File size: 64,308 Bytes
e7d3e35
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
# coding=utf-8
# Copyright 2018 The OpenAI Team Authors and HuggingFace Inc. team.
# Copyright (c) 2018, NVIDIA CORPORATION.  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 OpenAI GPT-2 model."""

import math
import os
from typing import Optional, Tuple, Union

import torch
import torch.utils.checkpoint
from torch import nn
from torch.cuda.amp import autocast
from torch.nn import CrossEntropyLoss
from transformers.activations import ACT2FN
from transformers.modeling_outputs import BaseModelOutputWithPastAndCrossAttentions, CausalLMOutputWithCrossAttentions
from transformers.pytorch_utils import Conv1D, find_pruneable_heads_and_indices, prune_conv1d_layer
from transformers.utils import add_code_sample_docstrings, add_start_docstrings, add_start_docstrings_to_model_forward
from transformers.utils.model_parallel_utils import assert_device_map, get_device_map

from m4.models import DecoupledEmbedding, DecoupledLinear
from m4.models.common import (
    expand_inputs_for_generation,
    prepare_inputs_for_generation,
    update_model_kwargs_for_generation,
)
from m4.models.custom_modules import VLOOMPreTrainedModelBase
from m4.models.perceiver.perceiver import PerceiverResampler
from m4.models.vgpt2.configuration_vgpt2 import VGPT2Config
from m4.training.utils import (
    compute_perceiver_tflops_per_batch_per_gpu,
    compute_tflops_per_batch_per_gpu,
    freeze_model,
)
from m4.utils import logging


logger = logging.get_logger(__name__)

_CHECKPOINT_FOR_DOC = "gpt2"
_CONFIG_FOR_DOC = "VGPT2Config"
_TOKENIZER_FOR_DOC = "GPT2Tokenizer"

GPT2_PRETRAINED_MODEL_ARCHIVE_LIST = [
    "gpt2",
    "gpt2-medium",
    "gpt2-large",
    "gpt2-xl",
    "distilgpt2",
    # See all GPT-2 models at https://huggingface.co/models?filter=gpt2
]


def load_tf_weights_in_gpt2(model, config, gpt2_checkpoint_path):
    """Load tf checkpoints in a pytorch model"""
    try:
        import re

        import tensorflow as tf
    except ImportError:
        logger.error(
            "Loading a TensorFlow model in PyTorch, requires TensorFlow to be installed. Please see "
            "https://www.tensorflow.org/install/ for installation instructions."
        )
        raise
    tf_path = os.path.abspath(gpt2_checkpoint_path)
    logger.info(f"Converting TensorFlow checkpoint from {tf_path}")
    # Load weights from TF model
    init_vars = tf.train.list_variables(tf_path)
    names = []
    arrays = []
    for name, shape in init_vars:
        logger.info(f"Loading TF weight {name} with shape {shape}")
        array = tf.train.load_variable(tf_path, name)
        names.append(name)
        arrays.append(array.squeeze())

    for name, array in zip(names, arrays):
        name = name[6:]  # skip "model/"
        name = name.split("/")
        pointer = model
        for m_name in name:
            if re.fullmatch(r"[A-Za-z]+\d+", m_name):
                scope_names = re.split(r"(\d+)", m_name)
            else:
                scope_names = [m_name]
            if scope_names[0] == "w" or scope_names[0] == "g":
                pointer = getattr(pointer, "weight")
            elif scope_names[0] == "b":
                pointer = getattr(pointer, "bias")
            elif scope_names[0] == "wpe" or scope_names[0] == "wte":
                pointer = getattr(pointer, scope_names[0])
                pointer = getattr(pointer, "weight")
            else:
                pointer = getattr(pointer, scope_names[0])
            if len(scope_names) >= 2:
                num = int(scope_names[1])
                pointer = pointer[num]
        try:
            assert (
                pointer.shape == array.shape
            ), f"Pointer shape {pointer.shape} and array shape {array.shape} mismatched"
        except AssertionError as e:
            e.args += (pointer.shape, array.shape)
            raise
        logger.info(f"Initialize PyTorch weight {name}")
        pointer.data = torch.from_numpy(array)
    return model


class GPT2Attention(nn.Module):
    def __init__(self, config, is_cross_attention=False, layer_idx=None):
        super().__init__()

        max_positions = config.max_position_embeddings
        self.register_buffer(
            "bias",
            torch.tril(torch.ones((max_positions, max_positions), dtype=torch.uint8)).view(
                1, 1, max_positions, max_positions
            ),
        )
        self.register_buffer("masked_bias", torch.tensor(-1e4))

        self.embed_dim = config.hidden_size
        self.num_heads = config.num_attention_heads
        self.head_dim = self.embed_dim // self.num_heads
        self.split_size = self.embed_dim
        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_attn_weights = config.scale_attn_weights
        self.is_cross_attention = is_cross_attention

        # Layer-wise attention scaling, reordering, and upcasting
        self.scale_attn_by_inverse_layer_idx = config.scale_attn_by_inverse_layer_idx
        self.layer_idx = layer_idx
        self.reorder_and_upcast_attn = config.reorder_and_upcast_attn

        if self.is_cross_attention:
            in_dim = self.embed_dim if not hasattr(config, "vision_embed_dim") else config.vision_embed_dim
            self.c_attn = Conv1D(2 * self.embed_dim, in_dim)
            self.q_attn = Conv1D(self.embed_dim, self.embed_dim)
        else:
            self.c_attn = Conv1D(3 * self.embed_dim, self.embed_dim)
        self.c_proj = Conv1D(self.embed_dim, self.embed_dim)

        self.attn_dropout = nn.Dropout(config.attn_pdrop)
        self.resid_dropout = nn.Dropout(config.resid_pdrop)

        self.pruned_heads = set()

    def prune_heads(self, heads):
        if len(heads) == 0:
            return
        heads, index = find_pruneable_heads_and_indices(heads, self.num_heads, self.head_dim, self.pruned_heads)
        index_attn = torch.cat([index, index + self.split_size, index + (2 * self.split_size)])

        # Prune conv1d layers
        self.c_attn = prune_conv1d_layer(self.c_attn, index_attn, dim=1)
        self.c_proj = prune_conv1d_layer(self.c_proj, index, dim=0)

        # Update hyper params
        self.split_size = (self.split_size // self.num_heads) * (self.num_heads - len(heads))
        self.num_heads = self.num_heads - len(heads)
        self.pruned_heads = self.pruned_heads.union(heads)

    def _attn(self, query, key, value, attention_mask=None, head_mask=None):
        attn_weights = torch.matmul(query, key.transpose(-1, -2))

        if self.scale_attn_weights:
            attn_weights = attn_weights / torch.tensor(
                value.size(-1) ** 0.5, dtype=attn_weights.dtype, device=attn_weights.device
            )

