File size: 34,324 Bytes
a3055fa
389b01f
a3055fa
 
 
 
 
 
06e4cba
a3055fa
 
 
 
 
 
25d8903
a3055fa
 
 
 
 
 
389b01f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a3055fa
 
 
 
 
 
 
 
 
389b01f
a3055fa
 
 
 
 
 
 
9244048
389b01f
9244048
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a3055fa
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
389b01f
 
a3055fa
389b01f
 
 
 
 
 
 
 
 
 
 
 
a3055fa
 
 
 
 
 
 
 
 
 
 
 
389b01f
a3055fa
 
389b01f
a3055fa
389b01f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a3055fa
389b01f
 
 
a3055fa
389b01f
 
 
 
a3055fa
389b01f
a3055fa
389b01f
 
a3055fa
389b01f
9244048
389b01f
 
 
 
9244048
 
389b01f
 
 
 
a3055fa
389b01f
a3055fa
389b01f
 
 
 
a3055fa
389b01f
 
 
 
 
a3055fa
389b01f
 
a3055fa
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
389b01f
 
 
 
 
a3055fa
 
 
 
 
 
 
 
 
9244048
a3055fa
 
 
 
 
 
 
 
 
 
 
 
 
9244048
a3055fa
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9244048
a3055fa
 
 
 
 
389b01f
 
 
a3055fa
 
389b01f
 
 
a3055fa
 
 
 
 
 
 
 
 
 
 
389b01f
a3055fa
 
 
06e4cba
a3055fa
06e4cba
a3055fa
 
 
 
 
 
 
9244048
a3055fa
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
389b01f
a3055fa
 
06e4cba
1765fef
 
 
a3055fa
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
389b01f
 
 
 
 
 
a3055fa
 
 
 
 
 
 
 
 
9244048
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
389b01f
9244048
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
389b01f
9244048
 
389b01f
9244048
 
 
 
 
 
 
 
 
389b01f
 
 
9244048
 
 
 
 
 
 
 
 
 
 
 
389b01f
 
 
9244048
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a3055fa
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9244048
 
 
 
 
 
a3055fa
 
06e4cba
a3055fa
 
 
06e4cba
 
9244048
a3055fa
06e4cba
a3055fa
 
 
 
 
06e4cba
a3055fa
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
06e4cba
 
a3055fa
 
 
 
06e4cba
 
9244048
06e4cba
 
a3055fa
06e4cba
 
 
 
a3055fa
 
 
 
 
 
06e4cba
a3055fa
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
06e4cba
a3055fa
 
06e4cba
a3055fa
 
06e4cba
 
 
a3055fa
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
import copy
import math
from typing import Optional, Tuple, Union

import torch
from torch import nn
from torch.nn import CrossEntropyLoss
from transformers.models.t5 import modeling_t5
from transformers.modeling_outputs import CausalLMOutputWithPast
from transformers.utils import (
    add_start_docstrings_to_model_forward,
    logging,
    replace_return_docstrings,
)

from .decoderonlyt5_config import DecoderOnlyT5Config


logger = logging.get_logger(__name__)
_CONFIG_FOR_DOC = "DecoderOnlyT5Config"


class DecoderOnlyT5LayerNorm(nn.Module):
    def __init__(self, hidden_size, eps=1e-6, use_scale=True, center_scale_at_zero=False):
        """
        Construct a layernorm module in the T5 style No bias and no subtraction of mean.
        """
        super().__init__()
        if use_scale:
            self.weight = nn.Parameter(torch.ones(hidden_size))
        else:
            assert not center_scale_at_zero
            self.weight = None
        self.center_scale_at_zero = center_scale_at_zero
        self.variance_epsilon = eps

    def forward(self, hidden_states):
        # https://github.com/google/flaxformer/blob/ea17eb012a1d340ddff017b7a534c2162aaec34c/flaxformer/components/layer_norm.py#L30

        # layer norm should always be calculated in float32
        mean2 = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True)
        hidden_states = hidden_states * torch.rsqrt(mean2 + self.variance_epsilon)

