File size: 67,823 Bytes
4c65bff
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
# coding=utf-8
# Copyright 2021 The Facebook, Inc and The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
""" TF 2.0 Blenderbot model."""


from __future__ import annotations

import os
import random
import warnings
from typing import List, Optional, Tuple, Union

import tensorflow as tf

from ...activations_tf import get_tf_activation
from ...modeling_tf_outputs import (
    TFBaseModelOutput,
    TFBaseModelOutputWithPastAndCrossAttentions,
    TFSeq2SeqLMOutput,
    TFSeq2SeqModelOutput,
)

# Public API
from ...modeling_tf_utils import (
    TFCausalLanguageModelingLoss,
    TFPreTrainedModel,
    keras_serializable,
    unpack_inputs,
)
from ...tf_utils import check_embeddings_within_bounds, shape_list, stable_softmax
from ...utils import (
    ContextManagers,
    add_code_sample_docstrings,
    add_end_docstrings,
    add_start_docstrings,
    add_start_docstrings_to_model_forward,
    logging,
    replace_return_docstrings,
)
from .configuration_blenderbot import BlenderbotConfig


logger = logging.get_logger(__name__)

_CHECKPOINT_FOR_DOC = "facebook/blenderbot-400M-distill"
_CONFIG_FOR_DOC = "BlenderbotConfig"


LARGE_NEGATIVE = -1e8


# Copied from transformers.models.bart.modeling_tf_bart.shift_tokens_right
def shift_tokens_right(input_ids: tf.Tensor, pad_token_id: int, decoder_start_token_id: int):
    pad_token_id = tf.cast(pad_token_id, input_ids.dtype)
    decoder_start_token_id = tf.cast(decoder_start_token_id, input_ids.dtype)
    start_tokens = tf.fill(
        (shape_list(input_ids)[0], 1), tf.convert_to_tensor(decoder_start_token_id, input_ids.dtype)
    )
    shifted_input_ids = tf.concat([start_tokens, input_ids[:, :-1]], -1)
    # replace possible -100 values in labels by `pad_token_id`
    shifted_input_ids = tf.where(
        shifted_input_ids == -100,
        tf.fill(shape_list(shifted_input_ids), tf.convert_to_tensor(pad_token_id, input_ids.dtype)),
        shifted_input_ids,
    )

    # "Verify that `labels` has only positive values and -100"
    assert_gte0 = tf.debugging.assert_greater_equal(shifted_input_ids, tf.constant(0, dtype=input_ids.dtype))

    # Make sure the assertion op is called by wrapping the result in an identity no-op
    with tf.control_dependencies([assert_gte0]):
        shifted_input_ids = tf.identity(shifted_input_ids)

    return shifted_input_ids


# Copied from transformers.models.bart.modeling_tf_bart._make_causal_mask
def _make_causal_mask(input_ids_shape: tf.TensorShape, past_key_values_length: int = 0):
    """
    Make causal mask used for bi-directional self-attention.
    """
    bsz = input_ids_shape[0]
    tgt_len = input_ids_shape[1]
    mask = tf.ones((tgt_len, tgt_len)) * LARGE_NEGATIVE
    mask_cond = tf.range(shape_list(mask)[-1])

    mask = tf.where(mask_cond < tf.reshape(mask_cond + 1, (shape_list(mask)[-1], 1)), 0.0, mask)

    if past_key_values_length > 0:
        mask = tf.concat([tf.zeros((tgt_len, past_key_values_length)), mask], axis=-1)

    return tf.tile(mask[None, None, :, :], (bsz, 1, 1, 1))


# Copied from transformers.models.bart.modeling_tf_bart._expand_mask
def _expand_mask(mask: tf.Tensor, tgt_len: Optional[int] = None):
    """
    Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
    """
    src_len = shape_list(mask)[1]
    tgt_len = tgt_len if tgt_len is not None else src_len
    one_cst = tf.constant(1.0)
    mask = tf.cast(mask, dtype=one_cst.dtype)
    expanded_mask = tf.tile(mask[:, None, None, :], (1, 1, tgt_len, 1))

    return (one_cst - expanded_mask) * LARGE_NEGATIVE


class TFBlenderbotLearnedPositionalEmbedding(tf.keras.layers.Embedding):
    """
    This module learns positional embeddings up to a fixed maximum size.
    """

    def __init__(self, num_embeddings: int, embedding_dim: int, **kwargs):
        super().__init__(num_embeddings, embedding_dim, **kwargs)

    def call(
        self, input_shape: tf.TensorShape, past_key_values_length: int = 0, position_ids: tf.Tensor | None = None
    ):
        """Input is expected to be of size [bsz x seqlen]."""
        if position_ids is None:
            seq_len = input_shape[1]
            position_ids = tf.range(seq_len, delta=1, name="range")
            position_ids += past_key_values_length

        return super().call(tf.cast(position_ids, dtype=tf.int32))


# Copied from transformers.models.bart.modeling_tf_bart.TFBartAttention with Bart->Blenderbot
class TFBlenderbotAttention(tf.keras.layers.Layer):
    """Multi-headed attention from "Attention Is All You Need"""

    def __init__(
        self,
        embed_dim: int,
        num_heads: int,
        dropout: float = 0.0,
        is_decoder: bool = False,
        bias: bool = True,
        **kwargs,
    ):
        super().__init__(**kwargs)
        self.embed_dim = embed_dim

        self.num_heads = num_heads
        self.dropout = tf.keras.layers.Dropout(dropout)
        self.head_dim = embed_dim // num_heads
        if (self.head_dim * num_heads) != self.embed_dim:
            raise ValueError(
                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim}"
                f" and `num_heads`: {num_heads})."
            )
        self.scaling = self.head_dim**-0.5
        self.is_decoder = is_decoder

        self.k_proj = tf.keras.layers.Dense(embed_dim, use_bias=bias, name="k_proj")
        self.q_proj = tf.keras.layers.Dense(embed_dim, use_bias=bias, name="q_proj")
        self.v_proj = tf.keras.layers.Dense(embed_dim, use_bias=bias, name="v_proj")
        self.out_proj = tf.keras.layers.Dense(embed_dim, use_bias=bias, name="out_proj")

    def _shape(self, tensor: tf.Tensor, seq_len: int, bsz: int):
        return tf.transpose(tf.reshape(tensor, (bsz, seq_len, self.num_heads, self.head_dim)), (0, 2, 1, 3))

    def call(
        self,
        hidden_states: tf.Tensor,
        key_value_states: tf.Tensor | None = None,
        past_key_value: Tuple[Tuple[tf.Tensor]] | None = None,
        attention_mask: tf.Tensor | None = None,
        layer_head_mask: tf.Tensor | None = None,
        training: Optional[bool] = False,
    ) -> Tuple[tf.Tensor, tf.Tensor | None]:
        """Input shape: Batch x Time x Channel"""

