File size: 46,400 Bytes
ace95dc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
85fc5eb
ace95dc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
85fc5eb
ace95dc
 
 
 
 
 
94df028
ace95dc
94df028
4edb1f9
ace95dc
 
 
94df028
ace95dc
94df028
ace95dc
 
 
 
 
 
 
 
 
4edb1f9
ace95dc
 
 
 
 
 
 
4edb1f9
 
 
 
 
 
 
 
 
 
ace95dc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
63ab0e0
 
 
 
8c5c0a2
 
 
 
 
 
 
 
 
ace95dc
8c5c0a2
ace95dc
8c5c0a2
ace95dc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
63ab0e0
ace95dc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
63ab0e0
ace95dc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f1e304d
ace95dc
 
 
f1e304d
ace95dc
 
 
 
 
 
 
 
f1e304d
ace95dc
f1e304d
 
 
ace95dc
f1e304d
 
 
 
 
 
ace95dc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4edb1f9
ace95dc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
94df028
ace95dc
 
 
94df028
ace95dc
 
 
 
94df028
ace95dc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
94df028
ace95dc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4edb1f9
 
ace95dc
 
94df028
ace95dc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
94df028
ace95dc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
94df028
ace95dc
 
 
 
94df028
 
 
ace95dc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4edb1f9
 
ace95dc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
94df028
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ace95dc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4edb1f9
ace95dc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
94df028
ace95dc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
94df028
ace95dc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4edb1f9
ace95dc
 
 
 
 
 
 
 
 
 
 
4edb1f9
ace95dc
 
 
4edb1f9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ace95dc
f1e304d
ace95dc
 
 
 
 
 
 
8c5c0a2
 
 
ace95dc
 
 
 
 
 
 
 
 
 
 
 
4edb1f9
ace95dc
 
 
 
 
4edb1f9
ace95dc
 
 
 
 
 
f1e304d
 
 
 
 
 
ace95dc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
63ab0e0
ace95dc
 
 
 
 
 
8c5c0a2
ace95dc
 
 
 
 
 
 
 
 
4edb1f9
ace95dc
 
 
 
 
 
 
 
 
 
 
8c5c0a2
ace95dc
 
 
 
 
 
 
 
 
 
 
 
4edb1f9
ace95dc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4edb1f9
ace95dc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8c5c0a2
ace95dc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f1e304d
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
from logging import warn
from transformers.models.roberta.modeling_roberta import *
import torch
import torch.nn as nn
from transformers.models.roberta.configuration_roberta import RobertaConfig
import sys

AUTO_MAP = {
        "AutoModel": "modeling_lsg_roberta.LSGRobertaModel",
        "AutoModelForCausalLM": "modeling_lsg_roberta.LSGRobertaForCausalLM",
        "AutoModelForMaskedLM": "modeling_lsg_roberta.LSGRobertaForMaskedLM",
        "AutoModelForMultipleChoice": "modeling_lsg_roberta.LSGRobertaForMultipleChoice",
        "AutoModelForQuestionAnswering": "modeling_lsg_roberta.LSGRobertaForQuestionAnswering",
        "AutoModelForSequenceClassification": "modeling_lsg_roberta.LSGRobertaForSequenceClassification",
        "AutoModelForTokenClassification": "modeling_lsg_roberta.LSGRobertaForTokenClassification"
    }

class LSGRobertaConfig(RobertaConfig):
    """
    This class overrides :class:`~transformers.RobertaConfig`. Please check the superclass for the appropriate
    documentation alongside usage examples.
    """

    base_model_prefix = "lsg"
    model_type = "roberta"

    def __init__(
        self,
        adaptive=True,
        base_model_prefix="lsg",
        block_size=128,
        lsh_num_pre_rounds=1,
        mask_first_token=False,
        num_global_tokens=1,
        pool_with_global=True,
        sparse_block_size=128,
        sparsity_factor=2,
        sparsity_type="norm",
        **kwargs
        ):
        """Constructs LSGRobertaConfig."""
        super().__init__(**kwargs)

        self.adaptive = adaptive
        self.auto_map = AUTO_MAP
        self.base_model_prefix = base_model_prefix
        self.block_size = block_size
        self.lsh_num_pre_rounds = lsh_num_pre_rounds
        self.mask_first_token = mask_first_token
        self.num_global_tokens = num_global_tokens
        self.pool_with_global = pool_with_global
        self.sparse_block_size = sparse_block_size
        self.sparsity_factor = sparsity_factor
        self.sparsity_type = sparsity_type

        if sparsity_type not in [None, "none", "norm", "lsh", "pooling", "stride", "block_stride", "bos_pooling"]:
            logger.warning(
                "[WARNING CONFIG]: sparsity_mode not in [None, 'none', 'norm', 'lsh', 'pooling', 'stride', 'block_stride', 'bos_pooling'], \
                    setting sparsity_type=None, computation will skip sparse attention")
            self.sparsity_type = None

        if self.sparsity_type in ["stride", "block_stride"]:
            if self.sparsity_factor > self.num_attention_heads:
                logger.warning(
                "[WARNING CONFIG]: sparsity_factor > num_attention_heads is not recommended for stride/block_stride sparsity"
            )
        
        if self.num_global_tokens < 1:
            logger.warning(
                "[WARNING CONFIG]: num_global_tokens < 1 is not compatible, setting num_global_tokens=1"
            )
            self.num_global_tokens = 1
        elif self.num_global_tokens > 512:
            logger.warning(
                "[WARNING CONFIG]: num_global_tokens > 512 is not allowed, setting num_global_tokens=512"
            )
            self.num_global_tokens = 512
        
        if self.sparsity_factor > 0:
            assert self.block_size % self.sparsity_factor == 0, "[ERROR CONFIG]: block_size must be divisible by sparsity_factor"
            assert self.block_size//self.sparsity_factor >= 1, "[ERROR CONFIG]: make sure block_size >= sparsity_factor"
            
        if self.mask_first_token and not pool_with_global:
            logger.warning(
                "[WARNING CONFIG]: pool_with_global==False is not compatible with mask_first_token==True. Setting pool_with_global to True.")
            self.pool_with_global = True
        
        if hasattr(self, "position_embedding_type"):
            if self.position_embedding_type != "absolute":
                logger.warning(
                "[WARNING CONFIG]: LSG Attention is not compatible with relative positional embedding and will skip its computation. Set position_embedding_type='absolute' to remove this warning.")
            
