test_flash_attn.py 95 KB
Newer Older
Tri Dao's avatar
Tri Dao committed
1
2
import math

Tri Dao's avatar
Tri Dao committed
3
import pytest
Tri Dao's avatar
Tri Dao committed
4
5
6
import torch
import torch.nn.functional as F
from einops import rearrange, repeat
Tri Dao's avatar
Tri Dao committed
7
8
9
10
11
12
13
from flash_attn import (
    flash_attn_func,
    flash_attn_kvpacked_func,
    flash_attn_qkvpacked_func,
    flash_attn_varlen_func,
    flash_attn_varlen_kvpacked_func,
    flash_attn_varlen_qkvpacked_func,
Tri Dao's avatar
Tri Dao committed
14
    flash_attn_with_kvcache,
Tri Dao's avatar
Tri Dao committed
15
)
16
from flash_attn.bert_padding import pad_input, unpad_input
Tri Dao's avatar
Tri Dao committed
17
from flash_attn.flash_attn_interface import _get_block_size_n
18
from flash_attn.layers.rotary import apply_rotary_emb
Tri Dao's avatar
Tri Dao committed
19
20

MAX_HEADDIM_SM8x = 192
Tri Dao's avatar
Tri Dao committed
21

Tri Dao's avatar
Tri Dao committed
22

Tri Dao's avatar
Tri Dao committed
23
24
25
26
is_sm75 = torch.cuda.get_device_capability("cuda") == (7, 5)
is_sm8x = torch.cuda.get_device_capability("cuda")[0] == 8
is_sm80 = torch.cuda.get_device_capability("cuda") == (8, 0)
is_sm90 = torch.cuda.get_device_capability("cuda") == (9, 0)
Tri Dao's avatar
Tri Dao committed
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
def attn_bias_from_alibi_slopes(
    slopes, seqlen_q, seqlen_k, query_padding_mask=None, key_padding_mask=None, causal=False
):
    batch, nheads = slopes.shape
    device = slopes.device
    slopes = rearrange(slopes, "b h -> b h 1 1")
    if causal:
        return torch.arange(-seqlen_k + 1, 1, device=device, dtype=torch.float32) * slopes
    else:
        row_idx = rearrange(torch.arange(seqlen_q, device=device, dtype=torch.long), "s -> s 1")
        col_idx = torch.arange(seqlen_k, device=device, dtype=torch.long)
        sk = (
            seqlen_k
            if key_padding_mask is None
            else rearrange(key_padding_mask.sum(-1), "b -> b 1 1 1")
        )
        sq = (
            seqlen_q
            if query_padding_mask is None
            else rearrange(query_padding_mask.sum(-1), "b -> b 1 1 1")
        )
        relative_pos = torch.abs(row_idx + sk - sq - col_idx)
        return -slopes * relative_pos.to(dtype=slopes.dtype)


Tri Dao's avatar
Tri Dao committed
54
55
56
def generate_random_padding_mask(max_seqlen, batch_size, device, mode="random"):
    assert mode in ["full", "random", "third"]
    if mode == "full":
Tri Dao's avatar
Tri Dao committed
57
        lengths = torch.full((batch_size, 1), max_seqlen, device=device, dtype=torch.int32)
Tri Dao's avatar
Tri Dao committed
58
    elif mode == "random":
59
60
61
        lengths = torch.randint(
            max(1, max_seqlen - 20), max_seqlen + 1, (batch_size, 1), device=device
        )
Tri Dao's avatar
Tri Dao committed
62
    elif mode == "third":
63
        lengths = torch.randint(max_seqlen // 3, max_seqlen + 1, (batch_size, 1), device=device)
Tri Dao's avatar
Tri Dao committed
64
65
66
    padding_mask = (
        repeat(torch.arange(max_seqlen, device=device), "s -> b s", b=batch_size) < lengths
    )
Tri Dao's avatar
Tri Dao committed
67
68
69
    return padding_mask


Tri Dao's avatar
Tri Dao committed
70
71
72
def generate_qkv(
    q, k, v, query_padding_mask=None, key_padding_mask=None, kvpacked=False, qkvpacked=False
):
Tri Dao's avatar
Tri Dao committed
73
74
    """
    Arguments:
Tri Dao's avatar
Tri Dao committed
75
76
77
        q: (batch_size, seqlen_q, nheads, d)
        k: (batch_size, seqlen_k, nheads_k, d)
        v: (batch_size, seqlen_k, nheads_k, d)
Tri Dao's avatar
Tri Dao committed
78
79
80
81
        query_padding_mask: (batch_size, seqlen), bool
        key_padding_mask: (batch_size, seqlen), bool
    """
    assert not (kvpacked and qkvpacked)
Tri Dao's avatar
Tri Dao committed
82
83
84
85
    batch_size, seqlen_q, nheads, d = q.shape
    _, seqlen_k, nheads_k, _ = k.shape
    assert k.shape == (batch_size, seqlen_k, nheads_k, d)
    assert v.shape == (batch_size, seqlen_k, nheads_k, d)
Tri Dao's avatar
Tri Dao committed
86
87
88

    if query_padding_mask is not None:
        q_unpad, indices_q, cu_seqlens_q, max_seqlen_q = unpad_input(q, query_padding_mask)
Tri Dao's avatar
Tri Dao committed
89
90
91
        output_pad_fn = lambda output_unpad: pad_input(
            output_unpad, indices_q, batch_size, seqlen_q
        )
Tri Dao's avatar
Tri Dao committed
92
    else:
Tri Dao's avatar
Tri Dao committed
93
94
95
96
        q_unpad = rearrange(q, "b s h d -> (b s) h d")
        cu_seqlens_q = torch.arange(
            0, (batch_size + 1) * seqlen_q, step=seqlen_q, dtype=torch.int32, device=q_unpad.device
        )
Tri Dao's avatar
Tri Dao committed
97
        max_seqlen_q = seqlen_q
Tri Dao's avatar
Tri Dao committed
98
99
100
        output_pad_fn = lambda output_unpad: rearrange(
            output_unpad, "(b s) h d -> b s h d", b=batch_size
        )
Tri Dao's avatar
Tri Dao committed
101
102
103
104
105

    if key_padding_mask is not None:
        k_unpad, indices_k, cu_seqlens_k, max_seqlen_k = unpad_input(k, key_padding_mask)
        v_unpad, _, _, _ = unpad_input(v, key_padding_mask)
    else:
Tri Dao's avatar
Tri Dao committed
106
107
108
109
110
        k_unpad = rearrange(k, "b s h d -> (b s) h d")
        v_unpad = rearrange(v, "b s h d -> (b s) h d")
        cu_seqlens_k = torch.arange(
            0, (batch_size + 1) * seqlen_k, step=seqlen_k, dtype=torch.int32, device=k_unpad.device
        )
Tri Dao's avatar
Tri Dao committed
111
        max_seqlen_k = seqlen_k
Tri Dao's avatar
Tri Dao committed
112
113
114

    if qkvpacked:
        assert (query_padding_mask == key_padding_mask).all()
Tri Dao's avatar
Tri Dao committed
115
        assert nheads == nheads_k
Tri Dao's avatar
Tri Dao committed
116
        qkv_unpad = torch.stack([q_unpad, k_unpad, v_unpad], dim=1)
Tri Dao's avatar
Tri Dao committed
117
        qkv = torch.stack([q, k, v], dim=2)
Tri Dao's avatar
Tri Dao committed
118
        if query_padding_mask is not None:
Tri Dao's avatar
Tri Dao committed
119
            dqkv_pad_fn = lambda dqkv_unpad: pad_input(dqkv_unpad, indices_q, batch_size, seqlen_q)
Tri Dao's avatar
Tri Dao committed
120
        else:
Tri Dao's avatar
Tri Dao committed
121
122
123
124
125
126
127
128
129
130
131
            dqkv_pad_fn = lambda dqkv_unpad: rearrange(
                dqkv_unpad, "(b s) t h d -> b s t h d", b=batch_size
            )
        return (
            qkv_unpad.detach().requires_grad_(),
            cu_seqlens_q,
            max_seqlen_q,
            qkv.detach().requires_grad_(),
            output_pad_fn,
            dqkv_pad_fn,
        )
Tri Dao's avatar
Tri Dao committed
132
133
    elif kvpacked:
        kv_unpad = torch.stack([k_unpad, v_unpad], dim=1)
Tri Dao's avatar
Tri Dao committed
134
        kv = torch.stack([k, v], dim=2)
Tri Dao's avatar
Tri Dao committed
135
136
        dq_pad_fn = output_pad_fn
        if key_padding_mask is not None:
Tri Dao's avatar
Tri Dao committed
137
            dkv_pad_fn = lambda dkv_unpad: pad_input(dkv_unpad, indices_k, batch_size, seqlen_k)
Tri Dao's avatar
Tri Dao committed
138
        else:
Tri Dao's avatar
Tri Dao committed
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
            dkv_pad_fn = lambda dkv_unpad: rearrange(
                dkv_unpad, "(b s) t h d -> b s t h d", b=batch_size
            )
        return (
            q_unpad.detach().requires_grad_(),
            kv_unpad.detach().requires_grad_(),
            cu_seqlens_q,
            cu_seqlens_k,
            max_seqlen_q,
            max_seqlen_k,
            q.detach().requires_grad_(),
            kv.detach().requires_grad_(),
            output_pad_fn,
            dq_pad_fn,
            dkv_pad_fn,
        )
Tri Dao's avatar
Tri Dao committed
155
156
157
    else:
        dq_pad_fn = output_pad_fn
        if key_padding_mask is not None:
Tri Dao's avatar
Tri Dao committed
158
            dk_pad_fn = lambda dk_unpad: pad_input(dk_unpad, indices_k, batch_size, seqlen_k)
Tri Dao's avatar
Tri Dao committed
159
        else:
Tri Dao's avatar
Tri Dao committed
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
            dk_pad_fn = lambda dk_unpad: rearrange(dk_unpad, "(b s) h d -> b s h d", b=batch_size)
        return (
            q_unpad.detach().requires_grad_(),
            k_unpad.detach().requires_grad_(),
            v_unpad.detach().requires_grad_(),
            cu_seqlens_q,
            cu_seqlens_k,
            max_seqlen_q,
            max_seqlen_k,
            q.detach().requires_grad_(),
            k.detach().requires_grad_(),
            v.detach().requires_grad_(),
            output_pad_fn,
            dq_pad_fn,
            dk_pad_fn,
        )
Tri Dao's avatar
Tri Dao committed
176
177


Tri Dao's avatar
Tri Dao committed
178
179
180
181
182
183
184
def construct_local_mask(
    seqlen_q,
    seqlen_k,
    window_size=(-1, -1),  # -1 means infinite window size
    query_padding_mask=None,
    key_padding_mask=None,
    device=None,
185
):
186
187
188
189
190
191
192
193
194
195
196
197
    row_idx = rearrange(torch.arange(seqlen_q, device=device, dtype=torch.long), "s -> s 1")
    col_idx = torch.arange(seqlen_k, device=device, dtype=torch.long)
    sk = (
        seqlen_k
        if key_padding_mask is None
        else rearrange(key_padding_mask.sum(-1), "b -> b 1 1 1")
    )
    sq = (
        seqlen_q
        if query_padding_mask is None
        else rearrange(query_padding_mask.sum(-1), "b -> b 1 1 1")
    )
Tri Dao's avatar
Tri Dao committed
198
199
200
201
202
203
204
205
    if window_size[0] < 0:
        return col_idx > row_idx + sk - sq + window_size[1]
    else:
        sk = torch.full_like(col_idx, seqlen_k) if key_padding_mask is None else sk
        return torch.logical_or(
            col_idx > torch.minimum(row_idx + sk - sq + window_size[1], sk),
            col_idx < row_idx + sk - sq - window_size[0],
        )
206
207


Tri Dao's avatar
Tri Dao committed
208
209
210
211
212
213
def attention_ref(
    q,
    k,
    v,
    query_padding_mask=None,
    key_padding_mask=None,
214
    attn_bias=None,
Tri Dao's avatar
Tri Dao committed
215
216
217
    dropout_p=0.0,
    dropout_mask=None,
    causal=False,
Tri Dao's avatar
Tri Dao committed
218
    window_size=(-1, -1),  # -1 means infinite window size
Nicolas Patry's avatar
Nicolas Patry committed
219
    softcap=0.0,
Tri Dao's avatar
Tri Dao committed
220
221
222
    upcast=True,
    reorder_ops=False,
):
Tri Dao's avatar
Tri Dao committed
223
224
225
    """
    Arguments:
        q: (batch_size, seqlen_q, nheads, head_dim)
Tri Dao's avatar
Tri Dao committed
226
227
        k: (batch_size, seqlen_k, nheads_k, head_dim)
        v: (batch_size, seqlen_k, nheads_k, head_dim)
Tri Dao's avatar
Tri Dao committed
228
229
        query_padding_mask: (batch_size, seqlen_q)
        key_padding_mask: (batch_size, seqlen_k)
230
        attn_bias: broadcastable to (batch_size, nheads, seqlen_q, seqlen_k)
Tri Dao's avatar
Tri Dao committed
231
232
        dropout_p: float
        dropout_mask: (batch_size, nheads, seqlen_q, seqlen_k)
Tri Dao's avatar
Tri Dao committed
233
234
        causal: whether to apply causal masking
        window_size: (int, int), left and right window size
Tri Dao's avatar
Tri Dao committed
235
236
        upcast: whether to cast all inputs to fp32, do all computation in fp32, then cast
            output back to fp16/bf16.
cao lei's avatar
cao lei committed
237
        reorder_ops: whether to change the order of operations (scaling k instead of scaling q, etc.)
Tri Dao's avatar
Tri Dao committed
238
239
240
241
242
243
            without changing the math. This is to estimate the numerical error from operation
            reordering.
    Output:
        output: (batch_size, seqlen_q, nheads, head_dim)
        attention: (batch_size, nheads, seqlen_q, seqlen_k), softmax after dropout
    """
Tri Dao's avatar
Tri Dao committed
244
245
    if causal:
        window_size = (window_size[0], 0)
Tri Dao's avatar
Tri Dao committed
246
247
248
249
    dtype_og = q.dtype
    if upcast:
        q, k, v = q.float(), k.float(), v.float()
    seqlen_q, seqlen_k = q.shape[1], k.shape[1]
Tri Dao's avatar
Tri Dao committed
250
251
    k = repeat(k, "b s h d -> b s (h g) d", g=q.shape[2] // k.shape[2])
    v = repeat(v, "b s h d -> b s (h g) d", g=q.shape[2] // v.shape[2])
Tri Dao's avatar
Tri Dao committed
252
253
    d = q.shape[-1]
    if not reorder_ops:
Tri Dao's avatar
Tri Dao committed
254
        scores = torch.einsum("bthd,bshd->bhts", q / math.sqrt(d), k)
Tri Dao's avatar
Tri Dao committed
255
    else:
Tri Dao's avatar
Tri Dao committed
256
        scores = torch.einsum("bthd,bshd->bhts", q, k / math.sqrt(d))
Nicolas Patry's avatar
Nicolas Patry committed
257
258
259
260
    if softcap > 0:
        scores /= softcap
        scores = scores.tanh()
        scores *= softcap
Tri Dao's avatar
Tri Dao committed
261
    if key_padding_mask is not None:
Tri Dao's avatar
Tri Dao committed
262
        scores.masked_fill_(rearrange(~key_padding_mask, "b s -> b 1 1 s"), float("-inf"))
Tri Dao's avatar
Tri Dao committed
263
264
265
266
267
268
269
270
    if window_size[0] >= 0 or window_size[1] >= 0:
        local_mask = construct_local_mask(
            seqlen_q,
            seqlen_k,
            window_size,
            query_padding_mask,
            key_padding_mask,
            q.device,
Tri Dao's avatar
Tri Dao committed
271
        )
Tri Dao's avatar
Tri Dao committed
272
        scores.masked_fill_(local_mask, float("-inf"))
273
274
275
    if attn_bias is not None:
        scores = scores + attn_bias
    attention = torch.softmax(scores, dim=-1).to(v.dtype)
Tri Dao's avatar
Tri Dao committed
276
277
278
279
280
281
282
    # Some rows might be completely masked out so we fill them with zero instead of NaN
    if window_size[0] >= 0 or window_size[1] >= 0:
        attention = attention.masked_fill(torch.all(local_mask, dim=-1, keepdim=True), 0.0)
    # We want to mask here so that the attention matrix doesn't have any NaNs
    # Otherwise we'll get NaN in dV
    if query_padding_mask is not None:
        attention = attention.masked_fill(rearrange(~query_padding_mask, "b s -> b 1 s 1"), 0.0)
Tri Dao's avatar
Tri Dao committed
283
284
285
286
287
    dropout_scaling = 1.0 / (1 - dropout_p)
    # attention_drop = attention.masked_fill(~dropout_mask, 0.0) * dropout_scaling
    # output = torch.einsum('bhts,bshd->bthd', attention_drop , v)
    if dropout_mask is not None:
        attention_drop = attention.masked_fill(~dropout_mask, 0.0)
Tri Dao's avatar
Tri Dao committed
288
289
    else:
        attention_drop = attention
Tri Dao's avatar
Tri Dao committed
290
    output = torch.einsum("bhts,bshd->bthd", attention_drop, v * dropout_scaling)
Tri Dao's avatar
Tri Dao committed
291
    if query_padding_mask is not None:
Tri Dao's avatar
Tri Dao committed
292
        output.masked_fill_(rearrange(~query_padding_mask, "b s -> b s 1 1"), 0.0)
Tri Dao's avatar
Tri Dao committed
293
294
295
    return output.to(dtype=dtype_og), attention.to(dtype=dtype_og)


Tri Dao's avatar
Tri Dao committed
296
297
298
299
300
def attention_kvpacked_ref(
    q,
    kv,
    query_padding_mask=None,
    key_padding_mask=None,
301
    attn_bias=None,
Tri Dao's avatar
Tri Dao committed
302
303
304
    dropout_p=0.0,
    dropout_mask=None,
    causal=False,
Tri Dao's avatar
Tri Dao committed
305
    window_size=(-1, -1),  # -1 means infinite window size
Tri Dao's avatar
Tri Dao committed
306
    softcap=0.0,
Tri Dao's avatar
Tri Dao committed
307
308
309
310
311
312
313
314
315
    upcast=True,
    reorder_ops=False,
):
    return attention_ref(
        q,
        kv[:, :, 0],
        kv[:, :, 1],
        query_padding_mask,
        key_padding_mask,
316
        attn_bias,
Tri Dao's avatar
Tri Dao committed
317
318
319
320
        dropout_p,
        dropout_mask,
        upcast=upcast,
        causal=causal,
Tri Dao's avatar
Tri Dao committed
321
        window_size=window_size,
Tri Dao's avatar
Tri Dao committed
322
        softcap=softcap,
Tri Dao's avatar
Tri Dao committed
323
324
        reorder_ops=reorder_ops,
    )
Tri Dao's avatar
Tri Dao committed
325
326


