File size: 39,072 Bytes
2e3ebcb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1c61b96
2e3ebcb
95b4916
2e3ebcb
95b4916
 
 
2e3ebcb
 
 
 
 
95b4916
 
 
2e3ebcb
 
 
 
 
95b4916
2e3ebcb
95b4916
2e3ebcb
 
77af1c7
2e3ebcb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
77af1c7
 
 
2e3ebcb
77af1c7
 
 
 
 
 
2e3ebcb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
95b4916
 
2e3ebcb
 
 
 
 
1c61b96
 
 
 
 
 
 
 
 
2e3ebcb
 
 
 
 
 
 
 
77af1c7
 
 
2e3ebcb
 
1c61b96
 
 
 
 
77af1c7
1c61b96
 
 
2e3ebcb
 
 
 
 
 
 
 
 
 
1c61b96
 
 
 
 
77af1c7
1c61b96
 
 
2e3ebcb
 
 
1c61b96
 
 
 
 
77af1c7
1c61b96
 
 
2e3ebcb
 
 
 
77af1c7
 
 
2e3ebcb
77af1c7
 
2e3ebcb
 
 
 
 
77af1c7
 
2e3ebcb
 
 
 
 
 
 
 
 
 
 
 
1c61b96
 
 
 
 
77af1c7
1c61b96
 
77af1c7
 
 
2e3ebcb
 
 
95b4916
2e3ebcb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
95b4916
2e3ebcb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
77af1c7
 
 
 
2e3ebcb
 
 
 
95b4916
2e3ebcb
 
 
 
 
 
 
95b4916
2e3ebcb
 
 
 
 
 
 
 
 
 
 
95b4916
2e3ebcb
 
95b4916
2e3ebcb
 
 
 
 
 
 
 
95b4916
2e3ebcb
 
 
77af1c7
95b4916
 
2e3ebcb
 
 
95b4916
2e3ebcb
 
 
95b4916
 
2e3ebcb
 
 
 
 
 
 
 
 
77af1c7
 
 
 
 
 
2e3ebcb
95b4916
2e3ebcb
 
 
 
 
 
 
 
95b4916
 
2e3ebcb
 
 
 
 
 
 
 
 
 
95b4916
 
2e3ebcb
95b4916
2e3ebcb
 
 
 
95b4916
 
 
 
77af1c7
 
 
 
95b4916
77af1c7
 
 
95b4916
2e3ebcb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
77af1c7
 
 
2e3ebcb
 
 
 
 
77af1c7
 
 
2e3ebcb
 
 
77af1c7
 
 
2e3ebcb
95b4916
 
 
2e3ebcb
 
 
 
 
 
95b4916
 
 
 
2e3ebcb
 
95b4916
 
 
 
 
 
 
 
2e3ebcb
 
95b4916
 
 
 
 
 
 
 
 
 
2e3ebcb
 
 
95b4916
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2e3ebcb
77af1c7
 
 
2e3ebcb
95b4916
2e3ebcb
95b4916
2e3ebcb
95b4916
 
 
 
 
 
 
 
2e3ebcb
95b4916
 
 
 
 
 
 
 
77af1c7
 
 
95b4916
 
 
77af1c7
 
 
95b4916
 
 
 
 
 
2e3ebcb
 
 
95b4916
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2e3ebcb
 
 
 
 
 
 
 
 
 
 
77af1c7
 
 
2e3ebcb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
77af1c7
 
 
2e3ebcb
 
77af1c7
 
 
2e3ebcb
77af1c7
 
 
2e3ebcb
 
 
 
 
 
 
 
 
 
 
 
 
77af1c7
 
 
2e3ebcb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
77af1c7
 
 
 
 
 
2e3ebcb
 
 
 
 
 
77af1c7
 
 
 
 
 
2e3ebcb
 
77af1c7
 
 
 
 
 
 
 
 
 
 
 
2e3ebcb
 
 
 
 
77af1c7
 
 
 
 
 
 
 
 
2e3ebcb
 
 
 
77af1c7
 
 
2e3ebcb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
77af1c7
 
 
2e3ebcb
 
 
 
 
 
77af1c7
 
 
2e3ebcb
 
 
 
 
 
 
77af1c7
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
# This implementation was adopted from https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/models/bert.py
# Commit id: abbc1311731867310635f9edc2a9ec18317c8c48

# Copyright (c) 2022, Tri Dao.
# This BERT implementation is based on our MLPerf 2.0 and MLPerf 2.1 BERT implementation.
# https://github.com/mlcommons/training_results_v2.0/blob/main/HazyResearch/benchmarks/bert/implementations/pytorch/modeling.py
# https://github.com/mlcommons/training_results_v2.1/blob/main/Azure-HazyResearch/benchmarks/bert/implementations/ND96amsr_A100_v4/modeling.py

