flash_attn_interface.py 27.4 KB
Newer Older
Tri Dao's avatar
Tri Dao committed
1
import flash_attn_2_cuda as flash_attn_cuda
Tri Dao's avatar
Tri Dao committed
2
3
import torch
import torch.nn as nn
Tri Dao's avatar
Tri Dao committed
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
from einops import rearrange


def _get_block_size(device, head_dim, is_dropout, is_causal):
    # This should match the block sizes in the CUDA kernel
    assert head_dim <= 256
    major, minor = torch.cuda.get_device_capability(device)
    is_sm8x = major == 8 and minor > 0  # Only include sm86 and sm89, exclude sm80 (A100)
    is_sm80 = major == 8 and minor == 0
    is_sm90 = major == 9 and minor == 0
    if head_dim <= 32:
        return 128, 128
    if head_dim <= 64:
        return (128, 128) if not is_dropout else (128, 64)
    elif head_dim <= 96:
        return (64, 64) if (is_sm8x and is_causal) else (128, 64)
    elif head_dim <= 128:
        if is_sm8x:
            return (64, 64) if (not is_dropout and is_causal) else (128, 32)
        else:
            return 128, (64 if not is_dropout else 32)
    elif head_dim <= 160:
        if is_sm8x:
            return (128, 64) if not is_causal else (64, 64)
        else:
            return 128, 32
    elif head_dim <= 192:
        return (128, 64) if not is_dropout else (64, 64)
    elif head_dim <= 224:
        return (128, 64) if (is_sm80 or is_sm90) else (64, 64)
    elif head_dim <= 256:
        return (128, 64) if is_sm80 else (64, 64)


def _flash_attn_forward(q, k, v, dropout_p, softmax_scale, causal, return_softmax):
Tri Dao's avatar
Tri Dao committed
39
40
    maybe_contiguous = lambda x: x.contiguous() if x.stride(-1) != 1 else x
    q, k, v = [maybe_contiguous(x) for x in (q, k, v)]
41
    out, q, k, v, out_padded, softmax_lse, S_dmask, rng_state = flash_attn_cuda.fwd(
Tri Dao's avatar
Tri Dao committed
42
43
        q, k, v, None, dropout_p, softmax_scale, causal, return_softmax, None
    )
44
    return out, q, k, v, out_padded, softmax_lse, S_dmask, rng_state
Tri Dao's avatar
Tri Dao committed
45
46


Tri Dao's avatar
Tri Dao committed
47
48
49
50
51
52
53
54
55
56
57
58
59
def _flash_attn_varlen_forward(
    q,
    k,
    v,
    cu_seqlens_q,
    cu_seqlens_k,
    max_seqlen_q,
    max_seqlen_k,
    dropout_p,
    softmax_scale,
    causal,
    return_softmax,
):
Tri Dao's avatar
Tri Dao committed
60
61
    maybe_contiguous = lambda x: x.contiguous() if x.stride(-1) != 1 else x
    q, k, v = [maybe_contiguous(x) for x in (q, k, v)]
62
    out, q, k, v, out_padded, softmax_lse, S_dmask, rng_state = flash_attn_cuda.varlen_fwd(
Tri Dao's avatar
Tri Dao committed
63
64
65
66
67
68
69
70
71
72
73
74
75
76
        q,
        k,
        v,
        None,
        cu_seqlens_q,
        cu_seqlens_k,
        max_seqlen_q,
        max_seqlen_k,
        dropout_p,
        softmax_scale,
        False,
        causal,
        return_softmax,
        None,
Tri Dao's avatar
Tri Dao committed
77
78
    )
    # if out.isnan().any() or softmax_lse.isnan().any():
Tri Dao's avatar
Tri Dao committed
79
    #     breakpoint()
80
    return out, q, k, v, out_padded, softmax_lse, S_dmask, rng_state
Tri Dao's avatar
Tri Dao committed
81
82


Tri Dao's avatar
Tri Dao committed
83
84
85
def _flash_attn_backward(
    dout, q, k, v, out, softmax_lse, dq, dk, dv, dropout_p, softmax_scale, causal, rng_state=None
):
Tri Dao's avatar
Tri Dao committed
86
87
88
    maybe_contiguous = lambda x: x.contiguous() if x.stride(-1) != 1 else x
    # dq, dk, dv are allocated by us so they should already be contiguous
    dout, q, k, v, out = [maybe_contiguous(x) for x in (dout, q, k, v, out)]
Tri Dao's avatar
Tri Dao committed
89
    dq, dk, dv, softmax_d, = flash_attn_cuda.bwd(
Tri Dao's avatar
Tri Dao committed
90
91
92
93
94
95
96
97
98
99
100
101
102
103
        dout,
        q,
        k,
        v,
        out,
        softmax_lse,
        dq,
        dk,
        dv,
        dropout_p,
        softmax_scale,
        causal,
        None,
        rng_state,
Tri Dao's avatar
Tri Dao committed
104
105
106
107
    )
    return dq, dk, dv, softmax_d


