incremental_state: Optional[
List[Dict[str, Dict[str, Optional[Tensor]]]]
] = None,
prev_self_attn_state: Optional[List[Tuple[torch.Tensor, torch.Tensor]]] = None,
prev_attn_state: Optional[List[torch.Tensor]] = None,
self_attn_mask: Optional[torch.Tensor] = None,
self_attn_padding_mask: Optional[torch.Tensor] = None,
need_attn: bool = False,
need_head_weights: bool = False,
):
"""
Args:
x_list_tensor (List[Tensor]): list of input tensors in different channels,
each tensor is of shape `(seq_len, batch, embed_dim)`
encoder_padding_mask (ByteTensor, optional): binary
ByteTensor of shape `(batch, src_len)` where padding
elements are indicated by ``1``.
incremental_state (optional): list of incremental_state dictionaries over
different channels (sequence generation mode)
prev_self_attn_state (List[Tuple[Tensor, Tensor]], optional): list of tuples
(self_attn_state, cross_channel_attn_state) over different channels
need_attn (bool, optional): return attention weights
need_head_weights (bool, optional): return attention weights
for each head (default: return average over heads).
Returns:
list of encoded output of shape `(seq_len, batch, embed_dim)`
"""
n_channels = len(x_list_tensor)
if need_head_weights:
need_attn = True
# incremental_state is a list of dictionaries over different channels
if incremental_state is not None:
assert isinstance(incremental_state, list)
assert len(incremental_state) == n_channels
# prev_self_attn_state is a list of tuples (self_attn_state, cross_channel_attn_state) over different channels
if prev_self_attn_state is not None:
assert isinstance(prev_self_attn_state, list)
assert len(prev_self_attn_state) == n_channels
for prev_self_attn_state_channel in prev_self_attn_state:
assert isinstance(prev_self_attn_state_channel, tuple)
assert len(prev_self_attn_state_channel) == 2
# Backup for other channels & cross channel attention
self_attn_mask_orin = self_attn_mask
self_attn_padding_mask_orin = self_attn_padding_mask
x_list = []
attn_list = []
for i, x in enumerate(x_list_tensor):
residual = x
if self.normalize_before:
x = self.self_attn_layer_norm(x)
if prev_self_attn_state is not None:
prev_key, prev_value = prev_self_attn_state[i][0][:2]
saved_state: Dict[str, Optional[Tensor]] = {
"prev_key": prev_key,
"prev_value": prev_value,
}
if len(prev_self_attn_state[i][0]) >= 3:
saved_state["prev_key_padding_mask"] = prev_self_attn_state[i][0][2]
assert incremental_state is not None
self.self_attn._set_input_buffer(incremental_state[i], saved_state)
_self_attn_input_buffer = self.self_attn._get_input_buffer(
incremental_state[i] if incremental_state is not None else None
)
if self.cross_self_attention and not (
incremental_state is not None
and _self_attn_input_buffer is not None
and "prev_key" in _self_attn_input_buffer
):
if self_attn_mask_orin is not None:
assert encoder_out is not None
self_attn_mask = torch.cat(
(
x.new_zeros(x.size(0), encoder_out.size(0)),
self_attn_mask_orin,
),
dim=1,
)
if self_attn_padding_mask_orin is not None:
if encoder_padding_mask is None:
assert encoder_out is not None
encoder_padding_mask = self_attn_padding_mask_orin.new_zeros(
encoder_out.size(1), encoder_out.size(0)
)
self_attn_padding_mask = torch.cat(
(encoder_padding_mask, self_attn_padding_mask_orin), dim=1
)
assert encoder_out is not None
y = torch.cat((encoder_out, x), dim=0)
else:
y = x
x, attn = self.self_attn(
query=x,
key=y,
value=y,
key_padding_mask=self_attn_padding_mask,
incremental_state=incremental_state[i]
if incremental_state is not None
else None,
need_weights=False,
attn_mask=self_attn_mask,
)
x = self.dropout_module(x)
x = self.residual_connection(x, residual)
if not self.normalize_before:
x = self.self_attn_layer_norm(x)
if self.encoder_attn is not None and encoder_out is not None:
residual = x
if self.normalize_before:
x = self.encoder_attn_layer_norm(x)
if prev_attn_state is not None:
prev_key, prev_value = prev_attn_state[:2]
saved_state: Dict[str, Optional[Tensor]] = {
"prev_key": prev_key,
"prev_value": prev_value,
}
if len(prev_attn_state) >= 3:
saved_state["prev_key_padding_mask"] = prev_attn_state[2]
assert incremental_state is not None
self.encoder_attn._set_input_buffer(
incremental_state[i], saved_state
)
