Spaces:
Sleeping
Sleeping
File size: 58,556 Bytes
19fe404 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 |
# Copyright 2023 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import math
from typing import Any, Dict, Optional
import torch
import torch.nn.functional as F
import torch.nn.init as init
from diffusers.models.activations import GEGLU, GELU, ApproximateGELU
from diffusers.models.attention import AdaLayerNorm, FeedForward
from diffusers.models.attention_processor import Attention
from diffusers.models.embeddings import SinusoidalPositionalEmbedding
from diffusers.models.lora import LoRACompatibleLinear
from diffusers.models.normalization import AdaLayerNorm, AdaLayerNormZero
from diffusers.utils import USE_PEFT_BACKEND
from diffusers.utils.import_utils import is_xformers_available
from diffusers.utils.torch_utils import maybe_allow_in_graph
from einops import rearrange, repeat
from torch import nn
from .motion_module import get_motion_module
if is_xformers_available():
import xformers
import xformers.ops
else:
xformers = None
@maybe_allow_in_graph
class GatedSelfAttentionDense(nn.Module):
r"""
A gated self-attention dense layer that combines visual features and object features.
Parameters:
query_dim (`int`): The number of channels in the query.
context_dim (`int`): The number of channels in the context.
n_heads (`int`): The number of heads to use for attention.
d_head (`int`): The number of channels in each head.
"""
def __init__(self, query_dim: int, context_dim: int, n_heads: int, d_head: int):
super().__init__()
# we need a linear projection since we need cat visual feature and obj feature
self.linear = nn.Linear(context_dim, query_dim)
self.attn = Attention(query_dim=query_dim, heads=n_heads, dim_head=d_head)
self.ff = FeedForward(query_dim, activation_fn="geglu")
self.norm1 = nn.LayerNorm(query_dim)
self.norm2 = nn.LayerNorm(query_dim)
self.register_parameter("alpha_attn", nn.Parameter(torch.tensor(0.0)))
self.register_parameter("alpha_dense", nn.Parameter(torch.tensor(0.0)))
self.enabled = True
def forward(self, x: torch.Tensor, objs: torch.Tensor) -> torch.Tensor:
if not self.enabled:
return x
n_visual = x.shape[1]
objs = self.linear(objs)
x = x + self.alpha_attn.tanh() * self.attn(self.norm1(torch.cat([x, objs], dim=1)))[:, :n_visual, :]
x = x + self.alpha_dense.tanh() * self.ff(self.norm2(x))
return x
def zero_module(module):
# Zero out the parameters of a module and return it.
for p in module.parameters():
p.detach().zero_()
return module
class KVCompressionCrossAttention(nn.Module):
r"""
A cross attention layer.
Parameters:
query_dim (`int`): The number of channels in the query.
cross_attention_dim (`int`, *optional*):
The number of channels in the encoder_hidden_states. If not given, defaults to `query_dim`.
heads (`int`, *optional*, defaults to 8): The number of heads to use for multi-head attention.
dim_head (`int`, *optional*, defaults to 64): The number of channels in each head.
dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
bias (`bool`, *optional*, defaults to False):
Set to `True` for the query, key, and value linear layers to contain a bias parameter.
"""
def __init__(
self,
query_dim: int,
cross_attention_dim: Optional[int] = None,
heads: int = 8,
dim_head: int = 64,
dropout: float = 0.0,
bias=False,
upcast_attention: bool = False,
upcast_softmax: bool = False,
added_kv_proj_dim: Optional[int] = None,
norm_num_groups: Optional[int] = None,
):
super().__init__()
inner_dim = dim_head * heads
cross_attention_dim = cross_attention_dim if cross_attention_dim is not None else query_dim
self.upcast_attention = upcast_attention
self.upcast_softmax = upcast_softmax
self.scale = dim_head**-0.5
self.heads = heads
# for slice_size > 0 the attention score computation
# is split across the batch axis to save memory
# You can set slice_size with `set_attention_slice`
self.sliceable_head_dim = heads
self._slice_size = None
self._use_memory_efficient_attention_xformers = True
self.added_kv_proj_dim = added_kv_proj_dim
if norm_num_groups is not None:
self.group_norm = nn.GroupNorm(num_channels=inner_dim, num_groups=norm_num_groups, eps=1e-5, affine=True)
else:
self.group_norm = None
self.to_q = nn.Linear(query_dim, inner_dim, bias=bias)
self.to_k = nn.Linear(cross_attention_dim, inner_dim, bias=bias)
self.to_v = nn.Linear(cross_attention_dim, inner_dim, bias=bias)
if self.added_kv_proj_dim is not None:
self.add_k_proj = nn.Linear(added_kv_proj_dim, cross_attention_dim)
self.add_v_proj = nn.Linear(added_kv_proj_dim, cross_attention_dim)
self.kv_compression = nn.Conv2d(
query_dim,
query_dim,
groups=query_dim,
kernel_size=2,
stride=2,
bias=True
)
self.kv_compression_norm = nn.LayerNorm(query_dim)
init.constant_(self.kv_compression.weight, 1 / 4)
if self.kv_compression.bias is not None:
init.constant_(self.kv_compression.bias, 0)
self.to_out = nn.ModuleList([])
self.to_out.append(nn.Linear(inner_dim, query_dim))
self.to_out.append(nn.Dropout(dropout))
def reshape_heads_to_batch_dim(self, tensor):
batch_size, seq_len, dim = tensor.shape
head_size = self.heads
tensor = tensor.reshape(batch_size, seq_len, head_size, dim // head_size)
tensor = tensor.permute(0, 2, 1, 3).reshape(batch_size * head_size, seq_len, dim // head_size)
return tensor
def reshape_batch_dim_to_heads(self, tensor):
batch_size, seq_len, dim = tensor.shape
head_size = self.heads
tensor = tensor.reshape(batch_size // head_size, head_size, seq_len, dim)
tensor = tensor.permute(0, 2, 1, 3).reshape(batch_size // head_size, seq_len, dim * head_size)
return tensor
def set_attention_slice(self, slice_size):
if slice_size is not None and slice_size > self.sliceable_head_dim:
raise ValueError(f"slice_size {slice_size} has to be smaller or equal to {self.sliceable_head_dim}.")
