File size: 8,358 Bytes
62dca4c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
import torch
from yunchang.kernels import AttnType, select_flash_attn_impl

from .utils import RingComm, update_out_and_lse


def ring_flash_attn_forward(
    process_group,
    q: torch.Tensor,
    k: torch.Tensor,
    v: torch.Tensor,
    softmax_scale,
    dropout_p=0,
    causal=True,
    window_size=(-1, -1),
    softcap=0.0,
    alibi_slopes=None,
    deterministic=False,
    attn_type: AttnType = AttnType.FA,
    attn_processor=None,
):
    comm = RingComm(process_group)

    out = None
    lse = None

    next_k, next_v = None, None

    for step in range(comm.world_size):
        if step + 1 != comm.world_size:
            next_k: torch.Tensor = comm.send_recv(k)
            next_v: torch.Tensor = comm.send_recv(v)
            comm.commit()

        if not causal or step <= comm.rank:
            fn = select_flash_attn_impl(
                attn_type, stage="fwd-only", attn_processor=attn_processor
            )
            block_out, block_lse = fn(
                q,
                k,
                v,
                dropout_p=dropout_p,
                softmax_scale=softmax_scale,
                causal=causal and step == 0,
                window_size=window_size,
                softcap=softcap,
                alibi_slopes=alibi_slopes,
                return_softmax=True and dropout_p > 0,
            )
            if attn_type == AttnType.SPARSE_SAGE:
                out, lse = block_out, block_lse
            else:
                out, lse = update_out_and_lse(out, lse, block_out, block_lse)

        if step + 1 != comm.world_size:
            comm.wait()
            k = next_k
            v = next_v

    out = out.to(q.dtype)
    if attn_type != AttnType.SPARSE_SAGE:
        lse = lse.squeeze(dim=-1).transpose(1, 2)
    return out, lse


def ring_flash_attn_backward(
    process_group,
    dout,
    q,
    k,
    v,
    out,
    softmax_lse,
    softmax_scale,
    dropout_p=0,
    causal=True,
    window_size=(-1, -1),
    softcap=0.0,
    alibi_slopes=None,
    deterministic=False,
    attn_type: AttnType = AttnType.FA,
):
    kv_comm = RingComm(process_group)
    d_kv_comm = RingComm(process_group)
    dq, dk, dv = None, None, None
    next_dk, next_dv = None, None

    block_dq_buffer = torch.empty(q.shape, dtype=q.dtype, device=q.device)
    block_dk_buffer = torch.empty(k.shape, dtype=k.dtype, device=k.device)
    block_dv_buffer = torch.empty(v.shape, dtype=v.dtype, device=v.device)

    next_dk, next_dv = None, None
    next_k, next_v = None, None

    for step in range(kv_comm.world_size):
        if step + 1 != kv_comm.world_size:
            next_k = kv_comm.send_recv(k)
            next_v = kv_comm.send_recv(v)
            kv_comm.commit()
        if step <= kv_comm.rank or not causal:
            bwd_causal = causal and step == 0
            fn = select_flash_attn_impl(attn_type, stage="bwd-only")
            fn(
                dout,
                q,
                k,
                v,
                out,
                softmax_lse,
                block_dq_buffer,
                block_dk_buffer,
                block_dv_buffer,
                dropout_p,
                softmax_scale,
                bwd_causal,
                window_size,
                softcap,
                alibi_slopes,
                deterministic,
                rng_state=None,
            )

            if dq is None:
                dq = block_dq_buffer.to(torch.float32)
                dk = block_dk_buffer.to(torch.float32)
                dv = block_dv_buffer.to(torch.float32)
            else:
                dq += block_dq_buffer
                d_kv_comm.wait()
                dk = block_dk_buffer + next_dk
                dv = block_dv_buffer + next_dv
        elif step != 0:
            d_kv_comm.wait()
            dk = next_dk
            dv = next_dv

        if step + 1 != kv_comm.world_size:
            kv_comm.wait()
            k = next_k
            v = next_v

        next_dk = d_kv_comm.send_recv(dk)
        next_dv = d_kv_comm.send_recv(dv)
        d_kv_comm.commit()

    d_kv_comm.wait()