        
class BaseSelfAttention(nn.Module):
    
    def init_modules(self, config):
        if config.hidden_size % config.num_attention_heads != 0 and not hasattr(
            config, "embedding_size"
        ):
            raise ValueError(
                "The hidden size (%d) is not a multiple of the number of attention "
                "heads (%d)" % (config.hidden_size, config.num_attention_heads)
            )

        self.num_attention_heads = config.num_attention_heads
        self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
        self.all_head_size = self.num_attention_heads * self.attention_head_size

        self.query = nn.Linear(config.hidden_size, self.all_head_size)
        self.key = nn.Linear(config.hidden_size, self.all_head_size)
        self.value = nn.Linear(config.hidden_size, self.all_head_size)

        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)

    def transpose_for_scores(self, x):
        new_x_shape = x.size()[:-1] + (
            self.num_attention_heads,
            self.attention_head_size,
        )
        x = x.view(*new_x_shape)
        return x.permute(0, 2, 1, 3)

    def reshape_output(self, context_layer):
        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
        return context_layer.view(*new_context_layer_shape)

    def project_QKV(self, hidden_states):

        query_layer = self.transpose_for_scores(self.query(hidden_states))
        key_layer = self.transpose_for_scores(self.key(hidden_states))
        value_layer = self.transpose_for_scores(self.value(hidden_states))
        return query_layer, key_layer, value_layer


class BaseAttentionProduct(nn.Module):

    def __init__(self, config):
        """
        Compute attention: softmax(Q @ K.T) @ V
        """
        super().__init__()
        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)

    def forward(self, query_layer, key_layer, value_layer, attention_mask=None):
        
        d = query_layer.shape[-1]

        # Take the dot product between "query" and "key" to get the raw attention scores.
        attention_scores = query_layer @ key_layer.transpose(-1, -2) / math.sqrt(d)

        del query_layer
        del key_layer

        if attention_mask is not None:
            # Apply the attention mask is (precomputed for all layers in RobertaModel forward() function)
            attention_scores = attention_scores + attention_mask
            del attention_mask

        # Normalize the attention scores to probabilities.
        attention_probs = nn.Softmax(dim=-1)(attention_scores)

        # This is actually dropping out entire tokens to attend to, which might
        # seem a bit unusual, but is taken from the original Transformer paper.
        context_layer = self.dropout(attention_probs) @ value_layer

        return context_layer


class CausalAttentionProduct(nn.Module):

    def __init__(self, config):
        """
        Compute attention: softmax(Q @ K.T) @ V
        """
        super().__init__()
        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
        self.block_size = config.block_size

    def forward(self, query_layer, key_layer, value_layer, attention_mask=None, causal_shape=None):
        
        d = query_layer.shape[-1]

        # Take the dot product between "query" and "key" to get the raw attention scores.
        attention_scores = query_layer @ key_layer.transpose(-1, -2) / math.sqrt(d)

        del query_layer
        del key_layer

        if attention_mask is not None:
            # Add causal mask
            causal_shape = (self.block_size, self.block_size) if causal_shape is None else causal_shape
            causal_mask = torch.tril(
                torch.ones(*causal_shape, device=attention_mask.device, dtype=attention_scores.dtype), 
                diagonal=-1
                ) 
            
            # Min value
            dtype_min = torch.tensor(
                        torch.finfo(attention_scores.dtype).min, device=attention_scores.device, dtype=attention_scores.dtype
                    )

            # Build causal + attention_mask
            causal_mask = torch.nn.functional.pad(causal_mask.T * dtype_min, (attention_mask.size()[-1] - self.block_size, 0), value=0)
            attention_mask = torch.max(attention_mask + causal_mask.unsqueeze(0).unsqueeze(0).unsqueeze(0), dtype_min)

            attention_scores = attention_scores + attention_mask
            del attention_mask
            del causal_mask

        # Normalize the attention scores to probabilities.
        attention_probs = nn.Softmax(dim=-1)(attention_scores)

        # This is actually dropping out entire tokens to attend to, which might
        # seem a bit unusual, but is taken from the original Transformer paper.
        context_layer = self.dropout(attention_probs) @ value_layer

        return context_layer


class LSGAttentionProduct(nn.Module):

    def __init__(self, config, block_size=None, sparse_block_size=None, sparsity_factor=4, is_causal=False):
        """
        Compute block or overlapping blocks attention products
        """
        super().__init__()
 
        self.block_size = block_size
        self.sparse_block_size = sparse_block_size
        self.sparsity_factor = sparsity_factor
        self.is_causal = is_causal

        if self.block_size is None:
            self.block_size = config.block_size

        if self.sparse_block_size is None:
            self.sparse_block_size = config.sparse_block_size