# Inspired by https://github.com/huggingface/transformers/blob/main/src/transformers/models/bert/modeling_bert.py

import logging
import re
from collections import OrderedDict
from collections.abc import Sequence
from functools import partial

import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.utils.checkpoint
from einops import rearrange
from transformers import PretrainedConfig
from transformers.modeling_utils import PreTrainedModel
from transformers.modeling_outputs import MaskedLMOutput
from transformers.models.xlm_roberta.modeling_xlm_roberta import XLMRobertaLMHead

from transformers.models.bert.modeling_bert import (
    BaseModelOutputWithPoolingAndCrossAttentions,
    BertForPreTrainingOutput,
)

from typing import Optional, Tuple, Union

from .xlm_padding import (
    index_first_axis,
    index_first_axis_residual,
    pad_input,
    unpad_input,
)
from .configuration_xlm_roberta import XLMRobertaFlashConfig
from .block import Block
from .embedding import XLMRobertaEmbeddings
from .mha import MHA
from .mlp import FusedMLP, Mlp
from .stochastic_depth import StochasticDepth


try:
    from flash_attn.ops.fused_dense import FusedDense
except ImportError:
    FusedDense = None

try:
    from flash_attn.ops.triton.layer_norm import layer_norm_fn
except ImportError:
    layer_norm_fn = None


try:
    from flash_attn.losses.cross_entropy import CrossEntropyLoss
except ImportError:
    CrossEntropyLoss = None


logger = logging.getLogger(__name__)


def create_mixer_cls(config, cross_attn=False, return_residual=False):
    use_flash_attn = getattr(config, "use_flash_attn", False)
    fused_bias_fc = getattr(config, "fused_bias_fc", False)
    rotary_kwargs = {}
    if config.position_embedding_type == "rotary":
        rotary_kwargs["rotary_emb_dim"] = getattr(
            config, "rotary_emb_dim", config.hidden_size
        )
        rotary_kwargs["rotary_emb_base"] = getattr(config, "rotary_emb_base", 10000.0)
        rotary_kwargs["rotary_emb_scale_base"] = getattr(
            config, "rotary_emb_scale_base", None
        )
        rotary_kwargs["rotary_emb_interleaved"] = getattr(
            config, "rotary_emb_interleaved", False
        )
    mixer_cls = partial(
        MHA,
        num_heads=config.num_attention_heads,
        cross_attn=cross_attn,
        dropout=config.attention_probs_dropout_prob,
        causal=False,
        fused_bias_fc=fused_bias_fc,
        use_flash_attn=use_flash_attn,
        return_residual=return_residual,
        **rotary_kwargs,
    )
    return mixer_cls


def create_mlp_cls(config, layer_idx=None, return_residual=False):
    inner_dim = config.intermediate_size
    fused_mlp = getattr(config, "fused_mlp", False)
    if fused_mlp:
        assert config.hidden_act in ["gelu_new", "gelu_fast", "gelu_pytorch_tanh"], (
            "fused_mlp only " "supports approximate gelu"
        )
    if not fused_mlp:
        approximate = (
            "tanh"
            if config.hidden_act in ["gelu_new", "gelu_fast", "gelu_pytorch_tanh"]
            else "none"
        )
        mlp_cls = partial(
            Mlp,
            hidden_features=inner_dim,
            activation=partial(F.gelu, approximate=approximate),
            return_residual=return_residual,
        )
    else:
        if FusedMLP is None:
            raise ImportError("fused_dense is not installed")
        mlp_checkpoint_lvl = getattr(config, "mlp_checkpoint_lvl", 0)
        # mlp_checkpoint_lvl could be a list, which contains the checkpoint_lvl for each layer
        if isinstance(mlp_checkpoint_lvl, Sequence):
            assert layer_idx is not None
            mlp_checkpoint_lvl = mlp_checkpoint_lvl[layer_idx]
        mlp_cls = partial(
            FusedMLP,
            hidden_features=inner_dim,
            checkpoint_lvl=mlp_checkpoint_lvl,
            return_residual=return_residual,
        )
    return mlp_cls


def create_block(config, layer_idx=None):
    last_layer_subset = getattr(config, "last_layer_subset", False)
    cross_attn = last_layer_subset and layer_idx == config.num_hidden_layers - 1
    # TD [2022-12-19]: For cross attention (last layer), we actually want to return the
    # residual x_kv, not residual x. But it's annoying to change the API (and it only affects
    # one layer) so we just choose not to return residual in this case.
    return_residual = not cross_attn
    mixer_cls = create_mixer_cls(config, cross_attn, return_residual=return_residual)
    mlp_cls = create_mlp_cls(config, layer_idx, return_residual=return_residual)
    norm_cls = partial(nn.LayerNorm, eps=config.layer_norm_eps)
    block = Block(
        config.hidden_size,
        mixer_cls,
        mlp_cls,
        norm_cls=norm_cls,
        prenorm=False,
        resid_dropout1=config.hidden_dropout_prob,
        resid_dropout2=config.hidden_dropout_prob,
        fused_dropout_add_ln=getattr(config, "fused_dropout_add_ln", False),
        return_residual=return_residual,
    )
    return block


