File size: 28,589 Bytes
f8b62b4
 
 
 
 
 
 
87b642a
 
 
 
 
 
 
 
 
 
32458be
 
87b642a
32458be
87b642a
 
 
 
6fb6577
8c27502
87b642a
 
 
 
5944ec8
3160695
87b642a
 
 
 
3160695
5944ec8
 
 
 
87b642a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
32458be
 
 
 
87b642a
 
 
 
 
463061d
 
 
 
87b642a
 
 
 
 
 
 
d4d5621
87b642a
 
 
463061d
87b642a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3b35eab
87b642a
 
 
 
 
 
 
 
d4d5621
87b642a
 
 
75d7a16
 
 
 
 
 
 
 
 
 
 
87b642a
 
 
 
 
 
d4d5621
87b642a
ca5f516
87b642a
 
 
 
 
 
 
3160695
87b642a
 
 
 
 
 
3160695
87b642a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3160695
87b642a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
80472cb
87b642a
 
 
45b2292
 
75d7a16
 
 
 
 
87b642a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8adf551
87b642a
32458be
 
 
 
 
 
 
 
 
8adf551
 
 
 
 
95ca1a8
 
 
87b642a
 
 
 
 
 
8adf551
87b642a
 
 
 
 
 
 
 
 
 
 
8adf551
 
 
87b642a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
32458be
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
87b642a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
bb281f0
7e06371
bb281f0
87b642a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0f43653
 
 
 
 
 
 
 
 
 
 
 
 
 
87b642a
 
 
0f43653
 
 
 
 
87b642a
 
 
 
 
4c68a4c
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
""" Implementation of BERT, using ALiBi and Flash Attention

The implementation was adopted from
https://github.com/Dao-AILab/flash-attention/blob/43950dda456e095969d842fca7a73c5bfe3cecd0/flash_attn/models/bert.py
and made modifications to use ALiBi.
"""

# 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
from collections.abc import Sequence
from functools import partial
from typing import Union, List, Optional
import warnings

import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from einops import rearrange
from transformers.modeling_utils import PreTrainedModel
from .configuration_bert import JinaBertConfig
from transformers.models.bert.modeling_bert import (
    BaseModelOutputWithPoolingAndCrossAttentions,
    BertForPreTrainingOutput,
)
from .bert_padding import (
    index_first_axis,
    index_first_axis_residual,
    pad_input,
    unpad_input,
)

from .block import Block
from .embedding import BertEmbeddings
from .mha import MHA
from .mlp import FusedMLP, Mlp

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

try:
    from tqdm.autonotebook import trange
except ImportError:
    trange = None

logger = logging.getLogger(__name__)


def create_mixer_cls(config, cross_attn=False, return_residual=False):
    use_flash_attn = config.use_flash_attn
    use_qk_norm = config.use_qk_norm
    fused_bias_fc = config.fused_bias_fc
    window_size = config.window_size
    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,
        use_alibi=True,
        window_size=window_size,
        qk_norm=use_qk_norm
    )
    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) and not getattr(module, "skip_init", False):
        nn.init.normal_(module.weight, std=initializer_range)
        if module.padding_idx is not None:
            nn.init.zeros_(module.weight[module.padding_idx])


class BertEncoder(nn.Module):
    def __init__(self, config: JinaBertConfig):
        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
        for block in self.layers:
            block.mixer.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:
                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:
                    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]:
                    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,
                }
                hidden_states = self.layers[-1](hidden_states_subset, mixer_kwargs=mixer_kwargs)
        return hidden_states


class BertPooler(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 BertPredictionHeadTransform(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 BertLMPredictionHead(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 = BertPredictionHeadTransform(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 BertPreTrainingHeads(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.predictions = BertLMPredictionHead(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 BertPreTrainedModel(PreTrainedModel):
    """An abstract class to handle weights initialization and
    a simple interface for dowloading and loading pretrained models.
    """
    config_class = JinaBertConfig
    base_model_prefix = "bert"
    supports_gradient_checkpointing = True