        # Layer-wise attention scaling
        if self.scale_attn_by_inverse_layer_idx:
            attn_weights = attn_weights / float(self.layer_idx + 1)

        if not self.is_cross_attention:
            # if only "normal" attention layer implements causal mask
            query_length, key_length = query.size(-2), key.size(-2)
            causal_mask = self.bias[:, :, key_length - query_length : key_length, :key_length].to(torch.bool)
            mask_value = torch.finfo(attn_weights.dtype).min
            # Need to be a tensor, otherwise we get error: `RuntimeError: expected scalar type float but found double`.
            # Need to be on the same device, otherwise `RuntimeError: ..., x and y to be on the same device`
            mask_value = torch.tensor(mask_value, dtype=attn_weights.dtype).to(attn_weights.device)
            attn_weights = torch.where(causal_mask, attn_weights, mask_value)

        if attention_mask is not None:
            # Apply the attention mask
            attn_weights = attn_weights + attention_mask

        attn_weights = nn.functional.softmax(attn_weights, dim=-1)

        # Downcast (if necessary) back to V's dtype (if in mixed-precision) -- No-Op otherwise
        attn_weights = attn_weights.type(value.dtype)
        attn_weights = self.attn_dropout(attn_weights)

        # Mask heads if we want to
        if head_mask is not None:
            attn_weights = attn_weights * head_mask

        attn_output = torch.matmul(attn_weights, value)

        return attn_output, attn_weights

    def _upcast_and_reordered_attn(self, query, key, value, attention_mask=None, head_mask=None):
        # Use `torch.baddbmm` (a bit more efficient w/ alpha param for scaling -- from Megatron-LM)
        bsz, num_heads, q_seq_len, dk = query.size()
        _, _, k_seq_len, _ = key.size()

        # Preallocate attn_weights for `baddbmm`
        attn_weights = torch.empty(bsz * num_heads, q_seq_len, k_seq_len, dtype=torch.float32, device=query.device)

        # Compute Scale Factor
        scale_factor = 1.0
        if self.scale_attn_weights:
            scale_factor /= float(value.size(-1)) ** 0.5

        if self.scale_attn_by_inverse_layer_idx:
            scale_factor /= float(self.layer_idx + 1)

        # Upcast (turn off autocast) and reorder (Scale K by 1 / root(dk))
        with autocast(enabled=False):
            q, k = query.reshape(-1, q_seq_len, dk), key.transpose(-1, -2).reshape(-1, dk, k_seq_len)
            attn_weights = torch.baddbmm(attn_weights, q.float(), k.float(), beta=0, alpha=scale_factor)
            attn_weights = attn_weights.reshape(bsz, num_heads, q_seq_len, k_seq_len)

        if not self.is_cross_attention:
            # if only "normal" attention layer implements causal mask
            query_length, key_length = query.size(-2), key.size(-2)
            causal_mask = self.bias[:, :, key_length - query_length : key_length, :key_length].bool()
            mask_value = torch.finfo(attn_weights.dtype).min
            # Need to be a tensor, otherwise we get error: `RuntimeError: expected scalar type float but found double`.
            # Need to be on the same device, otherwise `RuntimeError: ..., x and y to be on the same device`
            mask_value = torch.tensor(mask_value, dtype=attn_weights.dtype).to(attn_weights.device)
            attn_weights = torch.where(causal_mask, attn_weights, mask_value)

        if attention_mask is not None:
            # Apply the attention mask
            attn_weights = attn_weights + attention_mask

        attn_weights = nn.functional.softmax(attn_weights, dim=-1)

        # Downcast (if necessary) back to V's dtype (if in mixed-precision) -- No-Op if otherwise
        if attn_weights.dtype != torch.float32:
            raise RuntimeError("Error with upcasting, attn_weights does not have dtype torch.float32")
        attn_weights = attn_weights.type(value.dtype)
        attn_weights = self.attn_dropout(attn_weights)

        # Mask heads if we want to
        if head_mask is not None:
            attn_weights = attn_weights * head_mask

        attn_output = torch.matmul(attn_weights, value)

        return attn_output, attn_weights

    def _split_heads(self, tensor, num_heads, attn_head_size):
        """
        Splits hidden_size dim into attn_head_size and num_heads
        """
        new_shape = tensor.size()[:-1] + (num_heads, attn_head_size)
        tensor = tensor.view(new_shape)
        return tensor.permute(0, 2, 1, 3)  # (batch, head, seq_length, head_features)

    def _merge_heads(self, tensor, num_heads, attn_head_size):
        """
        Merges attn_head_size dim and num_attn_heads dim into hidden_size
        """
        tensor = tensor.permute(0, 2, 1, 3).contiguous()
        new_shape = tensor.size()[:-2] + (num_heads * attn_head_size,)
        return tensor.view(new_shape)

    def forward(
        self,
        hidden_states: Optional[Tuple[torch.FloatTensor]],
        layer_past: Optional[Tuple[torch.Tensor]] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        head_mask: Optional[torch.FloatTensor] = None,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
        use_cache: Optional[bool] = False,
        output_attentions: Optional[bool] = False,
    ) -> Tuple[Union[torch.Tensor, Tuple[torch.Tensor]], ...]:
        if encoder_hidden_states is not None:
            if not hasattr(self, "q_attn"):
                raise ValueError(
                    "If class is used as cross attention, the weights `q_attn` have to be defined. "
                    "Please make sure to instantiate class with `GPT2Attention(..., is_cross_attention=True)`."
                )

            query = self.q_attn(hidden_states)
            key, value = self.c_attn(encoder_hidden_states).split(self.split_size, dim=2)
            attention_mask = encoder_attention_mask
        else:
            query, key, value = self.c_attn(hidden_states).split(self.split_size, dim=2)

        query = self._split_heads(query, self.num_heads, self.head_dim)
        key = self._split_heads(key, self.num_heads, self.head_dim)
        value = self._split_heads(value, self.num_heads, self.head_dim)

        if layer_past is not None:
            past_key, past_value = layer_past
            key = torch.cat((past_key, key), dim=-2)
            value = torch.cat((past_value, value), dim=-2)

        if use_cache is True:
            present = (key, value)
        else:
            present = None

        if self.reorder_and_upcast_attn:
            attn_output, attn_weights = self._upcast_and_reordered_attn(query, key, value, attention_mask, head_mask)
        else:
            attn_output, attn_weights = self._attn(query, key, value, attention_mask, head_mask)

