File size: 33,185 Bytes
ad1a1b6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
# Copyright (c) Microsoft Corporation.
# Licensed under the MIT license.
#
# Copyright (c) 2022, Tri Dao, trid@cs.stanford.edu.
# Licensed under the BSD 3-Clause License.

from __future__ import annotations

import math
from dataclasses import dataclass, field
from typing import Any, Dict, Optional, Tuple, Union

import torch
import torch.nn as nn
from einops import rearrange, repeat
from transformers import PretrainedConfig, PreTrainedModel
from transformers.activations import ACT2FN
from transformers.modeling_outputs import CausalLMOutputWithPast

from .configuration_phi import PhiConfig

try:
    from flash_attn.bert_padding import pad_input, unpad_input
    from flash_attn.layers.rotary import RotaryEmbedding as FlashRotaryEmbedding
    from flash_attn.modules.mha import FlashCrossAttention, FlashSelfAttention
    from flash_attn.ops.fused_dense import FusedDense
except:
    pad_input, unpad_input = None, None
    FlashRotaryEmbedding = None
    FlashSelfAttention, FlashCrossAttention = None, None
    FusedDense = None


@dataclass
class InferenceParams:
    """Inference parameters passed to model to efficiently calculate
    and store context during inference.

    Reference:
        https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/utils/generation.py.

    Args:
        max_seqlen: Maximum sequence length.
        max_batch_size: Maximum batch size.
        seqlen_offset: Sequence length offset.
        batch_size_offset: Batch size offset.
        key_value_memory_dict: Key value memory dictionary.
        lengths_per_sample: Lengths per sample.

    """

    max_seqlen: int = field(metadata={"help": "Maximum sequence length."})

    max_batch_size: int = field(metadata={"help": "Maximum batch size."})

    seqlen_offset: int = field(default=0, metadata={"help": "Sequence length offset."})

    batch_size_offset: int = field(default=0, metadata={"help": "Batch size offset."})

    key_value_memory_dict: Dict[str, Any] = field(
        default_factory=dict, metadata={"help": "Key value memory dictionary."}
    )

    lengths_per_sample: torch.Tensor = field(default=None, metadata={"help": "Lengths per sample."})


class Embedding(nn.Module):
    """Token embedding with dropout."""

    def __init__(self, config: PretrainedConfig) -> None:
        super().__init__()

        self.wte = nn.Embedding(config.vocab_size, config.n_embd)
        self.drop = nn.Dropout(config.embd_pdrop)

    def forward(self, input_ids: torch.LongTensor) -> torch.FloatTensor:
        input_shape = input_ids.size()
        input_ids = input_ids.view(-1, input_shape[-1])

        hidden_states = self.wte(input_ids)
        hidden_states = self.drop(hidden_states)

        return hidden_states


def _apply_rotary_emb(
    x: torch.FloatTensor,
    cos: torch.FloatTensor,
    sin: torch.FloatTensor,
) -> torch.FloatTensor:
    _, seqlen, _, _ = x.shape
    _, rotary_dim = cos.shape
    rotary_dim *= 2

    x_rot = x[:, :, :, :rotary_dim]
    x_pass = x[:, :, :, rotary_dim:]

    x1, x2 = x_rot.chunk(2, dim=-1)
    c, s = rearrange(cos[:seqlen], "s d -> s 1 d"), rearrange(sin[:seqlen], "s d -> s 1 d")
    x1, x2, c, s = [t.to(dtype=torch.float32) for t in [x1, x2, c, s]]

    x_rot = torch.cat([x1 * c - x2 * s, x1 * s + x2 * c], axis=-1).to(x.dtype)

    return torch.cat([x_rot, x_pass], axis=-1)


def _apply_rotary_emb_kv(
    kv: torch.FloatTensor,
    cos: torch.FloatTensor,
    sin: torch.FloatTensor,
    cos_k: Optional[torch.FloatTensor] = None,
    sin_k: Optional[torch.FloatTensor] = None,
) -> torch.FloatTensor:
    _, seqlen, _, _, _ = kv.shape
    _, rotary_dim = cos.shape
    rotary_dim *= 2

    k_rot = kv[:, :, 0, :, :rotary_dim]
    k_pass = kv[:, :, 0, :, rotary_dim:]