        # if key_value_states are provided this layer is used as a cross-attention layer
        # for the decoder
        is_cross_attention = key_value_states is not None
        bsz, tgt_len, embed_dim = shape_list(hidden_states)

        # get query proj
        query_states = self.q_proj(hidden_states) * self.scaling
        # get key, value proj
        if is_cross_attention and past_key_value is not None:
            # reuse k,v, cross_attentions
            key_states = past_key_value[0]
            value_states = past_key_value[1]
        elif is_cross_attention:
            # cross_attentions
            key_states = self._shape(self.k_proj(key_value_states), -1, bsz)
            value_states = self._shape(self.v_proj(key_value_states), -1, bsz)
        elif past_key_value is not None:
            # reuse k, v, self_attention
            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
            key_states = tf.concat([past_key_value[0], key_states], axis=2)
            value_states = tf.concat([past_key_value[1], value_states], axis=2)
        else:
            # self_attention
            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)

        if self.is_decoder:
            # if cross_attention save Tuple(tf.Tensor, tf.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(tf.Tensor, tf.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_states, value_states)

        proj_shape = (bsz * self.num_heads, -1, self.head_dim)
        query_states = tf.reshape(self._shape(query_states, tgt_len, bsz), proj_shape)
        key_states = tf.reshape(key_states, proj_shape)
        value_states = tf.reshape(value_states, proj_shape)

        src_len = shape_list(key_states)[1]
        attn_weights = tf.matmul(query_states, key_states, transpose_b=True)

        tf.debugging.assert_equal(
            shape_list(attn_weights),
            [bsz * self.num_heads, tgt_len, src_len],
            message=(
                f"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is"
                f" {shape_list(attn_weights)}"
            ),
        )

        if attention_mask is not None:
            tf.debugging.assert_equal(
                shape_list(attention_mask),
                [bsz, 1, tgt_len, src_len],
                message=(
                    f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is"
                    f" {shape_list(attention_mask)}"
                ),
            )

            attention_mask = tf.cast(attention_mask, dtype=attn_weights.dtype)
            attn_weights = tf.reshape(attn_weights, (bsz, self.num_heads, tgt_len, src_len)) + attention_mask
            attn_weights = tf.reshape(attn_weights, (bsz * self.num_heads, tgt_len, src_len))

        attn_weights = stable_softmax(attn_weights, axis=-1)

        if layer_head_mask is not None:
            tf.debugging.assert_equal(
                shape_list(layer_head_mask),
                [self.num_heads],
                message=(
                    f"Head mask for a single layer should be of size {(self.num_heads)}, but is"
                    f" {shape_list(layer_head_mask)}"
                ),
            )

            attn_weights = tf.reshape(layer_head_mask, (1, -1, 1, 1)) * tf.reshape(
                attn_weights, (bsz, self.num_heads, tgt_len, src_len)
            )
            attn_weights = tf.reshape(attn_weights, (bsz * self.num_heads, tgt_len, src_len))

        attn_probs = self.dropout(attn_weights, training=training)
        attn_output = tf.matmul(attn_probs, value_states)

        tf.debugging.assert_equal(
            shape_list(attn_output),
            [bsz * self.num_heads, tgt_len, self.head_dim],
            message=(
                f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
                f" {shape_list(attn_output)}"
            ),
        )

        attn_output = tf.transpose(
            tf.reshape(attn_output, (bsz, self.num_heads, tgt_len, self.head_dim)), (0, 2, 1, 3)
        )
        attn_output = tf.reshape(attn_output, (bsz, tgt_len, embed_dim))

        attn_output = self.out_proj(attn_output)
        attn_weights: tf.Tensor = tf.reshape(attn_weights, (bsz, self.num_heads, tgt_len, src_len))

        return attn_output, attn_weights, past_key_value


# Copied from transformers.models.mbart.modeling_tf_mbart.TFMBartEncoderLayer with MBart->Blenderbot
class TFBlenderbotEncoderLayer(tf.keras.layers.Layer):
    def __init__(self, config: BlenderbotConfig, **kwargs):
        super().__init__(**kwargs)
        self.embed_dim = config.d_model
        self.self_attn = TFBlenderbotAttention(
            self.embed_dim, config.encoder_attention_heads, dropout=config.attention_dropout, name="self_attn"
        )
        self.self_attn_layer_norm = tf.keras.layers.LayerNormalization(epsilon=1e-5, name="self_attn_layer_norm")
        self.dropout = tf.keras.layers.Dropout(config.dropout)
        self.activation_fn = get_tf_activation(config.activation_function)
        self.activation_dropout = tf.keras.layers.Dropout(config.activation_dropout)
        self.fc1 = tf.keras.layers.Dense(config.encoder_ffn_dim, name="fc1")
        self.fc2 = tf.keras.layers.Dense(self.embed_dim, name="fc2")
        self.final_layer_norm = tf.keras.layers.LayerNormalization(epsilon=1e-5, name="final_layer_norm")

    def call(
        self,
        hidden_states: tf.Tensor,
        attention_mask: tf.Tensor,
        layer_head_mask: tf.Tensor,
        training: Optional[bool] = False,
    ):
        """
        Args:
            hidden_states (`tf.Tensor`): input to the layer of shape *(batch, seq_len, embed_dim)*
            attention_mask (`tf.Tensor`): attention mask of size
                *(batch, 1, tgt_len, src_len)* where padding elements are indicated by very large negative values.
            layer_head_mask (`tf.Tensor`): mask for attention heads in a given layer of size
                *(encoder_attention_heads,)*
        """
        residual = hidden_states
        hidden_states = self.self_attn_layer_norm(hidden_states)
        hidden_states, self_attn_weights, _ = self.self_attn(
            hidden_states=hidden_states, attention_mask=attention_mask, layer_head_mask=layer_head_mask
        )

        tf.debugging.assert_equal(
            shape_list(hidden_states),
            shape_list(residual),
            message=f"Self attn modified the shape of query {shape_list(residual)} to {shape_list(hidden_states)}",
        )

        hidden_states = self.dropout(hidden_states, training=training)
        hidden_states = residual + hidden_states

        residual = hidden_states
        hidden_states = self.final_layer_norm(hidden_states)
        hidden_states = self.activation_fn(self.fc1(hidden_states))
        hidden_states = self.activation_dropout(hidden_states, training=training)
        hidden_states = self.fc2(hidden_states)
        hidden_states = self.dropout(hidden_states, training=training)
        hidden_states = residual + hidden_states

        return hidden_states, self_attn_weights


# Copied from transformers.models.mbart.modeling_tf_mbart.TFMBartDecoderLayer with MBart->Blenderbot
class TFBlenderbotDecoderLayer(tf.keras.layers.Layer):
    def __init__(self, config: BlenderbotConfig, **kwargs):
        super().__init__(**kwargs)
        self.embed_dim = config.d_model
        self.self_attn = TFBlenderbotAttention(
            embed_dim=self.embed_dim,
            num_heads=config.decoder_attention_heads,
            dropout=config.attention_dropout,
            name="self_attn",
            is_decoder=True,
        )
        self.dropout = tf.keras.layers.Dropout(config.dropout)
        self.activation_fn = get_tf_activation(config.activation_function)
        self.activation_dropout = tf.keras.layers.Dropout(config.activation_dropout)

