Repository navigation
Expand file tree
/
Copy pathmodeling_siglip.py
More file actions
1717 lines (1402 loc) · 72.9 KB
/
Copy pathmodeling_siglip.py
File metadata and controls
1717 lines (1402 loc) · 72.9 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
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
# coding=utf-8
# Copyright 2024 Google AI and 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.
"""PyTorch Siglip model."""
import math
import warnings
from dataclasses import dataclass
from typing import Any, Optional, Tuple, Union
import numpy as np
import torch
import torch.utils.checkpoint
from torch import nn
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
from torch.nn.init import _calculate_fan_in_and_fan_out
from ...activations import ACT2FN
from ...modeling_attn_mask_utils import _prepare_4d_attention_mask
from ...modeling_outputs import BaseModelOutput, BaseModelOutputWithPooling, ImageClassifierOutput
from ...modeling_utils import PreTrainedModel
from ...utils import (
ModelOutput,
add_start_docstrings,
add_start_docstrings_to_model_forward,
is_flash_attn_2_available,
is_flash_attn_greater_or_equal_2_10,
logging,
replace_return_docstrings,
torch_int,
)
from .configuration_siglip import SiglipConfig, SiglipTextConfig, SiglipVisionConfig
if is_flash_attn_2_available():
from ...modeling_flash_attention_utils import _flash_attention_forward
logger = logging.get_logger(__name__)
# General docstring
_CONFIG_FOR_DOC = "SiglipConfig"
_CHECKPOINT_FOR_DOC = "google/siglip-base-patch16-224"
class WeightQuant(torch.autograd.Function):
@staticmethod
# @torch.compile # enable fused kernel for faster training
def forward(ctx, x):
dtype = x.dtype
x = x.float()
s = 1.0 / x.abs().mean().clamp_(min=1e-5)
x = (x * s).round().clamp(-1, 1) / s
return x.to(dtype)
@staticmethod
def backward(ctx, grad_output):
grad_input = grad_output.clone()
return grad_input
class ActQuant(torch.autograd.Function):
@staticmethod
# @torch.compile # enable fused kernel for faster training
def forward(ctx, x):
dtype = x.dtype
x = x.float()
s = 127 / x.abs().max(dim=-1, keepdim=True).values.clamp_(min=1e-5)
x = (x * s).round().clamp(-128, 127) / s
return x.to(dtype)
@staticmethod
def backward(ctx, grad_output):
grad_input = grad_output.clone()
return grad_input
def absmean(input):
dtype = input.dtype
input = input.float()
s = input.abs().mean().clamp(min=1e-5)
result = (input / s).round().clamp(-1, 1) * s
return result.type(dtype)
def quantize_to_int2(weight: torch.Tensor,):
weight = weight.float()
step = weight.abs().mean().clamp(min=1e-5)
q = (weight / step).round().clamp(-1, 1).to(torch.uint8) + 1
q = q.flatten()
n = q.numel()
remainder = n % 4
if remainder != 0:
padding = 4 - remainder
q = torch.cat([q, torch.zeros(padding, dtype=torch.uint8, device=q.device)])
q = q.view(-1, 4)
packed = (q[:, 0] | (q[:, 1] << 2) | (q[:, 2] << 4) | (q[:, 3] << 6))
return packed, step, weight.shape, n
# @torch.compile
def dequantize_from_int2(packed: torch.Tensor, step: float, orig_shape, n_elems: int):
q0 = (packed & 0x03)
q1 = ((packed >> 2) & 0x03)
q2 = ((packed >> 4) & 0x03)
q3 = ((packed >> 6) & 0x03)
q_all = torch.stack([q0, q1, q2, q3], dim=1).view(-1)
q_all = q_all[:n_elems] # 截去补齐部分
q_all = q_all.float() - 1
weight = q_all * step
return weight.view(orig_shape)
class BitLinear(nn.Linear):
def __init__(self,
*kargs,
weight_bits=8,
input_bits=8,
**kwargs
):
super(BitLinear, self).__init__(*kargs, **kwargs)
self.weight_bits = weight_bits
self.input_bits = input_bits
self.enable_qlora = False
def quantize_weights(self):
if self.weight_bits == 32:
return
packed, step, orig_shape, n_elems = quantize_to_int2(self.weight.data)
self.register_buffer("q_weight", packed)
self.register_buffer("w_step", torch.tensor(step, dtype=torch.float32))
self.orig_shape = orig_shape
self.n_elems = n_elems
qweight = dequantize_from_int2(
self.q_weight,
self.w_step.item(),