    k1, k2 = k_rot.chunk(2, dim=-1)
    c, s = rearrange(cos[:seqlen], "s d -> s 1 d"), rearrange(sin[:seqlen], "s d -> s 1 d")
    k1, k2, c, s = [t.to(dtype=torch.float32) for t in [k1, k2, c, s]]

    k_rot = torch.cat([k1 * c - k2 * s, k1 * s + k2 * c], axis=-1).to(kv.dtype)

    return torch.cat(
        [
            torch.cat([k_rot, k_pass], axis=-1).unsqueeze(2),
            kv[:, :, 1:2, :, :],
        ],
        axis=2,
    )


def _apply_rotary_emb_qkv(
    qkv: torch.FloatTensor,
    cos: torch.FloatTensor,
    sin: torch.FloatTensor,
    cos_k: Optional[torch.FloatTensor] = None,
    sin_k: Optional[torch.FloatTensor] = None,
) -> torch.FloatTensor:
    _, seqlen, _, _, _ = qkv.shape
    _, rotary_dim = cos.shape
    rotary_dim *= 2

    q_rot = qkv[:, :, 0, :, :rotary_dim]
    q_pass = qkv[:, :, 0, :, rotary_dim:]

    k_rot = qkv[:, :, 1, :, :rotary_dim]
    k_pass = qkv[:, :, 1, :, rotary_dim:]

    q1, q2 = q_rot.chunk(2, dim=-1)
    k1, k2 = k_rot.chunk(2, dim=-1)
    c, s = rearrange(cos[:seqlen], "s d -> s 1 d"), rearrange(sin[:seqlen], "s d -> s 1 d")
    q1, q2, k1, k2, c, s = [t.to(dtype=torch.float32) for t in [q1, q2, k1, k2, c, s]]

    q_rot = torch.cat([q1 * c - q2 * s, q1 * s + q2 * c], axis=-1).to(qkv.dtype)
    k_rot = torch.cat([k1 * c - k2 * s, k1 * s + k2 * c], axis=-1).to(qkv.dtype)

    return torch.cat(
        [
            torch.cat([q_rot, q_pass], axis=-1).unsqueeze(2),
            torch.cat([k_rot, k_pass], axis=-1).unsqueeze(2),
            qkv[:, :, 2:3, :, :],
        ],
        axis=2,
    )


class RotaryEmbedding(nn.Module):
    """Rotary positional embedding (RoPE).

    Reference:
        RoFormer: Enhanced Transformer with Rotary Position Embedding.
        https://arxiv.org/pdf/2104.09864.pdf.

    """

    def __init__(
        self,
        dim: int,
        base: int = 10000,
        scale_base: Optional[float] = None,
        pos_idx_in_fp32: bool = True,
        max_position_embeddings: int = 2048,
        device: Optional[str] = None,
        **kwargs,
    ) -> None:
        super().__init__()

        if scale_base is not None:
            raise NotImplementedError

        self.dim = dim
        self.base = float(base)
        self.scale_base = scale_base
        self.pos_idx_in_fp32 = pos_idx_in_fp32
        self.max_position_embeddings = max_position_embeddings
        self.device = device

        # Generate and save the inverse frequency buffer (non-trainable)
        inv_freq = self._compute_inv_freq(device)
        self.register_buffer("inv_freq", inv_freq, persistent=False)

        # Generate and save the scale buffer (non-trainable)
        scale = (
            (torch.arange(0, dim, 2, device=device, dtype=torch.float32) + 0.4 * dim) / (1.4 * dim)
            if scale_base is not None
            else None
        )
        self.register_buffer("scale", scale, persistent=False)

        # Initialize cached attributes since ONNX can't rely on dynamic initialization
        self._update_cos_sin_cache(max_position_embeddings, device=device, dtype=torch.float32)

    def _compute_inv_freq(self, device: Optional[str] = None) -> torch.FloatTensor:
        return 1.0 / (self.base ** (torch.arange(0, self.dim, 2, device=device, dtype=torch.float32) / self.dim))

    def _update_cos_sin_cache(
        self,
        seqlen: int,
        device: Optional[str] = None,
        dtype: Optional[torch.dtype] = None,
    ) -> None:
        self._seq_len_cached = seqlen