        # Shape of blocks
        self.local_shapes = (self.block_size*3, self.block_size)
        if self.sparse_block_size and self.sparsity_factor > 0:
            self.sparse_shapes = (self.sparse_block_size*3, self.block_size//self.sparsity_factor)

        if is_causal:
            self.attention = CausalAttentionProduct(config)
        else:
            self.attention = BaseAttentionProduct(config)
        
    def build_lsg_inputs(self, hidden_states, sparse_hidden_states, global_hidden_states, is_attn_mask=False):
        
        # Build local tokens
        local_hidden_states = self.reshape_to_local_block(hidden_states, is_attn_mask)
        del hidden_states

        # Build sparse tokens
        if sparse_hidden_states is not None:
            sparse_hidden_states = self.reshape_to_sparse_block(sparse_hidden_states, is_attn_mask)
        
        return self.cat_global_sparse_local_tokens(global_hidden_states, sparse_hidden_states, local_hidden_states)

    def forward(
        self, 
        query_layer, 
        key_layer, 
        value_layer, 
        attention_mask=None, 
        sparse_key=None,
        sparse_value=None, 
        sparse_mask=None, 
        global_key=None, 
        global_value=None, 
        global_mask=None
        ):

        # Input batch, heads, length, hidden_size
        n, h, t, d = query_layer.size()
        n_blocks = t // self.block_size
        assert t % self.block_size == 0

        key_layer = self.build_lsg_inputs(
            key_layer, 
            sparse_key, 
            global_key
            )
        del sparse_key
        del global_key

        value_layer = self.build_lsg_inputs(
            value_layer, 
            sparse_value, 
            global_value
            )
        del sparse_value
        del global_value

        attention_mask = self.build_lsg_inputs(
            attention_mask, 
            sparse_mask, 
            global_mask.transpose(-1, -2), 
            is_attn_mask=True
            ).transpose(-1, -2)
        del sparse_mask
        del global_mask

        # expect (..., t, d) shape
        # Compute attention
        context_layer = self.attention(
                query_layer=self.chunk(query_layer, n_blocks), 
                key_layer=key_layer,
                value_layer=value_layer,
                attention_mask=attention_mask
                )
                
        return context_layer.reshape(n, h, -1, d)
    
    def reshape_to_local_block(self, hidden_states, is_attn_mask=False):
        
        size, step = self.local_shapes
        s = (size - step) // 2

        # Pad before block reshaping
        if is_attn_mask:
            pad_value = torch.finfo(hidden_states.dtype).min 
            hidden_states = hidden_states.transpose(-1, -2)
        else: 
            pad_value = 0

        hidden_states = torch.nn.functional.pad(
            hidden_states.transpose(-1, -2), 
            pad=(s, s),
            value=pad_value
            ).transpose(-1, -2)

        # Make blocks
        hidden_states = hidden_states.unfold(-2, size=size, step=step).transpose(-1, -2)

        # Skip third block if causal
        if self.is_causal:
            return hidden_states[..., :size*2//3, :]

        return hidden_states

    def reshape_to_sparse_block(self, hidden_states, is_attn_mask=False):
        
        size, step = self.sparse_shapes

        # In case of odd case
        odd_offset = (step % 2)

        # n, h, t, d*2 + 1
        size = size*2 
        s = (size - step) // 2 + odd_offset

        # Pad before block reshaping
        if is_attn_mask:
            pad_value = torch.finfo(hidden_states.dtype).min  
            hidden_states = hidden_states.transpose(-1, -2)
        else: 
            pad_value = 0

        hidden_states = torch.nn.functional.pad(
            hidden_states.transpose(-1, -2), 
            pad=(s, s),
            value=pad_value
            ).transpose(-1, -2)

        # Make blocks
        hidden_states = hidden_states.unfold(-2, size=size, step=step).transpose(-1, -2)

        # Fix case where block_size == sparsify_factor
        if odd_offset: 
            hidden_states = hidden_states[..., :-1, :, :]

        # Indexes for selection
        u = (size - self.block_size * 3 // self.sparsity_factor) // 2 + odd_offset
        s = self.sparse_block_size

        # Skip right block if causal
        if self.is_causal:
            return hidden_states[..., u-s:u, :]

        u_ = u + odd_offset
        return torch.cat([hidden_states[..., u-s:u, :], hidden_states[..., -u_:-u_+s, :]], dim=-2)

    def cat_global_sparse_local_tokens(self, x_global, x_sparse=None, x_local=None, dim=-2):

        n, h, b, t, d = x_local.size()
        x_global = x_global.unsqueeze(-3).expand(-1, -1, b, -1, -1)
        if x_sparse is not None:
            return torch.cat([x_global, x_sparse, x_local], dim=dim)
        return torch.cat([x_global, x_local], dim=dim)

    def chunk(self, x, n_blocks):

        t, d = x.size()[-2:]
        return x.reshape(*x.size()[:-2], n_blocks, -1, d)


class LSGRobertaEmbeddings(RobertaEmbeddings):

    def __init__(self, config):
        super().__init__(config)

        self.num_global_tokens = config.num_global_tokens

        # Hardcoded but partially trained
        self.global_embeddings = nn.Embedding(512, embedding_dim=config.hidden_size, )

        self.block_size = config.block_size

    def forward(
        self, input_ids=None, token_type_ids=None, position_ids=None, inputs_embeds=None, past_key_values_length=0
    ):
        if position_ids is None:
            if input_ids is not None:
                # Create the position ids from the input token ids. Any padded tokens remain padded.
                position_ids = create_position_ids_from_input_ids(input_ids, self.padding_idx, past_key_values_length)
            else:
                position_ids = self.create_position_ids_from_inputs_embeds(inputs_embeds)

        if input_ids is not None:
            input_shape = input_ids.size()
        else:
            input_shape = inputs_embeds.size()[:-1]

        seq_length = input_shape[1]