# https://github.com/huggingface/transformers/blob/7032e0203262ebb2ebf55da8d2e01f873973e835/src/transformers/models/bert/modeling_bert.py#L748
def _init_weights(module, initializer_range=0.02):
    if isinstance(module, nn.Linear):
        nn.init.normal_(module.weight, std=initializer_range)
        if module.bias is not None:
            nn.init.zeros_(module.bias)
    elif isinstance(module, nn.Embedding):
        nn.init.normal_(module.weight, std=initializer_range)
        if module.padding_idx is not None:
            nn.init.zeros_(module.weight[module.padding_idx])


class XLMRobertaEncoder(nn.Module):
    def __init__(self, config: XLMRobertaFlashConfig):
        super().__init__()
        self.use_flash_attn = getattr(config, "use_flash_attn", False)
        self.layers = nn.ModuleList(
            [create_block(config, layer_idx=i) for i in range(config.num_hidden_layers)]
        )
        self._grad_checkpointing = False

    @property
    def gradient_checkpointing(self):
        return self._grad_checkpointing

    @gradient_checkpointing.setter
    def gradient_checkpointing(self, value):
        self._grad_checkpointing = value

    def forward(self, hidden_states, key_padding_mask=None, subset_mask=None):
        """If subset_mask is not None, we only want output for the subset of the sequence.
        This means that we only compute the last layer output for these tokens.
        subset_mask: (batch, seqlen), dtype=torch.bool
        """
        if key_padding_mask is None or not self.use_flash_attn:
            mixer_kwargs = (
                {"key_padding_mask": key_padding_mask.bool()}
                if key_padding_mask is not None
                else None
            )
            for layer in self.layers:
                if self._grad_checkpointing:
                    hidden_states = torch.utils.checkpoint.checkpoint(
                        layer,
                        hidden_states,
                        use_reentrant=False,
                        mixer_kwargs=mixer_kwargs,
                    )
                else:
                    hidden_states = layer(hidden_states, mixer_kwargs=mixer_kwargs)
            if subset_mask is not None:
                hidden_states = hidden_states[subset_mask]
        else:
            batch, seqlen = hidden_states.shape[:2]
            hidden_states, indices, cu_seqlens, max_seqlen_in_batch = unpad_input(
                hidden_states, key_padding_mask
            )
            mixer_kwargs = {"cu_seqlens": cu_seqlens, "max_seqlen": max_seqlen_in_batch}
            if subset_mask is None:
                for layer in self.layers:
                    if self._grad_checkpointing:
                        hidden_states = torch.utils.checkpoint.checkpoint(
                            layer,
                            hidden_states,
                            use_reentrant=False,
                            mixer_kwargs=mixer_kwargs,
                        )
                    else:
                        hidden_states = layer(hidden_states, mixer_kwargs=mixer_kwargs)
                hidden_states = pad_input(hidden_states, indices, batch, seqlen)
            else:
                for layer in self.layers[:-1]:
                    if self._grad_checkpointing:
                        hidden_states = torch.utils.checkpoint.checkpoint(
                            layer,
                            hidden_states,
                            use_reentrant=False,
                            mixer_kwargs=mixer_kwargs,
                        )
                    else:
                        hidden_states = layer(hidden_states, mixer_kwargs=mixer_kwargs)
                if key_padding_mask is not None:
                    subset_idx = torch.nonzero(
                        subset_mask[key_padding_mask], as_tuple=False
                    ).flatten()
                    subset_seqlens = (subset_mask & key_padding_mask).sum(
                        dim=-1, dtype=torch.int32
                    )
                    subset_cu_seqlens = F.pad(
                        torch.cumsum(subset_seqlens, dim=0, dtype=torch.torch.int32),
                        (1, 0),
                    )
                else:
                    subset_idx = torch.nonzero(subset_mask, as_tuple=False).flatten()
                    subset_seqlens = subset_mask.sum(dim=-1, dtype=torch.int32)
                    subset_cu_seqlens = F.pad(
                        torch.cumsum(subset_seqlens, dim=0, dtype=torch.torch.int32),
                        (1, 0),
                    )
                hidden_states_subset, hidden_states = index_first_axis_residual(
                    hidden_states, subset_idx
                )
                # It's ok to set max_seqlen_q to be much larger
                mixer_kwargs = {
                    "x_kv": hidden_states,
                    "cu_seqlens": subset_cu_seqlens,
                    "max_seqlen": max_seqlen_in_batch,
                    "cu_seqlens_k": cu_seqlens,
                    "max_seqlen_k": max_seqlen_in_batch,
                }
                if self._grad_checkpointing:
                    torch.utils.checkpoint.checkpoint(
                        self.layers[-1],
                        hidden_states_subset,
                        use_reentrant=False,
                        mixer_kwargs=mixer_kwargs,
                    )
                else:
                    hidden_states = self.layers[-1](
                        hidden_states_subset, mixer_kwargs=mixer_kwargs
                    )
        return hidden_states


class XLMRobertaPooler(nn.Module):
    def __init__(self, config):
        super().__init__()
        fused_bias_fc = getattr(config, "fused_bias_fc", False)
        if fused_bias_fc and FusedDense is None:
            raise ImportError("fused_dense is not installed")
        linear_cls = nn.Linear if not fused_bias_fc else FusedDense
        self.dense = linear_cls(config.hidden_size, config.hidden_size)
        self.activation = nn.Tanh()