self._slice_size = slice_size
def forward(self, hidden_states, encoder_hidden_states=None, attention_mask=None, num_frames: int = 16, height: int = 32, width: int = 32):
batch_size, sequence_length, _ = hidden_states.shape
encoder_hidden_states = encoder_hidden_states
if self.group_norm is not None:
hidden_states = self.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
query = self.to_q(hidden_states)
dim = query.shape[-1]
query = self.reshape_heads_to_batch_dim(query)
if self.added_kv_proj_dim is not None:
key = self.to_k(hidden_states)
value = self.to_v(hidden_states)
encoder_hidden_states_key_proj = self.add_k_proj(encoder_hidden_states)
encoder_hidden_states_value_proj = self.add_v_proj(encoder_hidden_states)
key = rearrange(key, "b (f h w) c -> (b f) c h w", f=num_frames, h=height, w=width)
key = self.kv_compression(key)
key = rearrange(key, "(b f) c h w -> b (f h w) c", f=num_frames)
key = self.kv_compression_norm(key)
key = key.to(query.dtype)
value = rearrange(value, "b (f h w) c -> (b f) c h w", f=num_frames, h=height, w=width)
value = self.kv_compression(value)
value = rearrange(value, "(b f) c h w -> b (f h w) c", f=num_frames)
value = self.kv_compression_norm(value)
value = value.to(query.dtype)
key = self.reshape_heads_to_batch_dim(key)
value = self.reshape_heads_to_batch_dim(value)
encoder_hidden_states_key_proj = self.reshape_heads_to_batch_dim(encoder_hidden_states_key_proj)
encoder_hidden_states_value_proj = self.reshape_heads_to_batch_dim(encoder_hidden_states_value_proj)
key = torch.concat([encoder_hidden_states_key_proj, key], dim=1)
value = torch.concat([encoder_hidden_states_value_proj, value], dim=1)
else:
encoder_hidden_states = encoder_hidden_states if encoder_hidden_states is not None else hidden_states
key = self.to_k(encoder_hidden_states)
value = self.to_v(encoder_hidden_states)
key = rearrange(key, "b (f h w) c -> (b f) c h w", f=num_frames, h=height, w=width)
key = self.kv_compression(key)
key = rearrange(key, "(b f) c h w -> b (f h w) c", f=num_frames)
key = self.kv_compression_norm(key)
key = key.to(query.dtype)
value = rearrange(value, "b (f h w) c -> (b f) c h w", f=num_frames, h=height, w=width)
value = self.kv_compression(value)
value = rearrange(value, "(b f) c h w -> b (f h w) c", f=num_frames)
value = self.kv_compression_norm(value)
value = value.to(query.dtype)
key = self.reshape_heads_to_batch_dim(key)
value = self.reshape_heads_to_batch_dim(value)
if attention_mask is not None:
if attention_mask.shape[-1] != query.shape[1]:
target_length = query.shape[1]
attention_mask = F.pad(attention_mask, (0, target_length), value=0.0)
attention_mask = attention_mask.repeat_interleave(self.heads, dim=0)
# attention, what we cannot get enough of
if self._use_memory_efficient_attention_xformers:
hidden_states = self._memory_efficient_attention_xformers(query, key, value, attention_mask)
# Some versions of xformers return output in fp32, cast it back to the dtype of the input
hidden_states = hidden_states.to(query.dtype)
else:
if self._slice_size is None or query.shape[0] // self._slice_size == 1:
hidden_states = self._attention(query, key, value, attention_mask)
else:
hidden_states = self._sliced_attention(query, key, value, sequence_length, dim, attention_mask)
# linear proj
hidden_states = self.to_out[0](hidden_states)
# dropout
hidden_states = self.to_out[1](hidden_states)
return hidden_states
def _attention(self, query, key, value, attention_mask=None):
if self.upcast_attention:
query = query.float()
key = key.float()
attention_scores = torch.baddbmm(
torch.empty(query.shape[0], query.shape[1], key.shape[1], dtype=query.dtype, device=query.device),
query,
key.transpose(-1, -2),
beta=0,
alpha=self.scale,
)
if attention_mask is not None:
attention_scores = attention_scores + attention_mask
if self.upcast_softmax:
attention_scores = attention_scores.float()
attention_probs = attention_scores.softmax(dim=-1)
# cast back to the original dtype
attention_probs = attention_probs.to(value.dtype)
# compute attention output
hidden_states = torch.bmm(attention_probs, value)
# reshape hidden_states
hidden_states = self.reshape_batch_dim_to_heads(hidden_states)
return hidden_states
def _sliced_attention(self, query, key, value, sequence_length, dim, attention_mask):
batch_size_attention = query.shape[0]
hidden_states = torch.zeros(
(batch_size_attention, sequence_length, dim // self.heads), device=query.device, dtype=query.dtype
)
slice_size = self._slice_size if self._slice_size is not None else hidden_states.shape[0]
for i in range(hidden_states.shape[0] // slice_size):
start_idx = i * slice_size
end_idx = (i + 1) * slice_size
query_slice = query[start_idx:end_idx]
key_slice = key[start_idx:end_idx]
if self.upcast_attention:
query_slice = query_slice.float()
key_slice = key_slice.float()
attn_slice = torch.baddbmm(
torch.empty(slice_size, query.shape[1], key.shape[1], dtype=query_slice.dtype, device=query.device),
query_slice,
key_slice.transpose(-1, -2),
beta=0,
alpha=self.scale,
)
if attention_mask is not None:
attn_slice = attn_slice + attention_mask[start_idx:end_idx]
if self.upcast_softmax:
attn_slice = attn_slice.float()
attn_slice = attn_slice.softmax(dim=-1)
# cast back to the original dtype
attn_slice = attn_slice.to(value.dtype)
attn_slice = torch.bmm(attn_slice, value[start_idx:end_idx])
hidden_states[start_idx:end_idx] = attn_slice
# reshape hidden_states
hidden_states = self.reshape_batch_dim_to_heads(hidden_states)
return hidden_states
def _memory_efficient_attention_xformers(self, query, key, value, attention_mask):
# TODO attention_mask
query = query.contiguous()
key = key.contiguous()
value = value.contiguous()
hidden_states = xformers.ops.memory_efficient_attention(query, key, value, attn_bias=attention_mask)
hidden_states = self.reshape_batch_dim_to_heads(hidden_states)
return hidden_states
@maybe_allow_in_graph
class TemporalTransformerBlock(nn.Module):
r"""
A Temporal Transformer block.