        # convert into float16 if necessary
        if self.weight is None:
            return hidden_states        
        if self.weight.dtype == torch.float16:
            hidden_states = hidden_states.to(torch.float16)
        if self.center_scale_at_zero:
            return (self.weight + 1.0) * hidden_states
        else:
            return self.weight * hidden_states
            


class DecoderOnlyT5LayerFF(modeling_t5.T5LayerFF):
    def __init__(self, config: DecoderOnlyT5Config):
        super(modeling_t5.T5LayerFF, self).__init__()
        if config.is_gated_act:
            self.DenseReluDense = modeling_t5.T5DenseGatedActDense(config)
        else:
            self.DenseReluDense = modeling_t5.T5DenseActDense(config)

        if not config.parallel_layers:
            self.layer_norm = modeling_t5.DecoderOnlyT5LayerNorm(
                config.d_model, eps=config.layer_norm_epsilon
            )
        else:
            self.layer_norm = nn.Identity()
        self.dropout = nn.Dropout(config.dropout_rate)


# LlamaRotaryEmbedding
class DecoderOnlyT5RotaryEmbedding(nn.Module):
    def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
        super().__init__()

        self.dim = dim
        self.max_position_embeddings = max_position_embeddings
        self.base = base
        inv_freq = 1.0 / (
            self.base ** (torch.arange(0, self.dim, 2).float().to(device) / self.dim)
        )
        self.register_buffer("inv_freq", inv_freq, persistent=False)

        # Build here to make `torch.jit.trace` work.
        self._set_cos_sin_cache(
            seq_len=max_position_embeddings,
            device=self.inv_freq.device,
            dtype=torch.get_default_dtype(),
        )

    def _set_cos_sin_cache(self, seq_len, device, dtype):
        self.max_seq_len_cached = seq_len
        t = torch.arange(
            self.max_seq_len_cached, device=device, dtype=self.inv_freq.dtype
        )

        freqs = torch.einsum("i,j->ij", t, self.inv_freq)
        # Different from paper, but it uses a different permutation in order to obtain the same calculation
        emb = torch.cat((freqs, freqs), dim=-1)
        self.register_buffer("cos_cached", emb.cos().to(dtype), persistent=False)
        self.register_buffer("sin_cached", emb.sin().to(dtype), persistent=False)

    def forward(self, x, seq_len=None):
        # x: [bs, num_attention_heads, seq_len, head_size]
        if seq_len > self.max_seq_len_cached:
            self._set_cos_sin_cache(seq_len=seq_len, device=x.device, dtype=x.dtype)

        return (
            self.cos_cached[:seq_len].to(dtype=x.dtype),
            self.sin_cached[:seq_len].to(dtype=x.dtype),
        )


def rotate_half(x):
    """Rotates half the hidden dims of the input."""
    x1 = x[..., : x.shape[-1] // 2]
    x2 = x[..., x.shape[-1] // 2 :]
    return torch.cat((-x2, x1), dim=-1)


def apply_rotary_pos_emb(q, k, cos, sin, position_ids, unsqueeze_dim=1):
    """Applies Rotary Position Embedding to the query and key tensors.

    Args:
        q (`torch.Tensor`): The query tensor.
        k (`torch.Tensor`): The key tensor.
        cos (`torch.Tensor`): The cosine part of the rotary embedding.
        sin (`torch.Tensor`): The sine part of the rotary embedding.
        position_ids (`torch.Tensor`):
            The position indices of the tokens corresponding to the query and key tensors. For example, this can be
            used to pass offsetted position ids when working with a KV-cache.
        unsqueeze_dim (`int`, *optional*, defaults to 1):
            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
    Returns:
        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
    """
    cos = cos[position_ids].unsqueeze(unsqueeze_dim)
    sin = sin[position_ids].unsqueeze(unsqueeze_dim)
    q_embed = (q * cos) + (rotate_half(q) * sin)
    k_embed = (k * cos) + (rotate_half(k) * sin)
    return q_embed, k_embed