Tri Dao's avatar
Tri Dao committed
327
328
329
def attention_qkvpacked_ref(
    qkv,
    key_padding_mask=None,
330
    attn_bias=None,
Tri Dao's avatar
Tri Dao committed
331
332
333
    dropout_p=0.0,
    dropout_mask=None,
    causal=False,
Tri Dao's avatar
Tri Dao committed
334
    window_size=(-1, -1),  # -1 means infinite window size
Tri Dao's avatar
Tri Dao committed
335
    softcap=0.0,
Tri Dao's avatar
Tri Dao committed
336
337
338
339
340
341
342
343
344
    upcast=True,
    reorder_ops=False,
):
    return attention_ref(
        qkv[:, :, 0],
        qkv[:, :, 1],
        qkv[:, :, 2],
        key_padding_mask,
        key_padding_mask,
345
        attn_bias,
Tri Dao's avatar
Tri Dao committed
346
347
348
349
        dropout_p,
        dropout_mask,
        upcast=upcast,
        causal=causal,
Tri Dao's avatar
Tri Dao committed
350
        window_size=window_size,
Tri Dao's avatar
Tri Dao committed
351
        softcap=softcap,
Tri Dao's avatar
Tri Dao committed
352
353
        reorder_ops=reorder_ops,
    )
Tri Dao's avatar
Tri Dao committed
354
355
356
357
358
359
360
361
362
363
364


def generate_sparsity_mask(seqlen, sparsity=0.3):
    repeats = seqlen // 16 // 2
    # mask = torch.stack([torch.tensor([1, 0] * repeats, dtype=torch.bool, device='cuda'),
    #                     torch.tensor([0, 1] * repeats, dtype=torch.bool, device='cuda')], dim=-1)
    # mask = torch.stack([torch.tensor([1, 1] * repeats, dtype=torch.bool, device='cuda'),
    #                     torch.tensor([1, 1] * repeats, dtype=torch.bool, device='cuda')], dim=-1)
    # mask = torch.stack([torch.tensor([1, 1] * repeats, dtype=torch.bool, device='cuda')], dim=-1)
    # mask = torch.stack([torch.tensor([1, 0] * repeats, dtype=torch.bool, device='cuda')], dim=-1)
    nrow, ncol = seqlen // 16, seqlen // 256
Tri Dao's avatar
Tri Dao committed
365
    mask = torch.rand(nrow, ncol, device="cuda") < sparsity
Tri Dao's avatar
Tri Dao committed
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
    return mask


def attention_blocksparse_ref(qkv, blockmask, attn_mask, dropout_p, dropout_mask):
    """
    Arguments:
        qkv: (batch_size, seqlen, 3, nheads, head_dim)
        blockmask: (seqlen / 16, seqlen / 256)
        attn_mask: (batch_size, seqlen)
        dropout_p: float
        dropout_mask: (batch_size, nheads, seqlen, seqlen)
    Output:
        output: (batch_size, seqlen, nheads, head_dim)
        attention: softmax after dropout
    """
    q, k, v = qkv.float().unbind(dim=2)
    d = qkv.shape[-1]
    seqlen = qkv.shape[1]
Tri Dao's avatar
Tri Dao committed
384
385
386
    scores = torch.einsum("bthd,bshd->bhts", q / math.sqrt(d), k)
    scores.masked_fill_(rearrange(~attn_mask, "b s -> b 1 1 s"), float("-inf"))
    blockmask = repeat(blockmask, "s_16 s_256 -> (s_16 16) (s_256 256)")
Tri Dao's avatar
Tri Dao committed
387
    blockmask = blockmask[:seqlen, :seqlen]
Tri Dao's avatar
Tri Dao committed
388
    scores.masked_fill_(rearrange(~blockmask, "t s -> 1 1 t s"), float("-inf"))
Tri Dao's avatar
Tri Dao committed
389
    attention = torch.softmax(scores, dim=-1)
Tri Dao's avatar
Tri Dao committed
390
391
    attention = attention.masked_fill(rearrange(~attn_mask, "b s -> b 1 s 1"), 0.0)
    attention = attention.masked_fill_(rearrange(~blockmask, "t s -> 1 1 t s"), 0.0)
Tri Dao's avatar
Tri Dao committed
392
    attention_drop = attention.masked_fill(~dropout_mask, 0.0) / (1 - dropout_p)
Tri Dao's avatar
Tri Dao committed
393
394
    output = torch.einsum("bhts,bshd->bthd", attention_drop, v)
    output.masked_fill_(rearrange(~attn_mask, "b s -> b s 1 1"), 0)
Tri Dao's avatar
Tri Dao committed
395
396
397
    return output.to(dtype=qkv.dtype), attention.to(dtype=qkv.dtype)


Tri Dao's avatar
Tri Dao committed
398
def convert_flash_attn_S_to_softmax(
Tri Dao's avatar
Tri Dao committed
399
400
401
402
403
404
405
406
407
    S,
    seqlen_q,
    seqlen_k,
    query_padding_mask,
    key_padding_mask,
    head_dim,
    is_dropout,
    causal=False,
    window_size=(-1, -1),  # -1 means infinite window size
Tri Dao's avatar
Tri Dao committed
408
):
Tri Dao's avatar
Tri Dao committed
409
410
    """FlashAttention stores the S matrix in a different way.
    Arguments:
Tri Dao's avatar
Tri Dao committed
411
        S: (batch_size, nheads, seqlen_q_rounded, seqlen_k_rounded)
412
413
        query_padding_mask: (batch_size, seqlen_q_rounded)
        key_padding_mask: (batch_size, seqlen_k_rounded)
Tri Dao's avatar
Tri Dao committed
414
    """
Tri Dao's avatar
Tri Dao committed
415
416
    if causal:
        window_size = (window_size[0], 0)
417
    seqlen_q_rounded, seqlen_k_rounded = S.shape[-2:]
Tri Dao's avatar
Tri Dao committed
418
    S_converted = S
Tri Dao's avatar
Tri Dao committed
419
420
421
422
423
424
425
426
    if window_size[0] >= 0 or window_size[1] >= 0:
        local_mask = construct_local_mask(
            seqlen_q,
            seqlen_k,
            window_size,
            query_padding_mask,
            key_padding_mask,
            S.device,
Tri Dao's avatar
Tri Dao committed
427
        )
Tri Dao's avatar
Tri Dao committed
428
429
        local_mask = F.pad(
            local_mask,
430
431
432
            (0, seqlen_k_rounded - seqlen_k, 0, seqlen_q_rounded - seqlen_q),
            value=True,
        )
Tri Dao's avatar
Tri Dao committed
433
        S_converted = S_converted.masked_fill(local_mask, 0.0)
Tri Dao's avatar
Tri Dao committed
434
435
436

    # Need to zero out things not in attention_mask in case S was initialized with random values
    # and some of those values aren't overwritten.
437
438
439
    seqlen_q_og = (
        query_padding_mask.shape[-1] if query_padding_mask is not None else seqlen_q_rounded
    )
Tri Dao's avatar
Tri Dao committed
440
    if query_padding_mask is not None:
441
        query_padding_mask = F.pad(query_padding_mask, (0, seqlen_q_rounded - seqlen_q_og))
Tri Dao's avatar
Tri Dao committed
442
        S_converted = S_converted.masked_fill(rearrange(~query_padding_mask, "b s -> b 1 s 1"), 0.0)
Tri Dao's avatar
Tri Dao committed
443
444
    seqlen_k_og = key_padding_mask.shape[-1] if key_padding_mask is not None else seqlen_k
    if key_padding_mask is not None:
445
        key_padding_mask = F.pad(key_padding_mask, (0, seqlen_k_rounded - seqlen_k_og))
Tri Dao's avatar
Tri Dao committed
446
        S_converted = S_converted.masked_fill(rearrange(~key_padding_mask, "b s -> b 1 1 s"), 0.0)
447
448
449
    S_converted = F.pad(S_converted, (0, 0, 0, seqlen_q_og - seqlen_q_rounded))
    S_converted = F.pad(S_converted, (0, seqlen_k_og - seqlen_k_rounded))
    return S_converted[:, :, :seqlen_q, :seqlen_k]
Tri Dao's avatar
Tri Dao committed
450
451


Tri Dao's avatar
Tri Dao committed
452
453
454
455
456
457
458
def normalize_flash_attn_S(
    attn_unnorm,
    q,
    k,
    v,
    query_padding_mask=None,
    key_padding_mask=None,
459
    attn_bias=None,
Tri Dao's avatar
Tri Dao committed
460
461
    is_dropout=False,
    causal=False,
Tri Dao's avatar
Tri Dao committed
462
    window_size=(-1, -1),  # -1 means infinite window size
Tri Dao's avatar
Tri Dao committed
463
):
Tri Dao's avatar
Tri Dao committed
464
465
466
467
468
    """
    Arguments:
        q: (batch_size, seqlen_q, nheads, head_dim)
        k, v: (batch_size, seqlen_k, nheads, head_dim)
        key_padding_mask: (batch_size, seqlen_q)
469
        attn_bias: broadcastable to (batch_size, nheads, seqlen_q, seqlen_k)
Tri Dao's avatar
Tri Dao committed
470
471
472
473
    Output:
        softmax_lse: (batch_size, nheads, seqlen_q)
        softmax_max: (batch_size, nheads, seqlen_q)
    """
Tri Dao's avatar
Tri Dao committed
474
475
    if causal:
        window_size = (window_size[0], 0)
Tri Dao's avatar
Tri Dao committed
476
477
478
    q, k, v = q.float(), k.float(), v.float()
    _, seqlen_q, _, head_dim = q.shape
    seqlen_k = k.shape[1]
Tri Dao's avatar
Tri Dao committed
479
    scores = torch.einsum("bthd,bshd->bhts", q / math.sqrt(head_dim), k)
Tri Dao's avatar
Tri Dao committed
480
    if key_padding_mask is not None:
Tri Dao's avatar
Tri Dao committed
481
        scores.masked_fill_(rearrange(~key_padding_mask, "b s -> b 1 1 s"), float("-inf"))
Tri Dao's avatar
Tri Dao committed
482
483
484
485
486
487
488
489
    if window_size[0] >= 0 or window_size[1] >= 0:
        local_mask = construct_local_mask(
            seqlen_q,
            seqlen_k,
            window_size,
            query_padding_mask,
            key_padding_mask,
            q.device,
Tri Dao's avatar
Tri Dao committed
490
        )
Tri Dao's avatar
Tri Dao committed
491
        scores.masked_fill_(local_mask, float("-inf"))
492
493
    if attn_bias is not None:
        scores = scores + attn_bias.to(dtype=scores.dtype)
Tri Dao's avatar
Tri Dao committed
494
    block_size_n = _get_block_size_n(scores.device, head_dim, is_dropout, causal)
Tri Dao's avatar
Tri Dao committed
495
    scores_block = scores.split(block_size_n, dim=-1)
Tri Dao's avatar
Tri Dao committed
496
    lse_block = torch.stack([torch.logsumexp(s, dim=-1) for s in scores_block], dim=-1)
Tri Dao's avatar
Tri Dao committed
497
    lse = torch.logsumexp(lse_block, dim=-1)
498
499
500
    # lse could be -inf (i.e. all values in scores are -inf), and we want to set those to inf
    # so that when we do torch.exp(m - lse), we get 0.0 instead of NaN.
    lse[lse == float("-inf")] = float("inf")
Tri Dao's avatar
Tri Dao committed
501
502
503
    scores_max_block = torch.stack([torch.amax(s, dim=-1) for s in scores_block], dim=-1)
    cummax_block = torch.cummax(scores_max_block.flip(-1), dim=-1).values.flip(-1).unbind(dim=-1)
    attn_unnorm_block = attn_unnorm.split(block_size_n, dim=-1)
Tri Dao's avatar
Tri Dao committed
504
505
    attn_norm = torch.cat(
        [
506
            a * rearrange(torch.exp(m - lse), "b h s -> b h s 1")
Tri Dao's avatar
Tri Dao committed
507
508
509
510
            for a, m in zip(attn_unnorm_block, cummax_block)
        ],
        dim=-1,
    )
Tri Dao's avatar
Tri Dao committed
511
    if query_padding_mask is not None:
Tri Dao's avatar
Tri Dao committed
512
        attn_norm.masked_fill_(rearrange(~query_padding_mask, "b s -> b 1 s 1"), 0.0)
Tri Dao's avatar
Tri Dao committed
513
514
515
    return attn_norm.to(dtype=attn_unnorm.dtype)


Tri Dao's avatar
Tri Dao committed
516
def get_dropout_fraction(
Tri Dao's avatar
Tri Dao committed
517
518
519
520
521
    dropout_mask,
    query_padding_mask=None,
    key_padding_mask=None,
    causal=False,
    window_size=(-1, -1),  # -1 means infinite window size
Tri Dao's avatar
Tri Dao committed
522
):
Tri Dao's avatar
Tri Dao committed
523
524
525
526
527
    """
    dropout_mask: (batch_size, nheads, seqlen_q, seqlen_k), bool. True means keep, False means drop.
    query_padding_mask: (batch_size, seqlen_q)
    key_padding_mask: (batch_size, seqlen_k)
    """
Tri Dao's avatar
Tri Dao committed
528
529
    if causal:
        window_size = (window_size[0], 0)
Tri Dao's avatar
Tri Dao committed
530
531
    batch_size, nheads, seqlen_q, seqlen_k = dropout_mask.shape
    dropped = ~dropout_mask
Tri Dao's avatar
Tri Dao committed
532
    valid = torch.ones_like(dropout_mask)
Tri Dao's avatar
Tri Dao committed
533
    if query_padding_mask is not None:
Tri Dao's avatar
Tri Dao committed
534
        dropped.masked_fill_(rearrange(~query_padding_mask, "b s -> b 1 s 1"), False)
Tri Dao's avatar
Tri Dao committed
535
        valid.masked_fill_(rearrange(~query_padding_mask, "b s -> b 1 s 1"), False)
Tri Dao's avatar
Tri Dao committed
536
    if key_padding_mask is not None:
Tri Dao's avatar
Tri Dao committed
537
        dropped.masked_fill_(rearrange(~key_padding_mask, "b s -> b 1 1 s"), False)
Tri Dao's avatar
Tri Dao committed
538
539
540
541
542
543
544
545
546
        valid.masked_fill_(rearrange(~key_padding_mask, "b s -> b 1 1 s"), False)
    if window_size[0] >= 0 or window_size[1] >= 0:
        local_mask = construct_local_mask(
            seqlen_q,
            seqlen_k,
            window_size,
            query_padding_mask,
            key_padding_mask,
            dropout_mask.device,
Tri Dao's avatar
Tri Dao committed
547
        )
Tri Dao's avatar
Tri Dao committed
548
549
        dropped.masked_fill_(local_mask, False)
        valid.masked_fill_(local_mask, False)
Tri Dao's avatar
Tri Dao committed
550
    dropped_total = dropped.sum()
Tri Dao's avatar
Tri Dao committed
551
    return dropped.sum() / valid.sum()
Tri Dao's avatar
Tri Dao committed
552
553


Tri Dao's avatar
Tri Dao committed
554
@pytest.mark.parametrize("dtype", ([torch.float16] if is_sm75 else [torch.float16, torch.bfloat16]))
Tri Dao's avatar
Tri Dao committed
555
# @pytest.mark.parametrize("dtype", [torch.float16])
556
@pytest.mark.parametrize("deterministic", [False, True])
557
# @pytest.mark.parametrize("deterministic", [False])
558
@pytest.mark.parametrize("alibi", [False, True])
559
# @pytest.mark.parametrize("alibi", [False])
Tri Dao's avatar
Tri Dao committed
560
@pytest.mark.parametrize("local", [False, True])
561
# @pytest.mark.parametrize("local", [False])
Tri Dao's avatar
Tri Dao committed
562
@pytest.mark.parametrize("causal", [False, True])
Tri Dao's avatar
Tri Dao committed
563
# @pytest.mark.parametrize("causal", [False])
Tri Dao's avatar
Tri Dao committed
564
@pytest.mark.parametrize("d", [32, 40, 59, 64, 80, 96, 111, 128, 160, 192, 224, 256])
Tri Dao's avatar
Tri Dao committed
565
# @pytest.mark.parametrize("d", [32, 64, 96, 128, 160, 192, 224, 256])
Tri Dao's avatar
Tri Dao committed
566
# @pytest.mark.parametrize('d', [32, 64, 96, 128])
Tri Dao's avatar
Tri Dao committed
567
# @pytest.mark.parametrize("d", [64])
Tri Dao's avatar
Tri Dao committed
568
# @pytest.mark.parametrize('seqlen', [128, 256, 384, 512, 768, 1024, 2048])
569
@pytest.mark.parametrize("seqlen", [97, 128, 200, 384, 768, 1024, 1025, 2048])
570
# @pytest.mark.parametrize("seqlen", [512])
Tri Dao's avatar
Tri Dao committed
571
@pytest.mark.parametrize("dropout_p", [0.0, 0.17])
Tri Dao's avatar
Tri Dao committed
572
# @pytest.mark.parametrize("dropout_p", [0.0])
573
def test_flash_attn_qkvpacked(seqlen, d, dropout_p, causal, local, alibi, deterministic, dtype):
Tri Dao's avatar
Tri Dao committed
574
    if seqlen >= 2048 and torch.cuda.get_device_properties("cuda").total_memory <= 16 * 2**30:
Tri Dao's avatar
Tri Dao committed
575
        pytest.skip()  # Reference implementation OOM
Tri Dao's avatar
Tri Dao committed
576
    device = "cuda"
Tri Dao's avatar
Tri Dao committed
577
578
    # set seed
    torch.random.manual_seed(0)
579
    batch_size = 4
Tri Dao's avatar
Tri Dao committed
580
    nheads = 9
Tri Dao's avatar
Tri Dao committed
581
    window_size = (-1, -1) if not local else torch.randint(0, seqlen, (2,))
Tri Dao's avatar
Tri Dao committed
582
583
584
    qkv = torch.randn(
        batch_size, seqlen, 3, nheads, d, device=device, dtype=dtype, requires_grad=True
    )
585
586
587
588
589
    if alibi:
        alibi_slopes = torch.rand(batch_size, nheads, device=device, dtype=torch.float32) * 0.3
        attn_bias = attn_bias_from_alibi_slopes(alibi_slopes, seqlen, seqlen, causal=causal)
    else:
        alibi_slopes, attn_bias = None, None
Tri Dao's avatar
Tri Dao committed
590
    out, lse, S_dmask = flash_attn_qkvpacked_func(
591
592
593
594
595
        qkv,
        dropout_p,
        causal=causal,
        window_size=window_size,
        alibi_slopes=alibi_slopes,
596
        deterministic=deterministic,
597
        return_attn_probs=True,
Tri Dao's avatar
Tri Dao committed
598
    )
Tri Dao's avatar
Tri Dao committed
599
600
    if dropout_p > 0.0:
        S_dmask_converted = convert_flash_attn_S_to_softmax(
Tri Dao's avatar
Tri Dao committed
601
602
603
604
605
606
607
608
609
            S_dmask,
            seqlen,
            seqlen,
            None,
            None,
            d,
            dropout_p > 0.0,
            causal=causal,
            window_size=window_size,
610
        )
Tri Dao's avatar
Tri Dao committed
611
612
        dropout_mask = S_dmask_converted >= 0
        attn_unnorm = S_dmask_converted.abs()
Tri Dao's avatar
Tri Dao committed
613
614
615
616
617
618
619
        attn = normalize_flash_attn_S(
            attn_unnorm,
            qkv[:, :, 0],
            qkv[:, :, 1],
            qkv[:, :, 2],
            None,
            None,
620
            attn_bias,
Tri Dao's avatar
Tri Dao committed
621
622
            dropout_p > 0.0,
            causal=causal,
Tri Dao's avatar
Tri Dao committed
623
            window_size=window_size,
Tri Dao's avatar
Tri Dao committed
624
        )
Tri Dao's avatar
Tri Dao committed
625
626
627
        dropout_fraction = get_dropout_fraction(
            dropout_mask, None, None, causal=causal, window_size=window_size
        ).item()
Tri Dao's avatar
Tri Dao committed
628
        print(f"Actual dropout fraction: {dropout_fraction}")
Tri Dao's avatar
Tri Dao committed
629
630
631
    else:
        dropout_mask = None

