forked from huggingface/diffusers
-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathnormalization.py
More file actions
658 lines (537 loc) 路 24.6 KB
/
Copy pathnormalization.py
File metadata and controls
658 lines (537 loc) 路 24.6 KB
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
# coding=utf-8
# Copyright 2024 HuggingFace Inc.
#
# 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 numbers
from typing import Dict, Optional, Tuple
import torch
import torch.nn as nn
import torch.nn.functional as F
from ..utils import is_torch_npu_available, is_torch_version
from .activations import get_activation
from .embeddings import CombinedTimestepLabelEmbeddings, PixArtAlphaCombinedTimestepSizeEmbeddings
class AdaLayerNorm(nn.Module):
r"""
Norm layer modified to incorporate timestep embeddings.
Parameters:
embedding_dim (`int`): The size of each embedding vector.
num_embeddings (`int`, *optional*): The size of the embeddings dictionary.
output_dim (`int`, *optional*):
norm_elementwise_affine (`bool`, defaults to `False):
norm_eps (`bool`, defaults to `False`):
chunk_dim (`int`, defaults to `0`):
"""
def __init__(
self,
embedding_dim: int,
num_embeddings: Optional[int] = None,
output_dim: Optional[int] = None,
norm_elementwise_affine: bool = False,
norm_eps: float = 1e-5,
chunk_dim: int = 0,
):
super().__init__()
self.chunk_dim = chunk_dim
output_dim = output_dim or embedding_dim * 2
if num_embeddings is not None:
self.emb = nn.Embedding(num_embeddings, embedding_dim)
else:
self.emb = None
self.silu = nn.SiLU()
self.linear = nn.Linear(embedding_dim, output_dim)
self.norm = nn.LayerNorm(output_dim // 2, norm_eps, norm_elementwise_affine)
def forward(
self, x: torch.Tensor, timestep: Optional[torch.Tensor] = None, temb: Optional[torch.Tensor] = None
) -> torch.Tensor:
if self.emb is not None:
temb = self.emb(timestep)
temb = self.linear(self.silu(temb))
if self.chunk_dim == 1:
# This is a bit weird why we have the order of "shift, scale" here and "scale, shift" in the
# other if-branch. This branch is specific to CogVideoX and OmniGen for now.
shift, scale = temb.chunk(2, dim=1)
shift = shift[:, None, :]
scale = scale[:, None, :]
else:
scale, shift = temb.chunk(2, dim=0)
x = self.norm(x) * (1 + scale) + shift
return x
class FP32LayerNorm(nn.LayerNorm):
def forward(self, inputs: torch.Tensor) -> torch.Tensor:
origin_dtype = inputs.dtype
return F.layer_norm(
inputs.float(),
self.normalized_shape,
self.weight.float() if self.weight is not None else None,
self.bias.float() if self.bias is not None else None,
self.eps,
).to(origin_dtype)
class SD35AdaLayerNormZeroX(nn.Module):
r"""
Norm layer adaptive layer norm zero (AdaLN-Zero).
Parameters:
embedding_dim (`int`): The size of each embedding vector.
num_embeddings (`int`): The size of the embeddings dictionary.
"""
def __init__(self, embedding_dim: int, norm_type: str = "layer_norm", bias: bool = True) -> None:
super().__init__()
self.silu = nn.SiLU()
self.linear = nn.Linear(embedding_dim, 9 * embedding_dim, bias=bias)
if norm_type == "layer_norm":
self.norm = nn.LayerNorm(embedding_dim, elementwise_affine=False, eps=1e-6)
else:
raise ValueError(f"Unsupported `norm_type` ({norm_type}) provided. Supported ones are: 'layer_norm'.")
def forward(
self,
hidden_states: torch.Tensor,
emb: Optional[torch.Tensor] = None,
) -> Tuple[torch.Tensor, ...]:
emb = self.linear(self.silu(emb))
shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp, shift_msa2, scale_msa2, gate_msa2 = emb.chunk(
9, dim=1
)
norm_hidden_states = self.norm(hidden_states)
hidden_states = norm_hidden_states * (1 + scale_msa[:, None]) + shift_msa[:, None]
norm_hidden_states2 = norm_hidden_states * (1 + scale_msa2[:, None]) + shift_msa2[:, None]
return hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp, norm_hidden_states2, gate_msa2
class AdaLayerNormZero(nn.Module):
r"""
Norm layer adaptive layer norm zero (adaLN-Zero).