    def forward(self, hidden_states, pool=True):
        # We "pool" the model by simply taking the hidden state corresponding
        # to the first token.
        first_token_tensor = hidden_states[:, 0] if pool else hidden_states
        pooled_output = self.dense(first_token_tensor)
        pooled_output = self.activation(pooled_output)
        return pooled_output


class XLMRobertaPredictionHeadTransform(nn.Module):
    def __init__(self, config):
        super().__init__()
        fused_bias_fc = getattr(config, "fused_bias_fc", False)
        if fused_bias_fc and FusedDense is None:
            raise ImportError("fused_dense is not installed")
        self.fused_dropout_add_ln = getattr(config, "fused_dropout_add_ln", False)
        if self.fused_dropout_add_ln and layer_norm_fn is None:
            raise ImportError("Triton is not installed")
        linear_cls = nn.Linear if not fused_bias_fc else FusedDense
        self.dense = linear_cls(config.hidden_size, config.hidden_size)
        approximate = (
            "tanh"
            if config.hidden_act in ["gelu_new", "gelu_fast", "gelu_pytorch_tanh"]
            else "none"
        )
        self.transform_act_fn = nn.GELU(approximate=approximate)
        self.layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        hidden_states = self.dense(hidden_states)
        hidden_states = self.transform_act_fn(hidden_states)
        if not self.fused_dropout_add_ln:
            hidden_states = self.layer_norm(hidden_states)
        else:
            hidden_states = layer_norm_fn(
                hidden_states,
                self.layer_norm.weight,
                self.layer_norm.bias,
                eps=self.layer_norm.eps,
            )
        return hidden_states


class XLMRobertaLMPredictionHead(nn.Module):
    def __init__(self, config):
        super().__init__()
        fused_bias_fc = getattr(config, "fused_bias_fc", False)
        if fused_bias_fc and FusedDense is None:
            raise ImportError("fused_dense is not installed")
        linear_cls = nn.Linear if not fused_bias_fc else FusedDense

        self.transform = XLMRobertaPredictionHeadTransform(config)

        # The output weights are the same as the input embeddings, but there is
        # an output-only bias for each token.
        self.decoder = linear_cls(config.hidden_size, config.vocab_size, bias=True)

    def forward(self, hidden_states):
        hidden_states = self.transform(hidden_states)
        hidden_states = self.decoder(hidden_states)
        return hidden_states


class XLMRobertaPreTrainingHeads(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.predictions = XLMRobertaLMPredictionHead(config)
        self.seq_relationship = nn.Linear(config.hidden_size, 2)

    def forward(self, sequence_output, pooled_output):
        prediction_scores = self.predictions(sequence_output)
        seq_relationship_score = self.seq_relationship(pooled_output)
        return prediction_scores, seq_relationship_score


class XLMRobertaPreTrainedModel(PreTrainedModel):
    """An abstract class to handle weights initialization and
    a simple interface for dowloading and loading pretrained models.
    """

    config_class = XLMRobertaFlashConfig
    base_model_prefix = "roberta"
    supports_gradient_checkpointing = True

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


class XLMRobertaModel(XLMRobertaPreTrainedModel):
    def __init__(self, config: XLMRobertaFlashConfig, add_pooling_layer=True):
        super().__init__(config)
        self.pad_vocab_size_multiple = getattr(config, "pad_vocab_size_multiple", 1)
        if config.vocab_size % self.pad_vocab_size_multiple != 0:
            config.vocab_size += self.pad_vocab_size_multiple - (
                config.vocab_size % self.pad_vocab_size_multiple
            )
        self.fused_dropout_add_ln = getattr(config, "fused_dropout_add_ln", False)
        if self.fused_dropout_add_ln and layer_norm_fn is None:
            raise ImportError("Triton is not installed")
        assert config.hidden_act in [
            "gelu",
            "gelu_new",
            "gelu_fast",
            "gelu_pytorch_tanh",
        ]

        self.embeddings = XLMRobertaEmbeddings(
            config.hidden_size,
            config.vocab_size,
            config.max_position_embeddings,
            config.type_vocab_size,
            padding_idx=config.pad_token_id,
        )
        self.emb_drop = nn.Dropout(config.hidden_dropout_prob)
        self.emb_ln = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.encoder = XLMRobertaEncoder(config)
        self.pooler = XLMRobertaPooler(config) if add_pooling_layer else None

        self.apply(partial(_init_weights, initializer_range=config.initializer_range))

    def forward(
        self,
        input_ids,
        position_ids=None,
        token_type_ids=None,
        attention_mask=None,
        masked_tokens_mask=None,
        return_dict=None,
        **kwargs,
    ):
        """If masked_tokens_mask is not None (i.e. last_layer_subset == True in XLMForPreTraining),
        we only want the output for the masked tokens. This means that we only compute the last
        layer output for these tokens.
        masked_tokens_mask: (batch, seqlen), dtype=torch.bool
        """

        if kwargs:
            for key, value in kwargs.items():
                if value is not None:
                    logger.warning(
                        'Flash attention implementation does not support kwargs: %s',
                        key,
                    )