Tri Dao's avatar
Tri Dao committed
632
    out_ref, attn_ref = attention_qkvpacked_ref(
633
        qkv, None, attn_bias, dropout_p, dropout_mask, causal=causal, window_size=window_size
Tri Dao's avatar
Tri Dao committed
634
    )
Tri Dao's avatar
Tri Dao committed
635
    out_pt, attn_pt = attention_qkvpacked_ref(
Tri Dao's avatar
Tri Dao committed
636
637
        qkv,
        None,
638
        attn_bias,
Tri Dao's avatar
Tri Dao committed
639
640
641
642
643
644
        dropout_p,
        dropout_mask,
        causal=causal,
        window_size=window_size,
        upcast=False,
        reorder_ops=True,
Tri Dao's avatar
Tri Dao committed
645
    )
Tri Dao's avatar
Tri Dao committed
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
    # v = qkv[:, :, 2].float()
    # qk = torch.einsum('bshd,bthd->bhst', qkv[:, :, 0], qkv[:, :, 1]).float()
    # if causal:
    #     causal_mask = torch.triu(torch.ones(seqlen, seqlen, dtype=torch.bool, device=qkv.device), 1)
    #     qk.masked_fill_(causal_mask, float('-inf'))
    # m = qk.amax(-1, keepdim=True)
    # s_tmp = torch.exp((qk - m) / math.sqrt(d))
    # p_tmp = torch.softmax(qk / math.sqrt(d), -1)
    # p_dropped = p_tmp if dropout_mask is None else p_tmp.masked_fill(~dropout_mask, 0)
    # lse_ref = torch.logsumexp(qk / math.sqrt(d), -1)
    # qk_max1 = torch.max(qk[:, :, 128:, 192:], -1, keepdim=True).values
    # qk_max2 = torch.max(qk[:, :, 128:, 128:], -1, keepdim=True).values
    # qk_max3 = torch.max(qk[:, :, 128:, 64:], -1, keepdim=True).values
    # qk_max4 = torch.max(qk[:, :, 128:, :], -1, keepdim=True).values
    # o1 = torch.einsum('bhst,bthd->bshd', torch.exp((qk[:, :, 128:, 192:] - qk_max1) / math.sqrt(d)), v[:, 192:])
    # o2 = torch.einsum('bhst,bthd->bshd', torch.exp((qk[:, :, 128:, 128:] - qk_max2) / math.sqrt(d)), v[:, 128:])
    # o3 = torch.einsum('bhst,bthd->bshd', torch.exp((qk[:, :, 128:, 64:] - qk_max3) / math.sqrt(d)), v[:, 64:])
    # o4 = torch.einsum('bhst,bthd->bshd', torch.exp((qk[:, :, 128:, :] - qk_max4) / math.sqrt(d)), v[:, :])
Tri Dao's avatar
Tri Dao committed
664
665
666
667
    print(f"Output max diff: {(out - out_ref).abs().max().item()}")
    print(f"Output mean diff: {(out - out_ref).abs().mean().item()}")
    print(f"Pytorch max diff: {(out_pt - out_ref).abs().max().item()}")
    print(f"Pytorch mean diff: {(out_pt - out_ref).abs().mean().item()}")
Tri Dao's avatar
Tri Dao committed
668
    if dropout_p > 0.0:
Tri Dao's avatar
Tri Dao committed
669
670
        print(f"Attention max diff: {(attn - attn_ref).abs().max().item()}")
        print(f"Attention Pytorch max diff: {(attn_pt - attn_ref).abs().max().item()}")
Tri Dao's avatar
Tri Dao committed
671
672
673
674
675

    g = torch.randn_like(out)
    # do_o = (g.float() * out.float()).sum(-1)
    # dv_tmp = torch.einsum('bhts,bthd->bshd', attn_pt[:, :, :64], g[:, :64])
    # dv_tmp1 = torch.einsum('bhts,bthd->bshd', attn_pt[:, :, 64:], g[:, 64:])
676
    if (d <= MAX_HEADDIM_SM8x or (d > 224 and dropout_p == 0)) or (is_sm80 or is_sm90):
Tri Dao's avatar
Tri Dao committed
677
678
679
680
681
682
683
684
685
686
687
        (dqkv,) = torch.autograd.grad(out, qkv, g)
        (dqkv_ref,) = torch.autograd.grad(out_ref, qkv, g)
        (dqkv_pt,) = torch.autograd.grad(out_pt, qkv, g)
        print(f"dQ max diff: {(dqkv[:, :, 0] - dqkv_ref[:, :, 0]).abs().max().item()}")
        print(f"dK max diff: {(dqkv[:, :, 1] - dqkv_ref[:, :, 1]).abs().max().item()}")
        print(f"dV max diff: {(dqkv[:, :, 2] - dqkv_ref[:, :, 2]).abs().max().item()}")
        print(f"dQKV mean diff: {(dqkv - dqkv_ref).abs().mean().item()}")
        print(f"dQ Pytorch max diff: {(dqkv_pt[:, :, 0] - dqkv_ref[:, :, 0]).abs().max().item()}")
        print(f"dK Pytorch max diff: {(dqkv_pt[:, :, 1] - dqkv_ref[:, :, 1]).abs().max().item()}")
        print(f"dV Pytorch max diff: {(dqkv_pt[:, :, 2] - dqkv_ref[:, :, 2]).abs().max().item()}")
        print(f"dQKV Pytorch mean diff: {(dqkv_pt - dqkv_ref).abs().mean().item()}")
Tri Dao's avatar
Tri Dao committed
688
689
690

    # Check that FlashAttention's numerical error is at most twice the numerical error
    # of a Pytorch implementation.
Tri Dao's avatar
Tri Dao committed
691
692
693
694
    assert (out - out_ref).abs().max().item() <= 2 * (out_pt - out_ref).abs().max().item()

    if dropout_p > 0.0:
        assert (attn - attn_ref).abs().max().item() <= 2 * (attn_pt - attn_ref).abs().max().item()
695
696
697
        # With alibi, many of the prob values are 0.0 & -0.0 so dropout_fraction isn't accurate
        if not alibi:
            assert abs(dropout_fraction - dropout_p) <= (0.01 if not local else 0.025)
Tri Dao's avatar
Tri Dao committed
698

699
    if (d <= MAX_HEADDIM_SM8x or (d > 224 and dropout_p == 0)) or (is_sm80 or is_sm90):
Tri Dao's avatar
Tri Dao committed
700
        assert (dqkv - dqkv_ref).abs().max().item() <= 2 * (dqkv_pt - dqkv_ref).abs().max().item()
Tri Dao's avatar
Tri Dao committed
701
702


Tri Dao's avatar
Tri Dao committed
703
@pytest.mark.parametrize("dtype", ([torch.float16] if is_sm75 else [torch.float16, torch.bfloat16]))
Tri Dao's avatar
Tri Dao committed
704
# @pytest.mark.parametrize('dtype', [torch.float16])
705
706
@pytest.mark.parametrize("deterministic", [False, True])
# @pytest.mark.parametrize("deterministic", [True])
707
708
@pytest.mark.parametrize("alibi", [False, True])
# @pytest.mark.parametrize("alibi", [True])
Tri Dao's avatar
Tri Dao committed
709
710
@pytest.mark.parametrize("local", [False, True])
# @pytest.mark.parametrize("local", [True])
Tri Dao's avatar
Tri Dao committed
711
@pytest.mark.parametrize("causal", [False, True])
Tri Dao's avatar
Tri Dao committed
712
# @pytest.mark.parametrize('causal', [False])
713
@pytest.mark.parametrize("d", [32, 59, 64, 80, 96, 128, 160, 192, 224, 256])
Tri Dao's avatar
Tri Dao committed
714
# @pytest.mark.parametrize("d", [32, 64, 96, 128, 160, 192, 224, 256])
Tri Dao's avatar
Tri Dao committed
715
# @pytest.mark.parametrize('d', [64])
716
@pytest.mark.parametrize("seqlen", [97, 128, 200, 257, 384, 512, 768, 1025, 2048])
Tri Dao's avatar
Tri Dao committed
717
# @pytest.mark.parametrize('seqlen', [128])
Tri Dao's avatar
Tri Dao committed
718
@pytest.mark.parametrize("dropout_p", [0.0, 0.17])
Tri Dao's avatar
Tri Dao committed
719
# @pytest.mark.parametrize('dropout_p', [0.0])
Tri Dao's avatar
Tri Dao committed
720
721
722
def test_flash_attn_varlen_qkvpacked(
    seqlen, d, dropout_p, causal, local, alibi, deterministic, dtype
):
Tri Dao's avatar
Tri Dao committed
723
    if seqlen >= 2048 and torch.cuda.get_device_properties("cuda").total_memory <= 16 * 2**30:
Tri Dao's avatar
Tri Dao committed
724
        pytest.skip()  # Reference implementation OOM
Tri Dao's avatar
Tri Dao committed
725
    device = "cuda"
Tri Dao's avatar
Tri Dao committed
726
727
    # set seed
    torch.random.manual_seed(0)
Tri Dao's avatar
Tri Dao committed
728
729
    batch_size = 5
    nheads = 6
Tri Dao's avatar
Tri Dao committed
730
    window_size = (-1, -1) if not local else torch.randint(0, seqlen, (2,))
Tri Dao's avatar
Tri Dao committed
731
732
733
    qkv = torch.randn(
        batch_size, seqlen, 3, nheads, d, device=device, dtype=dtype, requires_grad=True
    )
Tri Dao's avatar
Tri Dao committed
734

Tri Dao's avatar
Tri Dao committed
735
    key_padding_mask = generate_random_padding_mask(seqlen, batch_size, device, mode="random")
Tri Dao's avatar
Tri Dao committed
736
    # key_padding_mask = generate_random_padding_mask(seqlen, batch_size, device, mode='full')
737
738
739
740
741
742
743
    if alibi:
        alibi_slopes = torch.rand(batch_size, nheads, device=device, dtype=torch.float32) * 0.3
        attn_bias = attn_bias_from_alibi_slopes(
            alibi_slopes, seqlen, seqlen, key_padding_mask, key_padding_mask, causal=causal
        )
    else:
        alibi_slopes, attn_bias = None, None
Tri Dao's avatar
Tri Dao committed
744

Tri Dao's avatar
Tri Dao committed
745
746
    qkv_unpad, cu_seqlens, max_seqlen, qkv, output_pad_fn, dqkv_pad_fn = generate_qkv(
        *qkv.unbind(dim=2), key_padding_mask, key_padding_mask, qkvpacked=True
Tri Dao's avatar
Tri Dao committed
747
    )
Tri Dao's avatar
Tri Dao committed
748
749

    out_unpad, sm_lse, S_dmask = flash_attn_varlen_qkvpacked_func(
Tri Dao's avatar
Tri Dao committed
750
751
752
753
754
755
        qkv_unpad,
        cu_seqlens,
        max_seqlen,
        dropout_p,
        causal=causal,
        window_size=window_size,
756
        alibi_slopes=alibi_slopes,
757
        deterministic=deterministic,
Tri Dao's avatar
Tri Dao committed
758
        return_attn_probs=True,
Tri Dao's avatar
Tri Dao committed
759
    )
Tri Dao's avatar
Tri Dao committed
760
761
762
    out = output_pad_fn(out_unpad)
    if dropout_p > 0.0:
        S_dmask_converted = convert_flash_attn_S_to_softmax(
763
764
765
766
767
768
769
770
            S_dmask,
            seqlen,
            seqlen,
            key_padding_mask,
            key_padding_mask,
            d,
            dropout_p > 0.0,
            causal=causal,
Tri Dao's avatar
Tri Dao committed
771
            window_size=window_size,
772
        )
Tri Dao's avatar
Tri Dao committed
773
774
        dropout_mask = S_dmask_converted >= 0
        attn_unnorm = S_dmask_converted.abs()
Tri Dao's avatar
Tri Dao committed
775
776
777
778
779
780
781
        attn = normalize_flash_attn_S(
            attn_unnorm,
            qkv[:, :, 0],
            qkv[:, :, 1],
            qkv[:, :, 2],
            key_padding_mask,
            key_padding_mask,
782
            attn_bias,
Tri Dao's avatar
Tri Dao committed
783
784
            dropout_p > 0.0,
            causal=causal,
Tri Dao's avatar
Tri Dao committed
785
            window_size=window_size,
Tri Dao's avatar
Tri Dao committed
786
787
        )
        dropout_fraction = get_dropout_fraction(
Tri Dao's avatar
Tri Dao committed
788
            dropout_mask, key_padding_mask, key_padding_mask, causal=causal, window_size=window_size
Tri Dao's avatar
Tri Dao committed
789
790
        ).item()
        print(f"Actual dropout fraction: {dropout_fraction}")
Tri Dao's avatar
Tri Dao committed
791
792
793
    else:
        dropout_mask = None

Tri Dao's avatar
Tri Dao committed
794
    out_ref, attn_ref = attention_qkvpacked_ref(
795
796
797
798
799
800
801
        qkv,
        key_padding_mask,
        attn_bias,
        dropout_p,
        dropout_mask,
        causal=causal,
        window_size=window_size,
Tri Dao's avatar
Tri Dao committed
802
803
804
805
    )
    out_pt, attn_pt = attention_qkvpacked_ref(
        qkv,
        key_padding_mask,
806
        attn_bias,
Tri Dao's avatar
Tri Dao committed
807
808
809
        dropout_p,
        dropout_mask,
        causal=causal,
Tri Dao's avatar
Tri Dao committed
810
        window_size=window_size,
Tri Dao's avatar
Tri Dao committed
811
812
813
814
815
816
817
        upcast=False,
        reorder_ops=True,
    )
    print(f"Output max diff: {(out - out_ref).abs().max().item()}")
    print(f"Output mean diff: {(out - out_ref).abs().mean().item()}")
    print(f"Pytorch max diff: {(out_pt - out_ref).abs().max().item()}")
    print(f"Pytorch mean diff: {(out_pt - out_ref).abs().mean().item()}")
Tri Dao's avatar
Tri Dao committed
818
    if dropout_p > 0.0:
Tri Dao's avatar
Tri Dao committed
819
820
        print(f"Attention max diff: {(attn - attn_ref).abs().max().item()}")
        print(f"Attention Pytorch max diff: {(attn_pt - attn_ref).abs().max().item()}")
Tri Dao's avatar
Tri Dao committed
821
822

    g = torch.randn_like(out)
823
    if (d <= MAX_HEADDIM_SM8x or (d > 224 and dropout_p == 0)) or (is_sm80 or is_sm90):
Tri Dao's avatar
Tri Dao committed
824
        (dqkv_unpad,) = torch.autograd.grad(out, qkv_unpad, g)
Tri Dao's avatar
Tri Dao committed
825
        dqkv = dqkv_pad_fn(dqkv_unpad)
Tri Dao's avatar
Tri Dao committed
826
827
828
829
830
831
832
833
834
835
        (dqkv_ref,) = torch.autograd.grad(out_ref, qkv, g)
        (dqkv_pt,) = torch.autograd.grad(out_pt, qkv, g)
        print(f"dQ max diff: {(dqkv[:, :, 0] - dqkv_ref[:, :, 0]).abs().max().item()}")
        print(f"dK max diff: {(dqkv[:, :, 1] - dqkv_ref[:, :, 1]).abs().max().item()}")
        print(f"dV max diff: {(dqkv[:, :, 2] - dqkv_ref[:, :, 2]).abs().max().item()}")
        print(f"dQKV mean diff: {(dqkv - dqkv_ref).abs().mean().item()}")
        print(f"dQ Pytorch max diff: {(dqkv_pt[:, :, 0] - dqkv_ref[:, :, 0]).abs().max().item()}")
        print(f"dK Pytorch max diff: {(dqkv_pt[:, :, 1] - dqkv_ref[:, :, 1]).abs().max().item()}")
        print(f"dV Pytorch max diff: {(dqkv_pt[:, :, 2] - dqkv_ref[:, :, 2]).abs().max().item()}")
        print(f"dQKV Pytorch mean diff: {(dqkv_pt - dqkv_ref).abs().mean().item()}")
Tri Dao's avatar
Tri Dao committed
836
837
838

    # Check that FlashAttention's numerical error is at most twice the numerical error
    # of a Pytorch implementation.
Tri Dao's avatar
Tri Dao committed
839
    assert (out - out_ref).abs().max().item() <= 2 * (out_pt - out_ref).abs().max().item()
Tri Dao's avatar
Tri Dao committed
840

Tri Dao's avatar
Tri Dao committed
841
842
    if dropout_p > 0.0:
        assert (attn - attn_ref).abs().max().item() <= 2 * (attn_pt - attn_ref).abs().max().item()
843
844
845
        # With alibi, many of the prob values are 0.0 & -0.0 so dropout_fraction isn't accurate
        if not alibi:
            assert abs(dropout_fraction - dropout_p) <= (0.01 if not local else 0.025)
Tri Dao's avatar
Tri Dao committed
846

847
    if (d <= MAX_HEADDIM_SM8x or (d > 224 and dropout_p == 0)) or (is_sm80 or is_sm90):
Tri Dao's avatar
Tri Dao committed
848
        assert (dqkv - dqkv_ref).abs().max().item() <= 2 * (dqkv_pt - dqkv_ref).abs().max().item()
Tri Dao's avatar
Tri Dao committed
849
850


