File size: 23,408 Bytes
2646361
 
 
 
e3681c2
 
 
 
 
 
 
7ad815b
 
 
 
 
2646361
7ad815b
2646361
e3681c2
 
 
 
 
 
 
 
2646361
 
 
e3681c2
 
 
 
 
 
 
 
 
955fea2
2646361
 
 
 
 
 
 
 
955fea2
e3681c2
2646361
 
 
 
e3681c2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7ad815b
e3681c2
2646361
 
 
e3681c2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7ad815b
e3681c2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7ad815b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e3681c2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2646361
 
 
e3681c2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4434bf3
e3681c2
 
 
 
 
 
 
2646361
 
e3681c2
 
 
 
 
 
 
 
 
 
 
4434bf3
 
 
 
 
 
 
 
 
 
 
e3681c2
 
 
2646361
 
 
 
e3681c2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2646361
 
 
e3681c2
 
2646361
 
 
e3681c2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2646361
 
 
e3681c2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2646361
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
# This implementation was adapted from https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/layers/rotary.py
# Commit id: 3566596ad867ee415dd3c12616dd50c610176f6c
# Rotary varlen support from https://github.com/Dao-AILab/flash-attention/pull/556

# Copyright (c) 2023, Tri Dao.

import math
from typing import Optional, Tuple, Union

import torch
from einops import rearrange, repeat

if torch.cuda.is_available():
    try:
        from flash_attn.ops.triton.rotary import apply_rotary
    except ImportError:

        def apply_rotary(*args, **kwargs):
            raise RuntimeError("RoPE requires flash-attention to be installed")


def rotate_half(x, interleaved=False):
    if not interleaved:
        x1, x2 = x.chunk(2, dim=-1)
        return torch.cat((-x2, x1), dim=-1)
    else:
        x1, x2 = x[..., ::2], x[..., 1::2]
        return rearrange(
            torch.stack((-x2, x1), dim=-1), "... d two -> ... (d two)", two=2
        )


def apply_rotary_emb_torch(x, cos, sin, interleaved=False):
    """
    x: (batch_size, seqlen, nheads, headdim)
    cos, sin: (seqlen, rotary_dim / 2) or (batch_size, seqlen, rotary_dim / 2)
    """
    ro_dim = cos.shape[-1] * 2
    assert ro_dim <= x.shape[-1]
    cos, sin = (
        cos[: x.shape[1]],
        sin[: x.shape[1]],
    )
    cos = repeat(
        cos, "... d -> ... 1 (2 d)" if not interleaved else "... d -> ... 1 (d 2)"
    )
    sin = repeat(
        sin, "... d -> ... 1 (2 d)" if not interleaved else "... d -> ... 1 (d 2)"
    )
    return torch.cat(
        [
            x[..., :ro_dim] * cos + rotate_half(x[..., :ro_dim], interleaved) * sin,
            x[..., ro_dim:],
        ],
        dim=-1,
    )


class ApplyRotaryEmb(torch.autograd.Function):
    @staticmethod
    def forward(
        ctx,
        x,
        cos,
        sin,
        interleaved=False,
        inplace=False,
        seqlen_offsets: Union[int, torch.Tensor] = 0,
        cu_seqlens: Optional[torch.Tensor] = None,
        max_seqlen: Optional[int] = None,
    ):
        out = apply_rotary(
            x,
            cos,
            sin,
            seqlen_offsets=seqlen_offsets,
            cu_seqlens=cu_seqlens,
            max_seqlen=max_seqlen,
            interleaved=interleaved,
            inplace=inplace,
        )

        if isinstance(seqlen_offsets, int):
            ctx.save_for_backward(
                cos, sin, cu_seqlens
            )  # Can't save int with save_for_backward
            ctx.seqlen_offsets = seqlen_offsets
        else:
            ctx.save_for_backward(cos, sin, cu_seqlens, seqlen_offsets)
            ctx.seqlen_offsets = None
        ctx.interleaved = interleaved
        ctx.inplace = inplace
        ctx.max_seqlen = max_seqlen
        return out if not inplace else x