        self.self_attn_layer_norm = tf.keras.layers.LayerNormalization(epsilon=1e-5, name="self_attn_layer_norm")
        self.encoder_attn = TFBlenderbotAttention(
            self.embed_dim,
            config.decoder_attention_heads,
            dropout=config.attention_dropout,
            name="encoder_attn",
            is_decoder=True,
        )
        self.encoder_attn_layer_norm = tf.keras.layers.LayerNormalization(epsilon=1e-5, name="encoder_attn_layer_norm")
        self.fc1 = tf.keras.layers.Dense(config.decoder_ffn_dim, name="fc1")
        self.fc2 = tf.keras.layers.Dense(self.embed_dim, name="fc2")
        self.final_layer_norm = tf.keras.layers.LayerNormalization(epsilon=1e-5, name="final_layer_norm")

    def call(
        self,
        hidden_states: tf.Tensor,
        attention_mask: tf.Tensor | None = None,
        encoder_hidden_states: tf.Tensor | None = None,
        encoder_attention_mask: tf.Tensor | None = None,
        layer_head_mask: tf.Tensor | None = None,
        cross_attn_layer_head_mask: tf.Tensor | None = None,
        past_key_value: Tuple[tf.Tensor] | None = None,
        training: Optional[bool] = False,
    ) -> Tuple[tf.Tensor, tf.Tensor, Tuple[Tuple[tf.Tensor]]]:
        """
        Args:
            hidden_states (`tf.Tensor`): input to the layer of shape *(batch, seq_len, embed_dim)*
            attention_mask (`tf.Tensor`): attention mask of size
                *(batch, 1, tgt_len, src_len)* where padding elements are indicated by very large negative values.
            encoder_hidden_states (`tf.Tensor`):
                cross attention input to the layer of shape *(batch, seq_len, embed_dim)*
            encoder_attention_mask (`tf.Tensor`): encoder attention mask of size
                *(batch, 1, tgt_len, src_len)* where padding elements are indicated by very large negative values.
            layer_head_mask (`tf.Tensor`): mask for attention heads in a given layer of size
                *(decoder_attention_heads,)*
            cross_attn_layer_head_mask (`tf.Tensor`): mask for heads of the cross-attention module.
                *(decoder_attention_heads,)*
            past_key_value (`Tuple(tf.Tensor)`): cached past key and value projection states
        """
        residual = hidden_states
        hidden_states = self.self_attn_layer_norm(hidden_states)

        # Self Attention
        # decoder uni-directional self-attention cached key/values tuple is at positions 1,2
        self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None
        # add present self-attn cache to positions 1,2 of present_key_value tuple
        hidden_states, self_attn_weights, present_key_value = self.self_attn(
            hidden_states=hidden_states,
            past_key_value=self_attn_past_key_value,
            attention_mask=attention_mask,
            layer_head_mask=layer_head_mask,
        )
        hidden_states = self.dropout(hidden_states, training=training)
        hidden_states = residual + hidden_states

        # Cross-Attention Block
        cross_attn_present_key_value = None
        cross_attn_weights = None
        if encoder_hidden_states is not None:
            residual = hidden_states
            hidden_states = self.encoder_attn_layer_norm(hidden_states)

            # cross_attn cached key/values tuple is at positions 3,4 of present_key_value tuple
            cross_attn_past_key_value = past_key_value[-2:] if past_key_value is not None else None
            hidden_states, cross_attn_weights, cross_attn_present_key_value = self.encoder_attn(
                hidden_states=hidden_states,
                key_value_states=encoder_hidden_states,
                attention_mask=encoder_attention_mask,
                layer_head_mask=cross_attn_layer_head_mask,
                past_key_value=cross_attn_past_key_value,
            )
            hidden_states = self.dropout(hidden_states, training=training)
            hidden_states = residual + hidden_states

            # add cross-attn to positions 3,4 of present_key_value tuple
            present_key_value = present_key_value + cross_attn_present_key_value

        # Fully Connected
        residual = hidden_states
        hidden_states = self.final_layer_norm(hidden_states)
        hidden_states = self.activation_fn(self.fc1(hidden_states))
        hidden_states = self.activation_dropout(hidden_states, training=training)
        hidden_states = self.fc2(hidden_states)
        hidden_states = self.dropout(hidden_states, training=training)
        hidden_states = residual + hidden_states

        return (
            hidden_states,
            self_attn_weights,
            cross_attn_weights,
            present_key_value,
        )


class TFBlenderbotPreTrainedModel(TFPreTrainedModel):
    config_class = BlenderbotConfig
    base_model_prefix = "model"


BLENDERBOT_START_DOCSTRING = r"""
    This model inherits from [`TFPreTrainedModel`]. 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 [tf.keras.Model](https://www.tensorflow.org/api_docs/python/tf/keras/Model) subclass. Use it
    as a regular TF 2.0 Keras Model and refer to the TF 2.0 documentation for all matter related to general usage and
    behavior.

    <Tip>

    TensorFlow models and layers in `transformers` accept two formats as input:

    - having all inputs as keyword arguments (like PyTorch models), or
    - having all inputs as a list, tuple or dict in the first positional argument.

    The reason the second format is supported is that Keras methods prefer this format when passing inputs to models
    and layers. Because of this support, when using methods like `model.fit()` things should "just work" for you - just
    pass your inputs and labels in any format that `model.fit()` supports! If, however, you want to use the second
    format outside of Keras methods like `fit()` and `predict()`, such as when creating your own layers or models with
    the Keras `Functional` API, there are three possibilities you can use to gather all the input Tensors in the first
    positional argument:

    - a single Tensor with `input_ids` only and nothing else: `model(input_ids)`
    - a list of varying length with one or several input Tensors IN THE ORDER given in the docstring:
    `model([input_ids, attention_mask])` or `model([input_ids, attention_mask, token_type_ids])`
    - a dictionary with one or several input Tensors associated to the input names given in the docstring:
    `model({"input_ids": input_ids, "token_type_ids": token_type_ids})`

    Note that when creating models and layers with
    [subclassing](https://keras.io/guides/making_new_layers_and_models_via_subclassing/) then you don't need to worry
    about any of this, as you can just pass inputs like you would to any other Python function!