Parameters:
dim (`int`): The number of channels in the input and output.
num_attention_heads (`int`): The number of heads to use for multi-head attention.
attention_head_dim (`int`): The number of channels in each head.
dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
num_embeds_ada_norm (:
obj: `int`, *optional*): The number of diffusion steps used during training. See `Transformer2DModel`.
attention_bias (:
obj: `bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter.
only_cross_attention (`bool`, *optional*):
Whether to use only cross-attention layers. In this case two cross attention layers are used.
double_self_attention (`bool`, *optional*):
Whether to use two self-attention layers. In this case no cross attention layers are used.
upcast_attention (`bool`, *optional*):
Whether to upcast the attention computation to float32. This is useful for mixed precision training.
norm_elementwise_affine (`bool`, *optional*, defaults to `True`):
Whether to use learnable elementwise affine parameters for normalization.
norm_type (`str`, *optional*, defaults to `"layer_norm"`):
The normalization layer to use. Can be `"layer_norm"`, `"ada_norm"` or `"ada_norm_zero"`.
final_dropout (`bool` *optional*, defaults to False):
Whether to apply a final dropout after the last feed-forward layer.
attention_type (`str`, *optional*, defaults to `"default"`):
The type of attention to use. Can be `"default"` or `"gated"` or `"gated-text-image"`.
positional_embeddings (`str`, *optional*, defaults to `None`):
The type of positional embeddings to apply to.
num_positional_embeddings (`int`, *optional*, defaults to `None`):
The maximum number of positional embeddings to apply.
"""
def __init__(
self,
dim: int,
num_attention_heads: int,
attention_head_dim: int,
dropout=0.0,
cross_attention_dim: Optional[int] = None,
activation_fn: str = "geglu",
num_embeds_ada_norm: Optional[int] = None,
attention_bias: bool = False,
only_cross_attention: bool = False,
double_self_attention: bool = False,
upcast_attention: bool = False,
norm_elementwise_affine: bool = True,
norm_type: str = "layer_norm", # 'layer_norm', 'ada_norm', 'ada_norm_zero', 'ada_norm_single'
norm_eps: float = 1e-5,
final_dropout: bool = False,
attention_type: str = "default",
positional_embeddings: Optional[str] = None,
num_positional_embeddings: Optional[int] = None,
# kv compression
kvcompression: Optional[bool] = False,
# motion module kwargs
motion_module_type = "VanillaGrid",
motion_module_kwargs = None,
):
super().__init__()
self.only_cross_attention = only_cross_attention
self.use_ada_layer_norm_zero = (num_embeds_ada_norm is not None) and norm_type == "ada_norm_zero"
self.use_ada_layer_norm = (num_embeds_ada_norm is not None) and norm_type == "ada_norm"
self.use_ada_layer_norm_single = norm_type == "ada_norm_single"
self.use_layer_norm = norm_type == "layer_norm"
if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None:
raise ValueError(
f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to"
f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}."
)
if positional_embeddings and (num_positional_embeddings is None):
raise ValueError(
"If `positional_embedding` type is defined, `num_positition_embeddings` must also be defined."
)
if positional_embeddings == "sinusoidal":
self.pos_embed = SinusoidalPositionalEmbedding(dim, max_seq_length=num_positional_embeddings)
else:
self.pos_embed = None
# Define 3 blocks. Each block has its own normalization layer.
# 1. Self-Attn
if self.use_ada_layer_norm:
self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm)
elif self.use_ada_layer_norm_zero:
self.norm1 = AdaLayerNormZero(dim, num_embeds_ada_norm)
else:
self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
self.kvcompression = kvcompression
if kvcompression:
self.attn1 = KVCompressionCrossAttention(
query_dim=dim,
heads=num_attention_heads,
dim_head=attention_head_dim,
dropout=dropout,
bias=attention_bias,
cross_attention_dim=cross_attention_dim if only_cross_attention else None,
upcast_attention=upcast_attention,
)
else:
self.attn1 = Attention(
query_dim=dim,
heads=num_attention_heads,
dim_head=attention_head_dim,
dropout=dropout,
bias=attention_bias,
cross_attention_dim=cross_attention_dim if only_cross_attention else None,
upcast_attention=upcast_attention,
)
print(self.attn1)
self.attn_temporal = get_motion_module(
in_channels = dim,
motion_module_type = motion_module_type,
motion_module_kwargs = motion_module_kwargs,
)
# 2. Cross-Attn
if cross_attention_dim is not None or double_self_attention:
# We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
# I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
# the second cross attention block.
self.norm2 = (
AdaLayerNorm(dim, num_embeds_ada_norm)
if self.use_ada_layer_norm
else nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
)
self.attn2 = Attention(
query_dim=dim,
cross_attention_dim=cross_attention_dim if not double_self_attention else None,
heads=num_attention_heads,
dim_head=attention_head_dim,
dropout=dropout,
bias=attention_bias,
upcast_attention=upcast_attention,
) # is self-attn if encoder_hidden_states is none
else:
self.norm2 = None
self.attn2 = None
# 3. Feed-forward
if not self.use_ada_layer_norm_single:
self.norm3 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn, final_dropout=final_dropout)
# 4. Fuser
if attention_type == "gated" or attention_type == "gated-text-image":
self.fuser = GatedSelfAttentionDense(dim, cross_attention_dim, num_attention_heads, attention_head_dim)
# 5. Scale-shift for PixArt-Alpha.
if self.use_ada_layer_norm_single:
self.scale_shift_table = nn.Parameter(torch.randn(6, dim) / dim**0.5)
# let chunk size default to None
self._chunk_size = None
self._chunk_dim = 0
def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int):
# Sets chunk feed-forward
self._chunk_size = chunk_size
self._chunk_dim = dim
def forward(
self,
hidden_states: torch.FloatTensor,
attention_mask: Optional[torch.FloatTensor] = None,
encoder_hidden_states: Optional[torch.FloatTensor] = None,
encoder_attention_mask: Optional[torch.FloatTensor] = None,
timestep: Optional[torch.LongTensor] = None,
cross_attention_kwargs: Dict[str, Any] = None,
class_labels: Optional[torch.LongTensor] = None,
num_frames: int = 16,
height: int = 32,
width: int = 32,
) -> torch.FloatTensor:
# Notice that normalization is always applied before the real computation in the following blocks.
# 0. Self-Attention
batch_size = hidden_states.shape[0]
if self.use_ada_layer_norm:
norm_hidden_states = self.norm1(hidden_states, timestep)
elif self.use_ada_layer_norm_zero:
norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(
hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype
)
elif self.use_layer_norm:
norm_hidden_states = self.norm1(hidden_states)
elif self.use_ada_layer_norm_single:
shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
self.scale_shift_table[None] + timestep.reshape(batch_size, 6, -1)
).chunk(6, dim=1)
norm_hidden_states = self.norm1(hidden_states)
norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa
norm_hidden_states = norm_hidden_states.squeeze(1)
else:
raise ValueError("Incorrect norm used")
if self.pos_embed is not None:
norm_hidden_states = self.pos_embed(norm_hidden_states)
# 1. Retrieve lora scale.
lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
# 2. Prepare GLIGEN inputs
cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {}
gligen_kwargs = cross_attention_kwargs.pop("gligen", None)
norm_hidden_states = rearrange(norm_hidden_states, "b (f d) c -> (b f) d c", f=num_frames)
if self.kvcompression:
attn_output = self.attn1(
norm_hidden_states,
encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
attention_mask=attention_mask,
num_frames=1,
height=height,
width=width,
**cross_attention_kwargs,
)
else:
attn_output = self.attn1(
norm_hidden_states,
encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
attention_mask=attention_mask,
**cross_attention_kwargs,
)
attn_output = rearrange(attn_output, "(b f) d c -> b (f d) c", f=num_frames)
if self.use_ada_layer_norm_zero:
attn_output = gate_msa.unsqueeze(1) * attn_output
elif self.use_ada_layer_norm_single:
attn_output = gate_msa * attn_output
hidden_states = attn_output + hidden_states
if hidden_states.ndim == 4:
hidden_states = hidden_states.squeeze(1)
# 2.5 GLIGEN Control
if gligen_kwargs is not None:
hidden_states = self.fuser(hidden_states, gligen_kwargs["objs"])
# 2.75. Temp-Attention
if self.attn_temporal is not None:
attn_output = rearrange(hidden_states, "b (f h w) c -> b c f h w", f=num_frames, h=height, w=width)
attn_output = self.attn_temporal(attn_output)
hidden_states = rearrange(attn_output, "b c f h w -> b (f h w) c")
# 3. Cross-Attention
if self.attn2 is not None:
if self.use_ada_layer_norm:
norm_hidden_states = self.norm2(hidden_states, timestep)
elif self.use_ada_layer_norm_zero or self.use_layer_norm:
norm_hidden_states = self.norm2(hidden_states)
elif self.use_ada_layer_norm_single:
# For PixArt norm2 isn't applied here:
# https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L70C1-L76C103
norm_hidden_states = hidden_states
else:
raise ValueError("Incorrect norm")
if self.pos_embed is not None and self.use_ada_layer_norm_single is None:
norm_hidden_states = self.pos_embed(norm_hidden_states)
attn_output = self.attn2(
norm_hidden_states,
encoder_hidden_states=encoder_hidden_states,
attention_mask=encoder_attention_mask,
**cross_attention_kwargs,
)
hidden_states = attn_output + hidden_states
# 4. Feed-forward
if not self.use_ada_layer_norm_single:
norm_hidden_states = self.norm3(hidden_states)
if self.use_ada_layer_norm_zero:
norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
if self.use_ada_layer_norm_single:
norm_hidden_states = self.norm2(hidden_states)
norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp
if self._chunk_size is not None:
# "feed_forward_chunk_size" can be used to save memory
if norm_hidden_states.shape[self._chunk_dim] % self._chunk_size != 0:
raise ValueError(
f"`hidden_states` dimension to be chunked: {norm_hidden_states.shape[self._chunk_dim]} has to be divisible by chunk size: {self._chunk_size}. Make sure to set an appropriate `chunk_size` when calling `unet.enable_forward_chunking`."