    @staticmethod
    def backward(ctx, do):
        seqlen_offsets = ctx.seqlen_offsets
        if seqlen_offsets is None:
            cos, sin, cu_seqlens, seqlen_offsets = ctx.saved_tensors
        else:
            cos, sin, cu_seqlens = ctx.saved_tensors
        # TD [2023-09-02]: For some reason Triton (2.0.0.post1) errors with
        # "[CUDA]: invalid device context", and cloning makes it work. Idk why. Triton 2.1.0 works.
        if not ctx.interleaved and not ctx.inplace:
            do = do.clone()

        dx = apply_rotary(
            do,
            cos,
            sin,
            seqlen_offsets=seqlen_offsets,
            cu_seqlens=cu_seqlens,
            max_seqlen=ctx.max_seqlen,
            interleaved=ctx.interleaved,
            inplace=ctx.inplace,
            conjugate=True,
        )
        return dx, None, None, None, None, None, None, None


def apply_rotary_emb(
    x,
    cos,
    sin,
    interleaved=False,
    inplace=False,
    seqlen_offsets: Union[int, torch.Tensor] = 0,
    cu_seqlens: Optional[torch.Tensor] = None,
    max_seqlen: Optional[int] = None,
):
    """
    Arguments:
        x: (batch_size, seqlen, nheads, headdim) if cu_seqlens is None
            else (total_seqlen, nheads, headdim)
        cos, sin: (seqlen_rotary, rotary_dim / 2)
        interleaved: if True, rotate pairs of even and odd dimensions (GPT-J style) instead
            of 1st half and 2nd half (GPT-NeoX style).
        inplace: if True, apply rotary embedding in-place.
        seqlen_offsets: (batch_size,) or int. Each sequence in x is shifted by this amount.
            Most commonly used in inference when we have KV cache.
        cu_seqlens: (batch + 1,) or None
        max_seqlen: int
    Return:
        out: (batch_size, seqlen, nheads, headdim) if cu_seqlens is None
            else (total_seqlen, nheads, headdim)
    rotary_dim must be <= headdim
    Apply rotary embedding to the first rotary_dim of x.
    """
    return ApplyRotaryEmb.apply(
        x, cos, sin, interleaved, inplace, seqlen_offsets, cu_seqlens, max_seqlen
    )


# For backward compatibility
apply_rotary_emb_func = apply_rotary_emb


class ApplyRotaryEmbQKV_(torch.autograd.Function):
    @staticmethod
    def forward(
        ctx,
        qkv,
        cos,
        sin,
        cos_k=None,
        sin_k=None,
        interleaved=False,
        seqlen_offsets: Union[int, torch.Tensor] = 0,
        cu_seqlens: Optional[torch.Tensor] = None,
        max_seqlen: Optional[int] = None,
    ):
        # batch, seqlen, three, nheads, headdim = qkv.shape
        assert qkv.shape[-3] == 3
        if cos_k is None and sin_k is None and qkv.is_contiguous():

            if torch.cuda.is_available():
                # Call 1 kernel instead of 2 kernels
                # We need qkv to be contiguous so that when we reshape to combine (3, nheads)
                # dimensions, we get the same tensor
                qk = rearrange(qkv[..., :2, :, :], "... t h d -> ... (t h) d")
                # qk = qkv[:, :, :2].reshape(batch, seqlen, -1, headdim)
                apply_rotary(
                    qk,
                    cos,
                    sin,
                    seqlen_offsets=seqlen_offsets,
                    interleaved=interleaved,
                    inplace=True,
                    cu_seqlens=cu_seqlens,
                    max_seqlen=max_seqlen,
                )
            else:
                q_rot = apply_rotary_emb_torch(
                    qkv[:, :, 0],
                    cos,
                    sin,
                    interleaved=interleaved,
                )
                k_rot = apply_rotary_emb_torch(
                    qkv[:, :, 1],
                    cos,
                    sin,
                    interleaved=interleaved,
                )
                qkv = torch.stack((q_rot, k_rot, qkv[:, :, 2]), dim=2)
        else:
            cos_k = cos if cos_k is None else cos_k
            sin_k = sin if sin_k is None else sin_k
            q, k = qkv[..., 0, :, :], qkv[..., 1, :, :]
            apply_rotary(
                q,
                cos,
                sin,
                seqlen_offsets,
                interleaved=interleaved,
                inplace=True,
                cu_seqlens=cu_seqlens,
                max_seqlen=max_seqlen,
            )
            apply_rotary(
                k,
                cos_k,
                sin_k,
                seqlen_offsets,
                interleaved=interleaved,
                inplace=True,
                cu_seqlens=cu_seqlens,
                max_seqlen=max_seqlen,
            )
            ctx.save_for_backward(cos, sin, cos_k, sin_k)
        if isinstance(seqlen_offsets, int):
            ctx.save_for_backward(cos, sin, cos_k, sin_k, cu_seqlens)
            ctx.seqlen_offsets = seqlen_offsets
        else:
            ctx.save_for_backward(cos, sin, cos_k, sin_k, cu_seqlens, seqlen_offsets)
            ctx.seqlen_offsets = None
        ctx.max_seqlen = max_seqlen
        ctx.interleaved = interleaved
        return qkv