self.orig_shape,
self.n_elems,
).type(torch.bfloat16)
assert (torch.isclose(qweight, absmean(self.weight)) == False).sum() == 0
self.weight = None
self.enable_qlora = True
def forward(self, input):
weight = self.weight
if self.input_bits == 8:
input = ActQuant.apply(input)
if self.enable_qlora and self.weight_bits == 1:
weight = dequantize_from_int2(
self.q_weight,
self.w_step.item(),
self.orig_shape,
self.n_elems,
).type(input.dtype)
elif self.weight_bits == 1:
weight = WeightQuant.apply(weight)
out = nn.functional.linear(input, weight, self.bias)
return out
def _trunc_normal_(tensor, mean, std, a, b):
# Cut & paste from PyTorch official master until it's in a few official releases - RW
# Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf
def norm_cdf(x):
# Computes standard normal cumulative distribution function
return (1.0 + math.erf(x / math.sqrt(2.0))) / 2.0
if (mean < a - 2 * std) or (mean > b + 2 * std):
warnings.warn(
"mean is more than 2 std from [a, b] in nn.init.trunc_normal_. "
"The distribution of values may be incorrect.",
stacklevel=2,
)
# Values are generated by using a truncated uniform distribution and
# then using the inverse CDF for the normal distribution.
# Get upper and lower cdf values
l = norm_cdf((a - mean) / std)
u = norm_cdf((b - mean) / std)
# Uniformly fill tensor with values from [l, u], then translate to
# [2l-1, 2u-1].
tensor.uniform_(2 * l - 1, 2 * u - 1)
# Use inverse cdf transform for normal distribution to get truncated
# standard normal
tensor.erfinv_()
# Transform to proper mean, std
tensor.mul_(std * math.sqrt(2.0))
tensor.add_(mean)
# Clamp to ensure it's in the proper range
tensor.clamp_(min=a, max=b)
def trunc_normal_tf_(
tensor: torch.Tensor, mean: float = 0.0, std: float = 1.0, a: float = -2.0, b: float = 2.0
) -> torch.Tensor:
"""Fills the input Tensor with values drawn from a truncated
normal distribution. The values are effectively drawn from the
normal distribution :math:`\\mathcal{N}(\text{mean}, \text{std}^2)`
with values outside :math:`[a, b]` redrawn until they are within
the bounds. The method used for generating the random values works
best when :math:`a \\leq \text{mean} \\leq b`.
NOTE: this 'tf' variant behaves closer to Tensorflow / JAX impl where the
bounds [a, b] are applied when sampling the normal distribution with mean=0, std=1.0
and the result is subsequently scaled and shifted by the mean and std args.
Args:
tensor: an n-dimensional `torch.Tensor`
mean: the mean of the normal distribution
std: the standard deviation of the normal distribution
a: the minimum cutoff value
b: the maximum cutoff value
"""
with torch.no_grad():
_trunc_normal_(tensor, 0, 1.0, a, b)
tensor.mul_(std).add_(mean)
def variance_scaling_(tensor, scale=1.0, mode="fan_in", distribution="normal"):
fan_in, fan_out = _calculate_fan_in_and_fan_out(tensor)
if mode == "fan_in":
denom = fan_in
elif mode == "fan_out":
denom = fan_out
elif mode == "fan_avg":
denom = (fan_in + fan_out) / 2
variance = scale / denom
if distribution == "truncated_normal":
# constant is stddev of standard normal truncated to (-2, 2)
trunc_normal_tf_(tensor, std=math.sqrt(variance) / 0.87962566103423978)
elif distribution == "normal":
with torch.no_grad():
tensor.normal_(std=math.sqrt(variance))
elif distribution == "uniform":
bound = math.sqrt(3 * variance)
with torch.no_grad():
tensor.uniform_(-bound, bound)
else:
raise ValueError(f"invalid distribution {distribution}")
def lecun_normal_(tensor):
variance_scaling_(tensor, mode="fan_in", distribution="truncated_normal")
def default_flax_embed_init(tensor):
variance_scaling_(tensor, mode="fan_in", distribution="normal")
@dataclass
# Copied from transformers.models.clip.modeling_clip.CLIPVisionModelOutput with CLIP->Siglip
class SiglipVisionModelOutput(ModelOutput):
"""
Base class for vision model's outputs that also contains image embeddings of the pooling of the last hidden states.