Tri Dao's avatar
Tri Dao committed
851
@pytest.mark.parametrize("kvpacked", [True, False])
852
# @pytest.mark.parametrize("kvpacked", [False])
Tri Dao's avatar
Tri Dao committed
853
@pytest.mark.parametrize("dtype", ([torch.float16] if is_sm75 else [torch.float16, torch.bfloat16]))
854
# @pytest.mark.parametrize("dtype", [torch.bfloat16])
Tri Dao's avatar
Tri Dao committed
855
@pytest.mark.parametrize("mha_type", ["mha", "mqa", "gqa"])
856
# @pytest.mark.parametrize("mha_type", ["mha"])
857
858
@pytest.mark.parametrize("deterministic", [False, True])
# @pytest.mark.parametrize("deterministic", [True])
859
860
@pytest.mark.parametrize("alibi", [False, True])
# @pytest.mark.parametrize("alibi", [True])
Tri Dao's avatar
Tri Dao committed
861
862
@pytest.mark.parametrize("local", [False, True])
# @pytest.mark.parametrize("local", [True])
Tri Dao's avatar
Tri Dao committed
863
@pytest.mark.parametrize("causal", [False, True])
864
# @pytest.mark.parametrize("causal", [True])
865
@pytest.mark.parametrize("d", [32, 40, 59, 64, 96, 111, 128, 160, 192, 224, 256])
Tri Dao's avatar
Tri Dao committed
866
# @pytest.mark.parametrize("d", [32, 64, 96, 128, 160, 192, 224, 256])
Tri Dao's avatar
Tri Dao committed
867
868
869
# @pytest.mark.parametrize('d', [32, 40, 64, 80, 96, 128, 160, 192])
# @pytest.mark.parametrize('d', [32, 64, 96, 128, 160, 192])
# @pytest.mark.parametrize('d', [56, 80])
870
# @pytest.mark.parametrize("d", [64])
Tri Dao's avatar
Tri Dao committed
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
@pytest.mark.parametrize(
    "seqlen_q,seqlen_k",
    [
        (113, 203),
        (128, 217),
        (113, 211),
        (108, 256),
        (256, 512),
        (512, 256),
        (1024, 1024),
        (1023, 1024),
        (1024, 1023),
        (2048, 2048),
    ],
)
886
# @pytest.mark.parametrize('seqlen_q,seqlen_k', [(256, 128)])
Tri Dao's avatar
Tri Dao committed
887
@pytest.mark.parametrize("dropout_p", [0.0, 0.17])
888
# @pytest.mark.parametrize("dropout_p", [0.17])
Nicolas Patry's avatar
Nicolas Patry committed
889
@pytest.mark.parametrize("softcap", [0.0, 50.0])
Tri Dao's avatar
Tri Dao committed
890
def test_flash_attn_output(
Nicolas Patry's avatar
Nicolas Patry committed
891
    seqlen_q, seqlen_k, d, dropout_p, causal, local, alibi, deterministic, mha_type, dtype, kvpacked, softcap
Tri Dao's avatar
Tri Dao committed
892
):
Tri Dao's avatar
Tri Dao committed
893
894
895
896
    if (
        max(seqlen_q, seqlen_k) >= 2048
        and torch.cuda.get_device_properties("cuda").total_memory <= 16 * 2**30
    ):
Tri Dao's avatar
Tri Dao committed
897
        pytest.skip()  # Reference implementation OOM
Tri Dao's avatar
Tri Dao committed
898
    device = "cuda"
Tri Dao's avatar
Tri Dao committed
899
900
    # set seed
    torch.random.manual_seed(0)
901
    batch_size = 4
Tri Dao's avatar
Tri Dao committed
902
903
904
    nheads = 9
    nheads_k = nheads if mha_type == "mha" else (1 if mha_type == "mqa" else 3)
    assert nheads % nheads_k == 0
Tri Dao's avatar
Tri Dao committed
905
    window_size = (-1, -1) if not local else torch.randint(0, seqlen_k, (2,))
Tri Dao's avatar
Tri Dao committed
906
    q = torch.randn(batch_size, seqlen_q, nheads, d, device=device, dtype=dtype, requires_grad=True)
Nicolas Patry's avatar
Nicolas Patry committed
907
908
909
    if softcap > 0:
        # Ensure the values of qk are at least within softcap range.
        q = q * softcap
Tri Dao's avatar
Tri Dao committed
910
    if kvpacked:
Tri Dao's avatar
Tri Dao committed
911
912
913
        kv = torch.randn(
            batch_size, seqlen_k, 2, nheads_k, d, device=device, dtype=dtype, requires_grad=True
        )
Tri Dao's avatar
Tri Dao committed
914
    else:
Tri Dao's avatar
Tri Dao committed
915
916
917
918
919
920
        k = torch.randn(
            batch_size, seqlen_k, nheads_k, d, device=device, dtype=dtype, requires_grad=True
        )
        v = torch.randn(
            batch_size, seqlen_k, nheads_k, d, device=device, dtype=dtype, requires_grad=True
        )
921
922
923
924
925
    if alibi:
        alibi_slopes = torch.rand(batch_size, nheads, device=device, dtype=torch.float32) * 0.3
        attn_bias = attn_bias_from_alibi_slopes(alibi_slopes, seqlen_q, seqlen_k, causal=causal)
    else:
        alibi_slopes, attn_bias = None, None
Tri Dao's avatar
Tri Dao committed
926
927
928

    if kvpacked:
        out, lse, S_dmask = flash_attn_kvpacked_func(
929
930
931
932
933
            q,
            kv,
            dropout_p,
            causal=causal,
            window_size=window_size,
Nicolas Patry's avatar
Nicolas Patry committed
934
            softcap=softcap,
935
            alibi_slopes=alibi_slopes,
936
            deterministic=deterministic,
937
            return_attn_probs=True,
Tri Dao's avatar
Tri Dao committed
938
939
940
        )
    else:
        out, lse, S_dmask = flash_attn_func(
941
942
943
944
945
946
            q,
            k,
            v,
            dropout_p,
            causal=causal,
            window_size=window_size,
Nicolas Patry's avatar
Nicolas Patry committed
947
            softcap=softcap,
948
            alibi_slopes=alibi_slopes,
949
            deterministic=deterministic,
950
            return_attn_probs=True,
Tri Dao's avatar
Tri Dao committed
951
952
953
        )
    if dropout_p > 0.0:
        S_dmask_converted = convert_flash_attn_S_to_softmax(
Tri Dao's avatar
Tri Dao committed
954
955
956
957
958
959
960
961
962
            S_dmask,
            seqlen_q,
            seqlen_k,
            None,
            None,
            d,
            dropout_p > 0.0,
            causal=causal,
            window_size=window_size,
963
        )
Tri Dao's avatar
Tri Dao committed
964
965
966
967
968
969
970
971
        dropout_mask = S_dmask_converted >= 0
        attn_unnorm = S_dmask_converted.abs()
        if kvpacked:
            kv_rep = repeat(kv, "b s two h d -> b s two (h g) d", g=nheads // nheads_k)
            k_rep, v_rep = kv_rep.unbind(dim=2)
        else:
            k_rep = repeat(k, "b s h d -> b s (h g) d", g=nheads // nheads_k)
            v_rep = repeat(v, "b s h d -> b s (h g) d", g=nheads // nheads_k)
Tri Dao's avatar
Tri Dao committed
972
        attn = normalize_flash_attn_S(
Tri Dao's avatar
Tri Dao committed
973
974
975
976
977
978
            attn_unnorm,
            q,
            k_rep,
            v_rep,
            None,
            None,
979
            attn_bias,
Tri Dao's avatar
Tri Dao committed
980
981
982
            dropout_p > 0.0,
            causal=causal,
            window_size=window_size,
Tri Dao's avatar
Tri Dao committed
983
        )
Tri Dao's avatar
Tri Dao committed
984
985
986
        dropout_fraction = get_dropout_fraction(
            dropout_mask, None, None, causal=causal, window_size=window_size
        ).item()
Tri Dao's avatar
Tri Dao committed
987
        print(f"Actual dropout fraction: {dropout_fraction}")
Tri Dao's avatar
Tri Dao committed
988
989
    else:
        dropout_mask = None
Tri Dao's avatar
Tri Dao committed
990

Tri Dao's avatar
Tri Dao committed
991
    if kvpacked:
Tri Dao's avatar
Tri Dao committed
992
        out_ref, attn_ref = attention_kvpacked_ref(
Tri Dao's avatar
Tri Dao committed
993
994
995
996
            q,
            kv,
            None,
            None,
997
            attn_bias,
Tri Dao's avatar
Tri Dao committed
998
999
1000
1001
            dropout_p,
            dropout_mask,
            causal=causal,
            window_size=window_size,
Nicolas Patry's avatar
Nicolas Patry committed
1002
            softcap=softcap,
Tri Dao's avatar
Tri Dao committed
1003
1004
1005
1006
1007
1008
        )
        out_pt, attn_pt = attention_kvpacked_ref(
            q,
            kv,
            None,
            None,
1009
            attn_bias,
Tri Dao's avatar
Tri Dao committed
1010
1011
1012
            dropout_p,
            dropout_mask,
            causal=causal,
Tri Dao's avatar
Tri Dao committed
1013
            window_size=window_size,
Nicolas Patry's avatar
Nicolas Patry committed
1014
            softcap=softcap,
Tri Dao's avatar
Tri Dao committed
1015
1016
1017
            upcast=False,
            reorder_ops=True,
        )
Tri Dao's avatar
Tri Dao committed
1018
    else:
Tri Dao's avatar
Tri Dao committed
1019
        out_ref, attn_ref = attention_ref(
Tri Dao's avatar
Tri Dao committed
1020
1021
1022
1023
1024
            q,
            k,
            v,
            None,
            None,
1025
            attn_bias,
Tri Dao's avatar
Tri Dao committed
1026
1027
1028
1029
            dropout_p,
            dropout_mask,
            causal=causal,
            window_size=window_size,
Nicolas Patry's avatar
Nicolas Patry committed
1030
            softcap=softcap,
Tri Dao's avatar
Tri Dao committed
1031
1032
1033
1034
1035
1036
1037
        )
        out_pt, attn_pt = attention_ref(
            q,
            k,
            v,
            None,
            None,
1038
            attn_bias,
Tri Dao's avatar
Tri Dao committed
1039
1040
1041
            dropout_p,
            dropout_mask,
            causal=causal,
Tri Dao's avatar
Tri Dao committed
1042
            window_size=window_size,
Nicolas Patry's avatar
Nicolas Patry committed
1043
            softcap=softcap,
Tri Dao's avatar
Tri Dao committed
1044
1045
1046
1047
1048
1049
1050
1051
            upcast=False,
            reorder_ops=True,
        )

    print(f"Output max diff: {(out - out_ref).abs().max().item()}")
    print(f"Output mean diff: {(out - out_ref).abs().mean().item()}")
    print(f"Pytorch max diff: {(out_pt - out_ref).abs().max().item()}")
    print(f"Pytorch mean diff: {(out_pt - out_ref).abs().mean().item()}")
Tri Dao's avatar
Tri Dao committed
1052
    if dropout_p > 0.0:
Tri Dao's avatar
Tri Dao committed
1053
1054
        print(f"Attention max diff: {(attn - attn_ref).abs().max().item()}")
        print(f"Attention Pytorch max diff: {(attn_pt - attn_ref).abs().max().item()}")
Tri Dao's avatar
Tri Dao committed
1055
1056
1057

    g = torch.randn_like(out)
    do_o = (g.float() * out.float()).sum(-1)
1058
    if ((d <= MAX_HEADDIM_SM8x or (d > 224 and dropout_p == 0)) or (is_sm80 or is_sm90)) and softcap == 0.0:
Tri Dao's avatar
Tri Dao committed
1059
        if kvpacked:
Tri Dao's avatar
Tri Dao committed
1060
1061
1062
1063
            (
                dq,
                dkv,
            ) = torch.autograd.grad(out, (q, kv), g)
Tri Dao's avatar
Tri Dao committed
1064
            dk, dv = dkv.unbind(2)
Tri Dao's avatar
Tri Dao committed
1065
1066
1067
1068
            (
                dq_ref,
                dkv_ref,
            ) = torch.autograd.grad(out_ref, (q, kv), g)
Tri Dao's avatar
Tri Dao committed
1069
            dk_ref, dv_ref = dkv_ref.unbind(2)
Tri Dao's avatar
Tri Dao committed
1070
1071
1072
1073
            (
                dq_pt,
                dkv_pt,
            ) = torch.autograd.grad(out_pt, (q, kv), g)
Tri Dao's avatar
Tri Dao committed
1074
1075
            dk_pt, dv_pt = dkv_pt.unbind(2)
        else:
Tri Dao's avatar
Tri Dao committed
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
            (
                dq,
                dk,
                dv,
            ) = torch.autograd.grad(out, (q, k, v), g)
            (
                dq_ref,
                dk_ref,
                dv_ref,
            ) = torch.autograd.grad(out_ref, (q, k, v), g)
            (
                dq_pt,
                dk_pt,
                dv_pt,
            ) = torch.autograd.grad(out_pt, (q, k, v), g)
        print(f"dQ max diff: {(dq - dq_ref).abs().max().item()}")
        print(f"dK max diff: {(dk - dk_ref).abs().max().item()}")
        print(f"dV max diff: {(dv - dv_ref).abs().max().item()}")
        print(f"dQ mean diff: {(dq - dq_ref).abs().mean().item()}")
        print(f"dK mean diff: {(dk - dk_ref).abs().mean().item()}")
        print(f"dV mean diff: {(dv - dv_ref).abs().mean().item()}")
        print(f"dQ Pytorch max diff: {(dq_pt - dq_ref).abs().max().item()}")
        print(f"dK Pytorch max diff: {(dk_pt - dk_ref).abs().max().item()}")
        print(f"dV Pytorch max diff: {(dv_pt - dv_ref).abs().max().item()}")
        print(f"dQ Pytorch mean diff: {(dq_pt - dq_ref).abs().mean().item()}")
        print(f"dK Pytorch mean diff: {(dk_pt - dk_ref).abs().mean().item()}")
        print(f"dV Pytorch mean diff: {(dv_pt - dv_ref).abs().mean().item()}")
Tri Dao's avatar
Tri Dao committed
1103
1104
1105

    # Check that FlashAttention's numerical error is at most twice the numerical error
    # of a Pytorch implementation.
Tri Dao's avatar
Tri Dao committed
1106
1107
1108
1109
    assert (out - out_ref).abs().max().item() <= 2 * (out_pt - out_ref).abs().max().item()

    if dropout_p > 0.0:
        assert (attn - attn_ref).abs().max().item() <= 2 * (attn_pt - attn_ref).abs().max().item()
1110
1111
1112
        # With alibi, many of the prob values are 0.0 & -0.0 so dropout_fraction isn't accurate
        if not alibi:
            assert abs(dropout_fraction - dropout_p) <= (0.01 if not local else 0.025)
Tri Dao's avatar
Tri Dao committed
1113

1114
    if ((d <= MAX_HEADDIM_SM8x or (d > 224 and dropout_p == 0)) or (is_sm80 or is_sm90)) and softcap == 0.0:
Tri Dao's avatar
Tri Dao committed
1115
1116
1117
1118
1119
        assert (dq - dq_ref).abs().max().item() <= 2 * (dq_pt - dq_ref).abs().max().item()
        assert (dk - dk_ref).abs().max().item() <= 2 * (dk_pt - dk_ref).abs().max().item()
        assert (dv - dv_ref).abs().max().item() <= 2 * (dv_pt - dv_ref).abs().max().item()


Tri Dao's avatar
Tri Dao committed
1120
@pytest.mark.parametrize("kvpacked", [True, False])
Tri Dao's avatar
Tri Dao committed
1121
# @pytest.mark.parametrize('kvpacked', [False])
Tri Dao's avatar
Tri Dao committed
1122
@pytest.mark.parametrize("dtype", ([torch.float16] if is_sm75 else [torch.float16, torch.bfloat16]))
1123
# @pytest.mark.parametrize('dtype', [torch.float16])
Tri Dao's avatar
Tri Dao committed
1124
@pytest.mark.parametrize("mha_type", ["mha", "mqa", "gqa"])
Tri Dao's avatar
Tri Dao committed
1125
# @pytest.mark.parametrize('mha_type', ["mqa"])
1126
1127
@pytest.mark.parametrize("deterministic", [False, True])
# @pytest.mark.parametrize("deterministic", [True])
1128
1129
@pytest.mark.parametrize("alibi", [False, True])
# @pytest.mark.parametrize("alibi", [True])
Tri Dao's avatar
Tri Dao committed
1130
1131
@pytest.mark.parametrize("local", [False, True])
# @pytest.mark.parametrize("local", [True])
Tri Dao's avatar
Tri Dao committed
1132
@pytest.mark.parametrize("causal", [False, True])
Tri Dao's avatar
Tri Dao committed
1133
# @pytest.mark.parametrize('causal', [True])
1134
@pytest.mark.parametrize("d", [32, 59, 64, 80, 96, 111, 128, 160, 192, 224, 256])
Tri Dao's avatar
Tri Dao committed
1135
# @pytest.mark.parametrize("d", [32, 64, 96, 128, 160, 192, 224, 256])
1136
# @pytest.mark.parametrize('d', [64])
Tri Dao's avatar
Tri Dao committed
1137
1138
1139
@pytest.mark.parametrize(
    "seqlen_q,seqlen_k",
    [
1140
        (1, 147),
Tri Dao's avatar
Tri Dao committed
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
        (113, 203),
        (128, 217),
        (113, 211),
        (108, 256),
        (256, 512),
        (512, 256),
        (1024, 1024),
        (1023, 1024),
        (1024, 1023),
        (2048, 2048),
    ],
)
Tri Dao's avatar
Tri Dao committed
1153
# @pytest.mark.parametrize('seqlen_q,seqlen_k', [(128, 128)])
Tri Dao's avatar
Tri Dao committed
1154
@pytest.mark.parametrize("dropout_p", [0.0, 0.17])
Nicolas Patry's avatar
Nicolas Patry committed
1155
@pytest.mark.parametrize("softcap", [0.0, 50.0])
1156
# @pytest.mark.parametrize('dropout_p', [0.0])
Tri Dao's avatar
Tri Dao committed
1157
def test_flash_attn_varlen_output(
Nicolas Patry's avatar
Nicolas Patry committed
1158
    seqlen_q, seqlen_k, d, dropout_p, causal, local, alibi, deterministic, mha_type, dtype, kvpacked, softcap
Tri Dao's avatar
Tri Dao committed
1159
1160
1161
1162
1163
):
    if (
        max(seqlen_q, seqlen_k) >= 2048
        and torch.cuda.get_device_properties("cuda").total_memory <= 16 * 2**30
    ):
1164
        pytest.skip()  # Reference implementation OOM
Tri Dao's avatar
Tri Dao committed
1165
    device = "cuda"
1166
1167
    # set seed
    torch.random.manual_seed(0)
1168
    batch_size = 4
Tri Dao's avatar
Tri Dao committed
1169
1170
1171
    nheads = 9
    nheads_k = nheads if mha_type == "mha" else (1 if mha_type == "mqa" else 3)
    assert nheads % nheads_k == 0
Tri Dao's avatar
Tri Dao committed
1172
    window_size = (-1, -1) if not local else torch.randint(0, seqlen_k, (2,))
Tri Dao's avatar
Tri Dao committed
1173
    q = torch.randn(batch_size, seqlen_q, nheads, d, device=device, dtype=dtype, requires_grad=True)
Nicolas Patry's avatar
Nicolas Patry committed
1174
1175
1176
    if softcap > 0:
        # Ensure the values of qk are at least within softcap range.
        q = q * softcap
1177