    </Tip>

    Args:
        config ([`BlenderbotConfig`]): 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 [`~TFPreTrainedModel.from_pretrained`] method to load the model weights.
"""

BLENDERBOT_GENERATION_EXAMPLE = r"""
    Conversation example::

    ```py
    >>> from transformers import AutoTokenizer, TFBlenderbotForConditionalGeneration

    >>> mname = "facebook/blenderbot-400M-distill"
    >>> model = TFBlenderbotForConditionalGeneration.from_pretrained(mname)
    >>> tokenizer = AutoTokenizer.from_pretrained(mname)
    >>> UTTERANCE = "My friends are cool but they eat too many carbs."
    >>> print("Human: ", UTTERANCE)

    >>> inputs = tokenizer([UTTERANCE], return_tensors="tf")
    >>> reply_ids = model.generate(**inputs)
    >>> print("Bot: ", tokenizer.batch_decode(reply_ids, skip_special_tokens=True)[0])

    >>> REPLY = "I'm not sure"
    >>> print("Human: ", REPLY)
    >>> NEXT_UTTERANCE = (
    ...     "My friends are cool but they eat too many carbs.</s> <s>That's unfortunate. "
    ...     "Are they trying to lose weight or are they just trying to be healthier?</s> "
    ...     "<s> I'm not sure."
    ... )
    >>> inputs = tokenizer([NEXT_UTTERANCE], return_tensors="tf")
    >>> next_reply_ids = model.generate(**inputs)
    >>> print("Bot: ", tokenizer.batch_decode(next_reply_ids, skip_special_tokens=True)[0])
    ```
"""

BLENDERBOT_INPUTS_DOCSTRING = r"""
    Args:
        input_ids (`tf.Tensor` of shape `({0})`):
            Indices of input sequence tokens in the vocabulary.

            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
            [`PreTrainedTokenizer.__call__`] for details.

            [What are input IDs?](../glossary#input-ids)
        attention_mask (`tf.Tensor` of shape `({0})`, *optional*):
            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:

            - 1 for tokens that are **not masked**,
            - 0 for tokens that are **masked**.

            [What are attention masks?](../glossary#attention-mask)
        decoder_input_ids (`tf.Tensor` of shape `(batch_size, target_sequence_length)`, *optional*):
            Indices of decoder input sequence tokens in the vocabulary.

            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
            [`PreTrainedTokenizer.__call__`] for details.

            [What are decoder input IDs?](../glossary#decoder-input-ids)

            Blenderbot uses the `bos_token_id` as the starting token for `decoder_input_ids` generation. If
            `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
            `past_key_values`).
        decoder_attention_mask (`tf.Tensor` of shape `(batch_size, target_sequence_length)`, *optional*):
            will be made by default and ignore pad tokens. It is not recommended to set this for most use cases.
        decoder_position_ids (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
            Indices of positions of each decoder input sequence tokens in the position embeddings. Selected in the
            range `[0, config.max_position_embeddings - 1]`.
        head_mask (`tf.Tensor` of shape `(encoder_layers, encoder_attention_heads)`, *optional*):
            Mask to nullify selected heads of the attention modules in the encoder. Mask values selected in `[0, 1]`:

            - 1 indicates the head is **not masked**,
            - 0 indicates the head is **masked**.

        decoder_head_mask (`tf.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
            Mask to nullify selected heads of the attention modules in the decoder. Mask values selected in `[0, 1]`:

            - 1 indicates the head is **not masked**,
            - 0 indicates the head is **masked**.

        cross_attn_head_mask (`tf.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
            Mask to nullify selected heads of the cross-attention modules. Mask values selected in `[0, 1]`:

            - 1 indicates the head is **not masked**,
            - 0 indicates the head is **masked**.

        encoder_outputs (`tf.FloatTensor`, *optional*):
            hidden states at the output of the last layer of the encoder. Used in the cross-attention of the decoder.
            of shape `(batch_size, sequence_length, hidden_size)` is a sequence of
        past_key_values (`Tuple[Tuple[tf.Tensor]]` of length `config.n_layers`)
            contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
            If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
            don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
            `decoder_input_ids` of shape `(batch_size, sequence_length)`.
        use_cache (`bool`, *optional*, defaults to `True`):
            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
            `past_key_values`). Set to `False` during training, `True` during generation
        output_attentions (`bool`, *optional*):
            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
            tensors for more detail. This argument can be used only in eager mode, in graph mode the value in the
            config will be used instead.
        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. This argument can be used only in eager mode, in graph mode the value in the config will be
            used instead.
        return_dict (`bool`, *optional*):
            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. This argument can be used in
            eager mode, in graph mode the value will always be set to True.
        training (`bool`, *optional*, defaults to `False`):
            Whether or not to use the model in training mode (some modules like dropout modules have different
            behaviors between training and evaluation).
"""


@keras_serializable
class TFBlenderbotEncoder(tf.keras.layers.Layer):
    config_class = BlenderbotConfig
    """
    Transformer encoder consisting of *config.encoder_layers* self attention layers. Each layer is a
    [`TFBlenderbotEncoderLayer`].

    Args:
        config: BlenderbotConfig
    """

    def __init__(self, config: BlenderbotConfig, embed_tokens: Optional[tf.keras.layers.Embedding] = None, **kwargs):
        super().__init__(**kwargs)
        self.config = config
        self.dropout = tf.keras.layers.Dropout(config.dropout)
        self.layerdrop = config.encoder_layerdrop
        self.padding_idx = config.pad_token_id
        self.max_source_positions = config.max_position_embeddings
        self.embed_scale = tf.math.sqrt(float(config.d_model)) if config.scale_embedding else 1.0

        self.embed_tokens = embed_tokens
        self.embed_positions = TFBlenderbotLearnedPositionalEmbedding(
            config.max_position_embeddings,
            config.d_model,
            name="embed_positions",
        )
        self.layers = [TFBlenderbotEncoderLayer(config, name=f"layers.{i}") for i in range(config.encoder_layers)]
        self.layer_norm = tf.keras.layers.LayerNormalization(epsilon=1e-5, name="layer_norm")

    def get_embed_tokens(self):
        return self.embed_tokens

    def set_embed_tokens(self, embed_tokens):
        self.embed_tokens = embed_tokens

    @unpack_inputs
    def call(
        self,
        input_ids=None,
        inputs_embeds=None,
        attention_mask=None,
        head_mask=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
        training=False,
    ):
        """
        Args:
            input_ids (`tf.Tensor` of shape `(batch_size, sequence_length)`):
                Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you
                provide it.

                Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
                [`PreTrainedTokenizer.__call__`] for details.

                [What are input IDs?](../glossary#input-ids)
            attention_mask (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
                Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:

                - 1 for tokens that are **not masked**,
                - 0 for tokens that are **masked**.