Args:
image_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim)` *optional* returned when model is initialized with `with_projection=True`):
The image embeddings obtained by applying the projection layer to the pooler_output.
last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
Sequence of hidden-states at the output of the last layer of the model.
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
"""
image_embeds: Optional[torch.FloatTensor] = None
last_hidden_state: torch.FloatTensor = None
hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
@dataclass
# Copied from transformers.models.clip.modeling_clip.CLIPTextModelOutput with CLIP->Siglip
class SiglipTextModelOutput(ModelOutput):
"""
Base class for text model's outputs that also contains a pooling of the last hidden states.
Args:
text_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim)` *optional* returned when model is initialized with `with_projection=True`):
The text embeddings obtained by applying the projection layer to the pooler_output.
last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
Sequence of hidden-states at the output of the last layer of the model.
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
"""
text_embeds: Optional[torch.FloatTensor] = None
last_hidden_state: torch.FloatTensor = None
hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
@dataclass
# Copied from transformers.models.clip.modeling_clip.CLIPOutput with CLIP->Siglip
class SiglipOutput(ModelOutput):
"""
Args:
loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `return_loss` is `True`):
Contrastive loss for image-text similarity.
logits_per_image (`torch.FloatTensor` of shape `(image_batch_size, text_batch_size)`):
The scaled dot product scores between `image_embeds` and `text_embeds`. This represents the image-text
similarity scores.
logits_per_text (`torch.FloatTensor` of shape `(text_batch_size, image_batch_size)`):
The scaled dot product scores between `text_embeds` and `image_embeds`. This represents the text-image
similarity scores.
text_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim`):
The text embeddings obtained by applying the projection layer to the pooled output of [`SiglipTextModel`].
image_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim`):
The image embeddings obtained by applying the projection layer to the pooled output of [`SiglipVisionModel`].
text_model_output (`BaseModelOutputWithPooling`):
The output of the [`SiglipTextModel`].
vision_model_output (`BaseModelOutputWithPooling`):
The output of the [`SiglipVisionModel`].
"""
loss: Optional[torch.FloatTensor] = None
logits_per_image: torch.FloatTensor = None
logits_per_text: torch.FloatTensor = None
text_embeds: torch.FloatTensor = None
image_embeds: torch.FloatTensor = None
text_model_output: BaseModelOutputWithPooling = None
vision_model_output: BaseModelOutputWithPooling = None
def to_tuple(self) -> Tuple[Any]:
return tuple(
self[k] if k not in ["text_model_output", "vision_model_output"] else getattr(self, k).to_tuple()
for k in self.keys()
)
class SiglipVisionEmbeddings(nn.Module):
def __init__(self, config: SiglipVisionConfig):
super().__init__()
self.config = config
self.embed_dim = config.hidden_size
self.image_size = config.image_size
self.patch_size = config.patch_size
self.patch_embedding = nn.Conv2d(
in_channels=config.num_channels,
out_channels=self.embed_dim,
kernel_size=self.patch_size,
stride=self.patch_size,
padding="valid",
)
self.num_patches = (self.image_size // self.patch_size) ** 2
self.num_positions = self.num_patches
self.position_embedding = nn.Embedding(self.num_positions, self.embed_dim)
self.register_buffer("position_ids", torch.arange(self.num_positions).expand((1, -1)), persistent=False)
def interpolate_pos_encoding(self, embeddings: torch.Tensor, height: int, width: int) -> torch.Tensor:
"""
This method allows to interpolate the pre-trained position encodings, to be able to use the model on higher resolution
images. This method is also adapted to support torch.jit tracing and no class embeddings.