        attn_output = self._merge_heads(attn_output, self.num_heads, self.head_dim)
        attn_output = self.c_proj(attn_output)
        attn_output = self.resid_dropout(attn_output)

        outputs = (attn_output, present)
        if output_attentions:
            outputs += (attn_weights,)

        return outputs  # a, present, (attentions)


class GPT2MLP(nn.Module):
    def __init__(self, intermediate_size, config):
        super().__init__()
        embed_dim = config.hidden_size
        self.c_fc = Conv1D(intermediate_size, embed_dim)
        self.c_proj = Conv1D(embed_dim, intermediate_size)
        self.act = ACT2FN[config.activation_function]
        self.dropout = nn.Dropout(config.resid_pdrop)

    def forward(self, hidden_states: Optional[Tuple[torch.FloatTensor]]) -> torch.FloatTensor:
        hidden_states = self.c_fc(hidden_states)
        hidden_states = self.act(hidden_states)
        hidden_states = self.c_proj(hidden_states)
        hidden_states = self.dropout(hidden_states)
        return hidden_states


class GPT2Block(nn.Module):
    def __init__(self, config, layer_idx=None):
        super().__init__()
        hidden_size = config.hidden_size
        inner_dim = config.n_inner if config.n_inner is not None else 4 * hidden_size

        self.ln_1 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
        self.attn = GPT2Attention(config, layer_idx=layer_idx)
        self.ln_2 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)

        if config.add_cross_attention:
            self.crossattention = GPT2Attention(config, is_cross_attention=True, layer_idx=layer_idx)
            self.ln_cross_attn = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)

        self.mlp = GPT2MLP(inner_dim, config)

    def forward(
        self,
        hidden_states: Optional[Tuple[torch.FloatTensor]],
        layer_past: Optional[Tuple[torch.Tensor]] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        head_mask: Optional[torch.FloatTensor] = None,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
        use_cache: Optional[bool] = False,
        output_attentions: Optional[bool] = False,
    ) -> Union[Tuple[torch.Tensor], Optional[Tuple[torch.Tensor, Tuple[torch.FloatTensor, ...]]]]:
        residual = hidden_states
        hidden_states = self.ln_1(hidden_states)
        attn_outputs = self.attn(
            hidden_states,
            layer_past=layer_past,
            attention_mask=attention_mask,
            head_mask=head_mask,
            use_cache=use_cache,
            output_attentions=output_attentions,
        )
        attn_output = attn_outputs[0]  # output_attn: a, present, (attentions)
        outputs = attn_outputs[1:]
        # residual connection
        hidden_states = attn_output + residual

        if encoder_hidden_states is not None:
            # add one self-attention block for cross-attention
            if not hasattr(self, "crossattention"):
                raise ValueError(
                    f"If `encoder_hidden_states` are passed, {self} has to be instantiated with "
                    "cross-attention layers by setting `config.add_cross_attention=True`"
                )
            residual = hidden_states
            hidden_states = self.ln_cross_attn(hidden_states)
            cross_attn_outputs = self.crossattention(
                hidden_states,
                attention_mask=attention_mask,
                head_mask=head_mask,
                encoder_hidden_states=encoder_hidden_states,
                encoder_attention_mask=encoder_attention_mask,
                output_attentions=output_attentions,
            )
            attn_output = cross_attn_outputs[0]
            # residual connection
            hidden_states = residual + attn_output
            outputs = outputs + cross_attn_outputs[2:]  # add cross attentions if we output attention weights

        residual = hidden_states
        hidden_states = self.ln_2(hidden_states)
        feed_forward_hidden_states = self.mlp(hidden_states)
        # residual connection
        hidden_states = residual + feed_forward_hidden_states

        if use_cache:
            outputs = (hidden_states,) + outputs
        else:
            outputs = (hidden_states,) + outputs[1:]

        return outputs  # hidden_states, present, (attentions, cross_attentions)


class VGPT2GatedCrossAttentionBlock(nn.Module):
    def __init__(self, config, layer_idx=None):
        super().__init__()
        hidden_size = config.hidden_size
        inner_dim = config.n_inner if config.n_inner is not None else 4 * hidden_size

        self.ln_1 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
        self.cross_attn = GPT2Attention(config, is_cross_attention=True, layer_idx=layer_idx)
        self.mlp = GPT2MLP(inner_dim, config)
        self.ln_2 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
        self.act = nn.Tanh()

        if config.alpha_initializer == "zeros":
            if config.alpha_type == "vector":
                self.alpha_cross_attn = nn.Parameter(torch.zeros(1, 1, hidden_size))
                self.alpha_dense = nn.Parameter(torch.zeros(1, 1, hidden_size))
            elif config.alpha_type == "float":
                self.alpha_cross_attn = nn.Parameter(torch.zeros(1))
                self.alpha_dense = nn.Parameter(torch.zeros(1))
            else:
                raise ValueError(f"Unknown value for `alpha_type` ({config.alpha_type})")

        elif config.alpha_initializer == "ones":
            if config.alpha_type == "vector":
                self.alpha_cross_attn = nn.Parameter(torch.ones(1, 1, hidden_size))
                self.alpha_dense = nn.Parameter(torch.ones(1, 1, hidden_size))
            elif config.alpha_type == "float":
                self.alpha_cross_attn = nn.Parameter(torch.ones(1))
                self.alpha_dense = nn.Parameter(torch.ones(1))
            else:
                raise ValueError(f"Unknown value for `alpha_type` ({config.alpha_type})")

        elif config.alpha_initializer in {"normal", "gaussian", "random"}:
            if config.alpha_type == "vector":
                self.alpha_cross_attn = nn.Parameter(
                    torch.normal(mean=0.0, std=config.alphas_initializer_range, size=(1, 1, hidden_size))
                )
                self.alpha_dense = nn.Parameter(
                    torch.normal(mean=0.0, std=config.alphas_initializer_range, size=(1, 1, hidden_size))
                )
            elif config.alpha_type == "float":
                self.alpha_cross_attn = nn.Parameter(
                    torch.normal(mean=0.0, std=config.alphas_initializer_range, size=(1))
                )
                self.alpha_dense = nn.Parameter(torch.normal(mean=0.0, std=config.alphas_initializer_range, size=(1)))
            else:
                raise ValueError(f"Unknown value for `alpha_type` ({config.alpha_type})")

        else:
            raise NotImplementedError(f"Alpha initialization scheme {config.alpha_initializer} not yet implemented!")