        # fp32 is preferred since the output of `torch.arange` can be quite large
        # and bf16 would lose a lot of precision
        if self.pos_idx_in_fp32:
            t = torch.arange(seqlen, device=device, dtype=torch.float32)
            if self.inv_freq.dtype != torch.float32:
                inv_freq = self._compute_inv_freq(device=device)
            else:
                inv_freq = self.inv_freq
        else:
            t = torch.arange(seqlen, device=device, dtype=self.inv_freq.dtype)
            inv_freq = self.inv_freq

        # `torch.outer` is preferred since `torch.einsum` converts from fp32 to fp16 if used with AMP
        freqs = torch.outer(t, inv_freq)
        if self.scale is None:
            self._cos_cached = torch.cos(freqs).to(dtype)
            self._sin_cached = torch.sin(freqs).to(dtype)
        else:
            power = (
                torch.arange(seqlen, dtype=self.scale.dtype, device=self.scale.device) - seqlen // 2
            ) / self.scale_base
            scale = self.scale.to(device=power.device) ** rearrange(power, "s -> s 1")

            # Force the scale multiplication to happen in fp32
            self._cos_cached = (torch.cos(freqs) * scale).to(dtype)
            self._sin_cached = (torch.sin(freqs) * scale).to(dtype)
            self._cos_k_cached = (torch.cos(freqs) / scale).to(dtype)
            self._sin_k_cached = (torch.sin(freqs) / scale).to(dtype)

    def forward(
        self,
        qkv: torch.Tensor,
        kv: Optional[torch.Tensor] = None,
        seqlen_offset: int = 0,
        **kwargs,
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        seq_start = seqlen_offset
        seq_end = seq_start + qkv.shape[1]

        if (
            self._cos_cached.device != qkv.device
            or self._cos_cached.dtype != qkv.dtype
            or (self.training and self._cos_cached.is_inference())
        ):
            self._update_cos_sin_cache(self.max_position_embeddings, device=qkv.device, dtype=qkv.dtype)

        if kv is None:
            return _apply_rotary_emb_qkv(
                qkv,
                self._cos_cached[seq_start:seq_end],
                self._sin_cached[seq_start:seq_end],
            )
        else:
            q = _apply_rotary_emb(
                qkv,
                self._cos_cached[seq_start:seq_end],
                self._sin_cached[seq_start:seq_end],
            )
            kv = _apply_rotary_emb_kv(
                kv,
                self._cos_cached[seq_start:seq_end],
                self._sin_cached[seq_start:seq_end],
            )

            return q, kv


class MLP(nn.Module):
    """Multi-Layer Perceptron.

    Reference:
        Attention Is All You Need.
        https://arxiv.org/pdf/1706.03762.pdf.

    """

    def __init__(
        self,
        config: PretrainedConfig,
        n_inner: Optional[int] = None,
        act_fn: Optional[str] = None,
    ) -> None:
        super().__init__()

        act_fn = config.activation_function if act_fn is None else act_fn

        n_inner = getattr(config, "n_inner", None) if n_inner is None else n_inner
        n_inner = n_inner if n_inner is not None else 4 * config.n_embd

        self.fc1 = nn.Linear(config.n_embd, n_inner)
        self.fc2 = nn.Linear(n_inner, config.n_embd)
        self.act = ACT2FN[act_fn]

    def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:
        hidden_states = self.fc1(hidden_states)
        hidden_states = self.act(hidden_states)
        hidden_states = self.fc2(hidden_states)

        return hidden_states


class SelfAttention(nn.Module):
    """Self-attention layer (compatible with PyTorch).

    Reference:
        https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/modules/mha.py.

    """

    def __init__(
        self,
        causal: bool = True,
        softmax_scale: Optional[float] = None,
        attention_dropout: float = 0.0,
    ) -> None:
        super().__init__()

        self.causal = causal
        self.softmax_scale = softmax_scale
        self.drop = nn.Dropout(attention_dropout)

    @torch.autocast("cpu", enabled=False)
    @torch.autocast("cuda", enabled=False)
    def forward(
        self,
        qkv: torch.FloatTensor,
        causal: bool = None,
        key_padding_mask: Optional[torch.BoolTensor] = None,
        **kwargs,
    ) -> torch.FloatTensor:
        batch_size, seqlen = qkv.shape[0], qkv.shape[1]
        q, k, v = qkv.unbind(dim=2)

        q = q.to(torch.float32)
        k = k.to(torch.float32)

        causal = self.causal if causal is None else causal
        softmax_scale = self.softmax_scale or 1.0 / math.sqrt(q.shape[-1])