        # Setting the token_type_ids to the registered buffer in constructor where it is all zeros, which usually occurs
        # when its auto-generated, registered buffer helps users when tracing the model without passing token_type_ids, solves
        # issue #5664
        if token_type_ids is None:
            if hasattr(self, "token_type_ids"):
                buffered_token_type_ids = self.token_type_ids[:, :seq_length]
                buffered_token_type_ids_expanded = buffered_token_type_ids.expand(input_shape[0], seq_length)
                token_type_ids = buffered_token_type_ids_expanded
            else:
                token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=self.position_ids.device)

        if inputs_embeds is None:
            inputs_embeds = self.word_embeddings(input_ids)
        token_type_embeddings = self.token_type_embeddings(token_type_ids[:, :seq_length])

        embeddings = inputs_embeds + token_type_embeddings
        if self.position_embedding_type == "absolute":
            position_embeddings = self.position_embeddings(position_ids[:, :seq_length])
            embeddings += position_embeddings

        #if self.num_global_tokens < 0:
        n, t, d = embeddings.size()
        
        # Add global_tokens
        indexes = torch.arange(self.num_global_tokens, device=embeddings.device).reshape(1, -1)
        global_embeddings = self.global_embeddings(indexes) 
        embeddings = torch.cat([global_embeddings.expand(n, -1, d), embeddings], dim=-2)
        
        embeddings = self.LayerNorm(embeddings)
        embeddings = self.dropout(embeddings)
        return embeddings


class LSGAttention(RobertaAttention):

    def __init__(self, config):
        
        super().__init__(config)
        
        self.self = LSGSelfAttention(config)


class LSGSelfAttention(BaseSelfAttention):
    '''
    Compute local attention with overlapping blocs
    Use global attention for tokens with highest norm
    '''
    def __init__(self, config):
        super().__init__()

        self.init_modules(config)

        self.block_size = config.block_size
        self.sparse_block_size = config.sparse_block_size
        self.num_global_tokens = config.num_global_tokens
        self.sparsity_factor = config.sparsity_factor
        self.is_causal = config.is_decoder
        self.is_decoder = config.is_decoder

        self.attention = LSGAttentionProduct(
            config, 
            block_size=config.block_size, 
            sparse_block_size=config.sparse_block_size, 
            sparsity_factor=self.sparsity_factor, 
            is_causal=self.is_causal
            )

        if self.is_causal:
            self.causal_attention = CausalAttentionProduct(config)
        self.full_attention = BaseAttentionProduct(config)

        sparse_functions = {
            "norm": self.get_sparse_tokens_with_norm, 
            "pooling": self.get_sparse_tokens_with_pooling,
            "lsh": self.get_sparse_tokens_with_lsh,
            "stride": self.get_sparse_tokens_with_stride,
            "block_stride": self.get_sparse_tokens_with_block_stride,
            "bos_pooling": self.get_sparse_tokens_with_bos_pooling
            }
        
        self.sparsity_type = config.sparsity_type
        self.get_sparse_elements = sparse_functions.get(self.sparsity_type, lambda w, x, y, z: (None, None, None))
            
        if config.sparsity_type == "lsh":
            self.lsh_num_pre_rounds = config.lsh_num_pre_rounds

    def get_sparse_tokens_with_norm(self, queries, keys, values, mask):
        
        if self.sparsity_factor == 1:
            return keys, values, mask.expand(-1, keys.size()[1], -1, -1)

        with torch.no_grad():

            block_size = min(self.block_size, self.sparse_block_size)
            key_norm = keys.detach().norm(dim=-1, keepdim=True)
            key_norm = key_norm * ~mask.transpose(-1, -2).bool()
            key_norm = self.chunk(key_norm, block_size)

            n, h, b, t, d = key_norm.size()
            
            idx = key_norm.argsort(dim=-2) 
            del key_norm
            idx += (torch.arange(b, device=keys.device)*t).reshape(1, 1, b, 1, 1)

            split = (t - block_size // self.sparsity_factor, block_size // self.sparsity_factor)
            sparse_idx = idx.split(split, -2)[-1].reshape(n, h, -1, 1)
        
        d = keys.size()[-1]
        keys = keys.gather(dim=-2, index=sparse_idx.expand(-1, -1, -1, d))
        values = values.gather(dim=-2, index=sparse_idx.expand(-1, -1, -1, d))
        mask = mask.expand(-1, h, -1, -1).transpose(-1, -2).gather(dim=-2, index=sparse_idx).transpose(-1, -2)

        return keys, values, mask

    def get_sparse_tokens_with_pooling(self, queries, keys, values, mask):
        
        if self.sparsity_factor == 1:
            return keys, values, mask.expand(-1, keys.size()[1], -1, -1)

        keys = self.chunk(keys, self.sparsity_factor)
        values = self.chunk(values, self.sparsity_factor)

        n, h, b, t, d = keys.size()
        mask = mask.reshape(n, 1, b, 1, t)
        mask = ~mask.transpose(-1, -2).bool()

        keys = keys * mask
        values = values * mask

        mask = mask.sum(dim=-2)
        keys = keys.sum(dim=-2) / (mask + 1e-6)
        values = values.sum(dim=-2) / (mask + 1e-6)

        mask = (1. - mask.clamp(0, 1)) 
        mask *= torch.finfo(mask.dtype).min
        return keys.reshape(n, h, -1, d), values.reshape(n, h, -1, d), mask.expand(-1, h, -1, -1).transpose(-1, -2)

    def get_sparse_tokens_with_stride(self, queries, keys, values, mask):