Tri Dao's avatar
Tri Dao committed
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
def _flash_attn_varlen_backward(
    dout,
    q,
    k,
    v,
    out,
    softmax_lse,
    dq,
    dk,
    dv,
    cu_seqlens_q,
    cu_seqlens_k,
    max_seqlen_q,
    max_seqlen_k,
    dropout_p,
    softmax_scale,
    causal,
    rng_state=None,
):
Tri Dao's avatar
Tri Dao committed
127
128
129
    maybe_contiguous = lambda x: x.contiguous() if x.stride(-1) != 1 else x
    # dq, dk, dv are allocated by us so they should already be contiguous
    dout, q, k, v, out = [maybe_contiguous(x) for x in (dout, q, k, v, out)]
Tri Dao's avatar
Tri Dao committed
130
    dq, dk, dv, softmax_d, = flash_attn_cuda.varlen_bwd(
Tri Dao's avatar
Tri Dao committed
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
        dout,
        q,
        k,
        v,
        out,
        softmax_lse,
        dq,
        dk,
        dv,
        cu_seqlens_q,
        cu_seqlens_k,
        max_seqlen_q,
        max_seqlen_k,
        dropout_p,
        softmax_scale,
        False,
        causal,
        None,
        rng_state,
Tri Dao's avatar
Tri Dao committed
150
    )
Tri Dao's avatar
Tri Dao committed
151
    # if dk.isnan().any() or dk.isnan().any() or dv.isnan().any() or softmax_d.isnan().any():
Tri Dao's avatar
Tri Dao committed
152
    #     breakpoint()
Tri Dao's avatar
Tri Dao committed
153
    return dq, dk, dv, softmax_d
Tri Dao's avatar
Tri Dao committed
154
155


Tri Dao's avatar
Tri Dao committed
156
class FlashAttnQKVPackedFunc(torch.autograd.Function):
Tri Dao's avatar
Tri Dao committed
157
    @staticmethod
Tri Dao's avatar
Tri Dao committed
158
    def forward(ctx, qkv, dropout_p, softmax_scale, causal, return_softmax):
Tri Dao's avatar
Tri Dao committed
159
160
        if softmax_scale is None:
            softmax_scale = qkv.shape[-1] ** (-0.5)
161
        out, q, k, v, out_padded, softmax_lse, S_dmask, rng_state = _flash_attn_forward(
Tri Dao's avatar
Tri Dao committed
162
163
164
165
166
167
168
            qkv[:, :, 0],
            qkv[:, :, 1],
            qkv[:, :, 2],
            dropout_p,
            softmax_scale,
            causal=causal,
            return_softmax=return_softmax and dropout_p > 0,
Tri Dao's avatar
Tri Dao committed
169
        )
Tri Dao's avatar
Tri Dao committed
170
        ctx.save_for_backward(q, k, v, out_padded, softmax_lse, rng_state)
Tri Dao's avatar
Tri Dao committed
171
172
173
        ctx.dropout_p = dropout_p
        ctx.softmax_scale = softmax_scale
        ctx.causal = causal
Tri Dao's avatar
Tri Dao committed
174
        return out if not return_softmax else (out, softmax_lse, S_dmask)
Tri Dao's avatar
Tri Dao committed
175
176

    @staticmethod
Tri Dao's avatar
Tri Dao committed
177
    def backward(ctx, dout, *args):
Tri Dao's avatar
Tri Dao committed
178
179
180
        q, k, v, out, softmax_lse, rng_state = ctx.saved_tensors
        qkv_shape = q.shape[:-2] + (3, *q.shape[-2:])
        dqkv = torch.empty(qkv_shape, dtype=q.dtype, device=q.device)
Tri Dao's avatar
Tri Dao committed
181
        _flash_attn_backward(
Tri Dao's avatar
Tri Dao committed
182
183
184
185
186
187
188
189
190
191
192
193
194
            dout,
            q,
            k,
            v,
            out,
            softmax_lse,
            dqkv[:, :, 0],
            dqkv[:, :, 1],
            dqkv[:, :, 2],
            ctx.dropout_p,
            ctx.softmax_scale,
            ctx.causal,
            rng_state=rng_state,
Tri Dao's avatar
Tri Dao committed
195
        )
Tri Dao's avatar
Tri Dao committed
196
        dqkv = dqkv[..., : dout.shape[-1]]  # We could have padded the head dimension
Tri Dao's avatar
Tri Dao committed
197
198
199
200
201
202
203
204
        return dqkv, None, None, None, None