        return_dict = (
            return_dict if return_dict is not None else self.config.use_return_dict
        )

        hidden_states = self.embeddings(
            input_ids, position_ids=position_ids, token_type_ids=token_type_ids
        )
        # TD [2022-12:18]: Don't need to force residual in fp32
        # BERT puts embedding LayerNorm before embedding dropout.
        if not self.fused_dropout_add_ln:
            hidden_states = self.emb_ln(hidden_states)
        else:
            hidden_states = layer_norm_fn(
                hidden_states, self.emb_ln.weight, self.emb_ln.bias, eps=self.emb_ln.eps
            )
        hidden_states = self.emb_drop(hidden_states)

        if masked_tokens_mask is not None:
            batch_size, seqlen = input_ids.shape[:2]
            # We also need the first column for the CLS token
            first_col_mask = torch.zeros(
                batch_size, seqlen, dtype=torch.bool, device=input_ids.device
            )
            first_col_mask[:, 0] = True
            subset_mask = masked_tokens_mask | first_col_mask
        else:
            subset_mask = None

        sequence_output = self.encoder(
            hidden_states, key_padding_mask=attention_mask, subset_mask=subset_mask
        )

        if masked_tokens_mask is None:
            pooled_output = (
                self.pooler(sequence_output) if self.pooler is not None else None
            )
        else:
            # TD [2022-03-01]: the indexing here is very tricky.
            if attention_mask is not None:
                subset_idx = subset_mask[attention_mask]
                pool_input = sequence_output[first_col_mask[attention_mask][subset_idx]]
                sequence_output = sequence_output[
                    masked_tokens_mask[attention_mask][subset_idx]
                ]
            else:
                pool_input = sequence_output[first_col_mask[subset_mask]]
                sequence_output = sequence_output[masked_tokens_mask[subset_mask]]
            pooled_output = (
                self.pooler(pool_input, pool=False) if self.pooler is not None else None
            )

        if not return_dict:
            return sequence_output, pooled_output

        return BaseModelOutputWithPoolingAndCrossAttentions(
            last_hidden_state=sequence_output,
            pooler_output=pooled_output,
        )


class XLMRobertaForMaskedLM(XLMRobertaPreTrainedModel):
    _tied_weights_keys = ["lm_head.decoder.weight", "lm_head.decoder.bias"]

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

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

        self.roberta = XLMRobertaModel(config, add_pooling_layer=False)
        self.lm_head = XLMRobertaLMHead(config)

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

    def get_input_embeddings(self):
        return self.roberta.embeddings.word_embeddings

    def get_output_embeddings(self):
        return self.lm_head.decoder

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

    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = 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,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple[torch.Tensor], MaskedLMOutput]:
        r"""
        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Labels for computing the masked language modeling loss. Indices should be in `[-100, 0, ...,
            config.vocab_size]` (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]`
        kwargs (`Dict[str, any]`, optional, defaults to *{}*):
            Used to hide legacy arguments that have been deprecated.
        """
        return_dict = (
            return_dict if return_dict is not None else self.config.use_return_dict
        )

        outputs = self.roberta(
            input_ids,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
            encoder_hidden_states=encoder_hidden_states,
            encoder_attention_mask=encoder_attention_mask,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )
        sequence_output = outputs[0]
        prediction_scores = self.lm_head(sequence_output)

        masked_lm_loss = None
        if labels is not None:
            # move labels to correct device to enable model parallelism
            labels = labels.to(prediction_scores.device)
            loss_fct = CrossEntropyLoss()
            masked_lm_loss = loss_fct(
                prediction_scores.view(-1, self.config.vocab_size), labels.view(-1)
            )

        if not return_dict:
            output = (prediction_scores,) + outputs[2:]
            return (
                ((masked_lm_loss,) + output) if masked_lm_loss is not None else output
            )

        return MaskedLMOutput(
            loss=masked_lm_loss,
            logits=prediction_scores,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )


# class XLMRobertaForPreTraining(XLMRobertaPreTrainedModel):
#     def __init__(self, config: XLMRobertaFlashConfig):
#         super().__init__(config)
#         # If dense_seq_output, we only need to pass the hidden states for the masked out tokens
#         # (around 15%) to the classifier heads.
#         self.dense_seq_output = getattr(config, "dense_seq_output", False)
#         # If last_layer_subset, we only need the compute the last layer for a subset of tokens
#         # (e.g., the tokens we need to compute the masked LM loss and the next-sentence prediction).
#         self.last_layer_subset = getattr(config, "last_layer_subset", False)
#         if self.last_layer_subset:
#             assert self.dense_seq_output, "last_layer_subset requires dense_seq_output"
#         use_xentropy = getattr(config, "use_xentropy", False)
#         if use_xentropy and CrossEntropyLoss is None:
#             raise ImportError("xentropy_cuda is not installed")
#         loss_cls = (
#             nn.CrossEntropyLoss
#             if not use_xentropy
#             else partial(CrossEntropyLoss, inplace_backward=True)
#         )
#
#         self.xlm = XLMRobertaModel(config)
#         self.cls = XLMRobertaPreTrainingHeads(config)
#         self.mlm_loss = loss_cls(ignore_index=0)
#         self.nsp_loss = loss_cls(ignore_index=-1)
#
#         # Initialize weights and apply final processing
#         self.apply(partial(_init_weights, initializer_range=config.initializer_range))
#         self.tie_weights()
#
#     def tie_weights(self):
#         self.cls.predictions.decoder.weight = self.xlm.embeddings.word_embeddings.weight
#
#     def forward(
#         self,
#         input_ids,
#         position_ids=None,
#         token_type_ids=None,
#         attention_mask=None,
#         labels=None,
#         next_sentence_label=None,
#     ):
#         """
#         If labels are provided, they must be 0 for masked out tokens (as specified in the attention
#         mask).
#         Outputs:
#             if `labels` and `next_sentence_label` are not `None`:
#                 Outputs the total_loss which is the sum of the masked language modeling loss and the next
#                 sentence classification loss.
#             if `labels` or `next_sentence_label` is `None`:
#                 Outputs a tuple comprising
#                 - the masked language modeling logits of shape [batch_size, sequence_length, vocab_size], and
#                 - the next sentence classification logits of shape [batch_size, 2].
#
#         """
#         masked_tokens_mask = labels > 0 if (self.last_layer_subset and labels is not None) else None
#         outputs = self.xlm(
#             input_ids,
#             position_ids=position_ids,
#             token_type_ids=token_type_ids,
#             attention_mask=attention_mask.bool() if attention_mask is not None else None,
#             masked_tokens_mask=masked_tokens_mask,
#         )
#         sequence_output, pooled_output = outputs.last_hidden_state, outputs.pooler_output
#         if self.dense_seq_output and labels is not None:
#             masked_token_idx = torch.nonzero(labels.flatten() > 0, as_tuple=False).flatten()
#             if not self.last_layer_subset:
#                 sequence_output = index_first_axis(
#                     rearrange(sequence_output, "b s d -> (b s) d"), masked_token_idx
#                 )
#         prediction_scores, seq_relationship_score = self.cls(sequence_output, pooled_output)
#
#         total_loss = None
#         if labels is not None and next_sentence_label is not None:
#             if (
#                 self.dense_seq_output and labels is not None
#             ):  # prediction_scores are already flattened
#                 masked_lm_loss = self.mlm_loss(
#                     prediction_scores, labels.flatten()[masked_token_idx]
#                 )
#             else:
#                 masked_lm_loss = self.mlm_loss(
#                     rearrange(prediction_scores, "... v -> (...) v"),
#                     rearrange(labels, "... -> (...)"),
#                 )
#             next_sentence_loss = self.nsp_loss(
#                 rearrange(seq_relationship_score, "... t -> (...) t"),
#                 rearrange(next_sentence_label, "... -> (...)"),
#             )
#             total_loss = masked_lm_loss.float() + next_sentence_loss.float()
#
#         return BertForPreTrainingOutput(
#             loss=total_loss,
#             prediction_logits=prediction_scores,
#             seq_relationship_logits=seq_relationship_score,
#         )


def remap_state_dict(state_dict, config: PretrainedConfig):
    """
    Map the state_dict of a Huggingface BERT model to be flash_attn compatible.
    """

    # LayerNorm
    def key_mapping_ln_gamma_beta(key):
        key = re.sub(r"LayerNorm.gamma$", "LayerNorm.weight", key)
        key = re.sub(r"LayerNorm.beta$", "LayerNorm.bias", key)
        return key

    state_dict = OrderedDict(
        (key_mapping_ln_gamma_beta(k), v) for k, v in state_dict.items()
    )

    # Layers
    def key_mapping_layers(key):
        return re.sub(r"^bert.encoder.layer.", "bert.encoder.layers.", key)

    state_dict = OrderedDict((key_mapping_layers(k), v) for k, v in state_dict.items())

    # LayerNorm
    def key_mapping_ln(key):
        key = re.sub(r"^bert.embeddings.LayerNorm.", "bert.emb_ln.", key)
        key = re.sub(
            r"^bert.encoder.layers.(\d+).attention.output.LayerNorm.(weight|bias)",
            r"bert.encoder.layers.\1.norm1.\2",
            key,
        )
        key = re.sub(
            r"^bert.encoder.layers.(\d+).output.LayerNorm.(weight|bias)",
            r"bert.encoder.layers.\1.norm2.\2",
            key,
        )
        key = re.sub(
            r"^cls.predictions.transform.LayerNorm.(weight|bias)",
            r"cls.predictions.transform.layer_norm.\1",
            key,
        )
        return key

    state_dict = OrderedDict((key_mapping_ln(k), v) for k, v in state_dict.items())

    # MLP
    def key_mapping_mlp(key):
        key = re.sub(
            r"^bert.encoder.layers.(\d+).intermediate.dense.(weight|bias)",
            r"bert.encoder.layers.\1.mlp.fc1.\2",
            key,
        )
        key = re.sub(
            r"^bert.encoder.layers.(\d+).output.dense.(weight|bias)",
            r"bert.encoder.layers.\1.mlp.fc2.\2",
            key,
        )
        return key

    state_dict = OrderedDict((key_mapping_mlp(k), v) for k, v in state_dict.items())