    def forward(
        self,
        hidden_states: Optional[Tuple[torch.FloatTensor]],
        layer_past: Optional[Tuple[torch.Tensor]] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        head_mask: Optional[torch.FloatTensor] = None,
        image_hidden_states: Optional[torch.Tensor] = None,
        image_attention_mask: Optional[torch.FloatTensor] = None,
        use_cache: Optional[bool] = False,
        output_attentions: Optional[bool] = False,
    ) -> Union[Tuple[torch.Tensor], Optional[Tuple[torch.Tensor, Tuple[torch.FloatTensor, ...]]]]:
        if image_hidden_states is None:
            raise ValueError(
                "`image_hidden_states` is required for VGPT2 cross attention module which are visual features to be"
                " conditioned on."
            )
            # add one self-attention block for cross-attention

        # TODO(aps): Handle cross attention in the outputs
        # if not hasattr(self, "crossattention"):
        #     raise ValueError(
        #         f"If `image_hidden_states` are passed, {self} has to be instantiated with "
        #         "cross-attention layers by setting `config.add_cross_attention=True`"
        #     )
        residual = hidden_states

        hidden_states = self.ln_1(hidden_states)
        cross_attn_outputs = self.cross_attn(
            hidden_states,
            attention_mask=attention_mask,
            head_mask=head_mask,
            encoder_hidden_states=image_hidden_states,
            encoder_attention_mask=image_attention_mask,
            output_attentions=output_attentions,
        )
        attn_output = cross_attn_outputs[0]
        outputs = cross_attn_outputs[1:]
        # residual connection
        hidden_states = residual + self.act(self.alpha_cross_attn) * attn_output
        outputs = outputs + cross_attn_outputs[2:]  # add cross attentions if we output attention weights

        residual = hidden_states
        hidden_states = self.ln_2(hidden_states)
        feed_forward_hidden_states = self.mlp(hidden_states)
        # residual connection
        hidden_states = residual + self.act(self.alpha_dense) * feed_forward_hidden_states

        if use_cache:
            outputs = (hidden_states,) + outputs
        else:
            outputs = (hidden_states,) + outputs[1:]

        return outputs  # hidden_states, present, (attentions, cross_attentions)


class VGPT2PreTrainedModel(VLOOMPreTrainedModelBase):
    """
    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
    models.
    """

    config_class = VGPT2Config
    load_tf_weights = load_tf_weights_in_gpt2
    base_model_prefix = "transformer"
    is_parallelizable = True
    supports_gradient_checkpointing = True
    _no_split_modules = ["GPT2Block"]

    def __init__(self, *inputs, **kwargs):
        super().__init__(*inputs, **kwargs)

    def _init_weights(self, module):
        """Initialize the weights."""
        if isinstance(module, (nn.Linear, Conv1D)):
            # Slightly different from the TF version which uses truncated_normal for initialization
            # cf https://github.com/pytorch/pytorch/pull/5617
            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
            if module.bias is not None:
                module.bias.data.zero_()
        elif isinstance(module, nn.Embedding):
            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
            if module.padding_idx is not None:
                module.weight.data[module.padding_idx].zero_()
        elif isinstance(module, nn.LayerNorm):
            module.bias.data.zero_()
            module.weight.data.fill_(1.0)

        # Reinitialize selected weights subject to the OpenAI GPT-2 Paper Scheme:
        #   > A modified initialization which accounts for the accumulation on the residual path with model depth. Scale
        #   > the weights of residual layers at initialization by a factor of 1/√N where N is the # of residual layers.
        #   >   -- GPT-2 :: https://openai.com/blog/better-language-models/
        #
        # Reference (Megatron-LM): https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/model/gpt_model.py
        for name, p in module.named_parameters():
            if name == "c_proj.weight":
                # Special Scaled Initialization --> There are 2 Layer Norms per Transformer Block
                p.data.normal_(mean=0.0, std=(self.config.initializer_range / math.sqrt(2 * self.config.n_layer)))

    def _set_gradient_checkpointing(self, module, value=False):
        if isinstance(module, VGPT2Model):
            module.gradient_checkpointing = value

    @classmethod
    def override_vision_model_wrapper(cls, model, config, vision_model_name, vision_model_params, torch_dtype):
        # this can be called via from_pretrained from a class w/ head or w/o head so we extract the beheaded model version
        beheaded_model = model.transformer if hasattr(model, "transformer") else model
        cls.override_vision_model(beheaded_model, vision_model_name, vision_model_params, torch_dtype)
        beheaded_model.freeze_relevant_params(config)


GPT2_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 ([`VGPT2Config`]): 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.
"""

GPT2_INPUTS_DOCSTRING = r"""
    Args:
        input_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`):
            `input_ids_length` = `sequence_length` if `past_key_values` is `None` else
            `past_key_values[0][0].shape[-2]` (`sequence_length` of input past key value states). Indices of input
            sequence tokens in the vocabulary.
            If `past_key_values` is used, only `input_ids` that do not have their past calculated should be passed as
            `input_ids`.
            Indices can be obtained using [`GPT2Tokenizer`]. See [`PreTrainedTokenizer.encode`] and
            [`PreTrainedTokenizer.__call__`] for details.
            [What are input IDs?](../glossary#input-ids)
        past_key_values (`Tuple[Tuple[torch.Tensor]]` of length `config.n_layers`):
            Contains precomputed hidden-states (key and values in the attention blocks) as computed by the model (see
            `past_key_values` output below). Can be used to speed up sequential decoding. The `input_ids` which have
            their past given to this model should not be passed as `input_ids` as they have already been computed.
        attention_mask (`torch.FloatTensor` 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**.
            If `past_key_values` is used, `attention_mask` needs to contain the masking strategy that was used for
            `past_key_values`. In other words, the `attention_mask` always has to have the length:
            `len(past_key_values) + len(input_ids)`
            [What are attention masks?](../glossary#attention-mask)
        token_type_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`, *optional*):
            Segment token indices to indicate first and second portions of the inputs. Indices are selected in `[0,
            1]`:
            - 0 corresponds to a *sentence A* token,
            - 1 corresponds to a *sentence B* token.
            [What are token type IDs?](../glossary#token-type-ids)
        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)
        head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
            Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:
            - 1 indicates the head is **not masked**,
            - 0 indicates the head is **masked**.
        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
            model's internal embedding lookup matrix.
            If `past_key_values` is used, optionally only the last `inputs_embeds` have to be input (see
            `past_key_values`).
        use_cache (`bool`, *optional*):
            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
            `past_key_values`).
        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.
"""
PARALLELIZE_DOCSTRING = r"""
    This is an experimental feature and is a subject to change at a moment's notice.
    Uses a device map to distribute attention modules of the model across several devices. If no device map is given,
    it will evenly distribute blocks across all devices.
    Args:
        device_map (`Dict[int, list]`, optional, defaults to None):
            A dictionary that maps attention modules to devices. Note that the embedding module and LMHead are always
            automatically mapped to the first device (for esoteric reasons). That means that the first device should
            have fewer attention modules mapped to it than other devices. For reference, the gpt2 models have the
            following number of attention modules:
                - gpt2: 12
                - gpt2-medium: 24
                - gpt2-large: 36
                - gpt2-xl: 48
    Example:
    ```python
    # Here is an example of a device map on a machine with 4 GPUs using gpt2-xl, which has a total of 48 attention modules:
    model = GPT2LMHeadModel.from_pretrained("gpt2-xl")
    device_map = {
        0: [0, 1, 2, 3, 4, 5, 6, 7, 8],
        1: [9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21],
        2: [22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34],
        3: [35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47],
    }
    model.parallelize(device_map)
    ```
"""
DEPARALLELIZE_DOCSTRING = r"""
    Moves the model to cpu from a model parallel state.
    Example:
    ```python
    # On a 4 GPU machine with gpt2-large:
    model = GPT2LMHeadModel.from_pretrained("gpt2-large")
    device_map = {
        0: [0, 1, 2, 3, 4, 5, 6, 7],
        1: [8, 9, 10, 11, 12, 13, 14, 15],
        2: [16, 17, 18, 19, 20, 21, 22, 23],
        3: [24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35],
    }
    model.parallelize(device_map)  # Splits the model across several devices
    model.deparallelize()  # Put the model back on cpu and cleans memory by calling torch.cuda.empty_cache()
    ```
"""