        # Autocast is manually disabled to avoid `torch.einsum` performing the operation
        # using float16, which might lead to overflow
        scores = torch.einsum("bthd,bshd->bhts", q, k * softmax_scale)

        if key_padding_mask is not None:
            padding_mask = torch.full((batch_size, seqlen), -10000.0, dtype=scores.dtype, device=scores.device)
            padding_mask.masked_fill_(key_padding_mask, 0.0)

            scores = scores + rearrange(padding_mask, "b s -> b 1 1 s")

        if causal:
            causal_mask = torch.triu(torch.full((seqlen, seqlen), -10000.0, device=scores.device), 1)
            scores = scores + causal_mask.to(dtype=scores.dtype)

        attention = torch.softmax(scores, dim=-1).to(v.dtype)
        attention = self.drop(attention)

        output = torch.einsum("bhts,bshd->bthd", attention, v)

        return output


class CrossAttention(nn.Module):
    """Cross-attention layer (compatible with PyTorch).

    Reference:
        https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/modules/mha.py.

    """

    def __init__(
        self,
        causal: bool = True,
        softmax_scale: Optional[float] = None,
        attention_dropout: float = 0.0,
    ) -> None:
        super().__init__()

        self.causal = causal
        self.softmax_scale = softmax_scale
        self.drop = nn.Dropout(attention_dropout)

    @torch.autocast("cpu", enabled=False)
    @torch.autocast("cuda", enabled=False)
    def forward(
        self,
        q: torch.FloatTensor,
        kv: torch.FloatTensor,
        causal: bool = None,
        key_padding_mask: Optional[torch.BoolTensor] = None,
        **kwargs,
    ) -> torch.FloatTensor:
        batch_size, seqlen_q = q.shape[0], q.shape[1]
        seqlen_k = kv.shape[1]

        if kv.shape[3] != q.shape[2]:
            kv = repeat(kv, "... hkv d -> ... (hkv g) d", g=q.shape[2] // kv.shape[3])
        k, v = kv.unbind(dim=2)

        q = q.to(torch.float32)
        k = k.to(torch.float32)

        causal = self.causal if causal is None else causal
        softmax_scale = self.softmax_scale or 1.0 / math.sqrt(q.shape[-1])

        # Autocast is manually disabled to avoid `torch.einsum` performing the operation
        # using float16, which might lead to overflow
        scores = torch.einsum("bthd,bshd->bhts", q, k * softmax_scale)

        if key_padding_mask is not None:
            padding_mask = torch.full(
                (batch_size, seqlen_k),
                -10000.0,
                dtype=scores.dtype,
                device=scores.device,
            )
            padding_mask.masked_fill_(key_padding_mask, 0.0)

            scores = scores + rearrange(padding_mask, "b s -> b 1 1 s")

        if causal:
            rows = rearrange(torch.arange(seqlen_q, device=q.device, dtype=torch.long), "s -> s 1")
            cols = torch.arange(seqlen_k, device=k.device, dtype=torch.long)
            causal_mask = cols > rows + seqlen_k - seqlen_q

            scores = scores.masked_fill(causal_mask, -10000.0)

        attention = torch.softmax(scores, dim=-1).to(v.dtype)
        attention = self.drop(attention)

        output = torch.einsum("bhts,bshd->bthd", attention, v)

        return output


def _find_mha_dims(
    config: PretrainedConfig,
    n_head: Optional[int] = None,
    n_head_kv: Optional[int] = None,
    head_dim: Optional[int] = None,
) -> Tuple[int, int]:
    if n_head is None and head_dim is None:
        head_dim = config.n_embd // config.n_head
        n_head = config.n_head
    elif n_head is None or head_dim is None:
        raise ValueError("`n_head` and `head_dim` must be both specified or `None`.")