)
num_chunks = norm_hidden_states.shape[self._chunk_dim] // self._chunk_size
ff_output = torch.cat(
[
self.ff(hid_slice, scale=lora_scale)
for hid_slice in norm_hidden_states.chunk(num_chunks, dim=self._chunk_dim)
],
dim=self._chunk_dim,
)
else:
ff_output = self.ff(norm_hidden_states, scale=lora_scale)
if self.use_ada_layer_norm_zero:
ff_output = gate_mlp.unsqueeze(1) * ff_output
elif self.use_ada_layer_norm_single:
ff_output = gate_mlp * ff_output
hidden_states = ff_output + hidden_states
if hidden_states.ndim == 4:
hidden_states = hidden_states.squeeze(1)
return hidden_states
@maybe_allow_in_graph
class SelfAttentionTemporalTransformerBlock(nn.Module):
r"""
A Temporal Transformer block.
Parameters:
dim (`int`): The number of channels in the input and output.
num_attention_heads (`int`): The number of heads to use for multi-head attention.
attention_head_dim (`int`): The number of channels in each head.
dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
num_embeds_ada_norm (:
obj: `int`, *optional*): The number of diffusion steps used during training. See `Transformer2DModel`.
attention_bias (:
obj: `bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter.
only_cross_attention (`bool`, *optional*):
Whether to use only cross-attention layers. In this case two cross attention layers are used.
double_self_attention (`bool`, *optional*):
Whether to use two self-attention layers. In this case no cross attention layers are used.
upcast_attention (`bool`, *optional*):
Whether to upcast the attention computation to float32. This is useful for mixed precision training.
norm_elementwise_affine (`bool`, *optional*, defaults to `True`):
Whether to use learnable elementwise affine parameters for normalization.
norm_type (`str`, *optional*, defaults to `"layer_norm"`):
The normalization layer to use. Can be `"layer_norm"`, `"ada_norm"` or `"ada_norm_zero"`.
final_dropout (`bool` *optional*, defaults to False):
Whether to apply a final dropout after the last feed-forward layer.
attention_type (`str`, *optional*, defaults to `"default"`):
The type of attention to use. Can be `"default"` or `"gated"` or `"gated-text-image"`.
positional_embeddings (`str`, *optional*, defaults to `None`):
The type of positional embeddings to apply to.
num_positional_embeddings (`int`, *optional*, defaults to `None`):
The maximum number of positional embeddings to apply.
"""
def __init__(
self,
dim: int,
num_attention_heads: int,
attention_head_dim: int,
dropout=0.0,
cross_attention_dim: Optional[int] = None,
activation_fn: str = "geglu",
num_embeds_ada_norm: Optional[int] = None,
attention_bias: bool = False,
only_cross_attention: bool = False,
double_self_attention: bool = False,
upcast_attention: bool = False,
norm_elementwise_affine: bool = True,
norm_type: str = "layer_norm", # 'layer_norm', 'ada_norm', 'ada_norm_zero', 'ada_norm_single'
norm_eps: float = 1e-5,
final_dropout: bool = False,
attention_type: str = "default",
positional_embeddings: Optional[str] = None,
num_positional_embeddings: Optional[int] = None,
):
super().__init__()
self.only_cross_attention = only_cross_attention
self.use_ada_layer_norm_zero = (num_embeds_ada_norm is not None) and norm_type == "ada_norm_zero"
self.use_ada_layer_norm = (num_embeds_ada_norm is not None) and norm_type == "ada_norm"
self.use_ada_layer_norm_single = norm_type == "ada_norm_single"
self.use_layer_norm = norm_type == "layer_norm"
if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None:
raise ValueError(
f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to"
f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}."
)
if positional_embeddings and (num_positional_embeddings is None):
raise ValueError(
"If `positional_embedding` type is defined, `num_positition_embeddings` must also be defined."
)
if positional_embeddings == "sinusoidal":
self.pos_embed = SinusoidalPositionalEmbedding(dim, max_seq_length=num_positional_embeddings)
else:
self.pos_embed = None
# Define 3 blocks. Each block has its own normalization layer.
# 1. Self-Attn
if self.use_ada_layer_norm:
self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm)
elif self.use_ada_layer_norm_zero:
self.norm1 = AdaLayerNormZero(dim, num_embeds_ada_norm)
else:
self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
self.attn1 = Attention(
query_dim=dim,
heads=num_attention_heads,
dim_head=attention_head_dim,
dropout=dropout,
bias=attention_bias,
cross_attention_dim=cross_attention_dim if only_cross_attention else None,
upcast_attention=upcast_attention,
)
# 2. Cross-Attn
if cross_attention_dim is not None or double_self_attention:
# We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
# I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
# the second cross attention block.
self.norm2 = (
AdaLayerNorm(dim, num_embeds_ada_norm)
if self.use_ada_layer_norm
else nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
)
self.attn2 = Attention(
query_dim=dim,
cross_attention_dim=cross_attention_dim if not double_self_attention else None,
heads=num_attention_heads,
dim_head=attention_head_dim,
dropout=dropout,
bias=attention_bias,
upcast_attention=upcast_attention,
) # is self-attn if encoder_hidden_states is none
else:
self.norm2 = None
self.attn2 = None
# 3. Feed-forward
if not self.use_ada_layer_norm_single:
self.norm3 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn, final_dropout=final_dropout)
# 4. Fuser
if attention_type == "gated" or attention_type == "gated-text-image":
self.fuser = GatedSelfAttentionDense(dim, cross_attention_dim, num_attention_heads, attention_head_dim)