Adapted from:
- https://github.com/facebookresearch/dino/blob/de9ee3df6cf39fac952ab558447af1fa1365362a/vision_transformer.py#L174-L194, and
- https://github.com/facebookresearch/dinov2/blob/e1277af2ba9496fbadf7aec6eba56e8d882d1e35/dinov2/models/vision_transformer.py#L179-L211
"""
num_patches = embeddings.shape[1]
num_positions = self.position_embedding.weight.shape[0]
# always interpolate when tracing to ensure the exported model works for dynamic input shapes
if not torch.jit.is_tracing() and num_patches == num_positions and height == width:
return self.position_embedding(self.position_ids)
patch_pos_embed = self.position_embedding.weight.unsqueeze(0)
dim = embeddings.shape[-1]
new_height = height // self.patch_size
new_width = width // self.patch_size
sqrt_num_positions = torch_int(num_positions**0.5)
patch_pos_embed = patch_pos_embed.reshape(1, sqrt_num_positions, sqrt_num_positions, dim)
patch_pos_embed = patch_pos_embed.permute(0, 3, 1, 2)
patch_pos_embed = nn.functional.interpolate(
patch_pos_embed,
size=(new_height, new_width),
mode="bicubic",
align_corners=False,
)
patch_pos_embed = patch_pos_embed.permute(0, 2, 3, 1).view(1, -1, dim)
return patch_pos_embed
def forward(self, pixel_values: torch.FloatTensor, interpolate_pos_encoding=False) -> torch.Tensor:
_, _, height, width = pixel_values.shape
target_dtype = self.patch_embedding.weight.dtype
patch_embeds = self.patch_embedding(pixel_values.to(dtype=target_dtype)) # shape = [*, width, grid, grid]
embeddings = patch_embeds.flatten(2).transpose(1, 2)
if interpolate_pos_encoding:
embeddings = embeddings + self.interpolate_pos_encoding(embeddings, height, width)
else:
embeddings = embeddings + self.position_embedding(self.position_ids)
return embeddings
# Copied from transformers.models.clip.modeling_clip.CLIPTextEmbeddings with CLIP->Siglip
class SiglipTextEmbeddings(nn.Module):
def __init__(self, config: SiglipTextConfig):
super().__init__()
embed_dim = config.hidden_size
self.token_embedding = nn.Embedding(config.vocab_size, embed_dim)
self.position_embedding = nn.Embedding(config.max_position_embeddings, embed_dim)
# position_ids (1, len position emb) is contiguous in memory and exported when serialized
self.register_buffer(
"position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)), persistent=False
)
def forward(
self,
input_ids: Optional[torch.LongTensor] = None,
position_ids: Optional[torch.LongTensor] = None,
inputs_embeds: Optional[torch.FloatTensor] = None,
) -> torch.Tensor:
seq_length = input_ids.shape[-1] if input_ids is not None else inputs_embeds.shape[-2]
max_position_embedding = self.position_embedding.weight.shape[0]
if seq_length > max_position_embedding:
raise ValueError(
f"Sequence length must be less than max_position_embeddings (got `sequence length`: "
f"{seq_length} and max_position_embeddings: {max_position_embedding}"
)
if position_ids is None:
position_ids = self.position_ids[:, :seq_length]
if inputs_embeds is None:
inputs_embeds = self.token_embedding(input_ids)
position_embeddings = self.position_embedding(position_ids)
embeddings = inputs_embeds + position_embeddings
return embeddings
class SiglipAttention(nn.Module):
"""Multi-headed attention from 'Attention Is All You Need' paper"""
# Copied from transformers.models.clip.modeling_clip.CLIPAttention.__init__
def __init__(self, config):
super().__init__()
self.config = config
self.embed_dim = config.hidden_size
self.num_heads = config.num_attention_heads
self.head_dim = self.embed_dim // self.num_heads
if self.head_dim * self.num_heads != self.embed_dim:
raise ValueError(
f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:"
f" {self.num_heads})."