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


class BertModel(BertPreTrainedModel):
    def __init__(self, config: JinaBertConfig, 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 = BertEmbeddings(
            config.hidden_size,
            config.vocab_size,
            -1,                  # No 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 = BertEncoder(config)
        self.pooler = BertPooler(config) if add_pooling_layer else None
        self.task_type_embeddings = nn.Embedding(config.num_tasks, config.hidden_size)

        self.emb_pooler = config.emb_pooler
        self._name_or_path = config._name_or_path
        if self.emb_pooler is not None:
            from transformers import AutoTokenizer

            self.tokenizer = AutoTokenizer.from_pretrained(config._name_or_path)
        else:
            self.tokenizer = None

        # We now initialize the task embeddings to 0; We do not use task types during
        # pretraining. When we start using task types during embedding training,
        # we want the model to behave exactly as in pretraining (i.e. task types
        # have no effect).
        nn.init.zeros_(self.task_type_embeddings.weight)
        self.task_type_embeddings.skip_init = True
        # The following code should skip the embeddings layer
        self.apply(partial(_init_weights, initializer_range=config.initializer_range))

    def forward(
        self,
        input_ids,
        position_ids=None,
        token_type_ids=None,
        task_type_ids=None,
        attention_mask=None,
        masked_tokens_mask=None,
    ):
        """If masked_tokens_mask is not None (i.e. last_layer_subset == True in BertForPreTraining),
        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
        """
        hidden_states = self.embeddings(
            input_ids, position_ids=position_ids, token_type_ids=token_type_ids
        )
        if task_type_ids is not None:
            hidden_states = hidden_states + self.task_type_embeddings(task_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

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


    @torch.inference_mode()
    def encode(
        self: 'BertModel',
        sentences: Union[str, List[str]],
        batch_size: int = 32,
        show_progress_bar: Optional[bool] = None,
        output_value: str = 'sentence_embedding',
        convert_to_numpy: bool = True,
        convert_to_tensor: bool = False,
        device: Optional[torch.device] = None,
        normalize_embeddings: bool = False,
        **tokenizer_kwargs,
    ) -> Union[List[torch.Tensor], np.ndarray, torch.Tensor]:
        """
        Computes sentence embeddings
        Args:
            sentences(`str` or `List[str]`):
                Sentence or sentences to be encoded
            batch_size(`int`, *optional*, defaults to 32):
                Batch size for the computation
            show_progress_bar(`bool`, *optional*, defaults to None):
                Show a progress bar when encoding sentences.
                If set to None, progress bar is only shown when `logger.level == logging.INFO` or `logger.level == logging.DEBUG`.
            output_value(`str`, *optional*, defaults to 'sentence_embedding'):
                Default sentence_embedding, to get sentence embeddings.
                Can be set to token_embeddings to get wordpiece token embeddings.
                Set to None, to get all output values
            convert_to_numpy(`bool`, *optional*, defaults to True):
                If true, the output is a list of numpy vectors.
                Else, it is a list of pytorch tensors.
            convert_to_tensor(`bool`, *optional*, defaults to False):
                If true, you get one large tensor as return.
                Overwrites any setting from convert_to_numpy
            device(`torch.device`, *optional*, defaults to None):
                Which torch.device to use for the computation
            normalize_embeddings(`bool`, *optional*, defaults to False):
                If set to true, returned vectors will have length 1. In that case, the faster dot-product (util.dot_score) instead of cosine similarity can be used.
            tokenizer_kwargs(`Dict[str, Any]`, *optional*, defaults to {}):
                Keyword arguments for the tokenizer
        Returns:
            By default, a list of tensors is returned.
            If convert_to_tensor, a stacked tensor is returned.
            If convert_to_numpy, a numpy matrix is returned.
        """
        if self.emb_pooler is None:
            warnings.warn("No emb_pooler specified, defaulting to mean pooling.")
            self.emb_pooler = 'mean'
            from transformers import AutoTokenizer