    # Attention
    last_layer_subset = getattr(config, "last_layer_subset", False)
    for d in range(config.num_hidden_layers):
        Wq = state_dict.pop(f"bert.encoder.layers.{d}.attention.self.query.weight")
        Wk = state_dict.pop(f"bert.encoder.layers.{d}.attention.self.key.weight")
        Wv = state_dict.pop(f"bert.encoder.layers.{d}.attention.self.value.weight")
        bq = state_dict.pop(f"bert.encoder.layers.{d}.attention.self.query.bias")
        bk = state_dict.pop(f"bert.encoder.layers.{d}.attention.self.key.bias")
        bv = state_dict.pop(f"bert.encoder.layers.{d}.attention.self.value.bias")
        if not (last_layer_subset and d == config.num_hidden_layers - 1):
            state_dict[f"bert.encoder.layers.{d}.mixer.Wqkv.weight"] = torch.cat(
                [Wq, Wk, Wv], dim=0
            )
            state_dict[f"bert.encoder.layers.{d}.mixer.Wqkv.bias"] = torch.cat(
                [bq, bk, bv], dim=0
            )
        else:
            state_dict[f"bert.encoder.layers.{d}.mixer.Wq.weight"] = Wq
            state_dict[f"bert.encoder.layers.{d}.mixer.Wkv.weight"] = torch.cat(
                [Wk, Wv], dim=0
            )
            state_dict[f"bert.encoder.layers.{d}.mixer.Wq.bias"] = bq
            state_dict[f"bert.encoder.layers.{d}.mixer.Wkv.bias"] = torch.cat(
                [bk, bv], dim=0
            )

    def key_mapping_attn(key):
        return re.sub(
            r"^bert.encoder.layers.(\d+).attention.output.dense.(weight|bias)",
            r"bert.encoder.layers.\1.mixer.out_proj.\2",
            key,
        )

    state_dict = OrderedDict((key_mapping_attn(k), v) for k, v in state_dict.items())

    def key_mapping_decoder_bias(key):
        return re.sub(r"^cls.predictions.bias", "cls.predictions.decoder.bias", key)

    state_dict = OrderedDict(
        (key_mapping_decoder_bias(k), v) for k, v in state_dict.items()
    )

    # Word embedding
    pad_vocab_size_multiple = getattr(config, "pad_vocab_size_multiple", 1)
    if pad_vocab_size_multiple > 1:
        word_embeddings = state_dict["bert.embeddings.word_embeddings.weight"]
        state_dict["bert.embeddings.word_embeddings.weight"] = F.pad(
            word_embeddings, (0, 0, 0, config.vocab_size - word_embeddings.shape[0])
        )
        decoder_weight = state_dict["cls.predictions.decoder.weight"]
        state_dict["cls.predictions.decoder.weight"] = F.pad(
            decoder_weight, (0, 0, 0, config.vocab_size - decoder_weight.shape[0])
        )
        # If the vocab was padded, we want to set the decoder bias for those padded indices to be
        # strongly negative (i.e. the decoder shouldn't predict those indices).
        # TD [2022-05-09]: I don't think it affects the MLPerf training.
        decoder_bias = state_dict["cls.predictions.decoder.bias"]
        state_dict["cls.predictions.decoder.bias"] = F.pad(
            decoder_bias, (0, config.vocab_size - decoder_bias.shape[0]), value=-100.0
        )

    return state_dict


def inv_remap_state_dict(state_dict, config: PretrainedConfig):
    """
    Map the state_dict of a flash_attn model to be Huggingface BERT compatible.

    This function is meant to be the inverse of remap_state_dict.
    """
    # Word embedding
    pad_vocab_size_multiple = getattr(config, "pad_vocab_size_multiple", 1)
    if pad_vocab_size_multiple > 1:
        word_embeddings = state_dict["bert.embeddings.word_embeddings.weight"]
        decoder_weight = state_dict["cls.predictions.decoder.weight"]
        decoder_bias = state_dict["cls.predictions.decoder.bias"]
        # unpad embeddings
        state_dict["bert.embeddings.word_embeddings.weight"] = word_embeddings[
            : config.orig_vocab_size, :
        ]
        state_dict["cls.predictions.decoder.weight"] = decoder_weight[
            : config.orig_vocab_size, :
        ]
        state_dict["cls.predictions.decoder.bias"] = decoder_bias[
            : config.orig_vocab_size
        ]