)
self.scale = self.head_dim**-0.5
self.dropout = config.attention_dropout
# self.k_proj = nn.Linear(self.embed_dim, self.embed_dim)
# self.v_proj = nn.Linear(self.embed_dim, self.embed_dim)
# self.q_proj = nn.Linear(self.embed_dim, self.embed_dim)
# self.out_proj = nn.Linear(self.embed_dim, self.embed_dim)
self.vit_weight_bits = getattr(config, "vit_weight_bits", 32)
self.vit_weight_quant_method = getattr(config, "vit_weight_quant_method", "absmean")
self.vit_act_bits = getattr(config, "vit_act_bits", 32)
self.vit_act_quant_method = getattr(config, "vit_act_quant_method", "absmax_per_token")
print(f"siglip attention, weight bits: {self.vit_weight_bits}, weight quant method: {self.vit_weight_quant_method}, act bits: {self.vit_act_bits}, act quant method: {self.vit_act_quant_method}")
self.k_proj = BitLinear(self.embed_dim, self.embed_dim, weight_bits=self.vit_weight_bits, input_bits=self.vit_act_bits)
self.v_proj = BitLinear(self.embed_dim, self.embed_dim, weight_bits=self.vit_weight_bits, input_bits=self.vit_act_bits)
self.q_proj = BitLinear(self.embed_dim, self.embed_dim, weight_bits=self.vit_weight_bits, input_bits=self.vit_act_bits)
self.out_proj = BitLinear(self.embed_dim, self.embed_dim, weight_bits=self.vit_weight_bits, input_bits=self.vit_act_bits)
def forward(
self,
hidden_states: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
output_attentions: Optional[bool] = False,
) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
"""Input shape: Batch x Time x Channel"""
batch_size, q_len, _ = hidden_states.size()
query_states = self.q_proj(hidden_states)
key_states = self.k_proj(hidden_states)
value_states = self.v_proj(hidden_states)
query_states = query_states.view(batch_size, q_len, self.num_heads, self.head_dim).transpose(1, 2)
key_states = key_states.view(batch_size, q_len, self.num_heads, self.head_dim).transpose(1, 2)
value_states = value_states.view(batch_size, q_len, self.num_heads, self.head_dim).transpose(1, 2)
k_v_seq_len = key_states.shape[-2]
attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) * self.scale
if attn_weights.size() != (batch_size, self.num_heads, q_len, k_v_seq_len):
raise ValueError(
f"Attention weights should be of size {(batch_size, self.num_heads, q_len, k_v_seq_len)}, but is"
f" {attn_weights.size()}"
)
if attention_mask is not None:
if attention_mask.size() != (batch_size, 1, q_len, k_v_seq_len):
raise ValueError(
f"Attention mask should be of size {(batch_size, 1, q_len, k_v_seq_len)}, but is {attention_mask.size()}"
)
attn_weights = attn_weights + attention_mask
# upcast attention to fp32
attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
attn_weights = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)
attn_output = torch.matmul(attn_weights, value_states)
if attn_output.size() != (batch_size, self.num_heads, q_len, self.head_dim):
raise ValueError(
f"`attn_output` should be of size {(batch_size, self.num_heads, q_len, self.head_dim)}, but is"
f" {attn_output.size()}"
)
attn_output = attn_output.transpose(1, 2).contiguous()
attn_output = attn_output.reshape(batch_size, q_len, self.embed_dim)
attn_output = self.out_proj(attn_output)
return attn_output, attn_weights
class SiglipFlashAttention2(SiglipAttention):
"""
SiglipAttention flash attention module. This module inherits from `SiglipAttention` as the weights of the module stays
untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
flash attention and deal with padding tokens in case the input contains any of them.