Parameters:
embedding_dim (`int`): The size of each embedding vector.
num_embeddings (`int`): The size of the embeddings dictionary.
"""
def __init__(self, embedding_dim: int, num_embeddings: Optional[int] = None, norm_type="layer_norm", bias=True):
super().__init__()
if num_embeddings is not None:
self.emb = CombinedTimestepLabelEmbeddings(num_embeddings, embedding_dim)
else:
self.emb = None
self.silu = nn.SiLU()
self.linear = nn.Linear(embedding_dim, 6 * embedding_dim, bias=bias)
if norm_type == "layer_norm":
self.norm = nn.LayerNorm(embedding_dim, elementwise_affine=False, eps=1e-6)
elif norm_type == "fp32_layer_norm":
self.norm = FP32LayerNorm(embedding_dim, elementwise_affine=False, bias=False)
else:
raise ValueError(
f"Unsupported `norm_type` ({norm_type}) provided. Supported ones are: 'layer_norm', 'fp32_layer_norm'."
)
def forward(
self,
x: torch.Tensor,
timestep: Optional[torch.Tensor] = None,
class_labels: Optional[torch.LongTensor] = None,
hidden_dtype: Optional[torch.dtype] = None,
emb: Optional[torch.Tensor] = None,
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
if self.emb is not None:
emb = self.emb(timestep, class_labels, hidden_dtype=hidden_dtype)
emb = self.linear(self.silu(emb))
shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = emb.chunk(6, dim=1)
x = self.norm(x) * (1 + scale_msa[:, None]) + shift_msa[:, None]
return x, gate_msa, shift_mlp, scale_mlp, gate_mlp
class AdaLayerNormZeroSingle(nn.Module):
r"""
Norm layer adaptive layer norm zero (adaLN-Zero).
Parameters:
embedding_dim (`int`): The size of each embedding vector.
num_embeddings (`int`): The size of the embeddings dictionary.
"""
def __init__(self, embedding_dim: int, norm_type="layer_norm", bias=True):
super().__init__()
self.silu = nn.SiLU()
self.linear = nn.Linear(embedding_dim, 3 * embedding_dim, bias=bias)
if norm_type == "layer_norm":
self.norm = nn.LayerNorm(embedding_dim, elementwise_affine=False, eps=1e-6)
else:
raise ValueError(
f"Unsupported `norm_type` ({norm_type}) provided. Supported ones are: 'layer_norm', 'fp32_layer_norm'."
)
def forward(
self,
x: torch.Tensor,
emb: Optional[torch.Tensor] = None,
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
emb = self.linear(self.silu(emb))
shift_msa, scale_msa, gate_msa = emb.chunk(3, dim=1)
x = self.norm(x) * (1 + scale_msa[:, None]) + shift_msa[:, None]
return x, gate_msa
class LuminaRMSNormZero(nn.Module):
"""
Norm layer adaptive RMS normalization zero.
Parameters:
embedding_dim (`int`): The size of each embedding vector.
"""
def __init__(self, embedding_dim: int, norm_eps: float, norm_elementwise_affine: bool):
super().__init__()
self.silu = nn.SiLU()
self.linear = nn.Linear(
min(embedding_dim, 1024),
4 * embedding_dim,
bias=True,
)
self.norm = RMSNorm(embedding_dim, eps=norm_eps)
def forward(
self,
x: torch.Tensor,
emb: Optional[torch.Tensor] = None,
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
emb = self.linear(self.silu(emb))
scale_msa, gate_msa, scale_mlp, gate_mlp = emb.chunk(4, dim=1)
x = self.norm(x) * (1 + scale_msa[:, None])
return x, gate_msa, scale_mlp, gate_mlp
class AdaLayerNormSingle(nn.Module):
r"""
Norm layer adaptive layer norm single (adaLN-single).