        if self.sparsity_factor == 1:
            return keys, values, mask.expand(-1, keys.size()[1], -1, -1)

        n, h, t, d = keys.size()
        sparse_idx = torch.arange(t // self.sparsity_factor, device=keys.device) * self.sparsity_factor
        sparse_idx = sparse_idx.reshape(1, 1, -1, 1) + (torch.arange(h, device=keys.device) % self.sparsity_factor).reshape(1, h, 1, 1)
        sparse_idx = sparse_idx.expand(n, h, -1, 1)

        keys = keys.gather(dim=-2, index=sparse_idx.expand(-1, -1, -1, d))
        values = values.gather(dim=-2, index=sparse_idx.expand(-1, -1, -1, d))
        mask = mask.expand(-1, h, -1, -1).transpose(-1, -2).gather(dim=-2, index=sparse_idx).transpose(-1, -2)

        return keys, values, mask

    def get_sparse_tokens_with_block_stride(self, queries, keys, values, mask):

        if self.sparsity_factor == 1:
            return keys, values, mask.expand(-1, keys.size()[1], -1, -1)

        n, h, t, d = keys.size()

        t, b = self.block_size, t // self.block_size
        sparse_idx = torch.arange(t // self.sparsity_factor, device=keys.device)
        sparse_idx = sparse_idx.reshape(1, 1, 1, -1, 1) + torch.arange(h, device=keys.device).reshape(1, h, 1, 1, 1) * (t // self.sparsity_factor)
        sparse_idx = (sparse_idx % t) 
        sparse_idx = sparse_idx + torch.arange(b, device=keys.device).reshape(1, 1, -1, 1, 1) * t
        sparse_idx = sparse_idx.reshape(1, h, -1, 1).expand(n, h, -1, 1)

        keys = keys.gather(dim=-2, index=sparse_idx.expand(-1, -1, -1, d))
        values = values.gather(dim=-2, index=sparse_idx.expand(-1, -1, -1, d))
        mask = mask.expand(-1, h, -1, -1).transpose(-1, -2).gather(dim=-2, index=sparse_idx).transpose(-1, -2)

        return keys, values, mask
        
    def get_sparse_tokens_with_lsh(self, queries, keys, values, mask):
        
        if self.sparsity_factor == 1:
            return keys, values, mask.expand(-1, keys.size()[1], -1, -1)

        if self.sparsity_factor == self.sparse_block_size:
            return self.get_sparse_tokens_with_bos_pooling(queries, keys, values, mask)
        
        block_size = min(self.block_size, self.sparse_block_size)
        keys = self.chunk(keys, block_size)
        values = self.chunk(values, block_size)

        n, h, b, t, d = keys.size()
        mask = mask.reshape(n, 1, b, 1, t)
        mask = ~mask.transpose(-1, -2).bool()

        keys = keys * mask
        values = values * mask
        mask = mask.expand(-1, h, -1, -1, -1).float()

        extra_factor = 1
        
        for _ in range(self.lsh_num_pre_rounds):
            keys, values, mask = self.lsh_round(keys, values, mask, t*extra_factor)

        keys, values, mask = self.lsh_round(keys, values, mask, t//self.sparsity_factor)
        keys /= mask + 1e-8
        values /= mask + 1e-8

        mask = (1. - mask.clamp(0, 1)) 
        mask *= torch.finfo(mask.dtype).min

        return keys.reshape(n, h, -1, d), values.reshape(n, h, -1, d), mask.transpose(-1, -2).reshape(n, h, 1, -1)

    def lsh_round(self, keys, values, mask, output_size):

        with torch.no_grad():

            n_hashes = output_size // 2
            n, h, b, t, d = keys.size()
            binary_mask = mask.clamp(0, 1)

            indexes = (torch.nn.functional.normalize(keys, dim=-1) * binary_mask) @ torch.randn(1, h, 1, d, n_hashes, device=keys.device)
            indexes = torch.cat([indexes, -indexes], dim=-1).argmax(dim=-1, keepdim=True)

        n, h, b, t, d = keys.size()
        
        x_ = torch.zeros(n, h, b, output_size, d, device=keys.device)
        mask_ = torch.zeros(n, h, b, output_size, 1, device=keys.device)
        keys = torch.scatter_add(x_, dim=-2, index=indexes.expand(-1, -1, -1, -1, d), src=keys)
        values = torch.scatter_add(x_, dim=-2, index=indexes.expand(-1, -1, -1, -1, d), src=values)
        mask = torch.scatter_add(mask_, dim=-2, index=indexes, src=mask)

        return keys[..., :output_size, :], values[..., :output_size, :], mask[..., :output_size, :]

    def get_sparse_tokens_with_bos_pooling(self, queries, keys, values, mask):
        
        if self.sparsity_factor == 1:
            return keys, values, mask.expand(-1, keys.size()[1], -1, -1)

        queries = queries.unsqueeze(-3)
        mask = self.chunk(mask.transpose(-1, -2), self.sparsity_factor).transpose(-1, -2)
        keys = self.chunk(keys, self.sparsity_factor)
        values = self.chunk(values, self.sparsity_factor)
        
        n, h, b, t, d = keys.size()
        scores = (queries[..., :1, :] @ keys.transpose(-1, -2)) / math.sqrt(d)
        if mask is not None:
            scores = scores + mask

        scores = torch.softmax(scores, dim=-1)
        keys = scores @ keys 
        values = scores @ values
        mask = mask.mean(dim=-1)
        mask[mask != torch.finfo(mask.dtype).min] = 0
        
        return keys.reshape(n, h, -1, d), values.reshape(n, h, -1, d), mask.expand(-1, h, -1, -1).transpose(-1, -2)
    
    def forward(
        self,
        hidden_states,
        attention_mask=None,
        head_mask=None,
        encoder_hidden_states=None,
        encoder_attention_mask=None,
        past_key_value=None,
        output_attentions=False,
        ):

        query_layer = self.query(hidden_states)