# https://github.com/huggingface/transformers/blob/7ee995fd9c692761c4601ddbffa2ac2ec9f27b0b/src/transformers/models/llama/modeling_llama.py#L263
def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
    """
    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
    """
    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
    if n_rep == 1:
        return hidden_states
    hidden_states = hidden_states[:, :, None, :, :].expand(
        batch, num_key_value_heads, n_rep, slen, head_dim
    )
    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)


class DecoderOnlyT5Attention(modeling_t5.T5Attention):
    """
    Supports both multi-head and multi-query attention.
    https://arxiv.org/abs/1911.02150
    https://github.com/google/flaxformer/blob/ea17eb012a1d340ddff017b7a534c2162aaec34c/flaxformer/components/attention/dense_attention.py#L292
    """

    def __init__(self, config: DecoderOnlyT5Config, has_relative_attention_bias=False):
        super(modeling_t5.T5Attention, self).__init__()
        self.is_decoder = config.is_decoder
        assert not has_relative_attention_bias
        assert config.use_rotary_embedding
        self.d_model = config.d_model
        self.head_dim = config.d_kv
        self.num_heads = config.num_heads
        self.num_key_value_heads = 1 if config.multi_query_attention else self.n_heads
        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
        self.attention_dropout = config.dropout_rate
        self.inner_dim = self.num_heads * self.head_dim
        self.kv_inner_dim = self.num_key_value_heads * self.head_dim    
        self.rotary_emb = DecoderOnlyT5RotaryEmbedding(
            self.head_dim,
            max_position_embeddings=config.relative_attention_max_distance,
            base=config.rotary_embedding_max_timescale,
        )

        # Mesh TensorFlow initialization to avoid scaling before softmax
        self.q = nn.Linear(self.d_model, self.inner_dim, bias=False)
        self.k = nn.Linear(self.d_model, self.kv_inner_dim, bias=False)
        self.v = nn.Linear(self.d_model, self.kv_inner_dim, bias=False)
        self.o = nn.Linear(self.inner_dim, self.d_model, bias=False)

        self.pruned_heads = set()
        self.gradient_checkpointing = False

    def forward(
        self,
        hidden_states: torch.Tensor,
        key_value_states=None,
        position_bias=None,
        mask: Optional[torch.Tensor] = None,
        layer_head_mask=None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_value: Optional[Tuple[torch.Tensor]] = None,
        output_attentions: bool = False,
        use_cache: bool = False,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        assert key_value_states is None
        assert position_bias is None
        assert layer_head_mask is None

        bsz, q_len, _ = hidden_states.size()

        query_states = self.q(hidden_states)
        key_states = self.k(hidden_states)
        value_states = self.v(hidden_states)

        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)

        kv_seq_len = key_states.shape[-2]
        if past_key_value is not None:
            kv_seq_len += past_key_value[0].shape[-2]
        cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)

        if past_key_value is not None:
            # reuse k, v, self_attention
            key_states = torch.cat([past_key_value[0], key_states], dim=2)
            value_states = torch.cat([past_key_value[1], value_states], dim=2)

        past_key_value = (key_states, value_states) if use_cache else None

        key_states = repeat_kv(key_states, self.num_key_value_groups)
        value_states = repeat_kv(value_states, self.num_key_value_groups)

        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)

        if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
            raise ValueError(
                f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
                f" {attn_weights.size()}"
            )

        if mask is not None:
            if mask.size() != (bsz, 1, q_len, kv_seq_len):
                raise ValueError(
                    f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {mask.size()}"
                )
            attn_weights = attn_weights + mask

        # upcast attention to fp32
        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
        attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout)
        attn_output = torch.matmul(attn_weights, value_states)

        if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
            raise ValueError(
                f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
                f" {attn_output.size()}"
            )

        attn_output = attn_output.transpose(1, 2).contiguous()
        attn_output = attn_output.reshape(bsz, q_len, self.inner_dim)
        attn_output = self.o(attn_output)

        present_key_value_state = (
            (key_states, value_states) if (self.is_decoder and use_cache) else None
        )
        outputs = (attn_output,) + (present_key_value_state,) + (position_bias,)

        if output_attentions:
            outputs = outputs + (attn_weights,)
        return outputs