Tri Dao's avatar
Tri Dao committed
1178
    if kvpacked:
Tri Dao's avatar
Tri Dao committed
1179
1180
1181
        kv = torch.randn(
            batch_size, seqlen_k, 2, nheads_k, d, device=device, dtype=dtype, requires_grad=True
        )
1182
    else:
Tri Dao's avatar
Tri Dao committed
1183
1184
1185
1186
1187
1188
        k = torch.randn(
            batch_size, seqlen_k, nheads_k, d, device=device, dtype=dtype, requires_grad=True
        )
        v = torch.randn(
            batch_size, seqlen_k, nheads_k, d, device=device, dtype=dtype, requires_grad=True
        )
Tri Dao's avatar
Tri Dao committed
1189

Tri Dao's avatar
Tri Dao committed
1190
1191
    query_padding_mask = generate_random_padding_mask(seqlen_q, batch_size, device, mode="random")
    key_padding_mask = generate_random_padding_mask(seqlen_k, batch_size, device, mode="random")
Tri Dao's avatar
Tri Dao committed
1192
    # key_padding_mask = generate_random_padding_mask(seqlen_k, batch_size, device, mode='full')
1193
1194
1195
1196
1197
1198
1199
    if alibi:
        alibi_slopes = torch.rand(batch_size, nheads, device=device, dtype=torch.float32) * 0.3
        attn_bias = attn_bias_from_alibi_slopes(
            alibi_slopes, seqlen_q, seqlen_k, query_padding_mask, key_padding_mask, causal=causal
        )
    else:
        alibi_slopes, attn_bias = None, None
Tri Dao's avatar
Tri Dao committed
1200
1201

    if kvpacked:
Tri Dao's avatar
Tri Dao committed
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
        (
            q_unpad,
            kv_unpad,
            cu_seqlens_q,
            cu_seqlens_k,
            max_seqlen_q,
            max_seqlen_k,
            q,
            kv,
            output_pad_fn,
            dq_pad_fn,
            dkv_pad_fn,
        ) = generate_qkv(q, *kv.unbind(dim=2), query_padding_mask, key_padding_mask, kvpacked=True)
Tri Dao's avatar
Tri Dao committed
1215
        out_unpad, sm_lse, S_dmask = flash_attn_varlen_kvpacked_func(
Tri Dao's avatar
Tri Dao committed
1216
1217
1218
1219
1220
1221
1222
1223
            q_unpad,
            kv_unpad,
            cu_seqlens_q,
            cu_seqlens_k,
            max_seqlen_q,
            max_seqlen_k,
            dropout_p,
            causal=causal,
Tri Dao's avatar
Tri Dao committed
1224
            window_size=window_size,
Nicolas Patry's avatar
Nicolas Patry committed
1225
            softcap=softcap,
1226
            alibi_slopes=alibi_slopes,
1227
            deterministic=deterministic,
1228
            return_attn_probs=True,
Tri Dao's avatar
Tri Dao committed
1229
1230
        )
    else:
Tri Dao's avatar
Tri Dao committed
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
        (
            q_unpad,
            k_unpad,
            v_unpad,
            cu_seqlens_q,
            cu_seqlens_k,
            max_seqlen_q,
            max_seqlen_k,
            q,
            k,
            v,
            output_pad_fn,
            dq_pad_fn,
            dk_pad_fn,
        ) = generate_qkv(q, k, v, query_padding_mask, key_padding_mask, kvpacked=False)
Tri Dao's avatar
Tri Dao committed
1246
        out_unpad, sm_lse, S_dmask = flash_attn_varlen_func(
Tri Dao's avatar
Tri Dao committed
1247
1248
1249
1250
1251
1252
1253
1254
1255
            q_unpad,
            k_unpad,
            v_unpad,
            cu_seqlens_q,
            cu_seqlens_k,
            max_seqlen_q,
            max_seqlen_k,
            dropout_p,
            causal=causal,
Tri Dao's avatar
Tri Dao committed
1256
            window_size=window_size,
Nicolas Patry's avatar
Nicolas Patry committed
1257
            softcap=softcap,
1258
            alibi_slopes=alibi_slopes,
1259
            deterministic=deterministic,
1260
            return_attn_probs=True,
Tri Dao's avatar
Tri Dao committed
1261
        )
Tri Dao's avatar
Tri Dao committed
1262
1263
    out = output_pad_fn(out_unpad)
    if dropout_p > 0.0:
Tri Dao's avatar
Tri Dao committed
1264
        S_dmask_converted = convert_flash_attn_S_to_softmax(
1265
1266
1267
1268
1269
1270
1271
1272
            S_dmask,
            seqlen_q,
            seqlen_k,
            query_padding_mask,
            key_padding_mask,
            d,
            dropout_p > 0.0,
            causal=causal,
Tri Dao's avatar
Tri Dao committed
1273
            window_size=window_size,
1274
        )
Tri Dao's avatar
Tri Dao committed
1275
1276
1277
1278
1279
1280
1281
1282
        dropout_mask = S_dmask_converted >= 0
        attn_unnorm = S_dmask_converted.abs()
        if kvpacked:
            kv_rep = repeat(kv, "b s two h d -> b s two (h g) d", g=nheads // nheads_k)
            k_rep, v_rep = kv_rep.unbind(dim=2)
        else:
            k_rep = repeat(k, "b s h d -> b s (h g) d", g=nheads // nheads_k)
            v_rep = repeat(v, "b s h d -> b s (h g) d", g=nheads // nheads_k)
Tri Dao's avatar
Tri Dao committed
1283
1284
1285
1286
1287
1288
1289
        attn = normalize_flash_attn_S(
            attn_unnorm,
            q,
            k_rep,
            v_rep,
            query_padding_mask,
            key_padding_mask,
1290
            attn_bias,
Tri Dao's avatar
Tri Dao committed
1291
1292
            dropout_p > 0.0,
            causal=causal,
Tri Dao's avatar
Tri Dao committed
1293
            window_size=window_size,
Tri Dao's avatar
Tri Dao committed
1294
1295
        )
        dropout_fraction = get_dropout_fraction(
Tri Dao's avatar
Tri Dao committed
1296
1297
1298
1299
1300
            dropout_mask,
            query_padding_mask,
            key_padding_mask,
            causal=causal,
            window_size=window_size,
Tri Dao's avatar
Tri Dao committed
1301
1302
        ).item()
        print(f"Actual dropout fraction: {dropout_fraction}")
Tri Dao's avatar
Tri Dao committed
1303
1304
1305
1306
    else:
        dropout_mask = None

    if kvpacked:
Tri Dao's avatar
Tri Dao committed
1307
        out_ref, attn_ref = attention_kvpacked_ref(
Tri Dao's avatar
Tri Dao committed
1308
1309
1310
1311
            q,
            kv,
            query_padding_mask,
            key_padding_mask,
1312
            attn_bias,
Tri Dao's avatar
Tri Dao committed
1313
1314
1315
1316
            dropout_p,
            dropout_mask,
            causal=causal,
            window_size=window_size,
Nicolas Patry's avatar
Nicolas Patry committed
1317
            softcap=softcap,
Tri Dao's avatar
Tri Dao committed
1318
1319
1320
1321
1322
1323
        )
        out_pt, attn_pt = attention_kvpacked_ref(
            q,
            kv,
            query_padding_mask,
            key_padding_mask,
1324
            attn_bias,
Tri Dao's avatar
Tri Dao committed
1325
1326
1327
            dropout_p,
            dropout_mask,
            causal=causal,
Tri Dao's avatar
Tri Dao committed
1328
            window_size=window_size,
Nicolas Patry's avatar
Nicolas Patry committed
1329
            softcap=softcap,
Tri Dao's avatar
Tri Dao committed
1330
1331
1332
            upcast=False,
            reorder_ops=True,
        )
Tri Dao's avatar
Tri Dao committed
1333
    else:
Tri Dao's avatar
Tri Dao committed
1334
        out_ref, attn_ref = attention_ref(
Tri Dao's avatar
Tri Dao committed
1335
1336
1337
1338
1339
            q,
            k,
            v,
            query_padding_mask,
            key_padding_mask,
1340
            attn_bias,
Tri Dao's avatar
Tri Dao committed
1341
1342
1343
1344
            dropout_p,
            dropout_mask,
            causal=causal,
            window_size=window_size,
Nicolas Patry's avatar
Nicolas Patry committed
1345
            softcap=softcap,
Tri Dao's avatar
Tri Dao committed
1346
1347
1348
1349
1350
1351
1352
        )
        out_pt, attn_pt = attention_ref(
            q,
            k,
            v,
            query_padding_mask,
            key_padding_mask,
1353
            attn_bias,
Tri Dao's avatar
Tri Dao committed
1354
1355
1356
            dropout_p,
            dropout_mask,
            causal=causal,
Tri Dao's avatar
Tri Dao committed
1357
            window_size=window_size,
Nicolas Patry's avatar
Nicolas Patry committed
1358
            softcap=softcap,
Tri Dao's avatar
Tri Dao committed
1359
1360
1361
1362
1363
1364
1365
1366
            upcast=False,
            reorder_ops=True,
        )

    print(f"Output max diff: {(out - out_ref).abs().max().item()}")
    print(f"Output mean diff: {(out - out_ref).abs().mean().item()}")
    print(f"Pytorch max diff: {(out_pt - out_ref).abs().max().item()}")
    print(f"Pytorch mean diff: {(out_pt - out_ref).abs().mean().item()}")
Tri Dao's avatar
Tri Dao committed
1367
    if dropout_p > 0.0:
Tri Dao's avatar
Tri Dao committed
1368
1369
        print(f"Attention max diff: {(attn - attn_ref).abs().max().item()}")
        print(f"Attention Pytorch max diff: {(attn_pt - attn_ref).abs().max().item()}")
Tri Dao's avatar
Tri Dao committed
1370
1371

    g = torch.randn_like(out)
1372
    if ((d <= MAX_HEADDIM_SM8x or (d > 224 and dropout_p == 0)) or (is_sm80 or is_sm90)) and softcap == 0.0:
Tri Dao's avatar
Tri Dao committed
1373
        if kvpacked:
Tri Dao's avatar
Tri Dao committed
1374
1375
1376
1377
            (
                dq_unpad,
                dkv_unpad,
            ) = torch.autograd.grad(out, (q_unpad, kv_unpad), g)
Tri Dao's avatar
Tri Dao committed
1378
            dk, dv = dkv_pad_fn(dkv_unpad).unbind(2)
Tri Dao's avatar
Tri Dao committed
1379
1380
1381
1382
            (
                dq_ref,
                dkv_ref,
            ) = torch.autograd.grad(out_ref, (q, kv), g)
Tri Dao's avatar
Tri Dao committed
1383
            dk_ref, dv_ref = dkv_ref.unbind(2)
Tri Dao's avatar
Tri Dao committed
1384
1385
1386
1387
            (
                dq_pt,
                dkv_pt,
            ) = torch.autograd.grad(out_pt, (q, kv), g)
Tri Dao's avatar
Tri Dao committed
1388
1389
            dk_pt, dv_pt = dkv_pt.unbind(2)
        else:
Tri Dao's avatar
Tri Dao committed
1390
1391
1392
1393
1394
            (
                dq_unpad,
                dk_unpad,
                dv_unpad,
            ) = torch.autograd.grad(out, (q_unpad, k_unpad, v_unpad), g)
Tri Dao's avatar
Tri Dao committed
1395
1396
            dk = dk_pad_fn(dk_unpad)
            dv = dk_pad_fn(dv_unpad)
Tri Dao's avatar
Tri Dao committed
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
            (
                dq_ref,
                dk_ref,
                dv_ref,
            ) = torch.autograd.grad(out_ref, (q, k, v), g)
            (
                dq_pt,
                dk_pt,
                dv_pt,
            ) = torch.autograd.grad(out_pt, (q, k, v), g)
Tri Dao's avatar
Tri Dao committed
1407
        dq = dq_pad_fn(dq_unpad)
Tri Dao's avatar
Tri Dao committed
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
        print(f"dQ max diff: {(dq - dq_ref).abs().max().item()}")
        print(f"dK max diff: {(dk - dk_ref).abs().max().item()}")
        print(f"dV max diff: {(dv - dv_ref).abs().max().item()}")
        print(f"dQ mean diff: {(dq - dq_ref).abs().mean().item()}")
        print(f"dK mean diff: {(dk - dk_ref).abs().mean().item()}")
        print(f"dV mean diff: {(dv - dv_ref).abs().mean().item()}")
        print(f"dQ Pytorch max diff: {(dq_pt - dq_ref).abs().max().item()}")
        print(f"dK Pytorch max diff: {(dk_pt - dk_ref).abs().max().item()}")
        print(f"dV Pytorch max diff: {(dv_pt - dv_ref).abs().max().item()}")
        print(f"dQ Pytorch mean diff: {(dq_pt - dq_ref).abs().mean().item()}")
        print(f"dK Pytorch mean diff: {(dk_pt - dk_ref).abs().mean().item()}")
        print(f"dV Pytorch mean diff: {(dv_pt - dv_ref).abs().mean().item()}")
1420
1421
1422

    # Check that FlashAttention's numerical error is at most twice the numerical error
    # of a Pytorch implementation.
Tri Dao's avatar
Tri Dao committed
1423
    assert (out - out_ref).abs().max().item() <= 2 * (out_pt - out_ref).abs().max().item()
1424

Tri Dao's avatar
Tri Dao committed
1425
1426
    if dropout_p > 0.0:
        assert (attn - attn_ref).abs().max().item() <= 2 * (attn_pt - attn_ref).abs().max().item()
1427
1428
1429
        # With alibi, many of the prob values are 0.0 & -0.0 so dropout_fraction isn't accurate
        if not alibi:
            assert abs(dropout_fraction - dropout_p) <= (0.01 if not local else 0.025)
Tri Dao's avatar
Tri Dao committed
1430

1431
    if ((d <= MAX_HEADDIM_SM8x or (d > 224 and dropout_p == 0)) or (is_sm80 or is_sm90)) and softcap == 0.0:
1432
1433
1434
        assert (dq - dq_ref).abs().max().item() <= 3 * (dq_pt - dq_ref).abs().max().item()
        assert (dk - dk_ref).abs().max().item() <= 3 * (dk_pt - dk_ref).abs().max().item()
        assert (dv - dv_ref).abs().max().item() <= 3 * (dv_pt - dv_ref).abs().max().item()
Tri Dao's avatar
Tri Dao committed
1435

1436

Tri Dao's avatar
Tri Dao committed
1437
@pytest.mark.parametrize("dtype", ([torch.float16] if is_sm75 else [torch.float16, torch.bfloat16]))
1438
# @pytest.mark.parametrize("dtype", [torch.bfloat16])
Tri Dao's avatar
Tri Dao committed
1439
1440
@pytest.mark.parametrize("local", [False, True])
# @pytest.mark.parametrize("local", [True])
1441
@pytest.mark.parametrize("d", [32, 40, 59, 64, 80, 96, 111, 128, 160, 192, 224, 256])
Tri Dao's avatar
Tri Dao committed
1442
# @pytest.mark.parametrize("d", [32, 64, 96, 128, 160, 192, 224, 256])
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
# @pytest.mark.parametrize('d', [32, 40, 64, 80, 96, 128, 160, 192])
# @pytest.mark.parametrize('d', [32, 64, 96, 128, 160, 192])
# @pytest.mark.parametrize('d', [56, 80])
# @pytest.mark.parametrize("d", [64, 128])
@pytest.mark.parametrize("swap_sq_sk", [False, True])
# @pytest.mark.parametrize("swap_sq_sk", [True])
@pytest.mark.parametrize(
    "seqlen_q,seqlen_k",
    [
        (1, 239),
        (3, 799),
        (127, 512),
        (127, 513),
        (113, 203),
        (128, 217),
        (113, 211),
        (108, 256),
        (256, 512),
        (1023, 1024),
    ],
)
# @pytest.mark.parametrize('seqlen_q,seqlen_k', [(256, 128)])
Tri Dao's avatar
Tri Dao committed
1465
def test_flash_attn_causal(seqlen_q, seqlen_k, swap_sq_sk, d, local, dtype):
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
    if (
        max(seqlen_q, seqlen_k) >= 2048
        and torch.cuda.get_device_properties("cuda").total_memory <= 16 * 2**30
    ):
        pytest.skip()  # Reference implementation OOM
    if swap_sq_sk:
        seqlen_q, seqlen_k = seqlen_k, seqlen_q
    device = "cuda"
    causal = True
    # set seed
    torch.random.manual_seed(0)
1477
    batch_size = 8
1478
    nheads = 9
Tri Dao's avatar
Tri Dao committed
1479
    window_size = (-1, -1) if not local else torch.randint(0, seqlen_k, (2,))
1480
1481
1482
    q = torch.randn(batch_size, seqlen_q, nheads, d, device=device, dtype=dtype, requires_grad=True)
    k = torch.randn(batch_size, seqlen_k, nheads, d, device=device, dtype=dtype, requires_grad=True)
    v = torch.randn(batch_size, seqlen_k, nheads, d, device=device, dtype=dtype, requires_grad=True)
Tri Dao's avatar
Tri Dao committed
1483
1484
    out = flash_attn_func(q, k, v, 0.0, causal=causal, window_size=window_size)
    out_ref, attn_ref = attention_ref(
1485
        q, k, v, None, None, None, 0.0, None, causal=causal, window_size=window_size
Tri Dao's avatar
Tri Dao committed
1486
    )
1487
1488
1489
1490
1491
1492
    out_pt, attn_pt = attention_ref(
        q,
        k,
        v,
        None,
        None,
1493
        None,
1494
1495
1496
        0.0,
        None,
        causal=causal,
Tri Dao's avatar
Tri Dao committed
1497
        window_size=window_size,
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
        upcast=False,
        reorder_ops=True,
    )

    print(f"Output max diff: {(out - out_ref).abs().max().item()}")
    print(f"Output mean diff: {(out - out_ref).abs().mean().item()}")
    print(f"Pytorch max diff: {(out_pt - out_ref).abs().max().item()}")
    print(f"Pytorch mean diff: {(out_pt - out_ref).abs().mean().item()}")

    g = torch.randn_like(out)
    do_o = (g.float() * out.float()).sum(-1)
1509
    if (d <= MAX_HEADDIM_SM8x or d > 224) or (is_sm80 or is_sm90):
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
        (
            dq,
            dk,
            dv,
        ) = torch.autograd.grad(out, (q, k, v), g)
        (
            dq_ref,
            dk_ref,
            dv_ref,
        ) = torch.autograd.grad(out_ref, (q, k, v), g)
        (
            dq_pt,
            dk_pt,
            dv_pt,
        ) = torch.autograd.grad(out_pt, (q, k, v), g)
        print(f"dQ max diff: {(dq - dq_ref).abs().max().item()}")
        print(f"dK max diff: {(dk - dk_ref).abs().max().item()}")
        print(f"dV max diff: {(dv - dv_ref).abs().max().item()}")
        print(f"dQ mean diff: {(dq - dq_ref).abs().mean().item()}")
        print(f"dK mean diff: {(dk - dk_ref).abs().mean().item()}")
        print(f"dV mean diff: {(dv - dv_ref).abs().mean().item()}")
        print(f"dQ Pytorch max diff: {(dq_pt - dq_ref).abs().max().item()}")
        print(f"dK Pytorch max diff: {(dk_pt - dk_ref).abs().max().item()}")
        print(f"dV Pytorch max diff: {(dv_pt - dv_ref).abs().max().item()}")
        print(f"dQ Pytorch mean diff: {(dq_pt - dq_ref).abs().mean().item()}")
        print(f"dK Pytorch mean diff: {(dk_pt - dk_ref).abs().mean().item()}")
        print(f"dV Pytorch mean diff: {(dv_pt - dv_ref).abs().mean().item()}")