@add_start_docstrings(
    "The bare GPT2 Model transformer outputting raw hidden-states without any specific head on top.",
    GPT2_START_DOCSTRING,
)
class VGPT2Model(VGPT2PreTrainedModel):
    _keys_to_ignore_on_load_missing = ["attn.masked_bias"]

    def __init__(self, config, vision_model=None):
        super().__init__(config)

        self.embed_dim = config.hidden_size
        self.config = config

        self.wte = DecoupledEmbedding(
            num_embeddings=config.vocab_size,
            num_additional_embeddings=config.additional_vocab_size,
            embedding_dim=self.embed_dim,
            partially_freeze=config.freeze_text_layers,
        )
        self.wpe = nn.Embedding(config.max_position_embeddings, self.embed_dim)

        self.drop = nn.Dropout(config.embd_pdrop)
        self.h = nn.ModuleList([GPT2Block(config, layer_idx=i) for i in range(config.num_hidden_layers)])
        self.ln_f = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon)

        self.cross_layer_interval = config.cross_layer_interval
        num_cross_layers = config.num_hidden_layers // self.cross_layer_interval
        self.gated_cross_attn_layers = nn.ModuleList(
            [VGPT2GatedCrossAttentionBlock(config, layer_idx=i) for i in range(num_cross_layers)]
        )

        # Perceiver Resampler
        if config.use_resampler:
            self.perceiver_resampler = PerceiverResampler(
                self.config,
                self.config.vision_embed_dim,
                config.resampler_depth,
                config.resampler_n_heads,
                config.resampler_head_dim,
                config.resampler_n_latents,
            )
        # Model parallel
        self.model_parallel = False
        self.device_map = None
        self.gradient_checkpointing = False
        # will be vocab_size because of indices starting from 0
        self.image_token_idx = config.image_token_index

        # Load an uninitialized model and later in from_pretrained will load the pre-trained model -
        # this solves the losing of weights in `from_pretrained` on the main model
        self.vision_model = vision_model

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

        self.freeze_relevant_params(config)

    def freeze_relevant_params(self, config=None):
        if config is None:
            config = self.config

        if config.freeze_text_layers:
            self.freeze_text_layers()

        if config.freeze_vision_layers:
            freeze_model(self.vision_model)

    def freeze_text_layers(self):
        for module in [self.wpe, self.h, self.ln_f]:
            freeze_model(module)

    @add_start_docstrings(PARALLELIZE_DOCSTRING)
    # TODO(aps): Implement later for VGPT2
    def parallelize(self, device_map=None):
        # Check validity of device_map
        self.device_map = (
            get_device_map(len(self.h), range(torch.cuda.device_count())) if device_map is None else device_map
        )
        assert_device_map(self.device_map, len(self.h))
        self.model_parallel = True
        self.first_device = "cpu" if "cpu" in self.device_map.keys() else "cuda:" + str(min(self.device_map.keys()))
        self.last_device = "cuda:" + str(max(self.device_map.keys()))
        self.wte = self.wte.to(self.first_device)
        self.wpe = self.wpe.to(self.first_device)
        # Load onto devices
        for k, v in self.device_map.items():
            for block in v:
                cuda_device = "cuda:" + str(k)
                self.h[block] = self.h[block].to(cuda_device)
        # ln_f to last
        self.ln_f = self.ln_f.to(self.last_device)

    @add_start_docstrings(DEPARALLELIZE_DOCSTRING)
    # TODO(aps): Implement later for VGPT2
    def deparallelize(self):
        self.model_parallel = False
        self.device_map = None
        self.first_device = "cpu"
        self.last_device = "cpu"
        self.wte = self.wte.to("cpu")
        self.wpe = self.wpe.to("cpu")
        for index in range(len(self.h)):
            self.h[index] = self.h[index].to("cpu")
        self.ln_f = self.ln_f.to("cpu")
        torch.cuda.empty_cache()

    def get_input_embeddings(self):
        return self.wte

    def set_input_embeddings(self, new_embeddings):
        self.wte = new_embeddings

    def _prune_heads(self, heads_to_prune):
        """
        Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer}
        """
        for layer, heads in heads_to_prune.items():
            self.h[layer].attn.prune_heads(heads)