class DecoderOnlyT5LayerSelfAttention(modeling_t5.T5LayerSelfAttention):
    def __init__(self, config, has_relative_attention_bias=False):
        super(modeling_t5.T5LayerSelfAttention, self).__init__()
        self.SelfAttention = DecoderOnlyT5Attention(
            config, has_relative_attention_bias=has_relative_attention_bias
        )
        self.layer_norm = DecoderOnlyT5LayerNorm(
            config.d_model,
            eps=config.layer_norm_epsilon,
            use_scale=True,
            center_scale_at_zero=True,
        )
        self.dropout = nn.Dropout(config.dropout_rate)
        self.parallel_layers = config.parallel_layers

    def forward(
        self,
        hidden_states,
        attention_mask=None,
        position_bias=None,
        position_ids=None,
        layer_head_mask=None,
        past_key_value=None,
        use_cache=False,
        output_attentions=False,
    ):
        if not self.parallel_layers:
            x = self.layer_norm(hidden_states)
        else:
            x = hidden_states
        attention_output = self.SelfAttention(
            x,
            mask=attention_mask,
            position_bias=position_bias,
            position_ids=position_ids,
            layer_head_mask=layer_head_mask,
            past_key_value=past_key_value,
            use_cache=use_cache,
            output_attentions=output_attentions,
        )
        if not self.parallel_layers:
            # When parallel_layers is True, the residual connection is applied
            # in the decoder block instead of here.
            hidden_states = hidden_states + self.dropout(attention_output[0])
        else:
            hidden_states = attention_output[0]
        outputs = (hidden_states,) + attention_output[
            1:
        ]  # add attentions if we output them
        return outputs


class DecoderOnlyT5Block(modeling_t5.T5Block):
    def __init__(self, config, has_relative_attention_bias=False):
        super(modeling_t5.T5Block, self).__init__()
        self.is_decoder = config.is_decoder
        self.is_decoder_only = config.is_decoder_only
        self.layer = nn.ModuleList()
        self.layer.append(
            DecoderOnlyT5LayerSelfAttention(
                config, has_relative_attention_bias=has_relative_attention_bias
            )
        )
        if self.is_decoder:
            if config.is_decoder_only:
                self.layer.append(nn.Identity())
            else:
                self.layer.append(modeling_t5.T5LayerCrossAttention(config))
        self.parallel_layers = config.parallel_layers
        self.layer.append(DecoderOnlyT5LayerFF(config))

    def forward(
        self,
        hidden_states,
        attention_mask=None,
        position_bias=None,
        position_ids=None,
        encoder_hidden_states=None,
        layer_head_mask=None,
        past_key_value=None,
        use_cache=False,
        output_attentions=False,
        encoder_attention_mask=None,
        encoder_decoder_position_bias=None,
        cross_attn_layer_head_mask=None,
        return_dict=True,
    ):
        assert encoder_attention_mask is None
        assert encoder_decoder_position_bias is None
        assert cross_attn_layer_head_mask is None
        if past_key_value is not None:
            expected_num_past_key_values = 2 if encoder_hidden_states is None else 4

            if len(past_key_value) != expected_num_past_key_values:
                raise ValueError(
                    f"There should be {expected_num_past_key_values} past states. "
                    f"{'2 (past / key) for cross attention. ' if expected_num_past_key_values == 4 else ''}"
                    f"Got {len(past_key_value)} past key / value states"
                )
            self_attn_past_key_value = past_key_value[:2]
        else:
            self_attn_past_key_value = None

        ff_layer = self.layer[-1]
        if self.parallel_layers:
            # https://github.com/google/flaxformer/blob/ea17eb012a1d340ddff017b7a534c2162aaec34c/flaxformer/architectures/t5/t5_architecture.py#L563-L568
            x = self.layer[0].layer_norm(hidden_states)
            ff_output = ff_layer(x)
        else:
            x = hidden_states

        self_attention_outputs = self.layer[0](
            x,
            attention_mask=attention_mask,
            position_bias=position_bias,
            position_ids=position_ids,
            layer_head_mask=layer_head_mask,
            past_key_value=self_attn_past_key_value,
            use_cache=use_cache,
            output_attentions=output_attentions,
        )
        x, present_key_value_state = self_attention_outputs[:2]
        attention_outputs = self_attention_outputs[
            2:
        ]  # Keep self-attention outputs and relative position weights