class FlashAttnVarlenQKVPackedFunc(torch.autograd.Function):
    @staticmethod
    def forward(ctx, qkv, cu_seqlens, max_seqlen, dropout_p, softmax_scale, causal, return_softmax):
        if softmax_scale is None:
            softmax_scale = qkv.shape[-1] ** (-0.5)
205
        out, q, k, v, out_padded, softmax_lse, S_dmask, rng_state = _flash_attn_varlen_forward(
Tri Dao's avatar
Tri Dao committed
206
207
208
209
210
211
212
213
214
215
216
            qkv[:, 0],
            qkv[:, 1],
            qkv[:, 2],
            cu_seqlens,
            cu_seqlens,
            max_seqlen,
            max_seqlen,
            dropout_p,
            softmax_scale,
            causal=causal,
            return_softmax=return_softmax and dropout_p > 0,
Tri Dao's avatar
Tri Dao committed
217
        )
Tri Dao's avatar
Tri Dao committed
218
219
220
221
222
223
224
225
226
227
228
229
230
        ctx.save_for_backward(q, k, v, out_padded, softmax_lse, cu_seqlens, rng_state)
        ctx.dropout_p = dropout_p
        ctx.max_seqlen = max_seqlen
        ctx.softmax_scale = softmax_scale
        ctx.causal = causal
        return out if not return_softmax else (out, softmax_lse, S_dmask)

    @staticmethod
    def backward(ctx, dout, *args):
        q, k, v, out, softmax_lse, cu_seqlens, rng_state = ctx.saved_tensors
        qkv_shape = q.shape[:-2] + (3, *q.shape[-2:])
        dqkv = torch.empty(qkv_shape, dtype=q.dtype, device=q.device)
        _flash_attn_varlen_backward(
Tri Dao's avatar
Tri Dao committed
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
            dout,
            q,
            k,
            v,
            out,
            softmax_lse,
            dqkv[:, 0],
            dqkv[:, 1],
            dqkv[:, 2],
            cu_seqlens,
            cu_seqlens,
            ctx.max_seqlen,
            ctx.max_seqlen,
            ctx.dropout_p,
            ctx.softmax_scale,
            ctx.causal,
            rng_state=rng_state,
Tri Dao's avatar
Tri Dao committed
248
        )
Tri Dao's avatar
Tri Dao committed
249
        dqkv = dqkv[..., : dout.shape[-1]]  # We could have padded the head dimension
Tri Dao's avatar
Tri Dao committed
250
        return dqkv, None, None, None, None, None, None
Tri Dao's avatar
Tri Dao committed
251
252


Tri Dao's avatar
Tri Dao committed
253
class FlashAttnKVPackedFunc(torch.autograd.Function):
Tri Dao's avatar
Tri Dao committed
254
    @staticmethod
Tri Dao's avatar
Tri Dao committed
255
    def forward(ctx, q, kv, dropout_p, softmax_scale, causal, return_softmax):
Tri Dao's avatar
Tri Dao committed
256
        if softmax_scale is None:
Tri Dao's avatar
Tri Dao committed
257
            softmax_scale = q.shape[-1] ** (-0.5)
258
        out, q, k, v, out_padded, softmax_lse, S_dmask, rng_state = _flash_attn_forward(
Tri Dao's avatar
Tri Dao committed
259
260
261
262
263
264
265
            q,
            kv[:, :, 0],
            kv[:, :, 1],
            dropout_p,
            softmax_scale,
            causal=causal,
            return_softmax=return_softmax and dropout_p > 0,
Tri Dao's avatar
Tri Dao committed
266
        )
Tri Dao's avatar
Tri Dao committed
267
        ctx.save_for_backward(q, k, v, out_padded, softmax_lse, rng_state)
Tri Dao's avatar
Tri Dao committed
268
269
270
271
272
273
274
        ctx.dropout_p = dropout_p
        ctx.softmax_scale = softmax_scale
        ctx.causal = causal
        return out if not return_softmax else (out, softmax_lse, S_dmask)

    @staticmethod
    def backward(ctx, dout, *args):
Tri Dao's avatar
Tri Dao committed
275
        q, k, v, out, softmax_lse, rng_state = ctx.saved_tensors
Tri Dao's avatar
Tri Dao committed
276
        dq = torch.empty_like(q)
Tri Dao's avatar
Tri Dao committed
277
278
        kv_shape = k.shape[:-2] + (2, *k.shape[-2:])
        dkv = torch.empty(kv_shape, dtype=k.dtype, device=k.device)
Tri Dao's avatar
Tri Dao committed
279
        _flash_attn_backward(
Tri Dao's avatar
Tri Dao committed
280
281
282
283
284
285
286
287
288
289
290
291
292
            dout,
            q,
            k,
            v,
            out,
            softmax_lse,
            dq,
            dkv[:, :, 0],
            dkv[:, :, 1],
            ctx.dropout_p,
            ctx.softmax_scale,
            ctx.causal,
            rng_state=rng_state,
Tri Dao's avatar
Tri Dao committed
293
        )
Tri Dao's avatar
Tri Dao committed
294
295
        dq = dq[..., : dout.shape[-1]]  # We could have padded the head dimension
        dkv = dkv[..., : dout.shape[-1]]
Tri Dao's avatar
Tri Dao committed
296
        return dq, dkv, None, None, None, None
Tri Dao's avatar
Tri Dao committed
297
298