    @add_start_docstrings_to_model_forward(GPT2_INPUTS_DOCSTRING)
    @add_code_sample_docstrings(
        processor_class=_TOKENIZER_FOR_DOC,
        checkpoint=_CHECKPOINT_FOR_DOC,
        output_type=BaseModelOutputWithPastAndCrossAttentions,
        config_class=_CONFIG_FOR_DOC,
    )
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        token_type_ids: Optional[torch.LongTensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        head_mask: Optional[torch.FloatTensor] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        pixel_values: Optional[torch.FloatTensor] = None,
        image_embeddings: Optional[torch.FloatTensor] = None,
        image_attention_mask: Optional[torch.Tensor] = None,
        crossblock_head_mask: Optional[torch.Tensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple, BaseModelOutputWithPastAndCrossAttentions]:
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        use_cache = use_cache if use_cache is not None else self.config.use_cache
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        if input_ids is not None and inputs_embeds is not None:
            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
        elif input_ids is not None:
            input_shape = input_ids.size()
            input_ids = input_ids.view(-1, input_shape[-1])
            batch_size = input_ids.shape[0]
        elif inputs_embeds is not None:
            input_shape = inputs_embeds.size()[:-1]
            batch_size = inputs_embeds.shape[0]
        else:
            raise ValueError("You have to specify either input_ids or inputs_embeds")

        device = input_ids.device if input_ids is not None else inputs_embeds.device

        if token_type_ids is not None:
            token_type_ids = token_type_ids.view(-1, input_shape[-1])
        if position_ids is not None:
            position_ids = position_ids.view(-1, input_shape[-1])

        if past_key_values is None:
            past_length = 0
            past_key_values = tuple([None] * len(self.h))
        else:
            past_length = past_key_values[0][0].size(-2)
        if position_ids is None:
            position_ids = torch.arange(past_length, input_shape[-1] + past_length, dtype=torch.long, device=device)
            position_ids = position_ids.unsqueeze(0).view(-1, input_shape[-1])

        # GPT2Attention mask.
        if attention_mask is not None:
            if batch_size <= 0:
                raise ValueError("batch_size has to be defined and > 0")
            attention_mask = attention_mask.view(batch_size, -1)
            # We create a 3D attention mask from a 2D tensor mask.
            # Sizes are [batch_size, 1, 1, to_seq_length]
            # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
            # this attention mask is more simple than the triangular masking of causal attention
            # used in OpenAI GPT, we just need to prepare the broadcast dimension here.
            attention_mask = attention_mask[:, None, None, :]

            # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
            # masked positions, this operation will create a tensor which is 0.0 for
            # positions we want to attend and the dtype's smallest value for masked positions.
            # Since we are adding it to the raw scores before the softmax, this is
            # effectively the same as removing these entirely.
            attention_mask = attention_mask.to(dtype=self.dtype)  # fp16 compatibility
            attention_mask = (1.0 - attention_mask) * torch.finfo(self.dtype).min

        # If a 2D or 3D attention mask is provided for the cross-attention
        # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
        if pixel_values is not None and image_embeddings is not None:
            raise ValueError("You cannot specify both pixel_values and image_embeddings at the same time")
        elif pixel_values is not None:
            pixel_values = pixel_values.to(dtype=self.dtype, device=input_ids.device)  # fp16 compatibility
            batch_size, num_images = pixel_values.size(0), pixel_values.size(1)
            pixel_values = pixel_values.contiguous().view(batch_size * num_images, *pixel_values.shape[2:])
            # Get sequence from the vision encoder
            image_hidden_states = self.vision_model(pixel_values=pixel_values).last_hidden_state
        elif image_embeddings is not None:
            batch_size, num_images, image_seq_len, image_hidden_size = image_embeddings.size()
            image_hidden_states = image_embeddings.to(dtype=self.dtype, device=input_ids.device)
            image_hidden_states = image_hidden_states.view(batch_size * num_images, image_seq_len, image_hidden_size)

        if self.config.use_resampler:
            image_hidden_states = self.perceiver_resampler(image_hidden_states)
        image_seq_len, image_hidden_size = image_hidden_states.size(1), image_hidden_states.size(2)
        image_hidden_states = image_hidden_states.view(batch_size, num_images * image_seq_len, image_hidden_size)

        # Make image_attention_mask compatible with hidden states
        text_seq_len = image_attention_mask.size(1)
        image_attention_mask = image_attention_mask.unsqueeze(-1)
        image_attention_mask = image_attention_mask.repeat(1, 1, 1, image_seq_len)
        image_attention_mask = image_attention_mask.view(batch_size, text_seq_len, num_images * image_seq_len)
        if image_hidden_states is not None:
            image_batch_size, image_sequence_length, _ = image_hidden_states.size()
            image_hidden_shape = (image_batch_size, image_sequence_length)
            if image_attention_mask is None:
                image_attention_mask = torch.ones(image_hidden_shape, device=device)
            image_attention_mask = self.invert_attention_mask(image_attention_mask)
        else:
            image_attention_mask = None

        # Prepare head mask if needed
        # 1.0 in head_mask indicate we keep the head
        # attention_probs has shape bsz x n_heads x N x N
        # head_mask has shape n_layer x batch x n_heads x N x N
        head_mask = self.get_head_mask(head_mask, self.config.n_layer)

        if inputs_embeds is None:
            inputs_embeds = self.wte(input_ids)

        position_embeds = self.wpe(position_ids)
        hidden_states = inputs_embeds + position_embeds

        if token_type_ids is not None:
            token_type_embeds = self.wte(token_type_ids)
            hidden_states = hidden_states + token_type_embeds

        hidden_states = self.drop(hidden_states)

        output_shape = input_shape + (hidden_states.size(-1),)

        presents = () if use_cache else None
        all_self_attentions = () if output_attentions else None
        all_cross_attentions = () if output_attentions and self.config.add_cross_attention else None
        all_hidden_states = () if output_hidden_states else None
        for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)):
            # Model parallel
            if self.model_parallel:
                torch.cuda.set_device(hidden_states.device)
                # Ensure layer_past is on same device as hidden_states (might not be correct)
                if layer_past is not None:
                    layer_past = tuple(past_state.to(hidden_states.device) for past_state in layer_past)
                # Ensure that attention_mask is always on the same device as hidden_states
                if attention_mask is not None:
                    attention_mask = attention_mask.to(hidden_states.device)
                if isinstance(head_mask, torch.Tensor):
                    head_mask = head_mask.to(hidden_states.device)
            if output_hidden_states:
                all_hidden_states = all_hidden_states + (hidden_states,)

            def vblock(
                main_block,
                hidden_states,
                layer_past,
                attention_mask,
                layer_head_mask,
                use_cache,
                output_attentions,
                image_hidden_states,
                image_attention_mask,
                layer_idx,
                cross_layer_interval,
                gated_cross_attn_layers,
            ):
                # TODO(aps): Add cross attention values to respective lists
                # TODO(aps): Add xblock head mask support
                if layer_idx % cross_layer_interval == 0:
                    xblock = gated_cross_attn_layers[layer_idx // cross_layer_interval]
                    outputs = xblock(
                        hidden_states,
                        attention_mask=attention_mask,
                        image_hidden_states=image_hidden_states,
                        image_attention_mask=image_attention_mask,
                        use_cache=use_cache,
                        output_attentions=output_attentions,
                    )
                    hidden_states = outputs[0]