"""
is_causal = False
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
# TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
# flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignment, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
# Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
# Adapted from transformers.models.llama.modeling_llama.LlamaFlashAttention2.forward
def forward(
self,
hidden_states: torch.Tensor,
attention_mask: Optional[torch.LongTensor] = None,
output_attentions: bool = False,
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
output_attentions = False
batch_size, q_len, _ = hidden_states.size()
query_states = self.q_proj(hidden_states)
key_states = self.k_proj(hidden_states)
value_states = self.v_proj(hidden_states)
# Flash attention requires the input to have the shape
# batch_size x seq_length x head_dim x hidden_dim
# therefore we just need to keep the original shape
query_states = query_states.view(batch_size, q_len, self.num_heads, self.head_dim)
key_states = key_states.view(batch_size, q_len, self.num_heads, self.head_dim)
value_states = value_states.view(batch_size, q_len, self.num_heads, self.head_dim)
dropout_rate = self.dropout if self.training else 0.0
# In PEFT, usually we cast the layer norms in float32 for training stability reasons
# therefore the input hidden states gets silently casted in float32. Hence, we need
# cast them back in the correct dtype just to be sure everything works as expected.
# This might slowdown training & inference so it is recommended to not cast the LayerNorms
# in fp32.
input_dtype = query_states.dtype
if input_dtype == torch.float32:
if torch.is_autocast_enabled():
target_dtype = torch.get_autocast_gpu_dtype()
# Handle the case where the model is quantized
elif hasattr(self.config, "_pre_quantization_dtype"):
target_dtype = self.config._pre_quantization_dtype
else:
target_dtype = self.q_proj.weight.dtype
logger.warning_once(
f"The input hidden states seems to be silently casted in float32, this might be related to"
f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
f" {target_dtype}."
)
query_states = query_states.to(target_dtype)
key_states = key_states.to(target_dtype)
value_states = value_states.to(target_dtype)
attn_output = _flash_attention_forward(
query_states,
key_states,
value_states,
attention_mask,
q_len,
dropout=dropout_rate,
is_causal=self.is_causal,
use_top_left_mask=self._flash_attn_uses_top_left_mask,
)
attn_output = attn_output.reshape(batch_size, q_len, self.embed_dim).contiguous()
attn_output = self.out_proj(attn_output)
if not output_attentions:
attn_weights = None
return attn_output, attn_weights
class SiglipSdpaAttention(SiglipAttention):
"""
Siglip attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
`SiglipAttention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
SDPA API.
"""
is_causal = False
# Adapted from SiglipAttention.forward and transformers.models.llama.modeling_llama.LlamaSdpaAttention.forward
def forward(
self,
hidden_states: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
output_attentions: Optional[bool] = False,
) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
if output_attentions:
# TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
logger.warning_once(
"SiglipModel is using SiglipSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, "
'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
)
return super().forward(
hidden_states=hidden_states,
attention_mask=attention_mask,
output_attentions=output_attentions,
)
batch_size, q_len, _ = hidden_states.size()
query_states = self.q_proj(hidden_states)
key_states = self.k_proj(hidden_states)
value_states = self.v_proj(hidden_states)
query_states = query_states.view(batch_size, q_len, self.num_heads, self.head_dim).transpose(1, 2)
key_states = key_states.view(batch_size, q_len, self.num_heads, self.head_dim).transpose(1, 2)
value_states = value_states.view(batch_size, q_len, self.num_heads, self.head_dim).transpose(1, 2)
# SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
# Reference: https://github.com/pytorch/pytorch/issues/112577.
if query_states.device.type == "cuda" and attention_mask is not None:
query_states = query_states.contiguous()
key_states = key_states.contiguous()
value_states = value_states.contiguous()