                [What are attention masks?](../glossary#attention-mask)
            head_mask (`tf.Tensor` of shape `(encoder_layers, encoder_attention_heads)`, `optional):
                Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:

                - 1 indicates the head is **not masked**,
                - 0 indicates the head is **masked**.

            inputs_embeds (`tf.Tensor` 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.
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail. This argument can be used only in eager mode, in graph mode the value
                in the config will be used instead.
            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. This argument can be used only in eager mode, in graph mode the value in the config
                will be used instead.
            return_dict (`bool`, *optional*):
                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. This argument can be used
                in eager mode, in graph mode the value will always be set to True.
            training (`bool`, *optional*, defaults to `False`):
                Whether or not to use the model in training mode (some modules like dropout modules have different
                behaviors between training and evaluation).
        """
        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 = shape_list(input_ids)
        elif inputs_embeds is not None:
            input_shape = shape_list(inputs_embeds)[:-1]
        else:
            raise ValueError("You have to specify either input_ids or inputs_embeds")

        if inputs_embeds is None:
            # if `self.embed_tokens.load_weight_prefix` is set, runs the embedding operation with the correct name
            # scope, so that its weights are registered with the desired name for loading/storing. When `tf.name_scope`
            # is used with a name ending in `/`, that name replaces the current name scope.
            # (embeddings with tf.name_scope: self.embed_tokens.load_weight_prefix/self.embed_tokens.name/embeddings:0)
            context = []
            if hasattr(self.embed_tokens, "load_weight_prefix"):
                context.append(tf.name_scope(self.embed_tokens.load_weight_prefix + "/"))
            with ContextManagers(context):
                check_embeddings_within_bounds(input_ids, self.embed_tokens.input_dim)
                inputs_embeds = self.embed_tokens(input_ids) * self.embed_scale

        embed_pos = self.embed_positions(input_shape)
        hidden_states = inputs_embeds + embed_pos
        hidden_states = self.dropout(hidden_states, training=training)

        # check attention mask and invert
        if attention_mask is not None:
            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
            attention_mask = _expand_mask(attention_mask)
        else:
            attention_mask = None

        encoder_states = () if output_hidden_states else None
        all_attentions = () if output_attentions else None

        # check if head_mask has a correct number of layers specified if desired
        if head_mask is not None:
            tf.debugging.assert_equal(
                shape_list(head_mask)[0],
                len(self.layers),
                message=(
                    f"The head_mask should be specified for {len(self.layers)} layers, but it is for"
                    f" {shape_list(head_mask)[0]}."
                ),
            )

        # encoder layers
        for idx, encoder_layer in enumerate(self.layers):
            if output_hidden_states:
                encoder_states = encoder_states + (hidden_states,)
            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
            dropout_probability = random.uniform(0, 1)
            if training and (dropout_probability < self.layerdrop):  # skip the layer
                continue

            hidden_states, attn = encoder_layer(
                hidden_states,
                attention_mask,
                head_mask[idx] if head_mask is not None else None,
            )

            if output_attentions:
                all_attentions += (attn,)

        hidden_states = self.layer_norm(hidden_states)

        if output_hidden_states:
            encoder_states = encoder_states + (hidden_states,)

        if not return_dict:
            return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None)
        return TFBaseModelOutput(
            last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions
        )


@keras_serializable
class TFBlenderbotDecoder(tf.keras.layers.Layer):
    config_class = BlenderbotConfig
    """
    Transformer decoder consisting of *config.decoder_layers* layers. Each layer is a [`TFBlenderbotDecoderLayer`]

    Args:
        config: BlenderbotConfig
        embed_tokens: output embedding
    """

    def __init__(self, config: BlenderbotConfig, embed_tokens: Optional[tf.keras.layers.Embedding] = None, **kwargs):
        super().__init__(**kwargs)
        self.config = config
        self.padding_idx = config.pad_token_id
        self.embed_tokens = embed_tokens
        self.layerdrop = config.decoder_layerdrop
        self.embed_positions = TFBlenderbotLearnedPositionalEmbedding(
            config.max_position_embeddings,
            config.d_model,
            name="embed_positions",
        )
        self.embed_scale = tf.math.sqrt(float(config.d_model)) if config.scale_embedding else 1.0
        self.layers = [TFBlenderbotDecoderLayer(config, name=f"layers.{i}") for i in range(config.decoder_layers)]
        self.layer_norm = tf.keras.layers.LayerNormalization(epsilon=1e-5, name="layer_norm")

        self.dropout = tf.keras.layers.Dropout(config.dropout)

    def get_embed_tokens(self):
        return self.embed_tokens

    def set_embed_tokens(self, embed_tokens):
        self.embed_tokens = embed_tokens

    @unpack_inputs
    def call(
        self,
        input_ids=None,
        inputs_embeds=None,
        attention_mask=None,
        position_ids=None,
        encoder_hidden_states=None,
        encoder_attention_mask=None,
        head_mask=None,
        cross_attn_head_mask=None,
        past_key_values=None,
        use_cache=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
        training=False,
    ):
        r"""
        Args:
            input_ids (`tf.Tensor` of shape `(batch_size, sequence_length)`):
                Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you
                provide it.

                Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
                [`PreTrainedTokenizer.__call__`] for details.

                [What are input IDs?](../glossary#input-ids)
            attention_mask (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
                Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:

                - 1 for tokens that are **not masked**,
                - 0 for tokens that are **masked**.

                [What are attention masks?](../glossary#attention-mask)
            position_ids (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
                Indices of positions of each decoder input sequence tokens in the position embeddings. Selected in the
                range `[0, config.max_position_embeddings - 1]`.
            encoder_hidden_states (`tf.Tensor` of shape `(batch_size, encoder_sequence_length, hidden_size)`, *optional*):
                Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention
                of the decoder.
            encoder_attention_mask (`tf.Tensor` of shape `(batch_size, encoder_sequence_length)`, *optional*):
                Mask to avoid performing cross-attention on padding tokens indices of encoder input_ids. Mask values
                selected in `[0, 1]`:

                - 1 for tokens that are **not masked**,
                - 0 for tokens that are **masked**.

                [What are attention masks?](../glossary#attention-mask)
            head_mask (`tf.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
                Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:

                - 1 indicates the head is **not masked**,
                - 0 indicates the head is **masked**.

            cross_attn_head_mask (`tf.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
                Mask to nullify selected heads of the cross-attention modules. Mask values selected in `[0, 1]`:

                - 1 indicates the head is **not masked**,
                - 0 indicates the head is **masked**.

            past_key_values (`Tuple[Tuple[tf.Tensor]]` of length `config.n_layers` with each tuple having 2 tuples each of which has 2 tensors of shape `(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
                Contains precomputed key and value hidden-states of the attention blocks. Can be used to speed up
                decoding.