    return dq.to(torch.bfloat16), next_dk.to(q.dtype), next_dv.to(q.dtype)


class RingFlashAttnFunc(torch.autograd.Function):
    @staticmethod
    def forward(
        ctx,
        q,
        k,
        v,
        dropout_p,
        softmax_scale,
        causal,
        window_size,
        softcap,
        alibi_slopes,
        deterministic,
        return_softmax,
        group,
        attn_type,
        attn_processor,
    ):
        if softmax_scale is None:
            softmax_scale = q.shape[-1] ** (-0.5)

        assert alibi_slopes is None
        k = k.contiguous()
        v = v.contiguous()
        out, softmax_lse = ring_flash_attn_forward(
            group,
            q,
            k,
            v,
            softmax_scale=softmax_scale,
            dropout_p=dropout_p,
            causal=causal,
            window_size=window_size,
            softcap=softcap,
            alibi_slopes=alibi_slopes,
            deterministic=False,
            attn_type=attn_type,
            attn_processor=attn_processor,
        )
        # this should be out_padded
        ctx.save_for_backward(q, k, v, out, softmax_lse)
        ctx.dropout_p = dropout_p
        ctx.softmax_scale = softmax_scale
        ctx.causal = causal
        ctx.window_size = window_size
        ctx.softcap = softcap
        ctx.alibi_slopes = alibi_slopes
        ctx.deterministic = deterministic
        ctx.group = group
        ctx.attn_type = attn_type
        ctx.attn_processor = attn_processor
        return out if not return_softmax else (out, softmax_lse, None)

    @staticmethod
    def backward(ctx, dout, *args):
        q, k, v, out, softmax_lse = ctx.saved_tensors
        dq, dk, dv = ring_flash_attn_backward(
            ctx.group,
            dout,
            q,
            k,
            v,
            out,
            softmax_lse,
            softmax_scale=ctx.softmax_scale,
            dropout_p=ctx.dropout_p,
            causal=ctx.causal,
            window_size=ctx.window_size,
            softcap=ctx.softcap,
            alibi_slopes=ctx.alibi_slopes,
            deterministic=ctx.deterministic,
            attn_type=ctx.attn_type,
        )
        return (
            dq,
            dk,
            dv,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
        )


def ring_flash_attn_qkvpacked_func(
    qkv,
    dropout_p=0.0,
    softmax_scale=None,
    causal=False,
    window_size=(-1, -1),
    softcap=0.0,
    alibi_slopes=None,
    deterministic=False,
    return_attn_probs=False,
    group=None,
    attn_type: AttnType = AttnType.FA,
):
    return RingFlashAttnFunc.apply(
        qkv[:, :, 0],
        qkv[:, :, 1],
        qkv[:, :, 2],
        dropout_p,
        softmax_scale,
        causal,
        window_size,
        softcap,
        alibi_slopes,
        deterministic,
        return_attn_probs,
        group,
        attn_type,
    )


def ring_flash_attn_kvpacked_func(
    q,
    kv,
    dropout_p=0.0,
    softmax_scale=None,
    causal=False,
    window_size=(-1, -1),
    softcap=0.0,
    alibi_slopes=None,
    deterministic=False,
    return_attn_probs=False,
    group=None,
    attn_type: AttnType = AttnType.FA,
):
    return RingFlashAttnFunc.apply(
        q,
        kv[:, :, 0],
        kv[:, :, 1],
        dropout_p,
        softmax_scale,
        causal,
        window_size,
        softcap,
        alibi_slopes,
        deterministic,
        return_attn_probs,
        group,
        attn_type,
    )


def ring_flash_attn_func(
    q,
    k,
    v,
    dropout_p=0.0,
    softmax_scale=None,
    causal=False,
    window_size=(-1, -1),
    softcap=0.0,
    alibi_slopes=None,
    deterministic=False,
    return_attn_probs=False,
    group=None,
    attn_type: AttnType = AttnType.FA,
    attn_processor=None,
):
    return RingFlashAttnFunc.apply(
        q,
        k,
        v,
        dropout_p,
        softmax_scale,
        causal,
        window_size,
        softcap,
        alibi_slopes,
        deterministic,
        return_attn_probs,
        group,
        attn_type,
        attn_processor,
    )