    @staticmethod
    def backward(ctx, dqkv):
        seqlen_offsets = ctx.seqlen_offsets
        if seqlen_offsets is None:
            cos, sin, cos_k, sin_k, cu_seqlens, seqlen_offsets = ctx.saved_tensors
        else:
            cos, sin, cos_k, sin_k, cu_seqlens = ctx.saved_tensors
        if cos_k is None and sin_k is None and dqkv.is_contiguous():
            # Call 1 kernel instead of 2 kernels
            # We need dqkv to be contiguous so that when we reshape to combine (3, nheads)
            # dimensions, we get the same tensor
            dqk = rearrange(dqkv[..., :2, :, :], "... t h d -> ... (t h) d")
            apply_rotary(
                dqk,
                cos,
                sin,
                seqlen_offsets=seqlen_offsets,
                interleaved=ctx.interleaved,
                inplace=True,
                conjugate=True,
                cu_seqlens=cu_seqlens,
                max_seqlen=ctx.max_seqlen,
            )
        else:
            cos_k = cos if cos_k is None else cos_k
            sin_k = sin if sin_k is None else sin_k
            dq, dk = dqkv[..., 0, :, :], dqkv[..., 1, :, :]
            apply_rotary(
                dq,
                cos,
                sin,
                seqlen_offsets,
                interleaved=ctx.interleaved,
                inplace=True,
                conjugate=True,
                cu_seqlens=cu_seqlens,
                max_seqlen=ctx.max_seqlen,
            )
            apply_rotary(
                dk,
                cos_k,
                sin_k,
                seqlen_offsets,
                interleaved=ctx.interleaved,
                inplace=True,
                conjugate=True,
                cu_seqlens=cu_seqlens,
                max_seqlen=ctx.max_seqlen,
            )
        return dqkv, None, None, None, None, None, None, None, None


def apply_rotary_emb_qkv_(
    qkv,
    cos,
    sin,
    cos_k=None,
    sin_k=None,
    interleaved=False,
    seqlen_offsets: Union[int, torch.Tensor] = 0,
    cu_seqlens: Optional[torch.Tensor] = None,
    max_seqlen: Optional[int] = None,
):
    """
    Arguments:
        qkv: (batch_size, seqlen, 3, nheads, headdim) if cu_seqlens is None
            else (total_seqlen, 3, nheads, headdim)
        cos, sin: (seqlen, rotary_dim / 2)
        cos_k, sin_k: (seqlen, rotary_dim / 2), optional
        interleaved: if True, rotate pairs of even and odd dimensions (GPT-J style) instead of
            1st half and 2nd half (GPT-NeoX style).
        seqlen_offsets: (batch_size,) or int. Each sequence in Q and K is shifted by this amount.
            Most commonly used in inference when we have KV cache.
            cu_seqlens: (batch + 1,) or None
        max_seqlen: int
    Return:
        qkv: (batch_size, seqlen, 3, nheads, headdim) if cu_seqlens is None
            else (total_seqlen, 3, nheads, headdim)
    rotary_dim must be <= headdim
    Apply rotary embedding *inplace* to the first rotary_dim of Q and K.
    """
    return ApplyRotaryEmbQKV_.apply(
        qkv, cos, sin, cos_k, sin_k, interleaved, seqlen_offsets, cu_seqlens, max_seqlen
    )


class ApplyRotaryEmbKV_(torch.autograd.Function):
    @staticmethod
    def forward(
        ctx,
        kv,
        cos,
        sin,
        interleaved=False,
        seqlen_offsets: Union[int, torch.Tensor] = 0,
        cu_seqlens: Optional[torch.Tensor] = None,
        max_seqlen: Optional[int] = None,
    ):
        # batch, seqlen, two, nheads, headdim = kv.shape
        assert kv.shape[-3] == 2
        k = kv[..., 0, :, :]
        apply_rotary(
            k,
            cos,
            sin,
            seqlen_offsets=seqlen_offsets,
            interleaved=interleaved,
            inplace=True,
            cu_seqlens=cu_seqlens,
            max_seqlen=max_seqlen,
        )
        if isinstance(seqlen_offsets, int):
            ctx.save_for_backward(
                cos, sin, cu_seqlens
            )  # Can't save int with save_for_backward
            ctx.seqlen_offsets = seqlen_offsets
        else:
            ctx.save_for_backward(cos, sin, cu_seqlens, seqlen_offsets)
            ctx.seqlen_offsets = None
        ctx.max_seqlen = max_seqlen
        ctx.interleaved = interleaved
        return kv