x, attn = self.encoder_attn(
query=x,
key=encoder_out,
value=encoder_out,
key_padding_mask=encoder_padding_mask,
incremental_state=incremental_state[i]
if incremental_state is not None
else None,
static_kv=True,
need_weights=need_attn or (not self.training and self.need_attn),
need_head_weights=need_head_weights,
)
x = self.dropout_module(x)
x = self.residual_connection(x, residual)
if not self.normalize_before:
x = self.encoder_attn_layer_norm(x)
x_list.append(x)
attn_list.append(attn)
# Store attentions & new x(s) (bc the old x(s) are used in other channels)
x_list_new = []
# Here comes the cross channel attention
for i, x in enumerate(x_list):
residual = x
if self.normalize_before:
x = self.cross_channel_attn_layer_norm(x)
if prev_self_attn_state is not None:
prev_key, prev_value = prev_self_attn_state[i][1][:2]
saved_state: Dict[str, Optional[Tensor]] = {
"prev_key": prev_key,
"prev_value": prev_value,
}
if len(prev_self_attn_state[i][1]) >= 3:
saved_state["prev_key_padding_mask"] = prev_self_attn_state[i][1][2]
assert incremental_state is not None
self.cross_channel_attn._set_input_buffer(
incremental_state[i], saved_state
)
# The cross attention is computed with the concatenation of attentions from other channels
if len(x_list) > 1:
x_other = torch.cat(
[x_list[(i + j) % len(x_list)] for j in range(1, len(x_list))],
dim=0,
)
else:
# Self-attention when having only one channel
x_other = x_list[i]
x, attn = self.cross_channel_attn(
query=x,
key=x_other,
value=x_other,
key_padding_mask=self_attn_padding_mask_orin,
incremental_state=incremental_state[i]
if incremental_state is not None
else None,
need_weights=False,
attn_mask=self_attn_mask_orin,
)
x = self.dropout_module(x)
x = self.residual_connection(x, residual)
if not self.normalize_before:
x = self.cross_channel_attn_layer_norm(x)
x_list_new.append(x)
x_list = x_list_new
for i, x in enumerate(x_list):
residual = x
if self.normalize_before:
x = self.final_layer_norm(x)
x = self.activation_fn(self.fc1(x))
x = self.activation_dropout_module(x)
x = self.fc2(x)
x = self.dropout_module(x)
x = self.residual_connection(x, residual)
if not self.normalize_before:
x = self.final_layer_norm(x)
x_list[i] = x
# Trick for the checkpoint activation
x_list_tensor = torch.stack(x_list)
if self.onnx_trace and incremental_state is not None:
self_and_cross_attn_state_list = []
for i in range(n_channels):
self_and_cross_attn_state = []
for self_attn_module in [self.self_attn, self.cross_channel_attn]:
saved_state = self_attn_module._get_input_buffer(
incremental_state[i]
)
assert saved_state is not None
if self_attn_padding_mask is not None:
self_attn_module_state = [
saved_state["prev_key"],
saved_state["prev_value"],
saved_state["prev_key_padding_mask"],
]
else:
self_attn_module_state = [
saved_state["prev_key"],
saved_state["prev_value"],
]
self_and_cross_attn_state.append(self_attn_module_state)
self_and_cross_attn_state_list.append(tuple(self_and_cross_attn_state))
return x_list_tensor, attn_list, self_and_cross_attn_state_list
return x_list_tensor, attn_list, None
def make_generation_fast_(self, need_attn: bool = False, **kwargs):
self.need_attn = need_attn
# Rewrite fairseq.modules.TransformerDecoderLayer
# to be compatible with checkpoint_activations
# (avoid forwarding model multiple times)
class StandardTransformerDecoderLayer(nn.Module):
"""Rewrite fairseq.modules.TransformerDecoderLayer to avoid forwarding
model multiple times and be compatible with checkpoint_activations.
The input is expected to be a list of tensors from different channels,
each is forwarded to the same model (shared attention weights).
In the original paper each operation (multi-head attention, encoder
attention or FFN) is postprocessed with: `dropout -> add residual ->
layernorm`. In the tensor2tensor code they suggest that learning is more
robust when preprocessing each layer with layernorm and postprocessing with:
`dropout -> add residual`. We default to the approach in the paper, but the
tensor2tensor approach can be enabled by setting
*args.decoder_normalize_before* to ``True``.
Args:
args (argparse.Namespace): parsed command-line arguments
no_encoder_attn (bool, optional): whether to attend to encoder outputs
(default: False).