    if n_head_kv is None:
        n_head_kv = getattr(config, "n_head_kv", None) or n_head

    return n_head, n_head_kv, head_dim


def _update_kv_cache(kv: torch.FloatTensor, inference_params: InferenceParams, layer_idx: int) -> torch.FloatTensor:
    num_heads, head_dim = kv.shape[-2:]

    if layer_idx not in inference_params.key_value_memory_dict:
        kv_cache = torch.empty(
            inference_params.max_batch_size,
            inference_params.max_seqlen,
            2,
            num_heads,
            head_dim,
            dtype=kv.dtype,
            device=kv.device,
        )
        inference_params.key_value_memory_dict[layer_idx] = kv_cache
    else:
        kv_cache = inference_params.key_value_memory_dict[layer_idx]

    batch_start = inference_params.batch_size_offset
    batch_end = batch_start + kv.shape[0]

    sequence_start = inference_params.seqlen_offset
    sequence_end = sequence_start + kv.shape[1]

    kv_cache[batch_start:batch_end, sequence_start:sequence_end, ...] = kv
    kv = kv_cache[batch_start:batch_end, :sequence_end, ...]

    return kv


class MHA(nn.Module):
    """Multi-head attention layer."""

    def __init__(
        self,
        config: PretrainedConfig,
        dtype: Optional[torch.dtype] = None,
        device: Optional[str] = None,
        rotary_dim: Optional[int] = None,
        rotary_base: float = 10000.0,
        rotary_scale_base: Optional[float] = None,
        n_head: Optional[int] = None,
        n_head_kv: Optional[int] = None,
        head_dim: Optional[int] = None,
        bias: bool = True,
        causal: bool = True,
        softmax_scale: Optional[float] = None,
        layer_idx: Optional[int] = None,
        return_residual: bool = False,
        checkpointing: bool = False,
    ) -> None:
        super().__init__()

        # Rotary embedding
        self.rotary_dim = rotary_dim if rotary_dim is not None else getattr(config, "rotary_dim", 0)
        if self.rotary_dim > 0:
            rotary_cls = FlashRotaryEmbedding if config.flash_rotary else RotaryEmbedding
            if rotary_cls is None:
                rotary_cls = RotaryEmbedding

            rotary_kwargs = {}
            if rotary_cls is RotaryEmbedding:
                rotary_kwargs["max_position_embeddings"] = config.n_positions

            self.rotary_emb = rotary_cls(
                self.rotary_dim,
                base=rotary_base,
                scale_base=rotary_scale_base,
                device=device,
                **rotary_kwargs,
            )

        # MLP
        self.n_head, self.n_head_kv, self.head_dim = _find_mha_dims(
            config, n_head=n_head, n_head_kv=n_head_kv, head_dim=head_dim
        )
        op_size = self.head_dim * (self.n_head + 2 * self.n_head_kv)
        hidden_size = config.n_embd

        linear_cls = FusedDense if config.fused_dense else nn.Linear
        if linear_cls is None:
            linear_cls = nn.Linear

        self.Wqkv = linear_cls(hidden_size, op_size, bias=bias, device=device, dtype=dtype)
        self.out_proj = linear_cls(hidden_size, hidden_size, bias=bias, device=device, dtype=dtype)

        # Attention
        attn_cls = FlashSelfAttention if config.flash_attn else SelfAttention
        if attn_cls is None:
            attn_cls = SelfAttention

        cross_attn_cls = FlashCrossAttention if config.flash_attn else CrossAttention
        if cross_attn_cls is None:
            cross_attn_cls = CrossAttention

        self.inner_attn = attn_cls(
            causal=causal,
            softmax_scale=softmax_scale,
            attention_dropout=config.attn_pdrop,
        )
        self.inner_cross_attn = cross_attn_cls(
            causal=causal,
            softmax_scale=softmax_scale,
            attention_dropout=config.attn_pdrop,
        )

        self.flash_attn = config.flash_attn and attn_cls is FlashSelfAttention
        self.layer_idx = layer_idx
        self.return_residual = return_residual
        self.checkpointing = checkpointing

    def _forward_self_attn(
        self, x: torch.FloatTensor, key_padding_mask: Optional[torch.BoolTensor]
    ) -> torch.FloatTensor:
        qkv = self.Wqkv(x)
        qkv = rearrange(qkv, "... (three h d) -> ... three h d", three=3, d=self.head_dim)