As proposed in PixArt-Alpha (see: https://arxiv.org/abs/2310.00426; Section 2.3).
Parameters:
embedding_dim (`int`): The size of each embedding vector.
use_additional_conditions (`bool`): To use additional conditions for normalization or not.
"""
def __init__(self, embedding_dim: int, use_additional_conditions: bool = False):
super().__init__()
self.emb = PixArtAlphaCombinedTimestepSizeEmbeddings(
embedding_dim, size_emb_dim=embedding_dim // 3, use_additional_conditions=use_additional_conditions
)
self.silu = nn.SiLU()
self.linear = nn.Linear(embedding_dim, 6 * embedding_dim, bias=True)
def forward(
self,
timestep: torch.Tensor,
added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
batch_size: Optional[int] = None,
hidden_dtype: Optional[torch.dtype] = None,
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
# No modulation happening here.
added_cond_kwargs = added_cond_kwargs or {"resolution": None, "aspect_ratio": None}
embedded_timestep = self.emb(timestep, **added_cond_kwargs, batch_size=batch_size, hidden_dtype=hidden_dtype)
return self.linear(self.silu(embedded_timestep)), embedded_timestep
class AdaGroupNorm(nn.Module):
r"""
GroupNorm layer modified to incorporate timestep embeddings.
Parameters:
embedding_dim (`int`): The size of each embedding vector.
num_embeddings (`int`): The size of the embeddings dictionary.
num_groups (`int`): The number of groups to separate the channels into.
act_fn (`str`, *optional*, defaults to `None`): The activation function to use.
eps (`float`, *optional*, defaults to `1e-5`): The epsilon value to use for numerical stability.
"""
def __init__(
self, embedding_dim: int, out_dim: int, num_groups: int, act_fn: Optional[str] = None, eps: float = 1e-5
):
super().__init__()
self.num_groups = num_groups
self.eps = eps
if act_fn is None:
self.act = None
else:
self.act = get_activation(act_fn)
self.linear = nn.Linear(embedding_dim, out_dim * 2)
def forward(self, x: torch.Tensor, emb: torch.Tensor) -> torch.Tensor:
if self.act:
emb = self.act(emb)
emb = self.linear(emb)
emb = emb[:, :, None, None]
scale, shift = emb.chunk(2, dim=1)
x = F.group_norm(x, self.num_groups, eps=self.eps)
x = x * (1 + scale) + shift
return x
class AdaLayerNormContinuous(nn.Module):
r"""
Adaptive normalization layer with a norm layer (layer_norm or rms_norm).
Args:
embedding_dim (`int`): Embedding dimension to use during projection.
conditioning_embedding_dim (`int`): Dimension of the input condition.
elementwise_affine (`bool`, defaults to `True`):
Boolean flag to denote if affine transformation should be applied.
eps (`float`, defaults to 1e-5): Epsilon factor.
bias (`bias`, defaults to `True`): Boolean flag to denote if bias should be use.
norm_type (`str`, defaults to `"layer_norm"`):
Normalization layer to use. Values supported: "layer_norm", "rms_norm".