    @staticmethod
    def backward(ctx, dkv):
        seqlen_offsets = ctx.seqlen_offsets
        if seqlen_offsets is None:
            cos, sin, cu_seqlens, seqlen_offsets = ctx.saved_tensors
        else:
            cos, sin, cu_seqlens = ctx.saved_tensors
        apply_rotary(
            dkv[..., 0, :, :],
            cos,
            sin,
            seqlen_offsets=seqlen_offsets,
            interleaved=ctx.interleaved,
            inplace=True,
            conjugate=True,
            cu_seqlens=cu_seqlens,
            max_seqlen=ctx.max_seqlen,
        )
        return dkv, None, None, None, None, None, None


apply_rotary_emb_kv_ = ApplyRotaryEmbKV_.apply


def apply_rotary_emb_kv_(
    kv,
    cos,
    sin,
    interleaved=False,
    seqlen_offsets: Union[int, torch.Tensor] = 0,
    cu_seqlens: Optional[torch.Tensor] = None,
    max_seqlen: Optional[int] = None,
):
    """
    Arguments:
        kv: (batch_size, seqlen, 2, nheads, headdim) if cu_seqlens is None
            else (total_seqlen, 2, nheads, headdim)
        cos, sin: (seqlen, rotary_dim / 2)
        interleaved: if True, rotate pairs of even and odd dimensions (GPT-J style) instead of
            1st half and 2nd half (GPT-NeoX style).
        seqlen_offsets: (batch_size,) or int. Each sequence in Q and K is shifted by this amount.
            Most commonly used in inference when we have KV cache.
        cu_seqlens: (batch + 1,) or None
        max_seqlen: int
    Return:
        kv: (batch_size, seqlen, 2, nheads, headdim) if cu_seqlens is None
            else (total_seqlen, 2, nheads, headdim)
    rotary_dim must be <= headdim
    Apply rotary embedding *inplace* to the first rotary_dim of K.
    """
    return ApplyRotaryEmbKV_.apply(
        kv, cos, sin, interleaved, seqlen_offsets, cu_seqlens, max_seqlen
    )


class RotaryEmbedding(torch.nn.Module):
    """
    The rotary position embeddings from RoFormer_ (Su et. al).
    A crucial insight from the method is that the query and keys are
    transformed by rotation matrices which depend on the relative positions.

    Other implementations are available in the Rotary Transformer repo_ and in
    GPT-NeoX_, GPT-NeoX was an inspiration

    .. _RoFormer: https://arxiv.org/abs/2104.09864
    .. _repo: https://github.com/ZhuiyiTechnology/roformer
    .. _GPT-NeoX: https://github.com/EleutherAI/gpt-neox

    If scale_base is not None, this implements XPos (Sun et al., https://arxiv.org/abs/2212.10554).
    A recommended value for scale_base is 512: https://github.com/HazyResearch/flash-attention/issues/96
    Reference: https://github.com/sunyt32/torchscale/blob/main/torchscale/component/xpos_relative_position.py
    """

    def __init__(
        self,
        dim: int,
        base=10000.0,
        interleaved=False,
        scale_base=None,
        pos_idx_in_fp32=True,
        device=None,
    ):
        """
        interleaved: if True, rotate pairs of even and odd dimensions (GPT-J style) instead
            of 1st half and 2nd half (GPT-NeoX style).
        pos_idx_in_fp32: if True, the position indices [0.0, ..., seqlen - 1] are in fp32,
            otherwise they might be in lower precision.
            This option was added because previously (before 2023-07-02), when we construct
            the position indices, we use the dtype of self.inv_freq. In most cases this would
            be fp32, but if the model is trained in pure bf16 (not mixed precision), then
            self.inv_freq would be bf16, and the position indices are also in bf16.
            Because of the limited precision of bf16 (e.g. 1995.0 is rounded to 2000.0), the
            embeddings for some positions will coincide.
            To maintain compatibility with models previously trained in pure bf16,
            we add this option.
        """
        super().__init__()
        self.dim = dim
        self._base = float(base)
        self.pos_idx_in_fp32 = pos_idx_in_fp32
        # 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)
        self.interleaved = interleaved
        self.scale_base = scale_base
        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)

        self._seq_len_cached = 0
        self._cos_cached = None
        self._sin_cached = None
        self._cos_k_cached = None
        self._sin_k_cached = None