        if self.rotary_dim > 0:
            qkv = self.rotary_emb(qkv)

        if self.flash_attn:
            batch_size, seqlen = qkv.shape[0], qkv.shape[1]

            cu_seqlens, max_seqlen = None, None
            if key_padding_mask is not None:
                # If `key_padding_mask` is supplied, we need to unpad the input and retrieve
                # the `cu_seqlens` and `max_seqlen` to be used by `flash-attn`
                qkv, indices, cu_seqlens, max_seqlen = unpad_input(qkv, key_padding_mask)

            if self.checkpointing:
                attn_output = torch.utils.checkpoint.checkpoint(
                    self.inner_attn, qkv, cu_seqlens=cu_seqlens, max_seqlen=max_seqlen
                )
            else:
                attn_output = self.inner_attn(qkv, cu_seqlens=cu_seqlens, max_seqlen=max_seqlen).to(qkv.device)

            # If `key_padding_mask` is supplied, we need to pad the output back to the original shape
            return pad_input(attn_output, indices, batch_size, seqlen) if key_padding_mask is not None else attn_output

        if self.checkpointing:
            return torch.utils.checkpoint.checkpoint(self.inner_attn, qkv, key_padding_mask=key_padding_mask)

        return self.inner_attn(qkv, key_padding_mask=key_padding_mask)

    def _forward_cross_attn(
        self,
        x: torch.FloatTensor,
        past_key_values: Optional[InferenceParams],
        key_padding_mask: Optional[torch.BoolTensor],
    ) -> torch.FloatTensor:
        batch_size = x.shape[0]

        qkv = self.Wqkv(x)

        q = qkv[..., : self.n_head * self.head_dim]
        q = rearrange(q, "... (h d) -> ... h d", d=self.head_dim)

        kv = qkv[..., self.n_head * self.head_dim :]
        kv = rearrange(kv, "... (two hkv d) -> ... two hkv d", two=2, d=self.head_dim)

        seqlen_offset = past_key_values.seqlen_offset if past_key_values is not None else 0
        causal = None if seqlen_offset == 0 else False
        if self.rotary_dim > 0:
            q, kv = self.rotary_emb(q, kv=kv, seqlen_offset=seqlen_offset)

        if past_key_values is not None:
            kv = _update_kv_cache(kv, past_key_values, self.layer_idx)

        if self.flash_attn:
            batch_size, seqlen_q = q.shape[0], q.shape[1]
            seqlen_k = kv.shape[1]

            cu_seqlens_q, cu_seqlens_k, max_seqlen_q, max_seqlen_k = (
                None,
                None,
                None,
                None,
            )
            if key_padding_mask is not None:
                kv, _, cu_seqlens_k, max_seqlen_k = unpad_input(kv, key_padding_mask)

                if seqlen_q == 1:
                    key_padding_mask = torch.ones(batch_size, 1, device=q.device)
                elif seqlen_q != seqlen_k:
                    key_padding_mask = key_padding_mask[:, -seqlen_q:]

                q, indices_q, cu_seqlens_q, max_seqlen_q = unpad_input(q, key_padding_mask)

            if self.checkpointing:
                attn_output = torch.utils.checkpoint.checkpoint(
                    self.inner_cross_attn,
                    q,
                    kv,
                    causal=causal,
                    cu_seqlens=cu_seqlens_q,
                    max_seqlen=max_seqlen_q,
                    cu_seqlens_k=cu_seqlens_k,
                    max_seqlen_k=max_seqlen_k,
                )
            else:
                attn_output = self.inner_cross_attn(
                    q,
                    kv,
                    causal=causal,
                    cu_seqlens=cu_seqlens_q,
                    max_seqlen=max_seqlen_q,
                    cu_seqlens_k=cu_seqlens_k,
                    max_seqlen_k=max_seqlen_k,
                )

            return (
                pad_input(attn_output, indices_q, batch_size, max_seqlen_q)
                if key_padding_mask is not None
                else attn_output
            )

        if self.checkpointing:
            return torch.utils.checkpoint.checkpoint(
                self.inner_cross_attn,
                q,
                kv,
                key_padding_mask=key_padding_mask,
                causal=causal,
            )

        return self.inner_cross_attn(q, kv, key_padding_mask=key_padding_mask, causal=causal)

    def forward(
        self,
        x: torch.FloatTensor,
        past_key_values: Optional[InferenceParams] = None,
        attention_mask: Optional[Union[torch.LongTensor, torch.BoolTensor]] = None,
        **kwargs,
    ) -> Tuple[torch.FloatTensor, torch.FloatTensor]:
        # TODO: Need an alternative way for dynamic control flow: torch.any(~attention_mask.bool())
        if attention_mask is not None:
            attention_mask = attention_mask.bool()
        else:
            attention_mask = None