                outputs = main_block(
                    hidden_states,
                    layer_past=layer_past,
                    attention_mask=attention_mask,
                    head_mask=layer_head_mask,
                    use_cache=use_cache,
                    output_attentions=output_attentions,
                )

                return outputs

            if self.gradient_checkpointing and self.training:
                layer_past = None
                if use_cache:
                    logger.warning_once(
                        "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                    )
                    use_cache = False

                outputs = torch.utils.checkpoint.checkpoint(
                    vblock,
                    block,
                    hidden_states,
                    layer_past,
                    attention_mask,
                    head_mask[i],
                    use_cache,
                    output_attentions,
                    image_hidden_states,
                    image_attention_mask,
                    i,
                    self.cross_layer_interval,
                    self.gated_cross_attn_layers,
                )
            else:
                outputs = vblock(
                    block,
                    hidden_states,
                    layer_past=layer_past,
                    attention_mask=attention_mask,
                    layer_head_mask=head_mask[i],
                    use_cache=use_cache,
                    output_attentions=output_attentions,
                    image_hidden_states=image_hidden_states,
                    image_attention_mask=image_attention_mask,
                    layer_idx=i,
                    cross_layer_interval=self.cross_layer_interval,
                    gated_cross_attn_layers=self.gated_cross_attn_layers,
                )

            hidden_states = outputs[0]
            if use_cache is True:
                presents = presents + (outputs[1],)

            if output_attentions:
                all_self_attentions = all_self_attentions + (outputs[2 if use_cache else 1],)
                if self.config.add_cross_attention:
                    all_cross_attentions = all_cross_attentions + (outputs[3 if use_cache else 2],)

            # Model Parallel: If it's the last layer for that device, put things on the next device
            if self.model_parallel:
                for k, v in self.device_map.items():
                    if i == v[-1] and "cuda:" + str(k) != self.last_device:
                        hidden_states = hidden_states.to("cuda:" + str(k + 1))

        hidden_states = self.ln_f(hidden_states)

        hidden_states = hidden_states.view(output_shape)
        # Add last hidden state
        if output_hidden_states:
            all_hidden_states = all_hidden_states + (hidden_states,)

        if not return_dict:
            return tuple(
                v
                for v in [hidden_states, presents, all_hidden_states, all_self_attentions, all_cross_attentions]
                if v is not None
            )

        return BaseModelOutputWithPastAndCrossAttentions(
            last_hidden_state=hidden_states,
            past_key_values=presents,
            hidden_states=all_hidden_states,
            attentions=all_self_attentions,
            cross_attentions=all_cross_attentions,
        )


@add_start_docstrings(
    """
    The GPT2 Model transformer with a language modeling head on top (linear layer with weights tied to the input
    embeddings).
    """,
    GPT2_START_DOCSTRING,
)
class VGPT2LMHeadModel(VGPT2PreTrainedModel):
    _keys_to_ignore_on_load_missing = [r"attn.masked_bias", r"attn.bias", r"lm_head.weight"]

    def __init__(self, config, vision_model=None):
        super().__init__(config)
        self.transformer = VGPT2Model(config, vision_model=vision_model)
        self.lm_head = DecoupledLinear(
            in_features=config.n_embd,
            out_features=config.vocab_size,
            out_additional_features=config.additional_vocab_size,
            bias=False,
            partially_freeze=config.freeze_lm_head,
        )

        # Model parallel
        self.model_parallel = False
        self.device_map = None

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

    @add_start_docstrings(PARALLELIZE_DOCSTRING)
    def parallelize(self, device_map=None):
        self.device_map = (
            get_device_map(len(self.transformer.h), range(torch.cuda.device_count()))
            if device_map is None
            else device_map
        )
        assert_device_map(self.device_map, len(self.transformer.h))
        self.transformer.parallelize(self.device_map)
        self.lm_head = self.lm_head.to(self.transformer.first_device)
        self.model_parallel = True

    @add_start_docstrings(DEPARALLELIZE_DOCSTRING)
    def deparallelize(self):
        self.transformer.deparallelize()
        self.transformer = self.transformer.to("cpu")
        self.lm_head = self.lm_head.to("cpu")
        self.model_parallel = False
        torch.cuda.empty_cache()

    def get_output_embeddings(self):
        return self.lm_head

    def set_output_embeddings(self, new_embeddings):
        self.lm_head = new_embeddings

    def tie_weights(self):
        """
        Overwrite `transformers.modeling_utils.PreTrainedModel.tie_weights` to handle the case of DecoupledLinear and DecoupledEmbedding.
        """
        output_embeddings = self.get_output_embeddings()
        input_embeddings = self.get_input_embeddings()

        if getattr(self.config, "tie_word_embeddings", True):
            output_embeddings.weight = input_embeddings.weight
            if input_embeddings.num_additional_embeddings > 0:
                assert output_embeddings.out_additional_features == input_embeddings.num_additional_embeddings
                output_embeddings.additional_fc.weight = input_embeddings.additional_embedding.weight

        if hasattr(output_embeddings, "out_features") and hasattr(input_embeddings, "num_embeddings"):
            output_embeddings.out_features = input_embeddings.num_embeddings
            if hasattr(output_embeddings, "out_additional_features") and hasattr(
                input_embeddings, "num_additional_embeddings"
            ):
                output_embeddings.out_additional_features = input_embeddings.num_additional_embeddings

    def prepare_inputs_for_generation(self, input_ids, past=None, **kwargs):
        return prepare_inputs_for_generation(input_ids, past=past, **kwargs)

    @staticmethod
    def _expand_inputs_for_generation(
        *args,
        **model_kwargs,
    ):
        return expand_inputs_for_generation(*args, **model_kwargs)

    @staticmethod
    def _update_model_kwargs_for_generation(outputs, model_kwargs, is_encoder_decoder=False):
        return update_model_kwargs_for_generation(outputs, model_kwargs, is_encoder_decoder)

    @add_start_docstrings_to_model_forward(GPT2_INPUTS_DOCSTRING)
    @add_code_sample_docstrings(
        processor_class=_TOKENIZER_FOR_DOC,
        checkpoint=_CHECKPOINT_FOR_DOC,
        output_type=CausalLMOutputWithCrossAttentions,
        config_class=_CONFIG_FOR_DOC,
    )
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        token_type_ids: Optional[torch.LongTensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        head_mask: Optional[torch.FloatTensor] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        pixel_values: Optional[torch.FloatTensor] = None,
        image_embeddings: Optional[torch.FloatTensor] = None,
        image_attention_mask: Optional[torch.Tensor] = None,
        crossblock_head_mask: Optional[torch.Tensor] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple, CausalLMOutputWithCrossAttentions]:
        r"""
        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
            `labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100`
            are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]`
        """
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        transformer_outputs = self.transformer(
            input_ids,
            past_key_values=past_key_values,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
            pixel_values=pixel_values,
            image_embeddings=image_embeddings,
            image_attention_mask=image_attention_mask,
            crossblock_head_mask=crossblock_head_mask,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )
        hidden_states = transformer_outputs[0]