"""
def __init__(
self, args, no_encoder_attn=False, add_bias_kv=False, add_zero_attn=False
):
super().__init__()
self.embed_dim = args.decoder_embed_dim
self.dropout_module = FairseqDropout(
args.dropout, module_name=self.__class__.__name__
)
self.quant_noise = getattr(args, "quant_noise_pq", 0)
self.quant_noise_block_size = getattr(args, "quant_noise_pq_block_size", 8)
self.cross_self_attention = getattr(args, "cross_self_attention", False)
self.self_attn = self.build_self_attention(
self.embed_dim,
args,
add_bias_kv=add_bias_kv,
add_zero_attn=add_zero_attn,
)
self.activation_fn = utils.get_activation_fn(
activation=str(args.activation_fn)
if getattr(args, "activation_fn", None) is not None
else "relu"
)
activation_dropout_p = getattr(args, "activation_dropout", 0) or 0
if activation_dropout_p == 0:
# for backwards compatibility with models that use args.relu_dropout
activation_dropout_p = getattr(args, "relu_dropout", 0) or 0
self.activation_dropout_module = FairseqDropout(
float(activation_dropout_p), module_name=self.__class__.__name__
)
self.normalize_before = args.decoder_normalize_before
# use layerNorm rather than FusedLayerNorm for exporting.
# char_inputs can be used to determint this.
# TODO remove this once we update apex with the fix
export = getattr(args, "char_inputs", False)
self.self_attn_layer_norm = LayerNorm(self.embed_dim, export=export)
if no_encoder_attn:
self.encoder_attn = None
self.encoder_attn_layer_norm = None
else:
self.encoder_attn = self.build_encoder_attention(self.embed_dim, args)
self.encoder_attn_layer_norm = LayerNorm(self.embed_dim, export=export)
self.fc1 = self.build_fc1(
self.embed_dim,
args.decoder_ffn_embed_dim,
self.quant_noise,
self.quant_noise_block_size,
)
self.fc2 = self.build_fc2(
args.decoder_ffn_embed_dim,
self.embed_dim,
self.quant_noise,
self.quant_noise_block_size,
)
self.final_layer_norm = LayerNorm(self.embed_dim, export=export)
self.need_attn = True
self.onnx_trace = False
def build_fc1(self, input_dim, output_dim, q_noise, qn_block_size):
return quant_noise(nn.Linear(input_dim, output_dim), q_noise, qn_block_size)
def build_fc2(self, input_dim, output_dim, q_noise, qn_block_size):
return quant_noise(nn.Linear(input_dim, output_dim), q_noise, qn_block_size)
def build_self_attention(
self, embed_dim, args, add_bias_kv=False, add_zero_attn=False
):
return MultiheadAttention(
embed_dim,
args.decoder_attention_heads,
dropout=args.attention_dropout,
add_bias_kv=add_bias_kv,
add_zero_attn=add_zero_attn,
self_attention=not getattr(args, "cross_self_attention", False),
q_noise=self.quant_noise,
qn_block_size=self.quant_noise_block_size,
)
def build_encoder_attention(self, embed_dim, args):
return MultiheadAttention(
embed_dim,
args.decoder_attention_heads,
kdim=getattr(args, "encoder_embed_dim", None),
vdim=getattr(args, "encoder_embed_dim", None),
dropout=args.attention_dropout,
encoder_decoder_attention=True,
q_noise=self.quant_noise,
qn_block_size=self.quant_noise_block_size,
)
def prepare_for_onnx_export_(self):
self.onnx_trace = True
def residual_connection(self, x, residual):
return residual + x
def forward(
self,
x_list_tensor: List[torch.Tensor],
encoder_out: Optional[torch.Tensor] = None,
encoder_padding_mask: Optional[torch.Tensor] = None,
incremental_state: Optional[
List[Dict[str, Dict[str, Optional[Tensor]]]]
] = None,
prev_self_attn_state: Optional[List[Tuple[torch.Tensor, torch.Tensor]]] = None,
prev_attn_state: Optional[List[torch.Tensor]] = None,
self_attn_mask: Optional[torch.Tensor] = None,
self_attn_padding_mask: Optional[torch.Tensor] = None,
need_attn: bool = False,
need_head_weights: bool = False,
):
"""
Args:
x_list_tensor (List[Tensor]): list of input tensors in different channels,
each tensor is of shape `(seq_len, batch, embed_dim)`
encoder_padding_mask (ByteTensor, optional): binary
ByteTensor of shape `(batch, src_len)` where padding
elements are indicated by ``1``.
incremental_state (optional): list of incremental_state dictionaries over
different channels (sequence generation mode)
prev_self_attn_state (List[Tuple[Tensor, Tensor]], optional): list of tuples
(self_attn_state, cross_channel_attn_state) over different channels
need_attn (bool, optional): return attention weights
need_head_weights (bool, optional): return attention weights
for each head (default: return average over heads).