Tri Dao's avatar
Tri Dao committed
299
class FlashAttnVarlenKVPackedFunc(torch.autograd.Function):
Tri Dao's avatar
Tri Dao committed
300
    @staticmethod
Tri Dao's avatar
Tri Dao committed
301
302
303
304
305
306
307
308
309
310
311
312
313
    def forward(
        ctx,
        q,
        kv,
        cu_seqlens_q,
        cu_seqlens_k,
        max_seqlen_q,
        max_seqlen_k,
        dropout_p,
        softmax_scale,
        causal,
        return_softmax,
    ):
Tri Dao's avatar
Tri Dao committed
314
315
        if softmax_scale is None:
            softmax_scale = q.shape[-1] ** (-0.5)
316
        out, q, k, v, out_padded, softmax_lse, S_dmask, rng_state = _flash_attn_varlen_forward(
Tri Dao's avatar
Tri Dao committed
317
318
319
320
321
322
323
324
325
326
327
328
329
330
            q,
            kv[:, 0],
            kv[:, 1],
            cu_seqlens_q,
            cu_seqlens_k,
            max_seqlen_q,
            max_seqlen_k,
            dropout_p,
            softmax_scale,
            causal=causal,
            return_softmax=return_softmax and dropout_p > 0,
        )
        ctx.save_for_backward(
            q, k, v, out_padded, softmax_lse, cu_seqlens_q, cu_seqlens_k, rng_state
Tri Dao's avatar
Tri Dao committed
331
332
        )
        ctx.dropout_p = dropout_p
Tri Dao's avatar
Tri Dao committed
333
334
        ctx.max_seqlen_q = max_seqlen_q
        ctx.max_seqlen_k = max_seqlen_k
Tri Dao's avatar
Tri Dao committed
335
336
        ctx.softmax_scale = softmax_scale
        ctx.causal = causal
Tri Dao's avatar
Tri Dao committed
337
        return out if not return_softmax else (out, softmax_lse, S_dmask)
Tri Dao's avatar
Tri Dao committed
338
339

    @staticmethod
Tri Dao's avatar
Tri Dao committed
340
341
    def backward(ctx, dout, *args):
        q, k, v, out, softmax_lse, cu_seqlens_q, cu_seqlens_k, rng_state = ctx.saved_tensors
Tri Dao's avatar
Tri Dao committed
342
343
344
345
        dq = torch.empty_like(q)
        kv_shape = k.shape[:-2] + (2, *k.shape[-2:])
        dkv = torch.empty(kv_shape, dtype=k.dtype, device=k.device)
        _flash_attn_varlen_backward(
Tri Dao's avatar
Tri Dao committed
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
            dout,
            q,
            k,
            v,
            out,
            softmax_lse,
            dq,
            dkv[:, 0],
            dkv[:, 1],
            cu_seqlens_q,
            cu_seqlens_k,
            ctx.max_seqlen_q,
            ctx.max_seqlen_k,
            ctx.dropout_p,
            ctx.softmax_scale,
            ctx.causal,
            rng_state=rng_state,
Tri Dao's avatar
Tri Dao committed
363
        )
Tri Dao's avatar
Tri Dao committed
364
365
        dq = dq[..., : dout.shape[-1]]  # We could have padded the head dimension
        dkv = dkv[..., : dout.shape[-1]]
Tri Dao's avatar
Tri Dao committed
366
367
368
369
370
371
372
373
        return dq, dkv, None, None, None, None, None, None, None, None


class FlashAttnFunc(torch.autograd.Function):
    @staticmethod
    def forward(ctx, q, k, v, dropout_p, softmax_scale, causal, return_softmax):
        if softmax_scale is None:
            softmax_scale = q.shape[-1] ** (-0.5)
374
        out, q, k, v, out_padded, softmax_lse, S_dmask, rng_state = _flash_attn_forward(
Tri Dao's avatar
Tri Dao committed
375
376
377
378
379
380
381
            q,
            k,
            v,
            dropout_p,
            softmax_scale,
            causal=causal,
            return_softmax=return_softmax and dropout_p > 0,
Tri Dao's avatar
Tri Dao committed
382
383
384
385
386
387
388
389
390
391
        )
        ctx.save_for_backward(q, k, v, out_padded, softmax_lse, rng_state)
        ctx.dropout_p = dropout_p
        ctx.softmax_scale = softmax_scale
        ctx.causal = causal
        return out if not return_softmax else (out, softmax_lse, S_dmask)