"""
def __init__(
self,
embedding_dim: int,
conditioning_embedding_dim: int,
# NOTE: It is a bit weird that the norm layer can be configured to have scale and shift parameters
# because the output is immediately scaled and shifted by the projected conditioning embeddings.
# Note that AdaLayerNorm does not let the norm layer have scale and shift parameters.
# However, this is how it was implemented in the original code, and it's rather likely you should
# set `elementwise_affine` to False.
elementwise_affine=True,
eps=1e-5,
bias=True,
norm_type="layer_norm",
):
super().__init__()
self.silu = nn.SiLU()
self.linear = nn.Linear(conditioning_embedding_dim, embedding_dim * 2, bias=bias)
if norm_type == "layer_norm":
self.norm = LayerNorm(embedding_dim, eps, elementwise_affine, bias)
elif norm_type == "rms_norm":
self.norm = RMSNorm(embedding_dim, eps, elementwise_affine)
else:
raise ValueError(f"unknown norm_type {norm_type}")
def forward(self, x: torch.Tensor, conditioning_embedding: torch.Tensor) -> torch.Tensor:
# convert back to the original dtype in case `conditioning_embedding`` is upcasted to float32 (needed for hunyuanDiT)
emb = self.linear(self.silu(conditioning_embedding).to(x.dtype))
scale, shift = torch.chunk(emb, 2, dim=1)
x = self.norm(x) * (1 + scale)[:, None, :] + shift[:, None, :]
return x
class LuminaLayerNormContinuous(nn.Module):
def __init__(
self,
embedding_dim: int,
conditioning_embedding_dim: int,
# NOTE: It is a bit weird that the norm layer can be configured to have scale and shift parameters
# because the output is immediately scaled and shifted by the projected conditioning embeddings.
# Note that AdaLayerNorm does not let the norm layer have scale and shift parameters.
# However, this is how it was implemented in the original code, and it's rather likely you should
# set `elementwise_affine` to False.
elementwise_affine=True,
eps=1e-5,
bias=True,
norm_type="layer_norm",
out_dim: Optional[int] = None,
):
super().__init__()
# AdaLN
self.silu = nn.SiLU()
self.linear_1 = nn.Linear(conditioning_embedding_dim, embedding_dim, bias=bias)
if norm_type == "layer_norm":
self.norm = LayerNorm(embedding_dim, eps, elementwise_affine, bias)
elif norm_type == "rms_norm":
self.norm = RMSNorm(embedding_dim, eps=eps, elementwise_affine=elementwise_affine)
else:
raise ValueError(f"unknown norm_type {norm_type}")
self.linear_2 = None
if out_dim is not None:
self.linear_2 = nn.Linear(embedding_dim, out_dim, bias=bias)
def forward(
self,
x: torch.Tensor,
conditioning_embedding: torch.Tensor,
) -> torch.Tensor:
# convert back to the original dtype in case `conditioning_embedding`` is upcasted to float32 (needed for hunyuanDiT)
emb = self.linear_1(self.silu(conditioning_embedding).to(x.dtype))
scale = emb
x = self.norm(x) * (1 + scale)[:, None, :]
if self.linear_2 is not None:
x = self.linear_2(x)
return x
class CogView3PlusAdaLayerNormZeroTextImage(nn.Module):
r"""
Norm layer adaptive layer norm zero (adaLN-Zero).
Parameters:
embedding_dim (`int`): The size of each embedding vector.
num_embeddings (`int`): The size of the embeddings dictionary.