        # If this is instantiated as a cross-attention module, the keys
        # and values come from an encoder; the attention mask needs to be
        # such that the encoder's padding tokens are not attended to.
        is_cross_attention = encoder_hidden_states is not None

        if is_cross_attention and past_key_value is not None:
            # reuse k,v, cross_attentions
            key_layer = past_key_value[0]
            value_layer = past_key_value[1]
            attention_mask = encoder_attention_mask
        elif is_cross_attention:
            key_layer = self.transpose_for_scores(self.key(encoder_hidden_states))
            value_layer = self.transpose_for_scores(self.value(encoder_hidden_states))
            attention_mask = encoder_attention_mask
        elif past_key_value is not None:
            key_layer = self.transpose_for_scores(self.key(hidden_states))
            value_layer = self.transpose_for_scores(self.value(hidden_states))
            key_layer = torch.cat([past_key_value[0], key_layer], dim=2)
            value_layer = torch.cat([past_key_value[1], value_layer], dim=2)
        else:
            key_layer = self.transpose_for_scores(self.key(hidden_states))
            value_layer = self.transpose_for_scores(self.value(hidden_states))

        query_layer = self.transpose_for_scores(query_layer)

        if self.is_decoder:
            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
            # Further calls to cross_attention layer can then reuse all cross-attention
            # key/value_states (first "if" case)
            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
            # all previous decoder key/value_states. Further calls to uni-directional self-attention
            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
            # if encoder bi-directional self-attention `past_key_value` is always `None`
            past_key_value = (key_layer, value_layer)

            if is_cross_attention:
                outputs = self.cross_attention_forward(
                    query_layer=query_layer, 
                    key_layer=key_layer, 
                    value_layer=value_layer, 
                    attention_mask=attention_mask,
                    output_attentions=output_attentions
                    )
            else:
                outputs = self.causal_forward(
                    query_layer,
                    key_layer,
                    value_layer,
                    attention_mask=attention_mask,
                    output_attentions=output_attentions,
                )

            outputs = outputs + ((key_layer, value_layer),)
            
        else:
            outputs = self.not_causal_forward(
                query_layer,
                key_layer,
                value_layer, 
                attention_mask=attention_mask, 
                output_attentions=output_attentions
                )

        return outputs

    def causal_forward(
        self,
        query_layer,
        key_layer,
        value_layer,
        attention_mask=None,
        output_attentions=False,
        ):

        n, h, t, d = key_layer.size()

        # Cat global mask
        attention_mask = torch.nn.functional.pad(attention_mask, (self.num_global_tokens, 0), value=0)

        # Split input into global tokens and other tokens
        split = (self.num_global_tokens, t - self.num_global_tokens)
        global_query, query_layer = query_layer.split(split, dim=-2)

        # Use normal causal attention if local attention covers every tokens
        if t <= 2 * self.block_size + self.num_global_tokens:
            context_layer = self.causal_attention(
                query_layer=query_layer, 
                key_layer=key_layer, 
                value_layer=value_layer, 
                attention_mask=attention_mask,
                causal_shape=(t - self.num_global_tokens, t - self.num_global_tokens)
                )
            
            context_layer = torch.cat([global_query, context_layer], dim=-2)
            return (self.reshape_output(context_layer), )
        
        # Split K Q M on global and non global
        global_key, key_layer = key_layer.split(split, dim=-2)
        global_value, value_layer = value_layer.split(split, dim=-2)
        global_mask, attention_mask = attention_mask.split(split, dim=-1)
        
        n, h, t, d = key_layer.size()

        # Get sparse idx
        sparse_key, sparse_value, sparse_mask = (None, None, None)
        if self.sparse_block_size and self.sparsity_factor > 0:
            sparse_key, sparse_value, sparse_mask = self.get_sparse_elements(query_layer, key_layer, value_layer, attention_mask)
        
        # Expand masks on heads
        attention_mask = attention_mask.expand(-1, h, -1, -1)
        global_mask = global_mask.expand(-1, h, -1, -1)

        # Compute dot product attention
        context_layer = self.attention(
            query_layer, 
            key_layer, 
            value_layer, 
            attention_mask,
            sparse_key=sparse_key,
            sparse_value=sparse_value,
            sparse_mask=sparse_mask,
            global_key=global_key,
            global_value=global_value,
            global_mask=global_mask
            )

        # Merge pseudo global (causal) and local-sparse tokens
        context_layer = torch.cat([global_query, context_layer], dim=-2)
        context_layer = self.reshape_output(context_layer)

        return (context_layer,)

    def not_causal_forward(
        self,
        query_layer,
        key_layer,
        value_layer,
        attention_mask=None,
        output_attentions=False,
        ):

        n, h, t, d = query_layer.size()

        # Cat global mask
        attention_mask = torch.nn.functional.pad(attention_mask, (self.num_global_tokens, 0), value=0)
        
        # Use normal attention if local attention covers every tokens
        if t <= 2 * self.block_size + self.num_global_tokens:
            context_layer = self.full_attention(
                query_layer=query_layer, 
                key_layer=key_layer, 
                value_layer=value_layer, 
                attention_mask=attention_mask
                )
            return (self.reshape_output(context_layer), )

        # Split input into global tokens and other tokens
        split = (self.num_global_tokens, t - self.num_global_tokens)
        global_query, query_layer = query_layer.split(split, dim=-2)
        