    @staticmethod
    def backward(ctx, dout, *args):
        q, k, v, out, softmax_lse, rng_state = ctx.saved_tensors
Tri Dao's avatar
Tri Dao committed
392
393
        dq, dk, dv = torch.empty_like(q), torch.empty_like(k), torch.empty_like(v)
        _flash_attn_backward(
Tri Dao's avatar
Tri Dao committed
394
395
396
397
398
399
400
401
402
403
404
405
406
            dout,
            q,
            k,
            v,
            out,
            softmax_lse,
            dq,
            dk,
            dv,
            ctx.dropout_p,
            ctx.softmax_scale,
            ctx.causal,
            rng_state=rng_state,
Tri Dao's avatar
Tri Dao committed
407
        )
Tri Dao's avatar
Tri Dao committed
408
409
410
        dq = dq[..., : dout.shape[-1]]  # We could have padded the head dimension
        dk = dk[..., : dout.shape[-1]]
        dv = dv[..., : dout.shape[-1]]
Tri Dao's avatar
Tri Dao committed
411
        return dq, dk, dv, None, None, None, None, None, None, None, None
Tri Dao's avatar
Tri Dao committed
412
413


Tri Dao's avatar
Tri Dao committed
414
class FlashAttnVarlenFunc(torch.autograd.Function):
415
    @staticmethod
Tri Dao's avatar
Tri Dao committed
416
417
418
419
420
421
422
423
424
425
426
427
428
429
    def forward(
        ctx,
        q,
        k,
        v,
        cu_seqlens_q,
        cu_seqlens_k,
        max_seqlen_q,
        max_seqlen_k,
        dropout_p,
        softmax_scale,
        causal,
        return_softmax,
    ):
430
        if softmax_scale is None:
Tri Dao's avatar
Tri Dao committed
431
            softmax_scale = q.shape[-1] ** (-0.5)
432
        out, q, k, v, out_padded, softmax_lse, S_dmask, rng_state = _flash_attn_varlen_forward(
Tri Dao's avatar
Tri Dao committed
433
434
435
436
437
438
439
440
441
442
443
444
445
446
            q,
            k,
            v,
            cu_seqlens_q,
            cu_seqlens_k,
            max_seqlen_q,
            max_seqlen_k,
            dropout_p,
            softmax_scale,
            causal=causal,
            return_softmax=return_softmax and dropout_p > 0,
        )
        ctx.save_for_backward(
            q, k, v, out_padded, softmax_lse, cu_seqlens_q, cu_seqlens_k, rng_state
447
448
        )
        ctx.dropout_p = dropout_p
Tri Dao's avatar
Tri Dao committed
449
450
        ctx.max_seqlen_q = max_seqlen_q
        ctx.max_seqlen_k = max_seqlen_k
451
452
        ctx.softmax_scale = softmax_scale
        ctx.causal = causal
Tri Dao's avatar
Tri Dao committed
453
        return out if not return_softmax else (out, softmax_lse, S_dmask)
454
455
456

    @staticmethod
    def backward(ctx, dout, *args):
Tri Dao's avatar
Tri Dao committed
457
458
459
        q, k, v, out, softmax_lse, cu_seqlens_q, cu_seqlens_k, rng_state = ctx.saved_tensors
        dq, dk, dv = torch.empty_like(q), torch.empty_like(k), torch.empty_like(v)
        _flash_attn_varlen_backward(
Tri Dao's avatar
Tri Dao committed
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
            dout,
            q,
            k,
            v,
            out,
            softmax_lse,
            dq,
            dk,
            dv,
            cu_seqlens_q,
            cu_seqlens_k,
            ctx.max_seqlen_q,
            ctx.max_seqlen_k,
            ctx.dropout_p,
            ctx.softmax_scale,
            ctx.causal,
            rng_state=rng_state,
477
        )
Tri Dao's avatar
Tri Dao committed
478
479
480
        dq = dq[..., : dout.shape[-1]]  # We could have padded the head dimension
        dk = dk[..., : dout.shape[-1]]
        dv = dv[..., : dout.shape[-1]]
Tri Dao's avatar
Tri Dao committed
481
        return dq, dk, dv, None, None, None, None, None, None, None, None
482
483


Tri Dao's avatar
Tri Dao committed
484
485
486
def flash_attn_qkvpacked_func(
    qkv, dropout_p=0.0, softmax_scale=None, causal=False, return_attn_probs=False
):
Tri Dao's avatar
Tri Dao committed
487
    """dropout_p should be set to 0.0 during evaluation
Tri Dao's avatar
Tri Dao committed
488
489
490
    If Q, K, V are already stacked into 1 tensor, this function will be faster than
    calling flash_attn_func on Q, K, V since the backward pass avoids explicit concatenation
    of the gradients of Q, K, V.
491
492
    For multi-query and grouped-query attention (MQA/GQA), please see
    flash_attn_kvpacked_func and flash_attn_func.
Tri Dao's avatar
Tri Dao committed
493