Returns:
list of encoded output of shape `(seq_len, batch, embed_dim)`
"""
n_channels = len(x_list_tensor)
if need_head_weights:
need_attn = True
# incremental_state is a list of dictionaries over different channels
if incremental_state is not None:
assert isinstance(incremental_state, list)
assert len(incremental_state) == n_channels
# prev_self_attn_state is a list of self_attn_state over different channels
if prev_self_attn_state is not None:
assert isinstance(prev_self_attn_state, list)
assert len(prev_self_attn_state) == n_channels
x_list = []
attn_list = []
for i, x in enumerate(x_list_tensor):
residual = x
if self.normalize_before:
x = self.self_attn_layer_norm(x)
if prev_self_attn_state is not None:
prev_key, prev_value = prev_self_attn_state[i][:2]
saved_state: Dict[str, Optional[Tensor]] = {
"prev_key": prev_key,
"prev_value": prev_value,
}
if len(prev_self_attn_state[i]) >= 3:
saved_state["prev_key_padding_mask"] = prev_self_attn_state[2]
assert incremental_state is not None
self.self_attn._set_input_buffer(incremental_state[i], saved_state)
_self_attn_input_buffer = self.self_attn._get_input_buffer(
incremental_state
)
if self.cross_self_attention and not (
incremental_state is not None
and _self_attn_input_buffer is not None
and "prev_key" in _self_attn_input_buffer
):
if self_attn_mask is not None:
assert encoder_out is not None
self_attn_mask = torch.cat(
(x.new_zeros(x.size(0), encoder_out.size(0)), self_attn_mask),
dim=1,
)
if self_attn_padding_mask is not None:
if encoder_padding_mask is None:
assert encoder_out is not None
encoder_padding_mask = self_attn_padding_mask.new_zeros(
encoder_out.size(1), encoder_out.size(0)
)
self_attn_padding_mask = torch.cat(
(encoder_padding_mask, self_attn_padding_mask), dim=1
)
assert encoder_out is not None
y = torch.cat((encoder_out, x), dim=0)
else:
y = x
x, attn = self.self_attn(
query=x,
key=y,
value=y,
key_padding_mask=self_attn_padding_mask,
incremental_state=incremental_state[i]
if incremental_state is not None
else None,
need_weights=False,
attn_mask=self_attn_mask,
)
x = self.dropout_module(x)
x = self.residual_connection(x, residual)
if not self.normalize_before:
x = self.self_attn_layer_norm(x)
if self.encoder_attn is not None and encoder_out is not None:
residual = x
if self.normalize_before:
x = self.encoder_attn_layer_norm(x)
if prev_attn_state is not None:
prev_key, prev_value = prev_attn_state[:2]
saved_state: Dict[str, Optional[Tensor]] = {
"prev_key": prev_key,
"prev_value": prev_value,
}
if len(prev_attn_state) >= 3:
saved_state["prev_key_padding_mask"] = prev_attn_state[2]
assert incremental_state is not None
self.encoder_attn._set_input_buffer(incremental_state, saved_state)
x, attn = self.encoder_attn(
query=x,
key=encoder_out,
value=encoder_out,
key_padding_mask=encoder_padding_mask,
incremental_state=incremental_state[i]
if incremental_state is not None
else None,
static_kv=True,
need_weights=need_attn or (not self.training and self.need_attn),
need_head_weights=need_head_weights,
)
x = self.dropout_module(x)
x = self.residual_connection(x, residual)
if not self.normalize_before:
x = self.encoder_attn_layer_norm(x)
residual = x
if self.normalize_before:
x = self.final_layer_norm(x)
x = self.activation_fn(self.fc1(x))
x = self.activation_dropout_module(x)
x = self.fc2(x)
x = self.dropout_module(x)
x = self.residual_connection(x, residual)
if not self.normalize_before:
x = self.final_layer_norm(x)
x_list.append(x)
attn_list.append(attn)
# Trick for the checkpoint activation
x_list_tensor = torch.stack(x_list)
if self.onnx_trace and incremental_state is not None:
self_attn_state_list = []
for i in range(n_channels):
saved_state = self.self_attn._get_input_buffer(incremental_state[i])
assert saved_state is not None
if self_attn_padding_mask is not None:
self_attn_state = [
saved_state["prev_key"],
saved_state["prev_value"],
saved_state["prev_key_padding_mask"],
]
else:
self_attn_state = [
saved_state["prev_key"],
saved_state["prev_value"],
]
self_attn_state_list.append(self_attn_state)
return x_list_tensor, attn_list, self_attn_state_list
return x_list_tensor, attn_list, None
def make_generation_fast_(self, need_attn: bool = False, **kwargs):
self.need_attn = need_attn
|