        # clamp inf values to enable fp16 training
        if x.dtype == torch.float16:
            clamp_value = torch.where(
                torch.isinf(x).any(),
                torch.finfo(x.dtype).max - 1000,
                torch.finfo(x.dtype).max,
            )
            x = torch.clamp(x, min=-clamp_value, max=clamp_value)

        do_cross_attention = (
            self.is_decoder
            and not self.is_decoder_only
            and encoder_hidden_states is not None
        )
        assert not do_cross_attention

        if self.parallel_layers:
            # https://github.com/google/flaxformer/blob/ea17eb012a1d340ddff017b7a534c2162aaec34c/flaxformer/architectures/t5/t5_architecture.py#L534-L578
            x = x + ff_output
            x *= 2**-0.5
            hidden_states = hidden_states + self.layer[0].dropout(x)
        else:
            hidden_states = ff_layer(x)

        # clamp inf values to enable fp16 training
        if hidden_states.dtype == torch.float16:
            clamp_value = torch.where(
                torch.isinf(hidden_states).any(),
                torch.finfo(hidden_states.dtype).max - 1000,
                torch.finfo(hidden_states.dtype).max,
            )
            hidden_states = torch.clamp(
                hidden_states, min=-clamp_value, max=clamp_value
            )

        outputs = (hidden_states,)

        if use_cache:
            outputs = outputs + (present_key_value_state,) + attention_outputs
        else:
            outputs = outputs + attention_outputs

        return outputs  # hidden-states, present_key_value_states, (self-attention position bias), (self-attention weights), (cross-attention position bias), (cross-attention weights)


class DecoderOnlyT5Stack(modeling_t5.T5Stack):
    def __init__(self, config, embed_tokens=None):
        super(modeling_t5.T5Stack, self).__init__(config)

        self.embed_tokens = embed_tokens
        self.is_decoder = config.is_decoder

        self.block = nn.ModuleList(
            [
                DecoderOnlyT5Block(
                    config,
                    has_relative_attention_bias=(
                        config.has_relative_attention_bias and bool(i == 0)
                    ),
                )
                for i in range(config.num_layers)
            ]
        )
        self.final_layer_norm = DecoderOnlyT5LayerNorm(
            config.d_model,
            eps=config.layer_norm_epsilon,
            use_scale=False,
            center_scale_at_zero=False,
        )
        self.dropout = nn.Dropout(config.dropout_rate)

        # Initialize weights and apply final processing
        self.post_init()
        # Model parallel
        self.model_parallel = False
        self.device_map = None
        self.gradient_checkpointing = False

    def forward(
        self,
        input_ids=None,
        position_ids=None,
        attention_mask=None,
        encoder_hidden_states=None,
        encoder_attention_mask=None,
        inputs_embeds=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,
    ):
        # Model parallel
        if self.model_parallel:
            torch.cuda.set_device(self.first_device)
            self.embed_tokens = self.embed_tokens.to(self.first_device)
        use_cache = use_cache if use_cache is not None else self.config.use_cache
        output_attentions = (
            output_attentions
            if output_attentions is not None
            else self.config.output_attentions
        )
        output_hidden_states = (
            output_hidden_states
            if output_hidden_states is not None
            else self.config.output_hidden_states
        )
        return_dict = (
            return_dict if return_dict is not None else self.config.use_return_dict
        )

        if input_ids is not None and inputs_embeds is not None:
            err_msg_prefix = "decoder_" if self.is_decoder else ""
            raise ValueError(
                f"You cannot specify both {err_msg_prefix}input_ids and {err_msg_prefix}inputs_embeds at the same time"
            )
        elif input_ids is not None:
            input_shape = input_ids.size()
            input_ids = input_ids.view(-1, input_shape[-1])
        elif inputs_embeds is not None:
            input_shape = inputs_embeds.size()[:-1]
        else:
            err_msg_prefix = "decoder_" if self.is_decoder else ""
            raise ValueError(
                f"You have to specify either {err_msg_prefix}input_ids or {err_msg_prefix}inputs_embeds"
            )