Tri Dao's avatar
Tri Dao committed
494
    Arguments:
Tri Dao's avatar
Tri Dao committed
495
        qkv: (batch_size, seqlen, 3, nheads, headdim)
Tri Dao's avatar
Tri Dao committed
496
497
498
499
500
501
502
503
        dropout_p: float. Dropout probability.
        softmax_scale: float. The scaling of QK^T before applying softmax.
            Default to 1 / sqrt(headdim).
        causal: bool. Whether to apply causal attention mask (e.g., for auto-regressive modeling).
        return_attn_probs: bool. Whether to return the attention probabilities. This option is for
           testing only. The returned probabilities are not guaranteed to be correct
           (they might not have the right scaling).
    Return:
Tri Dao's avatar
Tri Dao committed
504
        out: (batch_size, seqlen, nheads, headdim).
Tri Dao's avatar
Tri Dao committed
505
506
507
508
509
510
511
        softmax_lse [optional, if return_attn_probs=True]: (batch_size, nheads, seqlen). The
            logsumexp of each row of the matrix QK^T * scaling (e.g., log of the softmax
            normalization factor).
        S_dmask [optional, if return_attn_probs=True]: (batch_size, nheads, seqlen, seqlen).
            The output of softmax (possibly with different scaling). It also encodes the dropout
            pattern (negative means that location was dropped, nonnegative means it was kept).
    """
Tri Dao's avatar
Tri Dao committed
512
    return FlashAttnQKVPackedFunc.apply(qkv, dropout_p, softmax_scale, causal, return_attn_probs)
Tri Dao's avatar
Tri Dao committed
513
514


Tri Dao's avatar
Tri Dao committed
515
516
517
def flash_attn_kvpacked_func(
    q, kv, dropout_p=0.0, softmax_scale=None, causal=False, return_attn_probs=False
):
Tri Dao's avatar
Tri Dao committed
518
    """dropout_p should be set to 0.0 during evaluation
Tri Dao's avatar
Tri Dao committed
519
520
521
522
    If K, V are already stacked into 1 tensor, this function will be faster than
    calling flash_attn_func on Q, K, V since the backward pass avoids explicit concatenation
    of the gradients of K, V.
    Supports multi-query and grouped-query attention (MQA/GQA) by passing in KV with fewer heads
523
    than Q. Note that the number of heads in Q must be divisible by the number of heads in KV.
Tri Dao's avatar
Tri Dao committed
524
525
526
    For example, if Q has 6 heads and K, V have 2 heads, head 0, 1, 2 of Q will attention to head
    0 of K, V, and head 3, 4, 5 of Q will attention to head 1 of K, V.

Tri Dao's avatar
Tri Dao committed
527
    Arguments:
Tri Dao's avatar
Tri Dao committed
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
        q: (batch_size, seqlen, nheads, headdim)
        kv: (batch_size, seqlen, 2, nheads_k, headdim)
        dropout_p: float. Dropout probability.
        softmax_scale: float. The scaling of QK^T before applying softmax.
            Default to 1 / sqrt(headdim).
        causal: bool. Whether to apply causal attention mask (e.g., for auto-regressive modeling).
        return_attn_probs: bool. Whether to return the attention probabilities. This option is for
           testing only. The returned probabilities are not guaranteed to be correct
           (they might not have the right scaling).
    Return:
        out: (batch_size, seqlen, nheads, headdim).
        softmax_lse [optional, if return_attn_probs=True]: (batch_size, nheads, seqlen). The
            logsumexp of each row of the matrix QK^T * scaling (e.g., log of the softmax
            normalization factor).
        S_dmask [optional, if return_attn_probs=True]: (batch_size, nheads, seqlen, seqlen).
            The output of softmax (possibly with different scaling). It also encodes the dropout
            pattern (negative means that location was dropped, nonnegative means it was kept).
    """
    return FlashAttnKVPackedFunc.apply(q, kv, dropout_p, softmax_scale, causal, return_attn_probs)


Tri Dao's avatar
Tri Dao committed
549
550
551
def flash_attn_func(
    q, k, v, dropout_p=0.0, softmax_scale=None, causal=False, return_attn_probs=False
):
Tri Dao's avatar
Tri Dao committed
552
553
    """dropout_p should be set to 0.0 during evaluation
    Supports multi-query and grouped-query attention (MQA/GQA) by passing in KV with fewer heads
554
    than Q. Note that the number of heads in Q must be divisible by the number of heads in KV.
Tri Dao's avatar
Tri Dao committed
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
    For example, if Q has 6 heads and K, V have 2 heads, head 0, 1, 2 of Q will attention to head
    0 of K, V, and head 3, 4, 5 of Q will attention to head 1 of K, V.

    Arguments:
        q: (batch_size, seqlen, nheads, headdim)
        k: (batch_size, seqlen, nheads_k, headdim)
        v: (batch_size, seqlen, nheads_k, headdim)
        dropout_p: float. Dropout probability.
        softmax_scale: float. The scaling of QK^T before applying softmax.
            Default to 1 / sqrt(headdim).
        causal: bool. Whether to apply causal attention mask (e.g., for auto-regressive modeling).
        return_attn_probs: bool. Whether to return the attention probabilities. This option is for
           testing only. The returned probabilities are not guaranteed to be correct
           (they might not have the right scaling).
    Return:
        out: (batch_size, seqlen, nheads, headdim).
        softmax_lse [optional, if return_attn_probs=True]: (batch_size, nheads, seqlen). The
            logsumexp of each row of the matrix QK^T * scaling (e.g., log of the softmax
            normalization factor).
        S_dmask [optional, if return_attn_probs=True]: (batch_size, nheads, seqlen, seqlen).
            The output of softmax (possibly with different scaling). It also encodes the dropout
            pattern (negative means that location was dropped, nonnegative means it was kept).
    """
    return FlashAttnFunc.apply(q, k, v, dropout_p, softmax_scale, causal, return_attn_probs)