        # Get global_attention
        bos = self.full_attention(
            query_layer=global_query, 
            key_layer=key_layer, 
            value_layer=value_layer, 
            attention_mask=attention_mask
            )
        
        # Split K Q M on global and non global
        global_key, key_layer = key_layer.split(split, dim=-2)
        global_value, value_layer = value_layer.split(split, dim=-2)
        global_mask, attention_mask = attention_mask.split(split, dim=-1)
        
        n, h, t, d = key_layer.size()

        # Get sparse idx
        sparse_key, sparse_value, sparse_mask = (None, None, None)

        if self.sparse_block_size and self.sparsity_factor > 0:
            sparse_key, sparse_value, sparse_mask = self.get_sparse_elements(query_layer, key_layer, value_layer, attention_mask)
        
        # Expand masks on heads
        attention_mask = attention_mask.expand(-1, h, -1, -1)
        global_mask = global_mask.expand(-1, h, -1, -1)

        # Compute dot product attention
        context_layer = self.attention(
            query_layer, 
            key_layer, 
            value_layer, 
            attention_mask,
            sparse_key=sparse_key, 
            sparse_value=sparse_value, 
            sparse_mask=sparse_mask,
            global_key=global_key,
            global_value=global_value,
            global_mask=global_mask
            )

        # Merge global and local-sparse tokens
        context_layer = torch.cat([bos, context_layer], dim=-2)
        context_layer = self.reshape_output(context_layer)
        
        return (context_layer,)

    def cross_attention_forward(
        self,
        query_layer,
        key_layer,
        value_layer,
        attention_mask=None,
        output_attentions=False,
        ):

        context_layer = self.full_attention(
            query_layer=query_layer, 
            key_layer=key_layer, 
            value_layer=value_layer, 
            attention_mask=attention_mask
        )
        return (self.reshape_output(context_layer), )

    def chunk(self, x, chunk_size):

        n, h, t, d = x.size()
        return x.reshape(n, h, -1, chunk_size, d)


class LSGRobertaLayer(RobertaLayer):
    
    def __init__(self, config):

        super().__init__(config)

        self.attention = LSGAttention(config)
        if self.add_cross_attention:
            assert self.is_decoder, f"{self} should be used as a decoder model if cross attention is added"
            self.crossattention = LSGAttention(config)


class LSGRobertaEncoder(RobertaEncoder):

    def __init__(self, config):

        super().__init__(config)

        self.layer = nn.ModuleList([LSGRobertaLayer(config) for _ in range(config.num_hidden_layers)])

        assert hasattr(config, "num_global_tokens")
        self.num_global_tokens = config.num_global_tokens
        self.pad_idx = config.pad_token_id

        assert hasattr(config, "block_size") and hasattr(config, "adaptive")
        self.block_size = config.block_size
        self.adaptive = config.adaptive
        self.mask_first_token = config.mask_first_token
        self.pool_with_global = config.pool_with_global

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.FloatTensor] = None,
        head_mask: Optional[torch.FloatTensor] = None,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
        past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = False,
        output_hidden_states: Optional[bool] = False,
        return_dict: Optional[bool] = True,
    ) -> Union[Tuple[torch.Tensor], BaseModelOutputWithPastAndCrossAttentions]:

        mask_value = torch.finfo(attention_mask.dtype).min
        n, _, __, t = attention_mask.size()
        
        if not (self.config.is_decoder and encoder_hidden_states is not None):
            b = self.block_size * 2
            pad = t % self.block_size
            
            # Check if t is multiple of block_size and pad
            if self.adaptive and t > b and pad > 0:
                pad_length = self.block_size - pad
                hidden_states = torch.nn.functional.pad(hidden_states.transpose(-1, -2), (0, pad_length), value=0.).transpose(-1, -2)
                attention_mask = torch.nn.functional.pad(attention_mask, (0, pad_length), value=mask_value)

            if self.mask_first_token:
                attention_mask[..., 0] = mask_value

        encoder_outputs = super().forward(
            hidden_states=hidden_states, 
            attention_mask=attention_mask, 
            head_mask=head_mask,
            encoder_hidden_states=encoder_hidden_states,
            encoder_attention_mask=encoder_attention_mask,
            past_key_values=past_key_values,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict
            )

        sequence_output = encoder_outputs[0]
        if self.pool_with_global:
            sequence_output[:, self.num_global_tokens] = sequence_output[:, 0]

        # Adapt sequence to initial shape
        sequence_output = sequence_output[..., self.num_global_tokens: t + self.num_global_tokens, :]

        if not return_dict:
            return (sequence_output, ) + encoder_outputs[1:]
        
        encoder_outputs.last_hidden_state = sequence_output 
        return encoder_outputs


class LSGRobertaPreTrainedModel(RobertaPreTrainedModel):
    """
    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
    models.
    """

    config_class = LSGRobertaConfig
    base_model_prefix = "roberta"
    supports_gradient_checkpointing = True
    _no_split_modules = []

    def _set_gradient_checkpointing(self, module, value=False):
        if isinstance(module, (RobertaEncoder, LSGRobertaEncoder)):
            module.gradient_checkpointing = value


class LSGRobertaModel(LSGRobertaPreTrainedModel, RobertaModel):
    """
    This class overrides :class:`~transformers.RobertaModel`. Please check the superclass for the appropriate
    documentation alongside usage examples.
    """

    def __init__(self, config, add_pooling_layer=True):
        