    # Check that FlashAttention's numerical error is at most twice the numerical error
    # of a Pytorch implementation.
    assert (out - out_ref).abs().max().item() <= 2 * (out_pt - out_ref).abs().max().item() + 1e-5

1542
    if (d <= MAX_HEADDIM_SM8x or d > 224) or (is_sm80 or is_sm90):
1543
1544
1545
1546
1547
1548
1549
        assert (dq - dq_ref).abs().max().item() <= 2 * (dq_pt - dq_ref).abs().max().item() + 1e-5
        assert (dk - dk_ref).abs().max().item() <= 2 * (dk_pt - dk_ref).abs().max().item() + 1e-5
        assert (dv - dv_ref).abs().max().item() <= 2 * (dv_pt - dv_ref).abs().max().item() + 1e-5


@pytest.mark.parametrize("dtype", ([torch.float16] if is_sm75 else [torch.float16, torch.bfloat16]))
# @pytest.mark.parametrize("dtype", [torch.bfloat16])
Tri Dao's avatar
Tri Dao committed
1550
1551
@pytest.mark.parametrize("local", [False, True])
# @pytest.mark.parametrize("local", [True])
1552
@pytest.mark.parametrize("d", [32, 40, 59, 64, 80, 96, 111, 128, 160, 192, 224, 256])
Tri Dao's avatar
Tri Dao committed
1553
# @pytest.mark.parametrize("d", [32, 64, 96, 128, 160, 192, 224, 256])
1554
1555
1556
# @pytest.mark.parametrize('d', [32, 40, 64, 80, 96, 128, 160, 192])
# @pytest.mark.parametrize('d', [32, 64, 96, 128, 160, 192])
# @pytest.mark.parametrize('d', [56, 80])
Tri Dao's avatar
Tri Dao committed
1557
# @pytest.mark.parametrize("d", [64])
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
@pytest.mark.parametrize("swap_sq_sk", [False, True])
# @pytest.mark.parametrize("swap_sq_sk", [True])
@pytest.mark.parametrize(
    "seqlen_q,seqlen_k",
    [
        (1, 239),
        (3, 799),
        (127, 512),
        (127, 513),
        (113, 203),
        (128, 217),
        (113, 211),
        (108, 256),
        (256, 512),
        (1023, 1024),
    ],
)
1575
1576
# TODO: add smaller page sizes when https://github.com/Dao-AILab/flash-attention/pull/824 is merged
@pytest.mark.parametrize("paged_kv_block_size", [None, 256, 512])
1577
# @pytest.mark.parametrize("seqlen_q,seqlen_k", [(256, 128)])
1578
1579
1580
def test_flash_attn_varlen_causal(
    seqlen_q, seqlen_k, swap_sq_sk, d, local, paged_kv_block_size, dtype
):
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
    if (
        max(seqlen_q, seqlen_k) >= 2048
        and torch.cuda.get_device_properties("cuda").total_memory <= 16 * 2**30
    ):
        pytest.skip()  # Reference implementation OOM
    if swap_sq_sk:
        seqlen_q, seqlen_k = seqlen_k, seqlen_q
    device = "cuda"
    causal = True
    # set seed
    torch.random.manual_seed(0)
1592
    batch_size = 8
1593
    nheads = 9
Tri Dao's avatar
Tri Dao committed
1594
    window_size = (-1, -1) if not local else torch.randint(0, seqlen_k, (2,))
1595
    q = torch.randn(batch_size, seqlen_q, nheads, d, device=device, dtype=dtype, requires_grad=True)
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608

    if paged_kv_block_size is None:
        k = torch.randn(
            batch_size, seqlen_k, nheads, d, device=device, dtype=dtype, requires_grad=True
        )
        v = torch.randn(
            batch_size, seqlen_k, nheads, d, device=device, dtype=dtype, requires_grad=True
        )
        block_table = None
    else:
        k, v, block_table, k_cache_paged, v_cache_paged, num_blocks = _generate_block_kvcache(
            seqlen_k, paged_kv_block_size, batch_size, nheads, d, device, dtype
        )
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
    query_padding_mask = generate_random_padding_mask(seqlen_q, batch_size, device, mode="random")
    key_padding_mask = generate_random_padding_mask(seqlen_k, batch_size, device, mode="random")
    (
        q_unpad,
        k_unpad,
        v_unpad,
        cu_seqlens_q,
        cu_seqlens_k,
        max_seqlen_q,
        max_seqlen_k,
        q,
        k,
        v,
        output_pad_fn,
        dq_pad_fn,
        dk_pad_fn,
    ) = generate_qkv(q, k, v, query_padding_mask, key_padding_mask, kvpacked=False)
    out_unpad = flash_attn_varlen_func(
        q_unpad,
1628
1629
        k_unpad if paged_kv_block_size is None else k_cache_paged,
        v_unpad if paged_kv_block_size is None else v_cache_paged,
1630
1631
1632
1633
1634
1635
        cu_seqlens_q,
        cu_seqlens_k,
        max_seqlen_q,
        max_seqlen_k,
        0.0,
        causal=causal,
Tri Dao's avatar
Tri Dao committed
1636
        window_size=window_size,
1637
        block_table=block_table,
1638
1639
1640
    )
    out = output_pad_fn(out_unpad)
    out_ref, attn_ref = attention_ref(
Tri Dao's avatar
Tri Dao committed
1641
1642
1643
1644
1645
        q,
        k,
        v,
        query_padding_mask,
        key_padding_mask,
1646
        None,
Tri Dao's avatar
Tri Dao committed
1647
1648
1649
1650
        0.0,
        None,
        causal=causal,
        window_size=window_size,
1651
1652
1653
1654
1655
1656
1657
    )
    out_pt, attn_pt = attention_ref(
        q,
        k,
        v,
        query_padding_mask,
        key_padding_mask,
1658
        None,
1659
1660
1661
        0.0,
        None,
        causal=causal,
Tri Dao's avatar
Tri Dao committed
1662
        window_size=window_size,
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
        upcast=False,
        reorder_ops=True,
    )

    print(f"Output max diff: {(out - out_ref).abs().max().item()}")
    print(f"Output mean diff: {(out - out_ref).abs().mean().item()}")
    print(f"Pytorch max diff: {(out_pt - out_ref).abs().max().item()}")
    print(f"Pytorch mean diff: {(out_pt - out_ref).abs().mean().item()}")

    g = torch.randn_like(out)
    do_o = (g.float() * out.float()).sum(-1)
1674
1675
    test_backward = (d <= MAX_HEADDIM_SM8x or d > 224 or is_sm80 or is_sm90) and block_table is None
    if test_backward:
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
        (
            dq_unpad,
            dk_unpad,
            dv_unpad,
        ) = torch.autograd.grad(out, (q_unpad, k_unpad, v_unpad), g)
        dq = dq_pad_fn(dq_unpad)
        dk = dk_pad_fn(dk_unpad)
        dv = dk_pad_fn(dv_unpad)
        (
            dq_ref,
            dk_ref,
            dv_ref,
        ) = torch.autograd.grad(out_ref, (q, k, v), g)
        (
            dq_pt,
            dk_pt,
            dv_pt,
        ) = torch.autograd.grad(out_pt, (q, k, v), g)
        print(f"dQ max diff: {(dq - dq_ref).abs().max().item()}")
        print(f"dK max diff: {(dk - dk_ref).abs().max().item()}")
        print(f"dV max diff: {(dv - dv_ref).abs().max().item()}")
        print(f"dQ mean diff: {(dq - dq_ref).abs().mean().item()}")
        print(f"dK mean diff: {(dk - dk_ref).abs().mean().item()}")
        print(f"dV mean diff: {(dv - dv_ref).abs().mean().item()}")
        print(f"dQ Pytorch max diff: {(dq_pt - dq_ref).abs().max().item()}")
        print(f"dK Pytorch max diff: {(dk_pt - dk_ref).abs().max().item()}")
        print(f"dV Pytorch max diff: {(dv_pt - dv_ref).abs().max().item()}")
        print(f"dQ Pytorch mean diff: {(dq_pt - dq_ref).abs().mean().item()}")
        print(f"dK Pytorch mean diff: {(dk_pt - dk_ref).abs().mean().item()}")
        print(f"dV Pytorch mean diff: {(dv_pt - dv_ref).abs().mean().item()}")

    # Check that FlashAttention's numerical error is at most twice the numerical error
    # of a Pytorch implementation.
    assert (out - out_ref).abs().max().item() <= 2 * (out_pt - out_ref).abs().max().item() + 1e-5

1711
    if test_backward:
1712
1713
1714
1715
1716
        assert (dq - dq_ref).abs().max().item() <= 2 * (dq_pt - dq_ref).abs().max().item() + 1e-5
        assert (dk - dk_ref).abs().max().item() <= 2 * (dk_pt - dk_ref).abs().max().item() + 1e-5
        assert (dv - dv_ref).abs().max().item() <= 2 * (dv_pt - dv_ref).abs().max().item() + 1e-5


Tri Dao's avatar
Tri Dao committed
1717
1718
@pytest.mark.parametrize("dtype", ([torch.float16] if is_sm75 else [torch.float16, torch.bfloat16]))
# @pytest.mark.parametrize("dtype", [torch.float16])
1719
1720
@pytest.mark.parametrize("deterministic", [False, True])
# @pytest.mark.parametrize("deterministic", [True])
1721
1722
@pytest.mark.parametrize("alibi", [False, True])
# @pytest.mark.parametrize("alibi", [True])
Tri Dao's avatar
Tri Dao committed
1723
@pytest.mark.parametrize("local", [False, True])
1724
# @pytest.mark.parametrize("local", [False])
Tri Dao's avatar
Tri Dao committed
1725
1726
1727
1728
1729
1730
1731
@pytest.mark.parametrize("causal", [False, True])
# @pytest.mark.parametrize("causal", [True])
@pytest.mark.parametrize("d", [32, 40, 59, 64, 80, 96, 111, 128, 160, 192, 224, 256])
# @pytest.mark.parametrize('d', [32, 64, 96, 128, 160, 192, 224, 256])
# @pytest.mark.parametrize('d', [32, 40, 64, 80, 96, 128, 160, 192])
# @pytest.mark.parametrize('d', [32, 64, 96, 128, 160, 192])
# @pytest.mark.parametrize('d', [56, 80])
1732
# @pytest.mark.parametrize("d", [64])
Tri Dao's avatar
Tri Dao committed
1733
1734
1735
1736
1737
1738
1739
@pytest.mark.parametrize("swap_sq_sk", [False, True])
# @pytest.mark.parametrize("swap_sq_sk", [False])
@pytest.mark.parametrize(
    "seqlen_q,seqlen_k",
    [
        (3, 1024),
        (1, 339),
1740
        (64, 800),
Tri Dao's avatar
Tri Dao committed
1741
1742
1743
1744
1745
1746
1747
1748
1749
        (3, 799),
        (64, 2048),
        (16, 20000),
        (16, 100000),
        (128, 128),
        (256, 256),
    ],
)
# @pytest.mark.parametrize('seqlen_q,seqlen_k', [(256, 128)])
Tri Dao's avatar
Tri Dao committed
1750
1751
1752
def test_flash_attn_splitkv(
    seqlen_q, seqlen_k, swap_sq_sk, d, causal, local, alibi, deterministic, dtype
):
Tri Dao's avatar
Tri Dao committed
1753
1754
1755
1756
1757
1758
1759
    if swap_sq_sk:
        seqlen_q, seqlen_k = seqlen_k, seqlen_q
    device = "cuda"
    # set seed
    torch.random.manual_seed(0)
    batch_size = 1
    nheads = 12
Tri Dao's avatar
Tri Dao committed
1760
    window_size = (-1, -1) if not local else torch.randint(0, seqlen_k, (2,))
Tri Dao's avatar
Tri Dao committed
1761
1762
1763
    q = torch.randn(batch_size, seqlen_q, nheads, d, device=device, dtype=dtype, requires_grad=True)
    k = torch.randn(batch_size, seqlen_k, nheads, d, device=device, dtype=dtype, requires_grad=True)
    v = torch.randn(batch_size, seqlen_k, nheads, d, device=device, dtype=dtype, requires_grad=True)
1764
1765
1766
1767
1768
    if alibi:
        alibi_slopes = torch.rand(batch_size, nheads, device=device, dtype=torch.float32) * 0.3
        attn_bias = attn_bias_from_alibi_slopes(alibi_slopes, seqlen_q, seqlen_k, causal=causal)
    else:
        alibi_slopes, attn_bias = None, None
Tri Dao's avatar
Tri Dao committed
1769
    out, lse, _ = flash_attn_func(
1770
1771
1772
1773
1774
1775
1776
        q,
        k,
        v,
        0.0,
        causal=causal,
        window_size=window_size,
        alibi_slopes=alibi_slopes,
1777
        deterministic=deterministic,
1778
        return_attn_probs=True,
Tri Dao's avatar
Tri Dao committed
1779
1780
    )
    out_ref, attn_ref = attention_ref(
1781
        q, k, v, None, None, attn_bias, 0.0, None, causal=causal, window_size=window_size
Tri Dao's avatar
Tri Dao committed
1782
    )
Tri Dao's avatar
Tri Dao committed
1783
1784
1785
1786
1787
1788
    out_pt, attn_pt = attention_ref(
        q,
        k,
        v,
        None,
        None,
1789
        attn_bias,
Tri Dao's avatar
Tri Dao committed
1790
1791
1792
        0.0,
        None,
        causal=causal,
Tri Dao's avatar
Tri Dao committed
1793
        window_size=window_size,
Tri Dao's avatar
Tri Dao committed
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
        upcast=False,
        reorder_ops=True,
    )

    print(f"Output max diff: {(out - out_ref).abs().max().item()}")
    print(f"Output mean diff: {(out - out_ref).abs().mean().item()}")
    print(f"Pytorch max diff: {(out_pt - out_ref).abs().max().item()}")
    print(f"Pytorch mean diff: {(out_pt - out_ref).abs().mean().item()}")

    g = torch.randn_like(out)
    do_o = (g.float() * out.float()).sum(-1)
1805
    if (d <= MAX_HEADDIM_SM8x or d > 224) or (is_sm80 or is_sm90):
Tri Dao's avatar
Tri Dao committed
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
        (
            dq,
            dk,
            dv,
        ) = torch.autograd.grad(out, (q, k, v), g)
        (
            dq_ref,
            dk_ref,
            dv_ref,
        ) = torch.autograd.grad(out_ref, (q, k, v), g)
        (
            dq_pt,
            dk_pt,
            dv_pt,
        ) = torch.autograd.grad(out_pt, (q, k, v), g)
        print(f"dQ max diff: {(dq - dq_ref).abs().max().item()}")
        print(f"dK max diff: {(dk - dk_ref).abs().max().item()}")
        print(f"dV max diff: {(dv - dv_ref).abs().max().item()}")
        print(f"dQ mean diff: {(dq - dq_ref).abs().mean().item()}")
        print(f"dK mean diff: {(dk - dk_ref).abs().mean().item()}")
        print(f"dV mean diff: {(dv - dv_ref).abs().mean().item()}")
        print(f"dQ Pytorch max diff: {(dq_pt - dq_ref).abs().max().item()}")
        print(f"dK Pytorch max diff: {(dk_pt - dk_ref).abs().max().item()}")
        print(f"dV Pytorch max diff: {(dv_pt - dv_ref).abs().max().item()}")
        print(f"dQ Pytorch mean diff: {(dq_pt - dq_ref).abs().mean().item()}")
        print(f"dK Pytorch mean diff: {(dk_pt - dk_ref).abs().mean().item()}")
        print(f"dV Pytorch mean diff: {(dv_pt - dv_ref).abs().mean().item()}")

    # Check that FlashAttention's numerical error is at most twice the numerical error
    # of a Pytorch implementation.
    assert (out - out_ref).abs().max().item() <= 2 * (out_pt - out_ref).abs().max().item() + 1e-5

1838
    mult = 2 if not alibi else 8
1839
    if (d <= MAX_HEADDIM_SM8x or d > 224) or (is_sm80 or is_sm90):
1840
1841
1842
        assert (dq - dq_ref).abs().max().item() <= mult * (dq_pt - dq_ref).abs().max().item() + 2e-4
        assert (dk - dk_ref).abs().max().item() <= mult * (dk_pt - dk_ref).abs().max().item() + 2e-4
        assert (dv - dv_ref).abs().max().item() <= mult * (dv_pt - dv_ref).abs().max().item() + 2e-4
Tri Dao's avatar
Tri Dao committed
1843