Tri Dao's avatar
Tri Dao committed
581
582
583
584
585
586
587
588
589
def flash_attn_varlen_qkvpacked_func(
    qkv,
    cu_seqlens,
    max_seqlen,
    dropout_p=0.0,
    softmax_scale=None,
    causal=False,
    return_attn_probs=False,
):
Tri Dao's avatar
Tri Dao committed
590
591
592
593
    """dropout_p should be set to 0.0 during evaluation
    If Q, K, V are already stacked into 1 tensor, this function will be faster than
    calling flash_attn_varlen_func on Q, K, V since the backward pass avoids explicit concatenation
    of the gradients of Q, K, V.
594
595
    For multi-query and grouped-query attention (MQA/GQA), please see
    flash_attn_varlen_kvpacked_func and flash_attn_varlen_func.
Tri Dao's avatar
Tri Dao committed
596
597
598
599
600
601

    Arguments:
        qkv: (total, 3, nheads, headdim), where total = total number of tokens in the batch.
        cu_seqlens: (batch_size + 1,), dtype torch.int32. The cumulative sequence lengths
           of the sequences in the batch, used to index into qkv.
        max_seqlen: int. Maximum sequence length in the batch.
Tri Dao's avatar
Tri Dao committed
602
603
604
605
606
607
608
609
        dropout_p: float. Dropout probability.
        softmax_scale: float. The scaling of QK^T before applying softmax.
            Default to 1 / sqrt(headdim).
        causal: bool. Whether to apply causal attention mask (e.g., for auto-regressive modeling).
        return_attn_probs: bool. Whether to return the attention probabilities. This option is for
           testing only. The returned probabilities are not guaranteed to be correct
           (they might not have the right scaling).
    Return:
Tri Dao's avatar
Tri Dao committed
610
        out: (total, nheads, headdim).
Tri Dao's avatar
Tri Dao committed
611
612
613
614
615
616
617
        softmax_lse [optional, if return_attn_probs=True]: (batch_size, nheads, seqlen). The
            logsumexp of each row of the matrix QK^T * scaling (e.g., log of the softmax
            normalization factor).
        S_dmask [optional, if return_attn_probs=True]: (batch_size, nheads, seqlen, seqlen).
            The output of softmax (possibly with different scaling). It also encodes the dropout
            pattern (negative means that location was dropped, nonnegative means it was kept).
    """
Tri Dao's avatar
Tri Dao committed
618
619
620
    return FlashAttnVarlenQKVPackedFunc.apply(
        qkv, cu_seqlens, max_seqlen, dropout_p, softmax_scale, causal, return_attn_probs
    )
Tri Dao's avatar
Tri Dao committed
621
622


Tri Dao's avatar
Tri Dao committed
623
624
625
626
627
628
629
630
631
632
633
634
def flash_attn_varlen_kvpacked_func(
    q,
    kv,
    cu_seqlens_q,
    cu_seqlens_k,
    max_seqlen_q,
    max_seqlen_k,
    dropout_p=0.0,
    softmax_scale=None,
    causal=False,
    return_attn_probs=False,
):
Tri Dao's avatar
Tri Dao committed
635
    """dropout_p should be set to 0.0 during evaluation
Tri Dao's avatar
Tri Dao committed
636
637
638
639
    If K, V are already stacked into 1 tensor, this function will be faster than
    calling flash_attn_func on Q, K, V since the backward pass avoids explicit concatenation
    of the gradients of K, V.
    Supports multi-query and grouped-query attention (MQA/GQA) by passing in KV with fewer heads
640
    than Q. Note that the number of heads in Q must be divisible by the number of heads in KV.
Tri Dao's avatar
Tri Dao committed
641
642
643
    For example, if Q has 6 heads and K, V have 2 heads, head 0, 1, 2 of Q will attention to head
    0 of K, V, and head 3, 4, 5 of Q will attention to head 1 of K, V.