# We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment
# in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling.
is_causal = True if self.is_causal and q_len > 1 else False
attn_output = torch.nn.functional.scaled_dot_product_attention(
query_states,
key_states,
value_states,
attn_mask=attention_mask,
dropout_p=self.dropout if self.training else 0.0,
is_causal=is_causal,
)
attn_output = attn_output.transpose(1, 2).contiguous()
attn_output = attn_output.view(batch_size, q_len, self.embed_dim)
attn_output = self.out_proj(attn_output)
return attn_output, None
SIGLIP_ATTENTION_CLASSES = {
"eager": SiglipAttention,
"flash_attention_2": SiglipFlashAttention2,
"sdpa": SiglipSdpaAttention,
}
# Copied from transformers.models.clip.modeling_clip.CLIPMLP with CLIP->Siglip
class SiglipMLP(nn.Module):
def __init__(self, config):
super().__init__()
self.config = config
self.activation_fn = ACT2FN[config.hidden_act]
# self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
# self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)
self.vit_weight_bits = getattr(config, "vit_weight_bits", 32)
self.vit_weight_quant_method = getattr(config, "vit_weight_quant_method", "absmean")
self.vit_act_bits = getattr(config, "vit_act_bits", 32)
self.vit_act_quant_method = getattr(config, "vit_act_quant_method", "absmax_per_token")
print(f"siglip mlp, weight bits: {self.vit_weight_bits}, weight quant method: {self.vit_weight_quant_method}, act bits: {self.vit_act_bits}, act quant method: {self.vit_act_quant_method}")
self.fc1 = BitLinear(config.hidden_size, config.intermediate_size, weight_bits=self.vit_weight_bits, input_bits=self.vit_act_bits)
self.fc2 = BitLinear(config.intermediate_size, config.hidden_size, weight_bits=self.vit_weight_bits, input_bits=self.vit_act_bits)
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
hidden_states = self.fc1(hidden_states)
hidden_states = self.activation_fn(hidden_states)
hidden_states = self.fc2(hidden_states)
return hidden_states
class SiglipEncoderLayer(nn.Module):
def __init__(self, config: SiglipConfig):
super().__init__()
self.embed_dim = config.hidden_size
self.self_attn = SIGLIP_ATTENTION_CLASSES[config._attn_implementation](config=config)
self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
self.mlp = SiglipMLP(config)
self.layer_norm2 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
# Ignore copy
def forward(
self,
hidden_states: torch.Tensor,
attention_mask: torch.Tensor,
output_attentions: Optional[bool] = False,
) -> Tuple[torch.FloatTensor]:
"""
Args:
hidden_states (`torch.FloatTensor`):
Input to the layer of shape `(batch, seq_len, embed_dim)`.
attention_mask (`torch.FloatTensor`):
Attention mask of shape `(batch, 1, q_len, k_v_seq_len)` where padding elements are indicated by very large negative values.
output_attentions (`bool`, *optional*, defaults to `False`):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
returned tensors for more detail.
"""
residual = hidden_states
hidden_states = self.layer_norm1(hidden_states)
hidden_states, attn_weights = self.self_attn(
hidden_states=hidden_states,
attention_mask=attention_mask,
output_attentions=output_attentions,
)
hidden_states = residual + hidden_states
residual = hidden_states
hidden_states = self.layer_norm2(hidden_states)
hidden_states = self.mlp(hidden_states)
hidden_states = residual + hidden_states
outputs = (hidden_states,)
if output_attentions:
outputs += (attn_weights,)
return outputs
class SiglipPreTrainedModel(PreTrainedModel):
"""
An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
models.
"""
config_class = SiglipConfig
base_model_prefix = "siglip"
supports_gradient_checkpointing = True
_no_split_modules = [
"SiglipTextEmbeddings",
"SiglipEncoderLayer",
"SiglipVisionEmbeddings",
"SiglipEncoderLayer",
"SiglipMultiheadAttentionPoolingHead",
]
_supports_flash_attn_2 = True
_supports_sdpa = True
def _init_weights(self, module):
"""Initialize the weights"""
if isinstance(module, SiglipVisionEmbeddings):
width = (
self.config.vision_config.hidden_size
if isinstance(self.config, SiglipConfig)
else self.config.hidden_size
)
nn.init.normal_(module.position_embedding.weight, std=1 / np.sqrt(width))
elif isinstance(module, nn.Embedding):
default_flax_embed_init(module.weight)
elif isinstance(module, SiglipAttention):
nn.init.xavier_uniform_(module.q_proj.weight)
nn.init.xavier_uniform_(module.k_proj.weight)
nn.init.xavier_uniform_(module.v_proj.weight)
nn.init.xavier_uniform_(module.out_proj.weight)
nn.init.zeros_(module.q_proj.bias)
nn.init.zeros_(module.k_proj.bias)
nn.init.zeros_(module.v_proj.bias)
nn.init.zeros_(module.out_proj.bias)
elif isinstance(module, SiglipMLP):
nn.init.xavier_uniform_(module.fc1.weight)
nn.init.xavier_uniform_(module.fc2.weight)
nn.init.normal_(module.fc1.bias, std=1e-6)
nn.init.normal_(module.fc2.bias, std=1e-6)
elif isinstance(module, SiglipMultiheadAttentionPoolingHead):
nn.init.xavier_uniform_(module.probe.data)
nn.init.xavier_uniform_(module.attention.in_proj_weight.data)
nn.init.zeros_(module.attention.in_proj_bias.data)
elif isinstance(module, SiglipModel):
logit_scale_init = torch.log(torch.tensor(1.0))
module.logit_scale.data.fill_(logit_scale_init)
module.logit_bias.data.zero_()
elif isinstance(module, SiglipForImageClassification):
nn.init.normal_(
module.classifier.weight,
std=self.config.vision_config.hidden_size**-0.5 * self.config.initializer_factor,
)
elif isinstance(module, (nn.Linear, nn.Conv2d)):
lecun_normal_(module.weight)
if module.bias is not None:
nn.init.zeros_(module.bias)
elif isinstance(module, nn.LayerNorm):
module.bias.data.zero_()
module.weight.data.fill_(1.0)
SIGLIP_START_DOCSTRING = r"""
This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
etc.)