    for d in range(config.num_hidden_layers):
        last_layer_subset = getattr(config, "last_layer_subset", False)
        if not last_layer_subset or d != (config.num_hidden_layers - 1):
            Wqkv_weights = state_dict.pop(f"bert.encoder.layers.{d}.mixer.Wqkv.weight")
            Wqkv_biases = state_dict.pop(f"bert.encoder.layers.{d}.mixer.Wqkv.bias")
            state_dict[
                f"bert.encoder.layers.{d}.attention.self.query.weight"
            ] = Wqkv_weights[: Wqkv_weights.shape[0] // 3, :]
            state_dict[
                f"bert.encoder.layers.{d}.attention.self.key.weight"
            ] = Wqkv_weights[
                Wqkv_weights.shape[0] // 3 : 2 * Wqkv_weights.shape[0] // 3, :
            ]
            state_dict[
                f"bert.encoder.layers.{d}.attention.self.value.weight"
            ] = Wqkv_weights[2 * Wqkv_weights.shape[0] // 3 :, :]
            state_dict[
                f"bert.encoder.layers.{d}.attention.self.query.bias"
            ] = Wqkv_biases[: Wqkv_biases.shape[0] // 3]
            state_dict[
                f"bert.encoder.layers.{d}.attention.self.key.bias"
            ] = Wqkv_biases[Wqkv_biases.shape[0] // 3 : 2 * Wqkv_biases.shape[0] // 3]
            state_dict[
                f"bert.encoder.layers.{d}.attention.self.value.bias"
            ] = Wqkv_biases[2 * Wqkv_biases.shape[0] // 3 :]
        else:
            Wq_weight = state_dict.pop(f"bert.encoder.layers.{d}.mixer.Wq.weight")
            Wkv_weights = state_dict.pop(f"bert.encoder.layers.{d}.mixer.Wkv.weight")
            Wq_bias = state_dict.pop(f"bert.encoder.layers.{d}.mixer.Wq.bias")
            Wkv_biases = state_dict.pop(f"bert.encoder.layers.{d}.mixer.Wkv.bias")
            state_dict[
                f"bert.encoder.layers.{d}.attention.self.query.weight"
            ] = Wq_weight
            state_dict[
                f"bert.encoder.layers.{d}.attention.self.key.weight"
            ] = Wkv_weights[: Wkv_weights.shape[0] // 2, :]
            state_dict[
                f"bert.encoder.layers.{d}.attention.self.value.weight"
            ] = Wkv_weights[Wkv_weights.shape[0] // 2 :, :]
            state_dict[f"bert.encoder.layers.{d}.attention.self.query.bias"] = Wq_bias
            state_dict[f"bert.encoder.layers.{d}.attention.self.key.bias"] = Wkv_biases[
                : Wkv_biases.shape[0] // 2
            ]
            state_dict[
                f"bert.encoder.layers.{d}.attention.self.value.bias"
            ] = Wkv_biases[Wkv_biases.shape[0] // 2 :]

    def inv_key_mapping_ln(key):
        key = re.sub(r"bert.emb_ln.", "bert.embeddings.LayerNorm.", key)
        key = re.sub(
            r"bert.encoder.layers.(\d+).norm1.(weight|bias)",
            r"bert.encoder.layers.\1.attention.output.LayerNorm.\2",
            key,
        )
        key = re.sub(
            r"bert.encoder.layers.(\d+).norm2.(weight|bias)",
            r"bert.encoder.layers.\1.output.LayerNorm.\2",
            key,
        )
        key = re.sub(
            r"cls.predictions.transform.layer_norm.(weight|bias)",
            r"cls.predictions.transform.LayerNorm.\1",
            key,
        )
        return key

    def inv_key_mapping_ln_gamma_beta(key):
        key = re.sub(r"LayerNorm.weight$", "LayerNorm.gamma", key)
        key = re.sub(r"LayerNorm.bias$", "LayerNorm.beta", key)
        return key

    def inv_key_mapping_layers(key):
        return re.sub(r"bert.encoder.layers.", "bert.encoder.layer.", key)

    def inv_key_mapping_mlp(key):
        key = re.sub(
            r"bert.encoder.layer.(\d+).mlp.fc1.(weight|bias)",
            r"bert.encoder.layer.\1.intermediate.dense.\2",
            key,
        )
        key = re.sub(
            r"bert.encoder.layer.(\d+).mlp.fc2.(weight|bias)",
            r"bert.encoder.layer.\1.output.dense.\2",
            key,
        )
        return key

    def inv_key_mapping_attn(key):
        return re.sub(
            r"bert.encoder.layer.(\d+).mixer.out_proj.(weight|bias)",
            r"bert.encoder.layer.\1.attention.output.dense.\2",
            key,
        )

    def inv_key_mapping_decoder_bias(key):
        return re.sub(r"cls.predictions.decoder.bias", "cls.predictions.bias", key)

    state_dict = OrderedDict(
        (inv_key_mapping_ln(key), value) for key, value in state_dict.items()
    )
    state_dict = OrderedDict(
        (inv_key_mapping_ln_gamma_beta(key), value) for key, value in state_dict.items()
    )
    state_dict = OrderedDict(
        (inv_key_mapping_layers(key), value) for key, value in state_dict.items()
    )
    state_dict = OrderedDict(
        (inv_key_mapping_mlp(key), value) for key, value in state_dict.items()
    )
    state_dict = OrderedDict(
        (inv_key_mapping_attn(key), value) for key, value in state_dict.items()
    )
    state_dict = OrderedDict(
        (inv_key_mapping_decoder_bias(key), value) for key, value in state_dict.items()
    )

    return state_dict