    @property
    def base(self):
        return self._base

    @base.setter
    def base(self, new_base):
        if new_base > 0:
            self._base = float(new_base)
        else:
            raise ValueError("Rotary base value must be positive")

    def _compute_inv_freq(self, device=None):
        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, device=None, dtype=None):
        # Reset the tables if the sequence length has changed,
        # if we're on a new device (possibly due to tracing for instance),
        # or if we're switching from inference mode to training
        if (
            seqlen > self._seq_len_cached
            or self._cos_cached is None
            or self._cos_cached.device != device
            or self._cos_cached.dtype != dtype
            or (self.training and self._cos_cached.is_inference())
        ):
            self._seq_len_cached = seqlen
            # We want fp32 here, not self.inv_freq.dtype, since the model could be loaded in bf16
            # And the output of arange can be quite large, so bf16 would lose a lot of precision.
            # However, for compatibility reason, we add an option to use the dtype of self.inv_freq.
            if self.pos_idx_in_fp32:
                t = torch.arange(seqlen, device=device, dtype=torch.float32)
                # We want fp32 here as well since inv_freq will be multiplied with t, and the output
                # will be large. Having it in bf16 will lose a lot of precision and cause the
                # cos & sin output to change significantly.
                # We want to recompute self.inv_freq if it was not loaded in fp32
                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
            # Don't do einsum, it converts fp32 to fp16 under AMP
            # freqs = torch.einsum("i,j->ij", t, self.inv_freq)
            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"
                )
                # We want the multiplication by scale 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: Union[int, torch.Tensor] = 0,
        cu_seqlens: Optional[torch.Tensor] = None,
        max_seqlen: Optional[int] = None,
    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
        """
        qkv: (batch, seqlen, 3, nheads, headdim) if kv is none,
             else it's just q of shape (batch, seqlen, nheads, headdim)
        kv: (batch, seqlen, 2, nheads, headdim)
        seqlen_offset: (batch_size,) or int. Each sequence in x is shifted by this amount.
            Most commonly used in inference when we have KV cache.
            If it's a tensor of shape (batch_size,), then to update the cos / sin cache, one
            should pass in max_seqlen, which will update the cos / sin cache up to that length.
        Apply rotary embedding *inplace* to qkv and / or kv.
        """
        if cu_seqlens is not None:
            assert max_seqlen is not None
        seqlen = qkv.shape[1] if max_seqlen is None else max_seqlen
        if max_seqlen is not None:
            self._update_cos_sin_cache(max_seqlen, device=qkv.device, dtype=qkv.dtype)
        elif isinstance(seqlen_offset, int):
            self._update_cos_sin_cache(
                seqlen + seqlen_offset, device=qkv.device, dtype=qkv.dtype
            )
        if kv is None:
            if self.scale is None:
                return apply_rotary_emb_qkv_(
                    qkv,
                    self._cos_cached,
                    self._sin_cached,
                    interleaved=self.interleaved,
                    seqlen_offsets=seqlen_offset,
                    cu_seqlens=cu_seqlens,
                    max_seqlen=max_seqlen,
                )
            else:
                return apply_rotary_emb_qkv_(
                    qkv,
                    self._cos_cached,
                    self._sin_cached,
                    self._cos_k_cached,
                    self._sin_k_cached,
                    interleaved=self.interleaved,
                    seqlen_offsets=seqlen_offset,
                    cu_seqlens=cu_seqlens,
                    max_seqlen=max_seqlen,
                )
        else:
            q = qkv
            q = apply_rotary_emb_func(
                q,
                self._cos_cached,
                self._sin_cached,
                interleaved=self.interleaved,
                inplace=True,
                seqlen_offsets=seqlen_offset,
                cu_seqlens=cu_seqlens,
                max_seqlen=max_seqlen,
            )
            if self.scale is None:
                kv = apply_rotary_emb_kv_(
                    kv,
                    self._cos_cached,
                    self._sin_cached,
                    interleaved=self.interleaved,
                    seqlen_offsets=seqlen_offset,
                    cu_seqlens=cu_seqlens,
                    max_seqlen=max_seqlen,
                )
            else:
                kv = apply_rotary_emb_kv_(
                    kv,
                    self._cos_k_cached,
                    self._sin_k_cached,
                    interleaved=self.interleaved,
                    seqlen_offsets=seqlen_offset,
                    cu_seqlens=cu_seqlens,
                    max_seqlen=max_seqlen,
                )
            return q, kv