        # MHA
        if self.n_head == self.n_head_kv:
            if past_key_values is None:
                # If `past_key_values` are not supplied, we run self-attention
                attn_output = self._forward_self_attn(x, attention_mask)
            else:
                # If `past_key_values` are supplied, it means that we might have cached values and
                # could take advantage of cross-attention
                attn_output = self._forward_cross_attn(x, past_key_values, attention_mask)
        # MQA / GQA
        else:
            # Regardless of `past_key_values` being supplied or not, it always use cross-attention
            # because `q` and `kv` lengths might be different
            attn_output = self._forward_cross_attn(x, past_key_values, attention_mask)

        output = rearrange(attn_output, "... h d -> ... (h d)")
        output = self.out_proj(output)

        return output if not self.return_residual else (output, x)


class ParallelBlock(nn.Module):
    """Parallel block.

    This block applies parallel mixer and MLP layers to the input (used in GPT-J and CodeGen).

    """

    def __init__(
        self,
        config: PretrainedConfig,
        block_idx: Optional[int] = None,
    ) -> None:
        super().__init__()

        self.ln = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
        self.resid_dropout = nn.Dropout(config.resid_pdrop)
        self.block_idx = block_idx

        self.mixer = MHA(config, layer_idx=block_idx)
        self.mlp = MLP(config)

    def forward(
        self,
        hidden_states: torch.FloatTensor,
        past_key_values: Optional[Union[torch.FloatTensor, InferenceParams]] = None,
        attention_mask: Optional[torch.BoolTensor] = None,
        **kwargs,
    ) -> torch.FloatTensor:
        residual = hidden_states
        hidden_states = self.ln(hidden_states)

        attn_outputs = self.mixer(
            hidden_states,
            past_key_values=past_key_values,
            attention_mask=attention_mask,
        )
        if isinstance(attn_outputs, tuple):
            attn_outputs = attn_outputs[0]

        attn_outputs = self.resid_dropout(attn_outputs)
        feed_forward_hidden_states = self.resid_dropout(self.mlp(hidden_states))

        hidden_states = attn_outputs + feed_forward_hidden_states + residual

        return hidden_states


class CausalLMHead(nn.Module):
    """Causal Language Modeling head.

    Reference:
        Improving Language Understanding by Generative Pre-Training.
        https://cdn.openai.com/research-covers/language-unsupervised/language_understanding_paper.pdf.

    """

    def __init__(self, config: PretrainedConfig) -> None:
        super().__init__()

        self.ln = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
        self.linear = nn.Linear(config.n_embd, config.vocab_size)

    def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:
        hidden_states = self.ln(hidden_states)
        logits = self.linear(hidden_states).to(torch.float32)

        return logits


class CausalLMLoss(nn.Module):
    """Causal Language Modeling loss.

    Reference:
        Improving Language Understanding by Generative Pre-Training.
        https://cdn.openai.com/research-covers/language-unsupervised/language_understanding_paper.pdf.

    """

    def __init__(self, shift_labels: bool = True) -> None:
        super().__init__()

        self.shift_labels = shift_labels
        self.loss_fct = nn.CrossEntropyLoss()

    def forward(self, logits: torch.FloatTensor, labels: torch.LongTensor) -> torch.FloatTensor:
        if self.shift_labels:
            logits = logits[..., :-1, :].contiguous()
            labels = labels[..., 1:].contiguous()

        loss = self.loss_fct(logits.view(-1, logits.size(-1)), labels.view(-1))

        return loss


class PhiPreTrainedModel(PreTrainedModel):
    """Phi pre-trained model."""