Tri Dao's avatar
Tri Dao committed
644
645
    Arguments:
        q: (total_q, nheads, headdim), where total_q = total number of query tokens in the batch.
Tri Dao's avatar
Tri Dao committed
646
        kv: (total_k, 2, nheads_k, headdim), where total_k = total number of key tokens in the batch.
Tri Dao's avatar
Tri Dao committed
647
648
649
650
651
652
653
654
655
656
657
658
659
660
        cu_seqlens_q: (batch_size + 1,), dtype torch.int32. The cumulative sequence lengths
           of the sequences in the batch, used to index into q.
        cu_seqlens_k: (batch_size + 1,), dtype torch.int32. The cumulative sequence lengths
           of the sequences in the batch, used to index into kv.
        max_seqlen_q: int. Maximum query sequence length in the batch.
        max_seqlen_k: int. Maximum key sequence length in the batch.
        dropout_p: float. Dropout probability.
        softmax_scale: float. The scaling of QK^T before applying softmax.
            Default to 1 / sqrt(headdim).
        causal: bool. Whether to apply causal attention mask (e.g., for auto-regressive modeling).
        return_attn_probs: bool. Whether to return the attention probabilities. This option is for
           testing only. The returned probabilities are not guaranteed to be correct
           (they might not have the right scaling).
    Return:
Tri Dao's avatar
Tri Dao committed
661
        out: (total, nheads, headdim).
Tri Dao's avatar
Tri Dao committed
662
663
664
665
666
667
668
        softmax_lse [optional, if return_attn_probs=True]: (batch_size, nheads, seqlen). The
            logsumexp of each row of the matrix QK^T * scaling (e.g., log of the softmax
            normalization factor).
        S_dmask [optional, if return_attn_probs=True]: (batch_size, nheads, seqlen, seqlen).
            The output of softmax (possibly with different scaling). It also encodes the dropout
            pattern (negative means that location was dropped, nonnegative means it was kept).
    """
Tri Dao's avatar
Tri Dao committed
669
    return FlashAttnVarlenKVPackedFunc.apply(
Tri Dao's avatar
Tri Dao committed
670
671
672
673
674
675
676
677
678
679
        q,
        kv,
        cu_seqlens_q,
        cu_seqlens_k,
        max_seqlen_q,
        max_seqlen_k,
        dropout_p,
        softmax_scale,
        causal,
        return_attn_probs,
Tri Dao's avatar
Tri Dao committed
680
    )
Tri Dao's avatar
Tri Dao committed
681

682

Tri Dao's avatar
Tri Dao committed
683
684
685
686
687
688
689
690
691
692
693
694
695
def flash_attn_varlen_func(
    q,
    k,
    v,
    cu_seqlens_q,
    cu_seqlens_k,
    max_seqlen_q,
    max_seqlen_k,
    dropout_p=0.0,
    softmax_scale=None,
    causal=False,
    return_attn_probs=False,
):
Tri Dao's avatar
Tri Dao committed
696
697
    """dropout_p should be set to 0.0 during evaluation
    Supports multi-query and grouped-query attention (MQA/GQA) by passing in K, V with fewer heads
698
    than Q. Note that the number of heads in Q must be divisible by the number of heads in KV.
Tri Dao's avatar
Tri Dao committed
699
700
    For example, if Q has 6 heads and K, V have 2 heads, head 0, 1, 2 of Q will attention to head
    0 of K, V, and head 3, 4, 5 of Q will attention to head 1 of K, V.
701
702

    Arguments:
Tri Dao's avatar
Tri Dao committed
703
704
705
706
707
708
709
710
711
        q: (total_q, nheads, headdim), where total_q = total number of query tokens in the batch.
        k: (total_k, nheads_k, headdim), where total_k = total number of key tokens in the batch.
        v: (total_k, nheads_k, headdim), where total_k = total number of key tokens in the batch.
        cu_seqlens_q: (batch_size + 1,), dtype torch.int32. The cumulative sequence lengths
           of the sequences in the batch, used to index into q.
        cu_seqlens_k: (batch_size + 1,), dtype torch.int32. The cumulative sequence lengths
           of the sequences in the batch, used to index into kv.
        max_seqlen_q: int. Maximum query sequence length in the batch.
        max_seqlen_k: int. Maximum key sequence length in the batch.
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
        dropout_p: float. Dropout probability.
        softmax_scale: float. The scaling of QK^T before applying softmax.
            Default to 1 / sqrt(headdim).
        causal: bool. Whether to apply causal attention mask (e.g., for auto-regressive modeling).
        return_attn_probs: bool. Whether to return the attention probabilities. This option is for
           testing only. The returned probabilities are not guaranteed to be correct
           (they might not have the right scaling).
    Return:
        out: (total, nheads, headdim).
        softmax_lse [optional, if return_attn_probs=True]: (batch_size, nheads, seqlen). The
            logsumexp of each row of the matrix QK^T * scaling (e.g., log of the softmax
            normalization factor).
        S_dmask [optional, if return_attn_probs=True]: (batch_size, nheads, seqlen, seqlen).
            The output of softmax (possibly with different scaling). It also encodes the dropout
            pattern (negative means that location was dropped, nonnegative means it was kept).
    """
Tri Dao's avatar
Tri Dao committed
728
    return FlashAttnVarlenFunc.apply(
Tri Dao's avatar
Tri Dao committed
729
730
731
732
733
734
735
736
737
738
739
        q,
        k,
        v,
        cu_seqlens_q,
        cu_seqlens_k,
        max_seqlen_q,
        max_seqlen_k,
        dropout_p,
        softmax_scale,
        causal,
        return_attn_probs,
Tri Dao's avatar
Tri Dao committed
740
    )