1844

1845
1846
# @pytest.mark.parametrize("dtype", ([torch.float16] if is_sm75 else [torch.float16, torch.bfloat16]))
@pytest.mark.parametrize("dtype", [torch.float16])
Tri Dao's avatar
Tri Dao committed
1847
@pytest.mark.parametrize("num_splits", [1, 0])
1848
# @pytest.mark.parametrize("num_splits", [1])
Tri Dao's avatar
Tri Dao committed
1849
@pytest.mark.parametrize("mha_type", ["mha", "mqa", "gqa"])
1850
# @pytest.mark.parametrize("mha_type", ["mha"])
Tri Dao's avatar
Tri Dao committed
1851
@pytest.mark.parametrize("new_kv", [False, True])
1852
1853
# @pytest.mark.parametrize("new_kv", [False])
@pytest.mark.parametrize("alibi", [False, True])
Tri Dao's avatar
Tri Dao committed
1854
# @pytest.mark.parametrize("alibi", [False])
Tri Dao's avatar
Tri Dao committed
1855
@pytest.mark.parametrize("local", [False, True])
1856
# @pytest.mark.parametrize("local", [False])
Tri Dao's avatar
Tri Dao committed
1857
@pytest.mark.parametrize("causal", [False, True])
1858
# @pytest.mark.parametrize("causal", [False])
1859
@pytest.mark.parametrize("seqlen_new_eq_seqlen_q", [True, False])
1860
1861
1862
1863
# @pytest.mark.parametrize("seqlen_new_eq_seqlen_q", [True])
@pytest.mark.parametrize("rotary_interleaved", [False, True])
# @pytest.mark.parametrize("rotary_interleaved", [False])
@pytest.mark.parametrize("rotary_fraction", [0.0, 0.5, 1.0])
1864
# @pytest.mark.parametrize("rotary_fraction", [0.0])
1865
1866
1867
@pytest.mark.parametrize("paged_kv_block_size", [None, 256])
# @pytest.mark.parametrize("paged_kv_block_size", [256, 512])
# @pytest.mark.parametrize("paged_kv_block_size", [256])
1868
@pytest.mark.parametrize("has_batch_idx", [False, True])
1869
# @pytest.mark.parametrize("has_batch_idx", [False])
Tri Dao's avatar
Tri Dao committed
1870
1871
@pytest.mark.parametrize("d", [32, 59, 64, 80, 128, 256])
# @pytest.mark.parametrize("d", [32, 64, 96, 128, 160, 192, 224, 256])
Tri Dao's avatar
Tri Dao committed
1872
1873
# @pytest.mark.parametrize('d', [32, 40, 64, 80, 96, 128, 160, 192])
# @pytest.mark.parametrize('d', [56, 80])
1874
# @pytest.mark.parametrize("d", [128])
Tri Dao's avatar
Tri Dao committed
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
@pytest.mark.parametrize(
    "seqlen_q,seqlen_k",
    [
        (1, 128),
        (1, 339),
        (3, 1024),
        (64, 800),
        (64, 256),
        (3, 799),
        (64, 2048),
        (16, 20000),
        (1, 128 * 1024),
        (16, 128 * 1024),
        (128, 128),
    ],
)
# @pytest.mark.parametrize('seqlen_q,seqlen_k', [(256, 128)])
1892
def test_flash_attn_kvcache(
1893
1894
1895
    seqlen_q,
    seqlen_k,
    d,
1896
    has_batch_idx,
Tri Dao's avatar
Tri Dao committed
1897
    paged_kv_block_size,
1898
1899
1900
1901
    rotary_fraction,
    rotary_interleaved,
    seqlen_new_eq_seqlen_q,
    causal,
Tri Dao's avatar
Tri Dao committed
1902
    local,
1903
    alibi,
1904
1905
1906
1907
    new_kv,
    mha_type,
    num_splits,
    dtype,
1908
):
Tri Dao's avatar
Tri Dao committed
1909
1910
    if seqlen_q > seqlen_k and new_kv:
        pytest.skip()
1911
1912
    if not new_kv and rotary_fraction > 0.0:
        pytest.skip()
Tri Dao's avatar
Tri Dao committed
1913
1914
    if has_batch_idx and paged_kv_block_size is not None:
        pytest.skip()
Tri Dao's avatar
Tri Dao committed
1915
1916
1917
1918
    device = "cuda"
    # set seed
    torch.random.manual_seed(0)
    batch_size = 2
1919
    batch_size_cache = batch_size if not has_batch_idx else batch_size * 2
Tri Dao's avatar
Tri Dao committed
1920
    nheads = 6
1921
1922
    # rotary_dim must be a multiple of 16, and must be <= d
    rotary_dim = math.floor(int(rotary_fraction * d) / 16) * 16
Tri Dao's avatar
Tri Dao committed
1923
1924
    nheads_k = nheads if mha_type == "mha" else (1 if mha_type == "mqa" else 3)
    assert nheads % nheads_k == 0
Tri Dao's avatar
Tri Dao committed
1925
    window_size = (-1, -1) if not local else torch.randint(0, seqlen_k, (2,))
Tri Dao's avatar
Tri Dao committed
1926
    q = torch.randn(batch_size, seqlen_q, nheads, d, device=device, dtype=dtype)
1927
    seqlen_new = seqlen_q if seqlen_new_eq_seqlen_q else torch.randint(1, seqlen_q + 1, (1,)).item()
Tri Dao's avatar
Tri Dao committed
1928
    if new_kv:
1929
1930
        k = torch.randn(batch_size, seqlen_new, nheads_k, d, device=device, dtype=dtype)
        v = torch.randn(batch_size, seqlen_new, nheads_k, d, device=device, dtype=dtype)
Tri Dao's avatar
Tri Dao committed
1931
1932
    else:
        k, v = None, None
Tri Dao's avatar
Tri Dao committed
1933
1934
1935
1936
1937
    if paged_kv_block_size is None:
        k_cache = torch.randn(batch_size_cache, seqlen_k, nheads_k, d, device=device, dtype=dtype)
        v_cache = torch.randn(batch_size_cache, seqlen_k, nheads_k, d, device=device, dtype=dtype)
        block_table = None
    else:
1938
1939
1940
1941
1942
1943
1944
1945
1946
        (
            k_cache,
            v_cache,
            block_table,
            k_cache_paged,
            v_cache_paged,
            num_blocks,
        ) = _generate_block_kvcache(
            seqlen_k, paged_kv_block_size, batch_size, nheads_k, d, device, dtype
Tri Dao's avatar
Tri Dao committed
1947
        )
1948
    cache_seqlens = torch.randint(
Tri Dao's avatar
Tri Dao committed
1949
        0 if new_kv else 1,
1950
        # If we don't use seqlen_q in the case of causal and rotary, cos/sin won't be long enough
1951
1952
1953
1954
1955
        (
            (seqlen_k - (seqlen_q if (causal or local) and rotary_dim > 1 else seqlen_new) + 1)
            if new_kv
            else (seqlen_k + 1)
        ),
1956
1957
1958
1959
        (batch_size,),
        dtype=torch.int32,
        device=device,
    )
1960
1961
1962
    arange = rearrange(torch.arange(seqlen_k, device=device), "s -> 1 s")
    cache_seqlens_expanded = rearrange(cache_seqlens, "b -> b 1")
    key_padding_mask = arange < cache_seqlens_expanded + (seqlen_new if new_kv else 0)
1963
    if has_batch_idx:
1964
1965
1966
        cache_batch_idx = torch.randperm(batch_size_cache, dtype=torch.int32, device=device)[
            :batch_size
        ]
1967
1968
    else:
        cache_batch_idx = None
1969
1970
1971
1972
1973
1974
1975
    if alibi:
        alibi_slopes = torch.rand(batch_size, nheads, device=device, dtype=torch.float32) * 0.3
        attn_bias = attn_bias_from_alibi_slopes(
            alibi_slopes, seqlen_q, seqlen_k, None, key_padding_mask, causal=causal
        )
    else:
        alibi_slopes, attn_bias = None, None
Tri Dao's avatar
Tri Dao committed
1976
    # cache_seqlens = torch.tensor([64], dtype=torch.int32, device=device)
1977
    if rotary_dim > 0:
Tri Dao's avatar
Tri Dao committed
1978
1979
1980
1981
1982
1983
1984
1985
1986
        angle = (
            torch.rand(
                seqlen_k if paged_kv_block_size is None else num_blocks * paged_kv_block_size,
                rotary_dim // 2,
                device=device,
            )
            * 2
            * math.pi
        )
1987
1988
        cos = torch.cos(angle).to(dtype=dtype)
        sin = torch.sin(angle).to(dtype=dtype)
Tri Dao's avatar
Tri Dao committed
1989
        if causal or local:
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
            q_ro = apply_rotary_emb(
                q, cos, sin, seqlen_offsets=cache_seqlens, interleaved=rotary_interleaved
            )
        else:
            q_ro = rearrange(
                apply_rotary_emb(
                    rearrange(q, "b s h d -> b 1 (s h) d"),
                    cos,
                    sin,
                    seqlen_offsets=cache_seqlens,
                    interleaved=rotary_interleaved,
                ),
                "b 1 (s h) d -> b s h d",
                s=seqlen_q,
            )
        # q_ro = q
        k_ro = apply_rotary_emb(
            k, cos, sin, seqlen_offsets=cache_seqlens, interleaved=rotary_interleaved
        )
    else:
        cos, sin = None, None
        q_ro, k_ro = q, k
Tri Dao's avatar
Tri Dao committed
2012
    # k_cache[:, 64:] = -1
2013
2014
2015
2016
2017
2018
    k_cache_ref = (
        k_cache if not has_batch_idx else k_cache[cache_batch_idx.to(dtype=torch.long)]
    ).clone()
    v_cache_ref = (
        v_cache if not has_batch_idx else v_cache[cache_batch_idx.to(dtype=torch.long)]
    ).clone()
Tri Dao's avatar
Tri Dao committed
2019
    if new_kv:
2020
2021
2022
        update_mask = torch.logical_and(
            cache_seqlens_expanded <= arange, arange < cache_seqlens_expanded + seqlen_new
        )
2023
        k_cache_ref[update_mask] = rearrange(k_ro, "b s ... -> (b s) ...")
Tri Dao's avatar
Tri Dao committed
2024
2025
2026
        v_cache_ref[update_mask] = rearrange(v, "b s ... -> (b s) ...")
    k_cache_rep = repeat(k_cache_ref, "b s h d -> b s (h g) d", g=nheads // nheads_k)
    v_cache_rep = repeat(v_cache_ref, "b s h d -> b s (h g) d", g=nheads // nheads_k)
2027
    out = flash_attn_with_kvcache(
2028
        q,
Tri Dao's avatar
Tri Dao committed
2029
2030
        k_cache if paged_kv_block_size is None else k_cache_paged,
        v_cache if paged_kv_block_size is None else v_cache_paged,
2031
2032
        k,
        v,
Tri Dao's avatar
Tri Dao committed
2033
2034
2035
2036
2037
        rotary_cos=cos,
        rotary_sin=sin,
        cache_seqlens=cache_seqlens,
        cache_batch_idx=cache_batch_idx,
        block_table=block_table,
2038
        causal=causal,
Tri Dao's avatar
Tri Dao committed
2039
        window_size=window_size,
2040
        rotary_interleaved=rotary_interleaved,
2041
        alibi_slopes=alibi_slopes,
2042
        num_splits=num_splits,
2043
    )
Tri Dao's avatar
Tri Dao committed
2044
2045
2046
2047
    # out = flash_attn_with_kvcache(
    #     q, k_cache, v_cache, cache_seqlens=cache_seqlens, causal=causal, window_size=window_size
    # )
    # out = flash_attn_with_kvcache(q, k_cache, v_cache, causal=causal, window_size=window_size)
Tri Dao's avatar
Tri Dao committed
2048
2049
2050
2051
2052
2053
    # qk = torch.einsum("bqhd,bkhd->bhqk", q, k_cache_ref)
    # m = qk.amax(-1, keepdim=True)
    # s_tmp = torch.exp((qk - m) / math.sqrt(d))
    # o1 = torch.einsum('bhst,bthd->bshd', s_tmp, v_cache_ref)
    # lse_ref = torch.logsumexp(qk / math.sqrt(d), -1)
    # probs = torch.softmax(qk, dim=-1)
2054
    out_ref, _ = attention_ref(
Tri Dao's avatar
Tri Dao committed
2055
2056
2057
2058
2059
        q_ro,
        k_cache_rep,
        v_cache_rep,
        None,
        key_padding_mask,
2060
        attn_bias,
Tri Dao's avatar
Tri Dao committed
2061
2062
2063
2064
        0.0,
        None,
        causal=causal,
        window_size=window_size,
2065
2066
    )
    out_pt, _ = attention_ref(
2067
        q_ro,
2068
2069
2070
2071
        k_cache_rep,
        v_cache_rep,
        None,
        key_padding_mask,
2072
        attn_bias,
2073
2074
2075
        0.0,
        None,
        causal=causal,
Tri Dao's avatar
Tri Dao committed
2076
        window_size=window_size,
2077
2078
2079
        upcast=False,
        reorder_ops=True,
    )
Tri Dao's avatar
Tri Dao committed
2080
2081
2082
2083
2084
2085
2086
2087
    print(f"Output max diff: {(out - out_ref).abs().max().item()}")
    print(f"Output mean diff: {(out - out_ref).abs().mean().item()}")
    print(f"Pytorch max diff: {(out_pt - out_ref).abs().max().item()}")
    print(f"Pytorch mean diff: {(out_pt - out_ref).abs().mean().item()}")

    # Check that FlashAttention's numerical error is at most twice the numerical error
    # of a Pytorch implementation.
    if new_kv:
Tri Dao's avatar
Tri Dao committed
2088
        if paged_kv_block_size is None:
2089
2090
2091
2092
2093
2094
            k_cache_select = (
                k_cache if not has_batch_idx else k_cache[cache_batch_idx.to(dtype=torch.long)]
            )
            v_cache_select = (
                v_cache if not has_batch_idx else v_cache[cache_batch_idx.to(dtype=torch.long)]
            )
Tri Dao's avatar
Tri Dao committed
2095
2096
        else:
            k_cache_select = rearrange(
2097
                k_cache_paged[block_table.to(dtype=torch.long).flatten()],
Tri Dao's avatar
Tri Dao committed
2098
2099
2100
2101
                "(b nblocks) block_size ... -> b (nblocks block_size) ...",
                b=batch_size,
            )[:, :seqlen_k]
            v_cache_select = rearrange(
2102
                v_cache_paged[block_table.to(dtype=torch.long).flatten()],
Tri Dao's avatar
Tri Dao committed
2103
2104
2105
                "(b nblocks) block_size ... -> b (nblocks block_size) ...",
                b=batch_size,
            )[:, :seqlen_k]
2106
2107
        assert torch.allclose(k_cache_select, k_cache_ref, rtol=1e-3, atol=1e-3)
        assert torch.equal(v_cache_select, v_cache_ref)
2108
2109
    mult = 3 if not alibi else 5
    assert (out - out_ref).abs().max().item() <= mult * (out_pt - out_ref).abs().max().item() + 1e-5
Tri Dao's avatar
Tri Dao committed
2110

Tri Dao's avatar
Tri Dao committed
2111

2112
2113
2114
2115
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
2131
2132
2133
2134
2135
2136
2137
2138
def _generate_block_kvcache(seqlen_k, paged_kv_block_size, batch_size, nheads_k, d, device, dtype):
    num_blocks = math.ceil(seqlen_k / paged_kv_block_size) * batch_size * 3
    k_cache_paged = torch.randn(
        num_blocks, paged_kv_block_size, nheads_k, d, device=device, dtype=dtype
    )
    v_cache_paged = torch.randn(
        num_blocks, paged_kv_block_size, nheads_k, d, device=device, dtype=dtype
    )
    block_table = rearrange(
        torch.randperm(num_blocks, dtype=torch.int32, device=device),
        "(b nblocks) -> b nblocks",
        b=batch_size,
    )
    k_cache = rearrange(
        # pytorch 1.12 doesn't have indexing with int32
        k_cache_paged[block_table.to(dtype=torch.long).flatten()],
        "(b nblocks) block_size ... -> b (nblocks block_size) ...",
        b=batch_size,
    )[:, :seqlen_k]
    v_cache = rearrange(
        v_cache_paged[block_table.to(dtype=torch.long).flatten()],
        "(b nblocks) block_size ... -> b (nblocks block_size) ...",
        b=batch_size,
    )[:, :seqlen_k]
    return k_cache, v_cache, block_table, k_cache_paged, v_cache_paged, num_blocks


2139
2140
# @pytest.mark.parametrize("dtype", ([torch.float16] if is_sm75 else [torch.float16, torch.bfloat16]))
@pytest.mark.parametrize("dtype", [torch.float16])
Tri Dao's avatar
Tri Dao committed
2141
@pytest.mark.parametrize("causal", [False, True])
2142
2143
# @pytest.mark.parametrize('causal', [True])
@pytest.mark.parametrize("d", [32, 40, 59, 64, 80, 96, 111, 128, 160, 192, 224, 256])
Tri Dao's avatar
Tri Dao committed
2144
# @pytest.mark.parametrize('d', [32, 56, 64, 80, 96, 128])
2145
# @pytest.mark.parametrize("d", [32, 64, 96, 128, 160, 192])
Tri Dao's avatar
Tri Dao committed
2146
# @pytest.mark.parametrize('d', [128])
2147
2148
2149
2150
2151
2152
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
@pytest.mark.parametrize(
    "seqlen_q,seqlen_k",
    [
        (1, 239),
        (239, 1),
        (3, 799),
        (799, 3),
        (1024, 128),
        (97, 97),
        (128, 128),
        (200, 200),
        (256, 256),
        (257, 257),
        (384, 384),
        (512, 512),
        (768, 768),
        (1024, 1024),
    ],
)
2166
@pytest.mark.parametrize("dropout_p", [0.0, 0.17])
2167
2168
# @pytest.mark.parametrize("dropout_p", [0.0])
def test_flash_attn_race_condition(seqlen_q, seqlen_k, d, dropout_p, causal, dtype):
Tri Dao's avatar
Tri Dao committed
2169
    device = "cuda"
Tri Dao's avatar
Tri Dao committed
2170
2171
    # set seed
    torch.random.manual_seed(0)
2172
    batch_size = 60  # Sometimes we need large batch size for the race conditions to trigger
Tri Dao's avatar
Tri Dao committed
2173
    nheads = 4
2174
2175
2176
2177
2178
    q = torch.randn(batch_size, seqlen_q, nheads, d, device=device, dtype=dtype, requires_grad=True)
    k = torch.randn(batch_size, seqlen_k, nheads, d, device=device, dtype=dtype, requires_grad=True)
    v = torch.randn(batch_size, seqlen_k, nheads, d, device=device, dtype=dtype, requires_grad=True)
    torch.random.manual_seed(42)
    out0, lse0, _ = flash_attn_func(q, k, v, dropout_p, causal=causal, return_attn_probs=True)
Tri Dao's avatar
Tri Dao committed
2179
    g = torch.randn_like(out0)
2180
    if (d <= MAX_HEADDIM_SM8x or (d > 224 and dropout_p == 0)) or (is_sm80 or is_sm90):
2181
2182
2183
2184
2185
        (
            dq0,
            dk0,
            dv0,
        ) = torch.autograd.grad(out0, (q, k, v), g)
2186
        # Numerical error if we just do any arithmetic on dq
2187
        dq_atol = 2 * ((dq0 + 0.3 - 0.3) - dq0).abs().max().item()
Tri Dao's avatar
Tri Dao committed
2188