# 5. Scale-shift for PixArt-Alpha.
if self.use_ada_layer_norm_single:
self.scale_shift_table = nn.Parameter(torch.randn(6, dim) / dim**0.5)
# let chunk size default to None
self._chunk_size = None
self._chunk_dim = 0
def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int):
# Sets chunk feed-forward
self._chunk_size = chunk_size
self._chunk_dim = dim
def forward(
self,
hidden_states: torch.FloatTensor,
attention_mask: Optional[torch.FloatTensor] = None,
encoder_hidden_states: Optional[torch.FloatTensor] = None,
encoder_attention_mask: Optional[torch.FloatTensor] = None,
timestep: Optional[torch.LongTensor] = None,
cross_attention_kwargs: Dict[str, Any] = None,
class_labels: Optional[torch.LongTensor] = None,
) -> torch.FloatTensor:
# Notice that normalization is always applied before the real computation in the following blocks.
# 0. Self-Attention
batch_size = hidden_states.shape[0]
if self.use_ada_layer_norm:
norm_hidden_states = self.norm1(hidden_states, timestep)
elif self.use_ada_layer_norm_zero:
norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(
hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype
)
elif self.use_layer_norm:
norm_hidden_states = self.norm1(hidden_states)
elif self.use_ada_layer_norm_single:
shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
self.scale_shift_table[None] + timestep.reshape(batch_size, 6, -1)
).chunk(6, dim=1)
norm_hidden_states = self.norm1(hidden_states)
norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa
norm_hidden_states = norm_hidden_states.squeeze(1)
else:
raise ValueError("Incorrect norm used")
if self.pos_embed is not None:
norm_hidden_states = self.pos_embed(norm_hidden_states)
# 1. Retrieve lora scale.
lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
# 2. Prepare GLIGEN inputs
cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {}
gligen_kwargs = cross_attention_kwargs.pop("gligen", None)
attn_output = self.attn1(
norm_hidden_states,
encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
attention_mask=attention_mask,
**cross_attention_kwargs,
)
if self.use_ada_layer_norm_zero:
attn_output = gate_msa.unsqueeze(1) * attn_output
elif self.use_ada_layer_norm_single:
attn_output = gate_msa * attn_output
hidden_states = attn_output + hidden_states
if hidden_states.ndim == 4:
hidden_states = hidden_states.squeeze(1)
# 2.5 GLIGEN Control
if gligen_kwargs is not None:
hidden_states = self.fuser(hidden_states, gligen_kwargs["objs"])
# 3. Cross-Attention
if self.attn2 is not None:
if self.use_ada_layer_norm:
norm_hidden_states = self.norm2(hidden_states, timestep)
elif self.use_ada_layer_norm_zero or self.use_layer_norm:
norm_hidden_states = self.norm2(hidden_states)
elif self.use_ada_layer_norm_single:
# For PixArt norm2 isn't applied here:
# https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L70C1-L76C103
norm_hidden_states = hidden_states
else:
raise ValueError("Incorrect norm")
if self.pos_embed is not None and self.use_ada_layer_norm_single is None:
norm_hidden_states = self.pos_embed(norm_hidden_states)
attn_output = self.attn2(
norm_hidden_states,
encoder_hidden_states=encoder_hidden_states,
attention_mask=encoder_attention_mask,
**cross_attention_kwargs,
)
hidden_states = attn_output + hidden_states
# 4. Feed-forward
if not self.use_ada_layer_norm_single:
norm_hidden_states = self.norm3(hidden_states)
if self.use_ada_layer_norm_zero:
norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
if self.use_ada_layer_norm_single:
norm_hidden_states = self.norm2(hidden_states)
norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp
if self._chunk_size is not None:
# "feed_forward_chunk_size" can be used to save memory
if norm_hidden_states.shape[self._chunk_dim] % self._chunk_size != 0:
raise ValueError(
f"`hidden_states` dimension to be chunked: {norm_hidden_states.shape[self._chunk_dim]} has to be divisible by chunk size: {self._chunk_size}. Make sure to set an appropriate `chunk_size` when calling `unet.enable_forward_chunking`."
)
num_chunks = norm_hidden_states.shape[self._chunk_dim] // self._chunk_size
ff_output = torch.cat(
[
self.ff(hid_slice, scale=lora_scale)
for hid_slice in norm_hidden_states.chunk(num_chunks, dim=self._chunk_dim)
],
dim=self._chunk_dim,
)
else:
ff_output = self.ff(norm_hidden_states, scale=lora_scale)
if self.use_ada_layer_norm_zero:
ff_output = gate_mlp.unsqueeze(1) * ff_output
elif self.use_ada_layer_norm_single:
ff_output = gate_mlp * ff_output
hidden_states = ff_output + hidden_states
if hidden_states.ndim == 4:
hidden_states = hidden_states.squeeze(1)
return hidden_states
@maybe_allow_in_graph
class KVCompressionTransformerBlock(nn.Module):
r"""
A Temporal Transformer block.
Parameters:
dim (`int`): The number of channels in the input and output.
num_attention_heads (`int`): The number of heads to use for multi-head attention.
attention_head_dim (`int`): The number of channels in each head.
dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
num_embeds_ada_norm (:
obj: `int`, *optional*): The number of diffusion steps used during training. See `Transformer2DModel`.
attention_bias (:
obj: `bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter.
only_cross_attention (`bool`, *optional*):
Whether to use only cross-attention layers. In this case two cross attention layers are used.
double_self_attention (`bool`, *optional*):
Whether to use two self-attention layers. In this case no cross attention layers are used.
upcast_attention (`bool`, *optional*):
Whether to upcast the attention computation to float32. This is useful for mixed precision training.
norm_elementwise_affine (`bool`, *optional*, defaults to `True`):
Whether to use learnable elementwise affine parameters for normalization.
norm_type (`str`, *optional*, defaults to `"layer_norm"`):
The normalization layer to use. Can be `"layer_norm"`, `"ada_norm"` or `"ada_norm_zero"`.
final_dropout (`bool` *optional*, defaults to False):
Whether to apply a final dropout after the last feed-forward layer.
attention_type (`str`, *optional*, defaults to `"default"`):
The type of attention to use. Can be `"default"` or `"gated"` or `"gated-text-image"`.
positional_embeddings (`str`, *optional*, defaults to `None`):
The type of positional embeddings to apply to.
num_positional_embeddings (`int`, *optional*, defaults to `None`):
The maximum number of positional embeddings to apply.