            self.tokenizer = AutoTokenizer.from_pretrained(self._name_or_path)
        if self.emb_pooler != 'mean':
            raise NotImplementedError

        is_training = self.training
        self.eval()

        if show_progress_bar is None:
            show_progress_bar = (
                logger.getEffectiveLevel() == logging.INFO
                or logger.getEffectiveLevel() == logging.DEBUG
            )

        if convert_to_tensor:
            convert_to_numpy = False

        if output_value != 'sentence_embedding':
            convert_to_tensor = False
            convert_to_numpy = False

        input_was_string = False
        if isinstance(sentences, str) or not hasattr(sentences, '__len__'):
            sentences = [sentences]
            input_was_string = True

        if device is not None:
            self.to(device)

        # TODO: Maybe use better length heuristic?
        permutation = np.argsort([-len(i) for i in sentences])
        inverse_permutation = np.argsort(permutation)
        sentences = [sentences[idx] for idx in permutation]

        tokenizer_kwargs['padding'] = tokenizer_kwargs.get('padding', True)
        tokenizer_kwargs['max_length'] = tokenizer_kwargs.get('max_length', 8192)
        tokenizer_kwargs['truncation'] = tokenizer_kwargs.get('truncation', True)

        all_embeddings = []

        if trange is not None:
            range_iter = trange(
                0,
                len(sentences),
                batch_size,
                desc="Encoding",
                disable=not show_progress_bar,
            )
        else:
            range_iter = range(0, len(sentences), batch_size)

        for i in range_iter:
            encoded_input = self.tokenizer(
                sentences[i : i + batch_size],
                return_tensors='pt',
                **tokenizer_kwargs,
            ).to(self.device)
            token_embs = self.forward(**encoded_input)[0]

            # Accumulate in fp32 to avoid overflow
            token_embs = token_embs.float()

            if output_value == 'token_embeddings':
                raise NotImplementedError
            elif output_value is None:
                raise NotImplementedError
            else:
                embeddings = self.mean_pooling(
                    token_embs, encoded_input['attention_mask']
                )

                if normalize_embeddings:
                    embeddings = torch.nn.functional.normalize(embeddings, p=2, dim=1)

                if convert_to_numpy:
                    embeddings = embeddings.cpu()
            all_embeddings.extend(embeddings)

        all_embeddings = [all_embeddings[idx] for idx in inverse_permutation]

        if convert_to_tensor:
            all_embeddings = torch.stack(all_embeddings)
        elif convert_to_numpy:
            all_embeddings = np.asarray([emb.numpy() for emb in all_embeddings])

        if input_was_string:
            all_embeddings = all_embeddings[0]

        self.train(is_training)
        return all_embeddings

    def mean_pooling(
        self, token_embeddings: torch.Tensor, attention_mask: torch.Tensor
    ):
        input_mask_expanded = (
            attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
        )
        return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(
            input_mask_expanded.sum(1), min=1e-9
        )

class BertForPreTraining(BertPreTrainedModel):
    def __init__(self, config: JinaBertConfig):
        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.bert = BertModel(config)
        self.cls = BertPreTrainingHeads(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.bert.embeddings.word_embeddings.weight

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

    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.bert(
            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)

        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]
            ).float()
        elif labels is not None:
            masked_lm_loss = self.mlm_loss(
                rearrange(prediction_scores, "... v -> (...) v"),
                rearrange(labels, "... -> (...)"),
            ).float()
        else:
            masked_lm_loss = 0
        if next_sentence_label is not None:
            next_sentence_loss = self.nsp_loss(
                rearrange(seq_relationship_score, "... t -> (...) t"),
                rearrange(next_sentence_label, "... -> (...)"),
            ).float()
        else:
            next_sentence_loss = 0

        total_loss = masked_lm_loss + next_sentence_loss

        return BertForPreTrainingOutput(
            loss=total_loss,
            prediction_logits=prediction_scores,
            seq_relationship_logits=seq_relationship_score,
        )