                If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those
                that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of
                all `decoder_input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`tf.Tensor` 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.
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail. This argument can be used only in eager mode, in graph mode the value
                in the config will be used instead.
            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. This argument can be used only in eager mode, in graph mode the value in the config
                will be used instead.
            return_dict (`bool`, *optional*):
                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. This argument can be used
                in eager mode, in graph mode the value will always be set to True.
            training (`bool`, *optional*, defaults to `False`):
                Whether or not to use the model in training mode (some modules like dropout modules have different
                behaviors between training and evaluation).
        """
        if input_ids is not None and inputs_embeds is not None:
            raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
        elif input_ids is not None:
            input_shape = shape_list(input_ids)
        elif inputs_embeds is not None:
            input_shape = shape_list(inputs_embeds)[:-1]
        else:
            raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")

        past_key_values_length = shape_list(past_key_values[0][0])[2] if past_key_values is not None else 0

        # embed positions
        if position_ids is None:
            positions = self.embed_positions(input_shape, past_key_values_length)
        else:
            positions = self.embed_positions(input_shape, position_ids=position_ids)

        if inputs_embeds is None:
            context = []
            if hasattr(self.embed_tokens, "load_weight_prefix"):
                context.append(tf.name_scope(self.embed_tokens.load_weight_prefix + "/"))
            with ContextManagers(context):
                check_embeddings_within_bounds(input_ids, self.embed_tokens.input_dim)
                inputs_embeds = self.embed_tokens(input_ids) * self.embed_scale

        hidden_states = inputs_embeds

        # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
        if input_shape[-1] > 1:
            combined_attention_mask = _make_causal_mask(input_shape, past_key_values_length=past_key_values_length)
        else:
            combined_attention_mask = _expand_mask(
                tf.ones((input_shape[0], input_shape[1] + past_key_values_length)), tgt_len=input_shape[-1]
            )

        if attention_mask is not None:
            combined_attention_mask = combined_attention_mask + _expand_mask(attention_mask, tgt_len=input_shape[-1])

        if encoder_hidden_states is not None and encoder_attention_mask is not None:
            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
            encoder_attention_mask = _expand_mask(encoder_attention_mask, tgt_len=input_shape[-1])

        hidden_states = hidden_states + positions
        hidden_states = self.dropout(hidden_states, training=training)

        # decoder layers
        all_hidden_states = () if output_hidden_states else None
        all_self_attns = () if output_attentions else None
        all_cross_attns = () if (output_attentions and encoder_hidden_states is not None) else None
        present_key_values = () if use_cache else None

        # check if head_mask and cross_attn_head_mask have a correct number of layers specified if desired
        for attn_mask_name, attn_mask in [("head_mask", head_mask), ("cross_attn_head_mask", cross_attn_head_mask)]:
            if attn_mask is not None:
                tf.debugging.assert_equal(
                    shape_list(attn_mask)[0],
                    len(self.layers),
                    message=(
                        f"The {attn_mask_name} should be specified for {len(self.layers)} layers, but it is for"
                        f" {shape_list(attn_mask)[0]}."
                    ),
                )
        for idx, decoder_layer in enumerate(self.layers):
            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
            if output_hidden_states:
                all_hidden_states += (hidden_states,)
            dropout_probability = random.uniform(0, 1)

            if training and (dropout_probability < self.layerdrop):
                continue

            past_key_value = past_key_values[idx] if past_key_values is not None else None

            hidden_states, layer_self_attn, layer_cross_attn, present_key_value = decoder_layer(
                hidden_states,
                attention_mask=combined_attention_mask,
                encoder_hidden_states=encoder_hidden_states,
                encoder_attention_mask=encoder_attention_mask,
                layer_head_mask=head_mask[idx] if head_mask is not None else None,
                cross_attn_layer_head_mask=cross_attn_head_mask[idx] if cross_attn_head_mask is not None else None,
                past_key_value=past_key_value,
            )

            if use_cache:
                present_key_values += (present_key_value,)

            if output_attentions:
                all_self_attns += (layer_self_attn,)

                if encoder_hidden_states is not None:
                    all_cross_attns += (layer_cross_attn,)

        hidden_states = self.layer_norm(hidden_states)

        if output_hidden_states:
            all_hidden_states += (hidden_states,)

        if not return_dict:
            return hidden_states, present_key_values, all_hidden_states, all_self_attns, all_cross_attns
        else:
            return TFBaseModelOutputWithPastAndCrossAttentions(
                last_hidden_state=hidden_states,
                past_key_values=present_key_values,
                hidden_states=all_hidden_states,
                attentions=all_self_attns,
                cross_attentions=all_cross_attns,
            )


@keras_serializable
class TFBlenderbotMainLayer(tf.keras.layers.Layer):
    config_class = BlenderbotConfig

    def __init__(self, config: BlenderbotConfig, **kwargs):
        super().__init__(**kwargs)

        self.config = config
        self.shared = tf.keras.layers.Embedding(
            input_dim=config.vocab_size,
            output_dim=config.d_model,
            embeddings_initializer=tf.keras.initializers.TruncatedNormal(stddev=self.config.init_std),
            name="model.shared",
        )
        # Additional attribute to specify the expected name scope of the layer (for loading/storing weights)
        self.shared.load_weight_prefix = "model.shared"

        self.encoder = TFBlenderbotEncoder(config, self.shared, name="encoder")
        self.decoder = TFBlenderbotDecoder(config, self.shared, name="decoder")

    def get_input_embeddings(self):
        return self.shared

    def set_input_embeddings(self, new_embeddings):
        self.shared = new_embeddings
        self.encoder.embed_tokens = self.shared
        self.decoder.embed_tokens = self.shared

    @unpack_inputs
    def call(
        self,
        input_ids=None,
        attention_mask=None,
        decoder_input_ids=None,
        decoder_attention_mask=None,
        decoder_position_ids=None,
        head_mask=None,
        decoder_head_mask=None,
        cross_attn_head_mask=None,
        encoder_outputs: Optional[Union[Tuple, TFBaseModelOutput]] = None,
        past_key_values=None,
        inputs_embeds=None,
        decoder_inputs_embeds=None,
        use_cache=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
        training=False,
        **kwargs,
    ):
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )

        if encoder_outputs is None:
            encoder_outputs = self.encoder(
                input_ids=input_ids,
                attention_mask=attention_mask,
                head_mask=head_mask,
                inputs_embeds=inputs_embeds,
                output_attentions=output_attentions,
                output_hidden_states=output_hidden_states,
                return_dict=return_dict,
                training=training,
            )
        # If the user passed a tuple for encoder_outputs, we wrap it in a TFBaseModelOutput when return_dict=True
        elif return_dict and not isinstance(encoder_outputs, TFBaseModelOutput):
            encoder_outputs = TFBaseModelOutput(
                last_hidden_state=encoder_outputs[0],
                hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None,
                attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None,
            )
        # If the user passed a TFBaseModelOutput for encoder_outputs, we wrap it in a tuple when return_dict=False
        elif not return_dict and not isinstance(encoder_outputs, tuple):
            encoder_outputs = encoder_outputs.to_tuple()

        decoder_outputs = self.decoder(
            decoder_input_ids,
            attention_mask=decoder_attention_mask,
            position_ids=decoder_position_ids,
            encoder_hidden_states=encoder_outputs[0],
            encoder_attention_mask=attention_mask,
            head_mask=decoder_head_mask,
            cross_attn_head_mask=cross_attn_head_mask,
            past_key_values=past_key_values,
            inputs_embeds=decoder_inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            training=training,
        )

        if not return_dict:
            return decoder_outputs + encoder_outputs

        return TFSeq2SeqModelOutput(
            last_hidden_state=decoder_outputs.last_hidden_state,
            past_key_values=decoder_outputs.past_key_values,
            decoder_hidden_states=decoder_outputs.hidden_states,
            decoder_attentions=decoder_outputs.attentions,
            cross_attentions=decoder_outputs.cross_attentions,
            encoder_last_hidden_state=encoder_outputs.last_hidden_state,
            encoder_hidden_states=encoder_outputs.hidden_states,
            encoder_attentions=encoder_outputs.attentions,
        )