This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
and behavior.
Parameters:
config ([`SiglipConfig`]): Model configuration class with all the parameters of the model.
Initializing with a config file does not load the weights associated with the model, only the
configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
"""
SIGLIP_TEXT_INPUTS_DOCSTRING = r"""
Args:
input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
it.
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
[`PreTrainedTokenizer.__call__`] for details.
[What are input IDs?](../glossary#input-ids)
attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
- 1 for tokens that are **not masked**,
- 0 for tokens that are **masked**.
[What are attention masks?](../glossary#attention-mask)
position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
config.max_position_embeddings - 1]`.
[What are position IDs?](../glossary#position-ids)
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
tensors for more detail.
output_hidden_states (`bool`, *optional*):
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
more detail.
return_dict (`bool`, *optional*):
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
"""
SIGLIP_VISION_INPUTS_DOCSTRING = r"""
Args:
pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
Pixel values. Padding will be ignored by default should you provide it. Pixel values can be obtained using
[`AutoImageProcessor`]. See [`CLIPImageProcessor.__call__`] for details.
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
tensors for more detail.
output_hidden_states (`bool`, *optional*):
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
more detail.
interpolate_pos_encoding (`bool`, *optional*, defaults to `False`):
Whether to interpolate the pre-trained position encodings.
return_dict (`bool`, *optional*):
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
"""
SIGLIP_INPUTS_DOCSTRING = r"""
Args:
input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
it.
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
[`PreTrainedTokenizer.__call__`] for details.
[What are input IDs?](../glossary#input-ids)
attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
- 1 for tokens that are **not masked**,
- 0 for tokens that are **masked**.
[What are attention masks?](../glossary#attention-mask)
position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
config.max_position_embeddings - 1]`.
[What are position IDs?](../glossary#position-ids)
pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
Pixel values. Padding will be ignored by default should you provide it. Pixel values can be obtained using
[`AutoImageProcessor`]. See [`CLIPImageProcessor.__call__`] for details.
return_loss (`bool`, *optional*):
Whether or not to return the contrastive loss.
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
tensors for more detail.
output_hidden_states (`bool`, *optional*):
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
more detail.
interpolate_pos_encoding (`bool`, *optional*, defaults to `False`):
Whether to interpolate the pre-trained position encodings.
return_dict (`bool`, *optional*):
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
"""
# Copied from transformers.models.altclip.modeling_altclip.AltCLIPEncoder with AltCLIP->Siglip
class SiglipEncoder(nn.Module):
"""
Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a
[`SiglipEncoderLayer`].
Args:
config: SiglipConfig
"""
def __init__(self, config: SiglipConfig):
super().__init__()
self.config = config
self.layers = nn.ModuleList([SiglipEncoderLayer(config) for _ in range(config.num_hidden_layers)])
self.gradient_checkpointing = False
# Ignore copy
def forward(
self,
inputs_embeds,
attention_mask: Optional[torch.Tensor] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
) -> Union[Tuple, BaseModelOutput]:
r"""
Args:
inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):