2189
2190
2191
    for i in range(250):
        torch.random.manual_seed(42)
        out, lse, _ = flash_attn_func(q, k, v, dropout_p, causal=causal, return_attn_probs=True)
Tri Dao's avatar
Tri Dao committed
2192
2193
        assert torch.equal(out, out0)
        assert torch.equal(lse, lse0)
Tri Dao's avatar
Tri Dao committed
2194

2195
        if (d <= MAX_HEADDIM_SM8x or (d > 224 and dropout_p == 0)) or (is_sm80 or is_sm90):
2196
2197
2198
2199
2200
2201
            (
                dq,
                dk,
                dv,
            ) = torch.autograd.grad(out, (q, k, v), g)
            dq_equal = torch.allclose(dq, dq0, atol=dq_atol)
2202
            if not dq_equal:
2203
2204
2205
                print(f"Iter {i}, {dq_atol = }, dQ max diff: {(dq - dq0).abs().max().item()}")
            assert torch.equal(dv, dv0)
            assert torch.equal(dk, dk0)
2206
            assert dq_equal
2207
2208


Tri Dao's avatar
Tri Dao committed
2209
2210
@pytest.mark.parametrize("dtype", [torch.float16])
@pytest.mark.parametrize("causal", [False, True])
2211
# @pytest.mark.parametrize('causal', [False])
Tri Dao's avatar
Tri Dao committed
2212
@pytest.mark.parametrize("d", [16, 32, 64])
2213
# @pytest.mark.parametrize('d', [16])
Tri Dao's avatar
Tri Dao committed
2214
@pytest.mark.parametrize("seqlen", [1, 2, 5, 17, 128])
2215
2216
# @pytest.mark.parametrize('seqlen', [2])
def test_flash_attn_bwd_overflow(seqlen, d, causal, dtype):
Tri Dao's avatar
Tri Dao committed
2217
    """We previously had a bug where not masking elements beyond seqlen_k caused NaN in dQ,
2218
2219
    in the case where seqlen % 128 != 0.
    """
Tri Dao's avatar
Tri Dao committed
2220
    device = "cuda"
2221
2222
2223
2224
2225
    # set seed
    torch.random.manual_seed(0)
    batch_size = 2
    nheads = 5
    q = torch.randn([batch_size, seqlen, nheads, d], dtype=dtype, device="cuda") * 5
Tri Dao's avatar
Tri Dao committed
2226
2227
2228
2229
    k, v = [
        torch.randn([batch_size, seqlen, nheads, d], dtype=dtype, device="cuda") * 3
        for _ in range(2)
    ]
2230
2231
2232
2233
2234
2235
2236
2237
2238
2239
2240
2241
2242
2243
2244
2245
    q.requires_grad_(True)
    k.requires_grad_(True)
    v.requires_grad_(True)
    out = flash_attn_func(q, k, v, causal=causal)
    g = torch.randn_like(out)
    out.backward(g)
    q_pt = q.detach().clone().requires_grad_(True)
    k_pt = k.detach().clone().requires_grad_(True)
    v_pt = v.detach().clone().requires_grad_(True)
    out_pt, _ = attention_ref(q_pt, k_pt, v_pt, causal=causal, upcast=False, reorder_ops=True)
    out_pt.backward(g)
    q_ref = q.detach().clone().requires_grad_(True)
    k_ref = k.detach().clone().requires_grad_(True)
    v_ref = v.detach().clone().requires_grad_(True)
    out_ref, attn_ref = attention_ref(q_ref, k_ref, v_ref, causal=causal)
    out_ref.backward(g)
Tri Dao's avatar
Tri Dao committed
2246
2247
2248
2249
2250
2251
    print(f"dQ max diff: {(q.grad - q_ref.grad).abs().max().item()}")
    print(f"dK max diff: {(k.grad - k_ref.grad).abs().max().item()}")
    print(f"dV max diff: {(v.grad - v_ref.grad).abs().max().item()}")
    print(f"dQ Pytorch max diff: {(q_pt.grad - q_ref.grad).abs().max().item()}")
    print(f"dK Pytorch max diff: {(k_pt.grad - k_ref.grad).abs().max().item()}")
    print(f"dV Pytorch max diff: {(v_pt.grad - v_ref.grad).abs().max().item()}")
2252
    assert (out - out_ref).abs().max().item() <= 2 * (out_pt - out_ref).abs().max().item()
Tri Dao's avatar
Tri Dao committed
2253
2254
2255
2256
2257
2258
2259
2260
2261
2262
2263
2264
    assert (q.grad - q_ref.grad).abs().max().item() <= 5 * (
        q_pt.grad - q_ref.grad
    ).abs().max().item() + 1e-3
    assert (k.grad - k_ref.grad).abs().max().item() <= 5 * (
        k_pt.grad - k_ref.grad
    ).abs().max().item() + 1e-3
    assert (v.grad - v_ref.grad).abs().max().item() <= 5 * (
        v_pt.grad - v_ref.grad
    ).abs().max().item() + 1e-3


@pytest.mark.parametrize("dtype", ([torch.float16] if is_sm75 else [torch.float16, torch.bfloat16]))
2265
# @pytest.mark.parametrize('dtype', [torch.bfloat16])
Tri Dao's avatar
Tri Dao committed
2266
@pytest.mark.parametrize("causal", [False, True])
2267
# @pytest.mark.parametrize('causal', [False])
Tri Dao's avatar
Tri Dao committed
2268
@pytest.mark.parametrize("d", [64, 128])
2269
# @pytest.mark.parametrize('d', [64])
Tri Dao's avatar
Tri Dao committed
2270
@pytest.mark.parametrize("seqlen", [97, 128, 200, 256])
2271
2272
# @pytest.mark.parametrize('seqlen', [128])
def test_flash_attn_bwd_transpose(seqlen, d, causal, dtype):
Tri Dao's avatar
Tri Dao committed
2273
    """We previously had a bug where we were using the wrong strides of dout, which shows up
2274
2275
    when dout is not contiguous.
    """
Tri Dao's avatar
Tri Dao committed
2276
    device = "cuda"
2277
2278
2279
2280
    # set seed
    torch.random.manual_seed(0)
    batch_size = 5
    nheads = 2
Tri Dao's avatar
Tri Dao committed
2281
2282
2283
2284
    q, k, v = [
        torch.randn([batch_size, seqlen, nheads, d], dtype=dtype, device="cuda", requires_grad=True)
        for _ in range(3)
    ]
2285
2286
2287
2288
2289
2290
2291
2292
2293
2294
2295
2296
2297
2298
2299
2300
    out = rearrange(flash_attn_func(q, k, v, causal=causal), "b s ... -> s b ...")
    # So g is not contiguous
    g = torch.randn(seqlen, 2 * batch_size, nheads, d, dtype=dtype, device="cuda")[:, ::2]
    out.backward(g)
    q_pt = q.detach().clone().requires_grad_(True)
    k_pt = k.detach().clone().requires_grad_(True)
    v_pt = v.detach().clone().requires_grad_(True)
    out_pt, attn_pt = attention_ref(q_pt, k_pt, v_pt, causal=causal, upcast=False, reorder_ops=True)
    out_pt = rearrange(out_pt, "b s ... -> s b ...")
    out_pt.backward(g)
    q_ref = q.detach().clone().requires_grad_(True)
    k_ref = k.detach().clone().requires_grad_(True)
    v_ref = v.detach().clone().requires_grad_(True)
    out_ref, attn_ref = attention_ref(q_ref, k_ref, v_ref, causal=causal)
    out_ref = rearrange(out_ref, "b s ... -> s b ...")
    out_ref.backward(g)
Tri Dao's avatar
Tri Dao committed
2301
2302
2303
2304
2305
2306
    print(f"dQ max diff: {(q.grad - q_ref.grad).abs().max().item()}")
    print(f"dK max diff: {(k.grad - k_ref.grad).abs().max().item()}")
    print(f"dV max diff: {(v.grad - v_ref.grad).abs().max().item()}")
    print(f"dQ Pytorch max diff: {(q_pt.grad - q_ref.grad).abs().max().item()}")
    print(f"dK Pytorch max diff: {(k_pt.grad - k_ref.grad).abs().max().item()}")
    print(f"dV Pytorch max diff: {(v_pt.grad - v_ref.grad).abs().max().item()}")
2307
    assert (out - out_ref).abs().max().item() <= 2 * (out_pt - out_ref).abs().max().item()
Tri Dao's avatar
Tri Dao committed
2308
2309
2310
2311
2312
2313
2314
2315
2316
2317
2318
2319
2320
    assert (q.grad - q_ref.grad).abs().max().item() <= 2 * (
        q_pt.grad - q_ref.grad
    ).abs().max().item()
    assert (k.grad - k_ref.grad).abs().max().item() <= 2 * (
        k_pt.grad - k_ref.grad
    ).abs().max().item()
    assert (v.grad - v_ref.grad).abs().max().item() <= 2 * (
        v_pt.grad - v_ref.grad
    ).abs().max().item()


@pytest.mark.parametrize("dtype", [torch.float16])
@pytest.mark.parametrize("causal", [False, True])
2321
# @pytest.mark.parametrize('causal', [False])
Tri Dao's avatar
Tri Dao committed
2322
@pytest.mark.parametrize("d", [16, 32, 64])
2323
2324
# @pytest.mark.parametrize('d', [16])
def test_flash_attn_bwd_varlen_overflow(d, causal, dtype):
Tri Dao's avatar
Tri Dao committed
2325
    """We previously had a bug where not masking elements beyond seqlen_k caused NaN in dQ,
2326
2327
    in the case where seqlen % 128 != 0 or varlen.
    """
Tri Dao's avatar
Tri Dao committed
2328
    device = "cuda"
2329
2330
2331
2332
2333
2334
2335
2336
2337
2338
2339
2340
2341
2342
2343
2344
2345
2346
2347
2348
2349
    # set seed
    torch.random.manual_seed(0)
    nheads = 5
    q_cuseqlen = torch.tensor([0, 76, 110, 256], device=device, dtype=torch.int32)
    k_cuseqlen = torch.tensor([0, 1, 2, 3], device=device, dtype=torch.int32)
    Mq = 256
    Mk = 3

    q = torch.randn([Mq, nheads, d], dtype=dtype, device=device) * 3
    k, v = [torch.randn([Mk, nheads, d], dtype=dtype, device=device) * 3 for _ in range(2)]
    q.requires_grad_(True)
    k.requires_grad_(True)
    v.requires_grad_(True)

    out = flash_attn_varlen_func(q, k, v, q_cuseqlen, k_cuseqlen, Mq, Mk, causal=causal)
    g = torch.randn_like(out)
    out.backward(g)

    assert not q.grad.isnan().any()
    assert not k.grad.isnan().any()
    assert not v.grad.isnan().any()
2350
2351
2352
2353
2354
2355
2356
2357
2358
2359
2360
2361
2362
2363
2364
2365
2366
2367
2368
2369
2370
2371
2372
2373
2374
2375
2376
2377
2378
2379
2380
2381
2382
2383
2384
2385
2386
2387
2388
2389
2390
2391
2392
2393
2394
2395
2396
2397
2398
2399
2400
2401


@pytest.mark.parametrize("dtype", ([torch.float16] if is_sm75 else [torch.float16, torch.bfloat16]))
# @pytest.mark.parametrize("dtype", [torch.bfloat16])
@pytest.mark.parametrize("local", [False, True])
# @pytest.mark.parametrize("local", [True])
@pytest.mark.parametrize("causal", [False, True])
# @pytest.mark.parametrize("causal", [True])
@pytest.mark.parametrize("d", [32, 40, 59, 64, 80, 96, 111, 128, 160, 192, 224, 256])
# @pytest.mark.parametrize("d", [32, 64, 96, 128, 160, 192, 224, 256])
# @pytest.mark.parametrize('d', [32, 40, 64, 80, 96, 128, 160, 192])
# @pytest.mark.parametrize('d', [32, 64, 96, 128, 160, 192])
# @pytest.mark.parametrize('d', [56, 80])
# @pytest.mark.parametrize("d", [64])
@pytest.mark.parametrize("swap_sq_sk", [False, True])
# @pytest.mark.parametrize("swap_sq_sk", [False])
@pytest.mark.parametrize(
    "seqlen_q,seqlen_k",
    [
        (1, 239),
        (3, 799),
        (127, 512),
        (127, 513),
        (113, 203),
        (128, 217),
        (113, 211),
        (108, 256),
        (256, 512),
        (1023, 1024),
    ],
)
# @pytest.mark.parametrize('seqlen_q,seqlen_k', [(256, 128)])
def test_flash_attn_deterministic(seqlen_q, seqlen_k, swap_sq_sk, d, causal, local, dtype):
    if (
        max(seqlen_q, seqlen_k) >= 2048
        and torch.cuda.get_device_properties("cuda").total_memory <= 16 * 2**30
    ):
        pytest.skip()  # Reference implementation OOM
    if swap_sq_sk:
        seqlen_q, seqlen_k = seqlen_k, seqlen_q
    device = "cuda"
    # set seed
    torch.random.manual_seed(0)
    batch_size = 4
    nheads = 9
    window_size = (-1, -1) if not local else torch.randint(0, seqlen_k, (2,))
    q = torch.randn(batch_size, seqlen_q, nheads, d, device=device, dtype=dtype, requires_grad=True)
    k = torch.randn(batch_size, seqlen_k, nheads, d, device=device, dtype=dtype, requires_grad=True)
    v = torch.randn(batch_size, seqlen_k, nheads, d, device=device, dtype=dtype, requires_grad=True)
    out = flash_attn_func(q, k, v, 0.0, causal=causal, window_size=window_size, deterministic=True)

    g = torch.randn_like(out)
2402
    if (d <= MAX_HEADDIM_SM8x or d > 224) or (is_sm80 or is_sm90):
2403
2404
2405
2406
2407
2408
2409
2410
2411
2412
2413
2414
2415
2416
2417
2418
2419
2420
2421
2422
2423
2424
2425
2426
2427
2428
2429
2430
2431
2432
2433
2434
2435
2436
2437
2438
2439
2440
2441
2442
2443
2444
2445
2446
2447
2448
2449
2450
2451
2452
2453
2454
2455
2456
2457
2458
2459
2460
2461
2462
2463
2464
2465
2466
2467
2468
2469
2470
2471
2472
2473
2474
2475
2476
2477
2478
2479
2480
2481
2482
2483
2484
2485
2486
2487
2488
2489
        dq0, dk0, dv0 = torch.autograd.grad(out, (q, k, v), g, retain_graph=True)
        for _ in range(50):
            dq, dk, dv = torch.autograd.grad(out, (q, k, v), g, retain_graph=True)
            assert torch.equal(dv, dv0)
            assert torch.equal(dk, dk0)
            assert torch.equal(dq, dq0)


@pytest.mark.parametrize("dtype", ([torch.float16] if is_sm75 else [torch.float16, torch.bfloat16]))
# @pytest.mark.parametrize("dtype", [torch.bfloat16])
@pytest.mark.parametrize("local", [False, True])
# @pytest.mark.parametrize("local", [True])
@pytest.mark.parametrize("causal", [False, True])
# @pytest.mark.parametrize("causal", [True])
@pytest.mark.parametrize("d", [32, 40, 59, 64, 80, 96, 111, 128, 160, 192, 224, 256])
# @pytest.mark.parametrize("d", [32, 64, 96, 128, 160, 192, 224, 256])
# @pytest.mark.parametrize('d', [32, 40, 64, 80, 96, 128, 160, 192])
# @pytest.mark.parametrize('d', [32, 64, 96, 128, 160, 192])
# @pytest.mark.parametrize('d', [56, 80])
# @pytest.mark.parametrize("d", [64])
@pytest.mark.parametrize("swap_sq_sk", [False, True])
# @pytest.mark.parametrize("swap_sq_sk", [True])
@pytest.mark.parametrize(
    "seqlen_q,seqlen_k",
    [
        (1, 239),
        (3, 799),
        (127, 512),
        (127, 513),
        (113, 203),
        (128, 217),
        (113, 211),
        (108, 256),
        (256, 512),
        (1023, 1024),
    ],
)
# @pytest.mark.parametrize("seqlen_q,seqlen_k", [(256, 128)])
def test_flash_attn_varlen_deterministic(seqlen_q, seqlen_k, swap_sq_sk, d, causal, local, dtype):
    if (
        max(seqlen_q, seqlen_k) >= 2048
        and torch.cuda.get_device_properties("cuda").total_memory <= 16 * 2**30
    ):
        pytest.skip()  # Reference implementation OOM
    if swap_sq_sk:
        seqlen_q, seqlen_k = seqlen_k, seqlen_q
    device = "cuda"
    # set seed
    torch.random.manual_seed(0)
    batch_size = 2
    nheads = 9
    window_size = (-1, -1) if not local else torch.randint(0, seqlen_k, (2,))
    q = torch.randn(batch_size, seqlen_q, nheads, d, device=device, dtype=dtype, requires_grad=True)
    k = torch.randn(batch_size, seqlen_k, nheads, d, device=device, dtype=dtype, requires_grad=True)
    v = torch.randn(batch_size, seqlen_k, nheads, d, device=device, dtype=dtype, requires_grad=True)
    query_padding_mask = generate_random_padding_mask(seqlen_q, batch_size, device, mode="random")
    key_padding_mask = generate_random_padding_mask(seqlen_k, batch_size, device, mode="random")
    (
        q_unpad,
        k_unpad,
        v_unpad,
        cu_seqlens_q,
        cu_seqlens_k,
        max_seqlen_q,
        max_seqlen_k,
        q,
        k,
        v,
        output_pad_fn,
        dq_pad_fn,
        dk_pad_fn,
    ) = generate_qkv(q, k, v, query_padding_mask, key_padding_mask, kvpacked=False)
    out = flash_attn_varlen_func(
        q_unpad,
        k_unpad,
        v_unpad,
        cu_seqlens_q,
        cu_seqlens_k,
        max_seqlen_q,
        max_seqlen_k,
        0.0,
        causal=causal,
        window_size=window_size,
        deterministic=True,
    )

    g = torch.randn_like(out)
2490
    if (d <= MAX_HEADDIM_SM8x or d > 224) or (is_sm80 or is_sm90):
2491
        dq0, dk0, dv0 = torch.autograd.grad(out, (q_unpad, k_unpad, v_unpad), g, retain_graph=True)
2492
2493
        for _ in range(50):
            dq, dk, dv = torch.autograd.grad(out, (q_unpad, k_unpad, v_unpad), g, retain_graph=True)
2494
2495
2496
            assert torch.equal(dv, dv0)
            assert torch.equal(dk, dk0)
            assert torch.equal(dq, dq0)