@add_start_docstrings(
    "The bare BLENDERBOT Model outputting raw hidden-states without any specific head on top.",
    BLENDERBOT_START_DOCSTRING,
)
class TFBlenderbotModel(TFBlenderbotPreTrainedModel):
    def __init__(self, config: BlenderbotConfig, *inputs, **kwargs):
        super().__init__(config, *inputs, **kwargs)

        self.model = TFBlenderbotMainLayer(config, name="model")

    def get_encoder(self):
        return self.model.encoder

    def get_decoder(self):
        return self.model.decoder

    @classmethod
    def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], *model_args, **kwargs):
        if pretrained_model_name_or_path == "facebook/blenderbot-90M":
            from ..blenderbot_small import TFBlenderbotSmallModel

            warnings.warn(
                "The checkpoint `facebook/blenderbot-90M` is deprecated. In the future, please use the identical"
                " checkpoint `facebook/small_blenderbot-90M` with"
                " `TFBlenderbotSmallForConditionalGeneration.from_pretrained('facebook/small_blenderbot-90M')`"
                " instead.",
                FutureWarning,
            )
            return TFBlenderbotSmallModel.from_pretrained(pretrained_model_name_or_path)

        return super().from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)

    @unpack_inputs
    @add_start_docstrings_to_model_forward(BLENDERBOT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
    @add_code_sample_docstrings(
        checkpoint=_CHECKPOINT_FOR_DOC,
        output_type=TFSeq2SeqModelOutput,
        config_class=_CONFIG_FOR_DOC,
    )
    def call(
        self,
        input_ids: tf.Tensor | None = None,
        attention_mask: tf.Tensor | None = None,
        decoder_input_ids: tf.Tensor | None = None,
        decoder_attention_mask: tf.Tensor | None = None,
        decoder_position_ids: tf.Tensor | None = None,
        head_mask: tf.Tensor | None = None,
        decoder_head_mask: tf.Tensor | None = None,
        cross_attn_head_mask: tf.Tensor | None = None,
        encoder_outputs: Optional[Union[Tuple, TFBaseModelOutput]] = None,
        past_key_values: List[tf.Tensor] | None = None,
        inputs_embeds: tf.Tensor | None = None,
        decoder_inputs_embeds: tf.Tensor | None = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        training: Optional[bool] = False,
        **kwargs,
    ) -> Union[Tuple[tf.Tensor], TFSeq2SeqModelOutput]:
        outputs = self.model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            decoder_input_ids=decoder_input_ids,
            decoder_attention_mask=decoder_attention_mask,
            decoder_position_ids=decoder_position_ids,
            head_mask=head_mask,
            decoder_head_mask=decoder_head_mask,
            cross_attn_head_mask=cross_attn_head_mask,
            encoder_outputs=encoder_outputs,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            decoder_inputs_embeds=decoder_inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            training=training,
        )

        return outputs

    # Copied from transformers.models.bart.modeling_tf_bart.TFBartModel.serving_output
    def serving_output(self, output):
        pkv = tf.tuple(output.past_key_values)[1] if self.config.use_cache else None
        dec_hs = tf.convert_to_tensor(output.decoder_hidden_states) if self.config.output_hidden_states else None
        dec_attns = tf.convert_to_tensor(output.decoder_attentions) if self.config.output_attentions else None
        cross_attns = tf.convert_to_tensor(output.cross_attentions) if self.config.output_attentions else None
        enc_hs = tf.convert_to_tensor(output.encoder_hidden_states) if self.config.output_hidden_states else None
        enc_attns = tf.convert_to_tensor(output.encoder_attentions) if self.config.output_attentions else None

        return TFSeq2SeqModelOutput(
            last_hidden_state=output.last_hidden_state,
            past_key_values=pkv,
            decoder_hidden_states=dec_hs,
            decoder_attentions=dec_attns,
            cross_attentions=cross_attns,
            encoder_last_hidden_state=output.encoder_last_hidden_state,
            encoder_hidden_states=enc_hs,
            encoder_attentions=enc_attns,
        )


# Copied from transformers.models.bart.modeling_tf_bart.BiasLayer
class BiasLayer(tf.keras.layers.Layer):
    """
    Bias as a layer. It is used for serialization purposes: `tf.keras.Model.save_weights` stores on a per-layer basis,
    so all weights have to be registered in a layer.
    """

    def __init__(self, shape, initializer, trainable, name, **kwargs):
        super().__init__(name=name, **kwargs)
        # Note: the name of this variable will NOT be scoped when serialized, i.e. it will not be in the format of
        # "outer_layer/inner_layer/.../name:0". Instead, it will be "name:0". For further details, see:
        # https://github.com/huggingface/transformers/pull/18833#issuecomment-1233090214
        self.bias = self.add_weight(name=name, shape=shape, initializer=initializer, trainable=trainable)

    def call(self, x):
        return x + self.bias


@add_start_docstrings(
    "The BLENDERBOT Model with a language modeling head. Can be used for summarization.",
    BLENDERBOT_START_DOCSTRING,
)
class TFBlenderbotForConditionalGeneration(TFBlenderbotPreTrainedModel, TFCausalLanguageModelingLoss):
    _keys_to_ignore_on_load_unexpected = [
        r"model.encoder.embed_tokens.weight",
        r"model.decoder.embed_tokens.weight",
    ]

    def __init__(self, config, *inputs, **kwargs):
        super().__init__(config, *inputs, **kwargs)
        self.model = TFBlenderbotMainLayer(config, name="model")
        self.use_cache = config.use_cache
        # final_bias_logits is registered as a buffer in pytorch, so not trainable for the sake of consistency.
        self.bias_layer = BiasLayer(
            name="final_logits_bias", shape=[1, config.vocab_size], initializer="zeros", trainable=False
        )

    def get_decoder(self):
        return self.model.decoder

    def get_encoder(self):
        return self.model.encoder

    def get_output_embeddings(self):
        return self.get_input_embeddings()

    def set_output_embeddings(self, value):
        self.set_input_embeddings(value)

    def get_bias(self):
        return {"final_logits_bias": self.bias_layer.bias}

    def set_bias(self, value):
        # Replaces the existing layers containing bias for correct (de)serialization.
        vocab_size = value["final_logits_bias"].shape[-1]
        self.bias_layer = BiasLayer(
            name="final_logits_bias", shape=[1, vocab_size], initializer="zeros", trainable=False
        )
        self.bias_layer.bias.assign(value["final_logits_bias"])