        # Set device for model parallelism
        if self.model_parallel:
            torch.cuda.set_device(self.transformer.first_device)
            hidden_states = hidden_states.to(self.lm_head.weight.device)

        lm_logits = self.lm_head(hidden_states)

        loss = None
        if labels is not None:
            # Shift so that tokens < n predict n
            if attention_mask is not None:
                shift_attention_mask = attention_mask[..., 1:]
                shift_logits = lm_logits[..., :-1, :][shift_attention_mask != 0].contiguous()
                shift_labels = labels[..., 1:][shift_attention_mask != 0].contiguous()
            else:
                shift_logits = lm_logits[..., :-1, :].contiguous()
                shift_labels = labels[..., 1:].contiguous()
            # Flatten the tokens
            loss_fct = CrossEntropyLoss()
            loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))

        if not return_dict:
            output = (lm_logits,) + transformer_outputs[1:]
            return ((loss,) + output) if loss is not None else output

        return CausalLMOutputWithCrossAttentions(
            loss=loss,
            logits=lm_logits,
            past_key_values=transformer_outputs.past_key_values,
            hidden_states=transformer_outputs.hidden_states,
            attentions=transformer_outputs.attentions,
            cross_attentions=transformer_outputs.cross_attentions,
        )

    @staticmethod
    def _reorder_cache(past: Tuple[Tuple[torch.Tensor]], beam_idx: torch.Tensor) -> Tuple[Tuple[torch.Tensor]]:
        """
        This function is used to re-order the `past_key_values` cache if [`~PreTrainedModel.beam_search`] or
        [`~PreTrainedModel.beam_sample`] is called. This is required to match `past_key_values` with the correct
        beam_idx at every generation step.
        """
        return tuple(
            tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past)
            for layer_past in past
        )

    def get_model_tflops_per_batch_per_gpu(self, hparams, data_param, tokenizer, max_num_images):
        config_vl_model = self.config

        language_embed_size = config_vl_model.n_embd
        num_language_layers = config_vl_model.n_layer
        ffn_inner_size = config_vl_model.n_inner

        vision_config = self.transformer.vision_model.config
        if hasattr(vision_config, "vision_config"):
            vision_config = vision_config.vision_config

        # Get vision model blocks infos
        vision_patch_size = vision_config.patch_size
        vision_hidden_size = vision_config.hidden_size
        num_vision_layers = vision_config.num_hidden_layers
        # The +1 is for the CLS token
        single_image_seq_len = (vision_config.image_size // vision_patch_size) ** 2 + 1
        vision_exp_factor = vision_config.intermediate_size // vision_hidden_size

        # Get language and cross-att blocks infos
        num_cross_attn_layers = num_language_layers // config_vl_model.cross_layer_interval
        language_seq_len = data_param.max_seq_len
        language_exp_factor = (ffn_inner_size // language_embed_size) if ffn_inner_size is not None else 4
        cross_att_exp_factor = (ffn_inner_size // language_embed_size) if ffn_inner_size is not None else 4
        k_v_cross_attn_seq_len = (
            (self.config.resampler_n_latents * max_num_images)
            if self.config.use_resampler
            else (single_image_seq_len * max_num_images)
        )

        language_tflops_per_batch_per_gpu = compute_tflops_per_batch_per_gpu(
            num_layers=num_language_layers,
            batch_size=hparams.batch_size_per_gpu,
            q_seq_len=language_seq_len,
            k_seq_len=language_seq_len,
            hidden_size=language_embed_size,
            kv_in_dim=language_embed_size,
            ff_exp_factor=language_exp_factor,
            grad_acc_size=hparams.grad_acc_size,
            swiglu=False,
            vocab_size=tokenizer.vocab_size,
            count_backward=True,  # Always True regardless of freezing, because gradients are computed for cross-attentions
            use_grad_checkpointing=hparams.gradient_checkpointing,
        )
        cross_attention_tflops_per_batch_per_gpu = compute_tflops_per_batch_per_gpu(
            num_layers=num_cross_attn_layers,
            batch_size=hparams.batch_size_per_gpu,
            q_seq_len=language_seq_len,
            k_seq_len=k_v_cross_attn_seq_len,
            hidden_size=language_embed_size,
            kv_in_dim=vision_hidden_size,
            ff_exp_factor=cross_att_exp_factor,
            grad_acc_size=hparams.grad_acc_size,
            swiglu=False,
            vocab_size=None,
            count_backward=True,
            use_grad_checkpointing=hparams.gradient_checkpointing,
        )
        vision_tflops_per_batch_per_gpu = compute_tflops_per_batch_per_gpu(
            num_layers=num_vision_layers,
            batch_size=hparams.batch_size_per_gpu * max_num_images,
            q_seq_len=single_image_seq_len,
            k_seq_len=single_image_seq_len,
            hidden_size=vision_hidden_size,
            kv_in_dim=vision_hidden_size,
            ff_exp_factor=vision_exp_factor,
            grad_acc_size=hparams.grad_acc_size,
            swiglu=False,
            vocab_size=None,
            count_backward=not hparams.model_params["freeze_vision_layers"],
            use_grad_checkpointing=hparams.gradient_checkpointing,
        )
        if self.config.use_resampler:
            perceiver_tflops_per_batch_per_gpu = compute_perceiver_tflops_per_batch_per_gpu(
                num_layers=self.config.resampler_depth,
                batch_size=hparams.batch_size_per_gpu * max_num_images,
                q_seq_len=self.config.resampler_n_latents,
                vision_embed_seq_len=single_image_seq_len,
                q_k_v_input_dim=vision_hidden_size,
                attention_hidden_size=self.config.resampler_n_heads * self.config.resampler_head_dim,
                ff_exp_factor=cross_att_exp_factor,
                count_backward=True,
                use_grad_checkpointing=hparams.gradient_checkpointing,
            )
            flop_count = (
                language_tflops_per_batch_per_gpu
                + cross_attention_tflops_per_batch_per_gpu
                + vision_tflops_per_batch_per_gpu
                + perceiver_tflops_per_batch_per_gpu
            )
        else:
            flop_count = (
                language_tflops_per_batch_per_gpu
                + cross_attention_tflops_per_batch_per_gpu
                + vision_tflops_per_batch_per_gpu
            )
        return flop_count