        LSGRobertaPreTrainedModel.__init__(self, config)

        self.embeddings = LSGRobertaEmbeddings(config)
        self.encoder = LSGRobertaEncoder(config)
        self.pooler = RobertaPooler(config) if add_pooling_layer else None

        if config.add_cross_attention:
            logger.warning(
                "Cross attention is computed using full attention since it is not LSG compatible."
            )
        
        self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"
        if self._use_flash_attention_2:
            logger.warning(
                    "[WARNING flash-attention]: LSG doesnt support flash-attention currently"
                )
            
        # Initialize weights and apply final processing
        self.post_init()

    def get_extended_attention_mask(self, attention_mask, input_shape, device=None):

        # Do not rely on original triangular mask from BERT/RoBERTa for causalLM
        if attention_mask.dim() == 3:
            extended_attention_mask = attention_mask[:, None, :, :]
        elif attention_mask.dim() == 2:
            extended_attention_mask = attention_mask[:, None, None, :]
        else:
            raise ValueError(
                f"Wrong shape for input_ids (shape {input_shape}) or attention_mask (shape {attention_mask.shape})"
            )

        extended_attention_mask = extended_attention_mask.to(dtype=self.dtype)  # fp16 compatibility
        extended_attention_mask = (1.0 - extended_attention_mask) * torch.finfo(extended_attention_mask.dtype).min

        return extended_attention_mask


class LSGRobertaForCausalLM(LSGRobertaPreTrainedModel, RobertaForCausalLM):

    _tied_weights_keys = ["lm_head.decoder.weight", "lm_head.decoder.bias"]

    def __init__(self, config):

        LSGRobertaPreTrainedModel.__init__(self, config)

        if not config.is_decoder:
            logger.warning("If you want to use `LSGRobertaLMHeadModel` as a standalone, add `is_decoder=True.`")

        self.roberta = LSGRobertaModel(config, add_pooling_layer=False)
        self.lm_head = RobertaLMHead(config)

        # Initialize weights and apply final processing
        self.post_init()


class LSGRobertaForMaskedLM(LSGRobertaPreTrainedModel, RobertaForMaskedLM):
    """
    This class overrides :class:`~transformers.RobertaForMaskedLM`. Please check the superclass for the appropriate
    documentation alongside usage examples.
    """

    _tied_weights_keys = ["lm_head.decoder.weight", "lm_head.decoder.bias"]

    def __init__(self, config):

        LSGRobertaPreTrainedModel.__init__(self, config)

        if config.is_decoder:
            logger.warning(
                "If you want to use `LSGRobertaForMaskedLM` make sure `config.is_decoder=False` for "
                "bi-directional self-attention."
            )

        self.roberta = LSGRobertaModel(config, add_pooling_layer=False)
        self.lm_head = RobertaLMHead(config)

        # Initialize weights and apply final processing
        self.post_init()


class LSGRobertaForSequenceClassification(LSGRobertaPreTrainedModel, RobertaForSequenceClassification):
    """
    This class overrides :class:`~transformers.RobertaForSequenceClassification`. Please check the superclass for the
    appropriate documentation alongside usage examples.
    """

    def __init__(self, config):
        
        LSGRobertaPreTrainedModel.__init__(self, config)

        self.num_labels = config.num_labels
        self.config = config

        self.roberta = LSGRobertaModel(config, add_pooling_layer=False)
        self.classifier = RobertaClassificationHead(config)

        # Initialize weights and apply final processing
        self.post_init()


class LSGRobertaForMultipleChoice(LSGRobertaPreTrainedModel, RobertaForMultipleChoice):
    """
    This class overrides :class:`~transformers.RobertaForMultipleChoice`. Please check the superclass for the
    appropriate documentation alongside usage examples.
    """

    def __init__(self, config):
        
        LSGRobertaPreTrainedModel.__init__(self, config)

        self.roberta = LSGRobertaModel(config)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)
        self.classifier = nn.Linear(config.hidden_size, 1)

        # Initialize weights and apply final processing
        self.post_init()


class LSGRobertaForTokenClassification(LSGRobertaPreTrainedModel, RobertaForTokenClassification):
    """
    This class overrides :class:`~transformers.RobertaForTokenClassification`. Please check the superclass for the
    appropriate documentation alongside usage examples.
    """

    _keys_to_ignore_on_load_unexpected = [r"pooler"]
    _keys_to_ignore_on_load_missing = [r"position_ids"]

    def __init__(self, config):
        
        LSGRobertaPreTrainedModel.__init__(self, config)

        self.num_labels = config.num_labels

        self.roberta = LSGRobertaModel(config, add_pooling_layer=False)
        classifier_dropout = (
            config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
        )
        self.dropout = nn.Dropout(classifier_dropout)
        self.classifier = nn.Linear(config.hidden_size, config.num_labels)

        # Initialize weights and apply final processing
        self.post_init()


class LSGRobertaForQuestionAnswering(LSGRobertaPreTrainedModel, RobertaForQuestionAnswering):
    """
    This class overrides :class:`~transformers.RobertaForQuestionAnswering`. Please check the superclass for the
    appropriate documentation alongside usage examples.
    """
    
    def __init__(self, config):
        
        LSGRobertaPreTrainedModel.__init__(self, config)
        
        self.num_labels = config.num_labels

        self.roberta = LSGRobertaModel(config, add_pooling_layer=False)
        self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)

        # Initialize weights and apply final processing
        self.post_init()


def str_to_class(classname):
    return getattr(sys.modules[__name__], classname)

# Register model in Auto API
try:
    LSGRobertaConfig.register_for_auto_class()
    for key, value in AUTO_MAP.items():
        str_to_class(value.split(".")[-1]).register_for_auto_class(key)
except:
    warn("AutoRegister isn't available, you'll have to manually copy modeling.py after .save_pretrained(...).")
    warn("Update to transformers >= 4.36.1 to fix.")