    @classmethod
    def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], *model_args, **kwargs):
        if pretrained_model_name_or_path == "facebook/blenderbot-90M":
            from ..blenderbot_small import TFBlenderbotSmallForConditionalGeneration

            warnings.warn(
                "The checkpoint `facebook/blenderbot-90M` is deprecated. In the future, please use the identical"
                " checkpoint `facebook/small_blenderbot-90M` with"
                " `TFBlenderbotSmallForConditionalGeneration.from_pretrained('facebook/small_blenderbot-90M')`"
                " instead.",
                FutureWarning,
            )
            return TFBlenderbotSmallForConditionalGeneration.from_pretrained(pretrained_model_name_or_path)

        return super().from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)

    @unpack_inputs
    @add_start_docstrings_to_model_forward(BLENDERBOT_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=TFSeq2SeqLMOutput, config_class=_CONFIG_FOR_DOC)
    @add_end_docstrings(BLENDERBOT_GENERATION_EXAMPLE)
    def call(
        self,
        input_ids: tf.Tensor | None = None,
        attention_mask: tf.Tensor | None = None,
        decoder_input_ids: tf.Tensor | None = None,
        decoder_attention_mask: tf.Tensor | None = None,
        decoder_position_ids: tf.Tensor | None = None,
        head_mask: tf.Tensor | None = None,
        decoder_head_mask: tf.Tensor | None = None,
        cross_attn_head_mask: tf.Tensor | None = None,
        encoder_outputs: Optional[Union[Tuple, TFBaseModelOutput]] = None,
        past_key_values: List[tf.Tensor] | None = None,
        inputs_embeds: tf.Tensor | None = None,
        decoder_inputs_embeds: tf.Tensor | None = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        labels: tf.Tensor | None = None,
        training: Optional[bool] = False,
    ) -> Union[Tuple[tf.Tensor], TFSeq2SeqLMOutput]:
        r"""
        labels (`tf.tensor` of shape `(batch_size, sequence_length)`, *optional*):
            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.

        Returns:

        """
        if labels is not None:
            labels = tf.where(
                labels == self.config.pad_token_id,
                tf.cast(tf.fill(shape_list(labels), -100), labels.dtype),
                labels,
            )
            use_cache = False
            if decoder_input_ids is None and decoder_inputs_embeds is None:
                decoder_input_ids = shift_tokens_right(
                    labels, self.config.pad_token_id, self.config.decoder_start_token_id
                )

        outputs = self.model(
            input_ids,
            attention_mask=attention_mask,
            decoder_input_ids=decoder_input_ids,
            encoder_outputs=encoder_outputs,
            decoder_attention_mask=decoder_attention_mask,
            decoder_position_ids=decoder_position_ids,
            head_mask=head_mask,
            decoder_head_mask=decoder_head_mask,
            cross_attn_head_mask=cross_attn_head_mask,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            decoder_inputs_embeds=decoder_inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            training=training,
        )
        lm_logits = tf.matmul(outputs[0], self.model.shared.weights, transpose_b=True)
        lm_logits = self.bias_layer(lm_logits)
        masked_lm_loss = None if labels is None else self.hf_compute_loss(labels, lm_logits)

        if not return_dict:
            output = (lm_logits,) + outputs[1:]
            return ((masked_lm_loss,) + output) if masked_lm_loss is not None else output
        return TFSeq2SeqLMOutput(
            loss=masked_lm_loss,
            logits=lm_logits,
            past_key_values=outputs.past_key_values,  # index 1 of d outputs
            decoder_hidden_states=outputs.decoder_hidden_states,  # index 2 of d outputs
            decoder_attentions=outputs.decoder_attentions,  # index 3 of d outputs
            cross_attentions=outputs.cross_attentions,  # index 4 of d outputs
            encoder_last_hidden_state=outputs.encoder_last_hidden_state,  # index 0 of encoder outputs
            encoder_hidden_states=outputs.encoder_hidden_states,  # 1 of e out
            encoder_attentions=outputs.encoder_attentions,  # 2 of e out
        )

    # Copied from transformers.models.bart.modeling_tf_bart.TFBartForConditionalGeneration.serving_output
    def serving_output(self, output):
        pkv = tf.tuple(output.past_key_values)[1] if self.config.use_cache else None
        dec_hs = tf.convert_to_tensor(output.decoder_hidden_states) if self.config.output_hidden_states else None
        dec_attns = tf.convert_to_tensor(output.decoder_attentions) if self.config.output_attentions else None
        cross_attns = tf.convert_to_tensor(output.cross_attentions) if self.config.output_attentions else None
        enc_hs = tf.convert_to_tensor(output.encoder_hidden_states) if self.config.output_hidden_states else None
        enc_attns = tf.convert_to_tensor(output.encoder_attentions) if self.config.output_attentions else None

        return TFSeq2SeqLMOutput(
            logits=output.logits,
            past_key_values=pkv,
            decoder_hidden_states=dec_hs,
            decoder_attentions=dec_attns,
            cross_attentions=cross_attns,
            encoder_last_hidden_state=output.encoder_last_hidden_state,
            encoder_hidden_states=enc_hs,
            encoder_attentions=enc_attns,
        )

    # Copied from transformers.models.bart.modeling_tf_bart.TFBartForConditionalGeneration.prepare_inputs_for_generation
    def prepare_inputs_for_generation(
        self,
        decoder_input_ids,
        past_key_values=None,
        attention_mask=None,
        decoder_attention_mask=None,
        head_mask=None,
        decoder_head_mask=None,
        cross_attn_head_mask=None,
        use_cache=None,
        encoder_outputs=None,
        **kwargs,
    ):
        # cut decoder_input_ids if past_key_values is used
        if past_key_values is not None:
            decoder_input_ids = decoder_input_ids[:, -1:]

        if decoder_attention_mask is not None:  # xla
            decoder_position_ids = tf.math.cumsum(decoder_attention_mask, axis=-1, exclusive=True)[:, -1:]
        elif past_key_values is not None:  # no xla + past_key_values
            decoder_position_ids = past_key_values[0][0].shape[2]
        else:  # no xla + no past_key_values
            decoder_position_ids = tf.range(decoder_input_ids.shape[1])

        return {
            "input_ids": None,  # encoder_outputs is defined. input_ids not needed
            "encoder_outputs": encoder_outputs,
            "past_key_values": past_key_values,
            "decoder_input_ids": decoder_input_ids,
            "attention_mask": attention_mask,
            "decoder_attention_mask": decoder_attention_mask,
            "decoder_position_ids": decoder_position_ids,
            "head_mask": head_mask,
            "decoder_head_mask": decoder_head_mask,
            "cross_attn_head_mask": cross_attn_head_mask,
            "use_cache": use_cache,  # change this to avoid caching (presumably for debugging)
        }