    config_class = PhiConfig
    base_model_prefix = "transformer"
    supports_gradient_checkpointing = False
    _no_split_modules = ["ParallelBlock"]

    def __init__(self, *inputs, **kwargs) -> None:
        super().__init__(*inputs, **kwargs)

    def _init_weights(self, module: nn.Module) -> None:
        if isinstance(module, (nn.Linear,)):
            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
            if module.bias is not None:
                module.bias.data.zero_()
        elif isinstance(module, nn.Embedding):
            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
            if module.padding_idx is not None:
                module.weight.data[module.padding_idx].zero_()
        elif isinstance(module, nn.LayerNorm):
            if module.bias is not None:
                module.bias.data.zero_()
            module.weight.data.fill_(1.0)

    def prepare_inputs_for_generation(
        self,
        input_ids: torch.LongTensor,
        past_key_values: Optional[Union[torch.FloatTensor, InferenceParams]] = None,
        attention_mask: Optional[Union[torch.LongTensor, torch.BoolTensor]] = None,
        **kwargs,
    ) -> Dict[str, Any]:
        if past_key_values is None or not (isinstance(past_key_values, InferenceParams)):
            past_key_values = InferenceParams(
                max_seqlen=self.config.n_positions,
                max_batch_size=input_ids.shape[0],
                seqlen_offset=0,
                batch_size_offset=0,
                key_value_memory_dict={},
                lengths_per_sample=None,
            )
        else:
            # Assume that `past_key_values` has cached all tokens up to the last token in `input_ids`
            past_key_values.seqlen_offset = len(input_ids[0]) - 1
            input_ids = input_ids[:, -1].unsqueeze(-1)

        return {
            "input_ids": input_ids,
            "past_key_values": past_key_values,
            "attention_mask": attention_mask,
        }


class PhiModel(PhiPreTrainedModel):
    """Phi model."""

    _keys_to_ignore_on_load_missing = [""]
    _keys_to_ignore_on_load_unexpected = [r"h\.\d+\.mlp.(fc_in|fc_out)\.(weight|bias)"]

    def __init__(self, config: PhiConfig) -> None:
        super().__init__(config)

        self.embd = Embedding(config)
        self.h = nn.ModuleList([ParallelBlock(config, block_idx=i) for i in range(config.n_layer)])
        self.gradient_checkpointing = False
        self.post_init()

    def get_input_embeddings(self) -> nn.Embedding:
        return self.embd.wte

    def set_input_embeddings(self, new_embeddings: nn.Embedding) -> None:
        self.embd.wte = new_embeddings

    def forward(
        self,
        input_ids: torch.LongTensor,
        past_key_values: Optional[Union[torch.FloatTensor, InferenceParams]] = None,
        attention_mask: Optional[torch.BoolTensor] = None,
    ) -> torch.FloatTensor:
        hidden_states = self.embd(input_ids)

        for layer in self.h:
            hidden_states = layer(
                hidden_states,
                past_key_values=past_key_values,
                attention_mask=attention_mask,
            )

        return hidden_states


class PhiForCausalLM(PhiPreTrainedModel):
    """Phi for Causal Language Modeling."""

    _keys_to_ignore_on_load_missing = [""]
    _keys_to_ignore_on_load_unexpected = [r"transformer\.h\.\d+\.mlp.(fc_in|fc_out)\.(weight|bias)"]

    def __init__(self, config: PhiConfig) -> None:
        super().__init__(config)

        self.transformer = PhiModel(config)
        self.lm_head = CausalLMHead(config)
        self.loss = CausalLMLoss()

        self.post_init()

    def get_output_embeddings(self) -> nn.Linear:
        return self.lm_head.linear

    def set_output_embeddings(self, new_embeddings: nn.Linear) -> None:
        self.lm_head.linear = new_embeddings

    def forward(
        self,
        input_ids: torch.LongTensor,
        past_key_values: Optional[Union[torch.FloatTensor, InferenceParams]] = None,
        attention_mask: Optional[torch.BoolTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        **kwargs,
    ) -> CausalLMOutputWithPast:
        hidden_states = self.transformer(input_ids, past_key_values=past_key_values, attention_mask=attention_mask)
        lm_logits = self.lm_head(hidden_states)

        loss = None
        if labels is not None:
            loss = self.loss(lm_logits, labels)

        return CausalLMOutputWithPast(loss=loss, logits=lm_logits, past_key_values=past_key_values)