"""
def __init__(
self,
dim: int,
num_attention_heads: int,
attention_head_dim: int,
dropout=0.0,
cross_attention_dim: Optional[int] = None,
activation_fn: str = "geglu",
num_embeds_ada_norm: Optional[int] = None,
attention_bias: bool = False,
only_cross_attention: bool = False,
double_self_attention: bool = False,
upcast_attention: bool = False,
norm_elementwise_affine: bool = True,
norm_type: str = "layer_norm", # 'layer_norm', 'ada_norm', 'ada_norm_zero', 'ada_norm_single'
norm_eps: float = 1e-5,
final_dropout: bool = False,
attention_type: str = "default",
positional_embeddings: Optional[str] = None,
num_positional_embeddings: Optional[int] = None,
kvcompression: Optional[bool] = False,
):
super().__init__()
self.only_cross_attention = only_cross_attention
self.use_ada_layer_norm_zero = (num_embeds_ada_norm is not None) and norm_type == "ada_norm_zero"
self.use_ada_layer_norm = (num_embeds_ada_norm is not None) and norm_type == "ada_norm"
self.use_ada_layer_norm_single = norm_type == "ada_norm_single"
self.use_layer_norm = norm_type == "layer_norm"
if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None:
raise ValueError(
f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to"
f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}."
)
if positional_embeddings and (num_positional_embeddings is None):
raise ValueError(
"If `positional_embedding` type is defined, `num_positition_embeddings` must also be defined."
)
if positional_embeddings == "sinusoidal":
self.pos_embed = SinusoidalPositionalEmbedding(dim, max_seq_length=num_positional_embeddings)
else:
self.pos_embed = None
# Define 3 blocks. Each block has its own normalization layer.
# 1. Self-Attn
if self.use_ada_layer_norm:
self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm)
elif self.use_ada_layer_norm_zero:
self.norm1 = AdaLayerNormZero(dim, num_embeds_ada_norm)
else:
self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
self.kvcompression = kvcompression
if kvcompression:
self.attn1 = KVCompressionCrossAttention(
query_dim=dim,
heads=num_attention_heads,
dim_head=attention_head_dim,
dropout=dropout,
bias=attention_bias,
cross_attention_dim=cross_attention_dim if only_cross_attention else None,
upcast_attention=upcast_attention,
)
else:
self.attn1 = Attention(
query_dim=dim,
heads=num_attention_heads,
dim_head=attention_head_dim,
dropout=dropout,
bias=attention_bias,
cross_attention_dim=cross_attention_dim if only_cross_attention else None,
upcast_attention=upcast_attention,
)
print(self.attn1)
# 2. Cross-Attn
if cross_attention_dim is not None or double_self_attention:
# We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
# I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
# the second cross attention block.
self.norm2 = (
AdaLayerNorm(dim, num_embeds_ada_norm)
if self.use_ada_layer_norm
else nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
)
self.attn2 = Attention(
query_dim=dim,
cross_attention_dim=cross_attention_dim if not double_self_attention else None,
heads=num_attention_heads,
dim_head=attention_head_dim,
dropout=dropout,
bias=attention_bias,
upcast_attention=upcast_attention,
) # is self-attn if encoder_hidden_states is none
else:
self.norm2 = None
self.attn2 = None
# 3. Feed-forward
if not self.use_ada_layer_norm_single:
self.norm3 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn, final_dropout=final_dropout)
# 4. Fuser
if attention_type == "gated" or attention_type == "gated-text-image":
self.fuser = GatedSelfAttentionDense(dim, cross_attention_dim, num_attention_heads, attention_head_dim)
# 5. Scale-shift for PixArt-Alpha.
if self.use_ada_layer_norm_single:
self.scale_shift_table = nn.Parameter(torch.randn(6, dim) / dim**0.5)
# let chunk size default to None
self._chunk_size = None
self._chunk_dim = 0
def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int):
# Sets chunk feed-forward
self._chunk_size = chunk_size
self._chunk_dim = dim
def forward(
self,
hidden_states: torch.FloatTensor,
attention_mask: Optional[torch.FloatTensor] = None,
encoder_hidden_states: Optional[torch.FloatTensor] = None,
encoder_attention_mask: Optional[torch.FloatTensor] = None,
timestep: Optional[torch.LongTensor] = None,
cross_attention_kwargs: Dict[str, Any] = None,
class_labels: Optional[torch.LongTensor] = None,
num_frames: int = 16,
height: int = 32,
width: int = 32,
use_reentrant: bool = False,
) -> torch.FloatTensor:
# Notice that normalization is always applied before the real computation in the following blocks.
# 0. Self-Attention
batch_size = hidden_states.shape[0]
if self.use_ada_layer_norm:
norm_hidden_states = self.norm1(hidden_states, timestep)
elif self.use_ada_layer_norm_zero:
norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(
hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype
)
elif self.use_layer_norm:
norm_hidden_states = self.norm1(hidden_states)
elif self.use_ada_layer_norm_single:
shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
self.scale_shift_table[None] + timestep.reshape(batch_size, 6, -1)
).chunk(6, dim=1)
norm_hidden_states = self.norm1(hidden_states)
norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa
norm_hidden_states = norm_hidden_states.squeeze(1)
else:
raise ValueError("Incorrect norm used")
if self.pos_embed is not None:
norm_hidden_states = self.pos_embed(norm_hidden_states)
# 1. Retrieve lora scale.
lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
# 2. Prepare GLIGEN inputs
cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {}
gligen_kwargs = cross_attention_kwargs.pop("gligen", None)
if self.kvcompression:
attn_output = self.attn1(
norm_hidden_states,
encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
attention_mask=attention_mask,
num_frames=num_frames,
height=height,
width=width,
**cross_attention_kwargs,
)
else:
attn_output = self.attn1(
norm_hidden_states,
encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
attention_mask=attention_mask,
**cross_attention_kwargs,
)
if self.use_ada_layer_norm_zero:
attn_output = gate_msa.unsqueeze(1) * attn_output
elif self.use_ada_layer_norm_single:
attn_output = gate_msa * attn_output
hidden_states = attn_output + hidden_states
if hidden_states.ndim == 4:
hidden_states = hidden_states.squeeze(1)
# 2.5 GLIGEN Control
if gligen_kwargs is not None:
hidden_states = self.fuser(hidden_states, gligen_kwargs["objs"])
# 3. Cross-Attention
if self.attn2 is not None:
if self.use_ada_layer_norm:
norm_hidden_states = self.norm2(hidden_states, timestep)
elif self.use_ada_layer_norm_zero or self.use_layer_norm:
norm_hidden_states = self.norm2(hidden_states)
elif self.use_ada_layer_norm_single:
# For PixArt norm2 isn't applied here:
# https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L70C1-L76C103
norm_hidden_states = hidden_states
else:
raise ValueError("Incorrect norm")
if self.pos_embed is not None and self.use_ada_layer_norm_single is None:
norm_hidden_states = self.pos_embed(norm_hidden_states)
attn_output = self.attn2(
norm_hidden_states,
encoder_hidden_states=encoder_hidden_states,
attention_mask=encoder_attention_mask,
**cross_attention_kwargs,
)
hidden_states = attn_output + hidden_states
# 4. Feed-forward
if not self.use_ada_layer_norm_single:
norm_hidden_states = self.norm3(hidden_states)
if self.use_ada_layer_norm_zero:
norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
if self.use_ada_layer_norm_single:
norm_hidden_states = self.norm2(hidden_states)
norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp
if self._chunk_size is not None:
# "feed_forward_chunk_size" can be used to save memory
if norm_hidden_states.shape[self._chunk_dim] % self._chunk_size != 0:
raise ValueError(
f"`hidden_states` dimension to be chunked: {norm_hidden_states.shape[self._chunk_dim]} has to be divisible by chunk size: {self._chunk_size}. Make sure to set an appropriate `chunk_size` when calling `unet.enable_forward_chunking`."
)
num_chunks = norm_hidden_states.shape[self._chunk_dim] // self._chunk_size
ff_output = torch.cat(
[
self.ff(hid_slice, scale=lora_scale)
for hid_slice in norm_hidden_states.chunk(num_chunks, dim=self._chunk_dim)
],
dim=self._chunk_dim,
)
else:
ff_output = self.ff(norm_hidden_states, scale=lora_scale)
if self.use_ada_layer_norm_zero:
ff_output = gate_mlp.unsqueeze(1) * ff_output
elif self.use_ada_layer_norm_single:
ff_output = gate_mlp * ff_output
hidden_states = ff_output + hidden_states
if hidden_states.ndim == 4:
hidden_states = hidden_states.squeeze(1)
return hidden_states
class FeedForward(nn.Module):
r"""
A feed-forward layer.
Parameters:
dim (`int`): The number of channels in the input.
dim_out (`int`, *optional*): The number of channels in the output. If not given, defaults to `dim`.
mult (`int`, *optional*, defaults to 4): The multiplier to use for the hidden dimension.
dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
final_dropout (`bool` *optional*, defaults to False): Apply a final dropout.
"""
def __init__(
self,
dim: int,
dim_out: Optional[int] = None,
mult: int = 4,
dropout: float = 0.0,
activation_fn: str = "geglu",
final_dropout: bool = False,
):
super().__init__()
inner_dim = int(dim * mult)
dim_out = dim_out if dim_out is not None else dim
linear_cls = LoRACompatibleLinear if not USE_PEFT_BACKEND else nn.Linear
if activation_fn == "gelu":
act_fn = GELU(dim, inner_dim)
if activation_fn == "gelu-approximate":
act_fn = GELU(dim, inner_dim, approximate="tanh")
elif activation_fn == "geglu":
act_fn = GEGLU(dim, inner_dim)
elif activation_fn == "geglu-approximate":
act_fn = ApproximateGELU(dim, inner_dim)
self.net = nn.ModuleList([])
# project in
self.net.append(act_fn)
# project dropout
self.net.append(nn.Dropout(dropout))
# project out
self.net.append(linear_cls(inner_dim, dim_out))
# FF as used in Vision Transformer, MLP-Mixer, etc. have a final dropout
if final_dropout:
self.net.append(nn.Dropout(dropout))
def forward(self, hidden_states: torch.Tensor, scale: float = 1.0) -> torch.Tensor:
compatible_cls = (GEGLU,) if USE_PEFT_BACKEND else (GEGLU, LoRACompatibleLinear)
for module in self.net:
if isinstance(module, compatible_cls):
hidden_states = module(hidden_states, scale)
else:
hidden_states = module(hidden_states)
return hidden_states
|