        if position_ids is None:
            seq_length = input_ids.shape[1]
            past_key_values_length = (
                0 if past_key_values is None else past_key_values[0][0].shape[2]
            )
            device = input_ids.device if input_ids is not None else inputs_embeds.device
            position_ids = torch.arange(
                past_key_values_length,
                seq_length + past_key_values_length,
                dtype=torch.long,
                device=device,
            ).unsqueeze(0)

        if inputs_embeds is None:
            if self.embed_tokens is None:
                raise ValueError(
                    "You have to initialize the model with valid token embeddings"
                )
            inputs_embeds = self.embed_tokens(input_ids)

        batch_size, seq_length = input_shape

        # required mask seq length can be calculated via length of past
        mask_seq_length = (
            past_key_values[0][0].shape[2] + seq_length
            if past_key_values is not None
            else seq_length
        )

        if use_cache is True:
            if not self.is_decoder:
                raise ValueError(
                    f"`use_cache` can only be set to `True` if {self} is used as a decoder"
                )

        if attention_mask is None:
            attention_mask = torch.ones(
                batch_size, mask_seq_length, device=inputs_embeds.device
            )

        # initialize past_key_values with `None` if past does not exist
        if past_key_values is None:
            past_key_values = [None] * len(self.block)

        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
        # ourselves in which case we just need to make it broadcastable to all heads.
        extended_attention_mask = self.get_extended_attention_mask(
            attention_mask, input_shape
        )

        if self.gradient_checkpointing and self.training:
            if use_cache:
                logger.warning_once(
                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                )
                use_cache = False

        # Prepare head mask if needed
        head_mask = self.get_head_mask(head_mask, self.config.num_layers)
        cross_attn_head_mask = self.get_head_mask(
            cross_attn_head_mask, self.config.num_layers
        )
        present_key_value_states = () if use_cache else None
        all_hidden_states = () if output_hidden_states else None
        all_attentions = () if output_attentions else None
        all_cross_attentions = () if (output_attentions and self.is_decoder) else None
        position_bias = None

        hidden_states = self.dropout(inputs_embeds)

        for i, (layer_module, past_key_value) in enumerate(
            zip(self.block, past_key_values)
        ):
            layer_head_mask = head_mask[i]
            cross_attn_layer_head_mask = cross_attn_head_mask[i]
            # Model parallel
            if self.model_parallel:
                torch.cuda.set_device(hidden_states.device)
                # Ensure that attention_mask is always on the same device as hidden_states
                if attention_mask is not None:
                    attention_mask = attention_mask.to(hidden_states.device)
                if position_bias is not None:
                    position_bias = position_bias.to(hidden_states.device)                
                if layer_head_mask is not None:
                    layer_head_mask = layer_head_mask.to(hidden_states.device)

            if output_hidden_states:
                all_hidden_states = all_hidden_states + (hidden_states,)

            if self.gradient_checkpointing and self.training:
                layer_outputs = self._gradient_checkpointing_func(
                    layer_module.forward,
                    hidden_states,
                    extended_attention_mask,
                    position_bias,
                    None,
                    None,
                    None,
                    layer_head_mask,
                    cross_attn_layer_head_mask,
                    None,  # past_key_value is always None with gradient checkpointing
                    use_cache,
                    output_attentions,
                )
            else:
                layer_outputs = layer_module(
                    hidden_states,
                    attention_mask=extended_attention_mask,
                    position_bias=position_bias,
                    position_ids=position_ids,
                    encoder_hidden_states=None,
                    encoder_attention_mask=None,
                    encoder_decoder_position_bias=None,
                    layer_head_mask=layer_head_mask,
                    cross_attn_layer_head_mask=cross_attn_layer_head_mask,
                    past_key_value=past_key_value,
                    use_cache=use_cache,
                    output_attentions=output_attentions,
                )

            # layer_outputs is a tuple with:
            # hidden-states, key-value-states, (self-attention position bias), (self-attention weights), (cross-attention position bias), (cross-attention weights)
            if use_cache is False:
                layer_outputs = layer_outputs[:1] + (None,) + layer_outputs[1:]

            hidden_states, present_key_value_state = layer_outputs[:2]