"""
def __init__(self, embedding_dim: int, dim: int):
super().__init__()
self.silu = nn.SiLU()
self.linear = nn.Linear(embedding_dim, 12 * dim, bias=True)
self.norm_x = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-5)
self.norm_c = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-5)
def forward(
self,
x: torch.Tensor,
context: torch.Tensor,
emb: Optional[torch.Tensor] = None,
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
emb = self.linear(self.silu(emb))
(
shift_msa,
scale_msa,
gate_msa,
shift_mlp,
scale_mlp,
gate_mlp,
c_shift_msa,
c_scale_msa,
c_gate_msa,
c_shift_mlp,
c_scale_mlp,
c_gate_mlp,
) = emb.chunk(12, dim=1)
normed_x = self.norm_x(x)
normed_context = self.norm_c(context)
x = normed_x * (1 + scale_msa[:, None]) + shift_msa[:, None]
context = normed_context * (1 + c_scale_msa[:, None]) + c_shift_msa[:, None]
return x, gate_msa, shift_mlp, scale_mlp, gate_mlp, context, c_gate_msa, c_shift_mlp, c_scale_mlp, c_gate_mlp
class CogVideoXLayerNormZero(nn.Module):
def __init__(
self,
conditioning_dim: int,
embedding_dim: int,
elementwise_affine: bool = True,
eps: float = 1e-5,
bias: bool = True,
) -> None:
super().__init__()
self.silu = nn.SiLU()
self.linear = nn.Linear(conditioning_dim, 6 * embedding_dim, bias=bias)
self.norm = nn.LayerNorm(embedding_dim, eps=eps, elementwise_affine=elementwise_affine)
def forward(
self, hidden_states: torch.Tensor, encoder_hidden_states: torch.Tensor, temb: torch.Tensor
) -> Tuple[torch.Tensor, torch.Tensor]:
shift, scale, gate, enc_shift, enc_scale, enc_gate = self.linear(self.silu(temb)).chunk(6, dim=1)
hidden_states = self.norm(hidden_states) * (1 + scale)[:, None, :] + shift[:, None, :]
encoder_hidden_states = self.norm(encoder_hidden_states) * (1 + enc_scale)[:, None, :] + enc_shift[:, None, :]
return hidden_states, encoder_hidden_states, gate[:, None, :], enc_gate[:, None, :]
if is_torch_version(">=", "2.1.0"):
LayerNorm = nn.LayerNorm
else:
# Has optional bias parameter compared to torch layer norm
# TODO: replace with torch layernorm once min required torch version >= 2.1
class LayerNorm(nn.Module):
r"""
LayerNorm with the bias parameter.
Args:
dim (`int`): Dimensionality to use for the parameters.
eps (`float`, defaults to 1e-5): Epsilon factor.
elementwise_affine (`bool`, defaults to `True`):
Boolean flag to denote if affine transformation should be applied.
bias (`bias`, defaults to `True`): Boolean flag to denote if bias should be use.
"""
def __init__(self, dim, eps: float = 1e-5, elementwise_affine: bool = True, bias: bool = True):
super().__init__()
self.eps = eps
if isinstance(dim, numbers.Integral):
dim = (dim,)
self.dim = torch.Size(dim)
if elementwise_affine:
self.weight = nn.Parameter(torch.ones(dim))
self.bias = nn.Parameter(torch.zeros(dim)) if bias else None
else:
self.weight = None
self.bias = None
def forward(self, input):
return F.layer_norm(input, self.dim, self.weight, self.bias, self.eps)
class RMSNorm(nn.Module):
r"""
RMS Norm as introduced in https://arxiv.org/abs/1910.07467 by Zhang et al.
Args:
dim (`int`): Number of dimensions to use for `weights`. Only effective when `elementwise_affine` is True.
eps (`float`): Small value to use when calculating the reciprocal of the square-root.
elementwise_affine (`bool`, defaults to `True`):
Boolean flag to denote if affine transformation should be applied.
bias (`bool`, defaults to False): If also training the `bias` param.
"""
def __init__(self, dim, eps: float, elementwise_affine: bool = True, bias: bool = False):
super().__init__()
self.eps = eps
self.elementwise_affine = elementwise_affine
if isinstance(dim, numbers.Integral):
dim = (dim,)
self.dim = torch.Size(dim)
self.weight = None
self.bias = None
if elementwise_affine:
self.weight = nn.Parameter(torch.ones(dim))
if bias:
self.bias = nn.Parameter(torch.zeros(dim))
def forward(self, hidden_states):
if is_torch_npu_available():
import torch_npu
if self.weight is not None:
# convert into half-precision if necessary
if self.weight.dtype in [torch.float16, torch.bfloat16]:
hidden_states = hidden_states.to(self.weight.dtype)
hidden_states = torch_npu.npu_rms_norm(hidden_states, self.weight, epsilon=self.eps)[0]
if self.bias is not None:
hidden_states = hidden_states + self.bias
elif is_torch_version(">=", "2.4"):
if self.weight is not None:
# convert into half-precision if necessary
if self.weight.dtype in [torch.float16, torch.bfloat16]:
hidden_states = hidden_states.to(self.weight.dtype)
hidden_states = nn.functional.rms_norm(
hidden_states, normalized_shape=(hidden_states.shape[-1],), weight=self.weight, eps=self.eps
)
if self.bias is not None:
hidden_states = hidden_states + self.bias
else:
input_dtype = hidden_states.dtype
variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True)
hidden_states = hidden_states * torch.rsqrt(variance + self.eps)
if self.weight is not None:
# convert into half-precision if necessary
if self.weight.dtype in [torch.float16, torch.bfloat16]:
hidden_states = hidden_states.to(self.weight.dtype)
hidden_states = hidden_states * self.weight
if self.bias is not None:
hidden_states = hidden_states + self.bias
else:
hidden_states = hidden_states.to(input_dtype)
return hidden_states
# TODO: (Dhruv) This can be replaced with regular RMSNorm in Mochi once `_keep_in_fp32_modules` is supported
# for sharded checkpoints, see: https://github.com/huggingface/diffusers/issues/10013
class MochiRMSNorm(nn.Module):
def __init__(self, dim, eps: float, elementwise_affine: bool = True):
super().__init__()
self.eps = eps
if isinstance(dim, numbers.Integral):
dim = (dim,)
self.dim = torch.Size(dim)
if elementwise_affine:
self.weight = nn.Parameter(torch.ones(dim))
else:
self.weight = None
def forward(self, hidden_states):
input_dtype = hidden_states.dtype
variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True)
hidden_states = hidden_states * torch.rsqrt(variance + self.eps)
if self.weight is not None:
hidden_states = hidden_states * self.weight
hidden_states = hidden_states.to(input_dtype)
return hidden_states
class GlobalResponseNorm(nn.Module):
r"""
Global response normalization as introduced in ConvNeXt-v2 (https://arxiv.org/abs/2301.00808).
Args:
dim (`int`): Number of dimensions to use for the `gamma` and `beta`.
"""
# Taken from https://github.com/facebookresearch/ConvNeXt-V2/blob/3608f67cc1dae164790c5d0aead7bf2d73d9719b/models/utils.py#L105
def __init__(self, dim):
super().__init__()
self.gamma = nn.Parameter(torch.zeros(1, 1, 1, dim))
self.beta = nn.Parameter(torch.zeros(1, 1, 1, dim))
def forward(self, x):
gx = torch.norm(x, p=2, dim=(1, 2), keepdim=True)
nx = gx / (gx.mean(dim=-1, keepdim=True) + 1e-6)
return self.gamma * (x * nx) + self.beta + x
class LpNorm(nn.Module):
def __init__(self, p: int = 2, dim: int = -1, eps: float = 1e-12):
super().__init__()
self.p = p
self.dim = dim
self.eps = eps
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
return F.normalize(hidden_states, p=self.p, dim=self.dim, eps=self.eps)
def get_normalization(
norm_type: str = "batch_norm",
num_features: Optional[int] = None,
eps: float = 1e-5,
elementwise_affine: bool = True,
bias: bool = True,
) -> nn.Module:
if norm_type == "rms_norm":
norm = RMSNorm(num_features, eps=eps, elementwise_affine=elementwise_affine, bias=bias)
elif norm_type == "layer_norm":
norm = nn.LayerNorm(num_features, eps=eps, elementwise_affine=elementwise_affine, bias=bias)
elif norm_type == "batch_norm":
norm = nn.BatchNorm2d(num_features, eps=eps, affine=elementwise_affine)
else:
raise ValueError(f"{norm_type=} is not supported.")
return norm