            # We share the position biases between the layers - the first layer store them
            # layer_outputs = hidden-states, key-value-states (self-attention position bias), (self-attention weights),
            # (cross-attention position bias), (cross-attention weights)
            position_bias = layer_outputs[2]
            # append next layer key value states
            if use_cache:
                present_key_value_states = present_key_value_states + (
                    present_key_value_state,
                )

            if output_attentions:
                all_attentions = all_attentions + (layer_outputs[3],)
                if self.is_decoder:
                    all_cross_attentions = all_cross_attentions + (layer_outputs[5],)

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

        hidden_states = self.final_layer_norm(hidden_states)
        hidden_states = self.dropout(hidden_states)

        # Add last layer
        if output_hidden_states:
            all_hidden_states = all_hidden_states + (hidden_states,)

        if not return_dict:
            return tuple(
                v
                for v in [
                    hidden_states,
                    present_key_value_states,
                    all_hidden_states,
                    all_attentions,
                    all_cross_attentions,
                ]
                if v is not None
            )
        return modeling_t5.BaseModelOutputWithPastAndCrossAttentions(
            last_hidden_state=hidden_states,
            past_key_values=present_key_value_states,
            hidden_states=all_hidden_states,
            attentions=all_attentions,
            cross_attentions=all_cross_attentions,
        )


class DecoderOnlyT5Model(modeling_t5.T5ForConditionalGeneration):
    def __init__(self, config: DecoderOnlyT5Config):
        super(modeling_t5.T5ForConditionalGeneration, self).__init__(config)
        self.model_dim = config.d_model

        self.shared = nn.Embedding(config.vocab_size, config.d_model)
        assert (
            self.config.num_layers == 0
        ), "Decoder only model cannot have encoder layers"
        self.encoder = None

        decoder_config = copy.deepcopy(config)
        decoder_config.is_decoder = True
        decoder_config.is_encoder_decoder = False
        decoder_config.num_layers = config.num_decoder_layers
        self.decoder = DecoderOnlyT5Stack(decoder_config, self.shared)

        self.lm_head = nn.Linear(config.d_model, config.vocab_size, bias=False)

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

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

    def _tie_weights(self):
        if not self.config.tie_word_embeddings:
            return
        if self.decoder:
            self._tie_or_clone_weights(self.decoder.embed_tokens, self.shared)

    @add_start_docstrings_to_model_forward(modeling_t5.T5_INPUTS_DOCSTRING)
    @replace_return_docstrings(
        output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
    )
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple, CausalLMOutputWithPast]:
        r"""
        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
            Labels for computing the sequence classification/regression loss. Indices should be in `[-100, 0, ...,
            config.vocab_size - 1]`. All labels set to `-100` are ignored (masked), the loss is only computed for
            labels in `[0, ..., config.vocab_size]`

        Returns:

        Examples:

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

        if self.model_parallel:
            torch.cuda.set_device(self.decoder.first_device)

        # Set device for model parallelism
        if self.model_parallel:
            torch.cuda.set_device(self.decoder.first_device)
            if input_ids is not None:
                input_ids = input_ids.to(self.decoder.first_device)
            if attention_mask is not None:
                attention_mask = attention_mask.to(self.decoder.first_device)

        # Decode
        outputs = self.decoder(
            input_ids=input_ids,
            position_ids=position_ids,
            attention_mask=attention_mask,
            inputs_embeds=inputs_embeds,
            past_key_values=past_key_values,
            encoder_hidden_states=None,
            encoder_attention_mask=None,
            head_mask=None,
            cross_attn_head_mask=None,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        sequence_output = outputs[0]

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

        if self.config.tie_word_embeddings:
            # Rescale output before projecting on vocab
            # See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/transformer/transformer.py#L586
            sequence_output = sequence_output * (self.model_dim**-0.5)

        lm_logits = self.lm_head(sequence_output)

        loss = None
        if labels is not None:
            loss_fct = CrossEntropyLoss(ignore_index=-100)
            # move labels to correct device to enable PP
            labels = labels.to(lm_logits.device)
            loss = loss_fct(lm_logits.view(-1, lm_logits.size(-1)), labels.view(-1))
            # TODO(thom): Add z_loss https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/layers.py#L666

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

        return CausalLMOutputWithPast(
            loss=loss,
            logits=lm_logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )