-
Notifications
You must be signed in to change notification settings - Fork 224
/
attention_kernels.cu
320 lines (304 loc) · 13.3 KB
/
attention_kernels.cu
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
/* Copyright 2023 CMU, Facebook, LANL, MIT, NVIDIA, and Stanford (alphabetical)
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
#include "device.h"
#include "kernels/attention_kernels.h"
#include "kernels/device.h"
namespace FlexFlow {
namespace Kernels {
namespace MultiHeadAttention {
MHAPerDeviceState init_kernel(PerDeviceFFHandle const &handle,
Allocator allocator,
int num_samples,
int num_heads,
int qSize,
int kSize,
int vSize,
int qProjSize,
int kProjSize,
int vProjSize,
int oProjSize,
int qoSeqLength,
int kvSeqLength,
bool add_bias_kv) {
cudaStream_t stream;
ffAttnDescriptor_t attnDesc;
ffSeqDataDescriptor_t qDesc;
ffSeqDataDescriptor_t kDesc;
ffSeqDataDescriptor_t vDesc;
ffSeqDataDescriptor_t oDesc;
void *reserveSpace;
void *dropoutStates;
int *devQoSeqArray;
int *devKvSeqArray;
size_t reserveSpaceSize;
size_t dropoutStateSize;
size_t weightSize;
checkCUDA(get_legion_stream(&stream));
checkCUDNN(cudnnSetStream(handle.dnn, stream));
checkCUDNN(cudnnCreateAttnDescriptor(&attnDesc));
checkCUDNN(cudnnCreateSeqDataDescriptor(&qDesc));
checkCUDNN(cudnnCreateSeqDataDescriptor(&kDesc));
checkCUDNN(cudnnCreateSeqDataDescriptor(&vDesc));
checkCUDNN(cudnnCreateSeqDataDescriptor(&oDesc));
// Currently do not support adding bias to key/value projection
assert(!add_bias_kv);
cudnnAttnQueryMap_t attnMode = CUDNN_ATTN_QUERYMAP_ALL_TO_ONE;
// Assume no beam search for now
int maxBeamSize = 1;
cudnnMathType_t math_type;
if (handle.allowTensorOpMathConversion) {
math_type = CUDNN_TENSOR_OP_MATH_ALLOW_CONVERSION;
} else {
math_type = CUDNN_TENSOR_OP_MATH;
}
checkCUDNN(cudnnSetAttnDescriptor(attnDesc,
attnMode,
num_heads,
1.0f /*smScalar*/,
CUDNN_DATA_FLOAT,
CUDNN_DATA_FLOAT,
math_type,
NULL /*attnDropoutDesc*/,
NULL /*postDropoutDesc*/,
qSize,
kSize,
vSize,
qProjSize,
kProjSize,
vProjSize,
oProjSize,
qoSeqLength,
kvSeqLength,
num_samples,
maxBeamSize));
size_t workSpaceSize;
checkCUDNN(cudnnGetMultiHeadAttnBuffers(
handle.dnn, attnDesc, &weightSize, &workSpaceSize, &reserveSpaceSize));
assert(workSpaceSize <= handle.workSpaceSize);
int dimA[CUDNN_SEQDATA_DIM_COUNT];
cudnnSeqDataAxis_t axes[CUDNN_SEQDATA_DIM_COUNT];
assert(CUDNN_SEQDATA_DIM_COUNT == 4);
axes[3] = CUDNN_SEQDATA_VECT_DIM; // 3 = nbDims-1
axes[2] = CUDNN_SEQDATA_BEAM_DIM;
axes[1] = CUDNN_SEQDATA_TIME_DIM;
axes[0] = CUDNN_SEQDATA_BATCH_DIM;
int *qoSeqArray = (int *)malloc(sizeof(int) * num_samples);
int *kvSeqArray = (int *)malloc(sizeof(int) * num_samples);
for (int i = 0; i < num_samples; i++) {
qoSeqArray[i] = qoSeqLength;
kvSeqArray[i] = kvSeqLength;
}
// Set qDesc
{
dimA[CUDNN_SEQDATA_BEAM_DIM] = 1;
dimA[CUDNN_SEQDATA_BATCH_DIM] = num_samples;
dimA[CUDNN_SEQDATA_TIME_DIM] = qoSeqLength;
dimA[CUDNN_SEQDATA_VECT_DIM] = qSize;
checkCUDNN(cudnnSetSeqDataDescriptor(qDesc,
CUDNN_DATA_FLOAT,
CUDNN_SEQDATA_DIM_COUNT,
dimA,
axes,
num_samples,
qoSeqArray,
NULL));
}
// Set kDesc
{
dimA[CUDNN_SEQDATA_BEAM_DIM] = 1;
dimA[CUDNN_SEQDATA_BATCH_DIM] = num_samples;
dimA[CUDNN_SEQDATA_TIME_DIM] = kvSeqLength;
dimA[CUDNN_SEQDATA_VECT_DIM] = kSize;
checkCUDNN(cudnnSetSeqDataDescriptor(kDesc,
CUDNN_DATA_FLOAT,
CUDNN_SEQDATA_DIM_COUNT,
dimA,
axes,
num_samples,
kvSeqArray,
NULL));
}
// Set vDesc
{
dimA[CUDNN_SEQDATA_BEAM_DIM] = 1;
dimA[CUDNN_SEQDATA_BATCH_DIM] = num_samples;
dimA[CUDNN_SEQDATA_TIME_DIM] = kvSeqLength;
dimA[CUDNN_SEQDATA_VECT_DIM] = vSize;
checkCUDNN(cudnnSetSeqDataDescriptor(vDesc,
CUDNN_DATA_FLOAT,
CUDNN_SEQDATA_DIM_COUNT,
dimA,
axes,
num_samples,
kvSeqArray,
NULL));
}
// Set oDesc
{
dimA[CUDNN_SEQDATA_BEAM_DIM] = 1;
dimA[CUDNN_SEQDATA_BATCH_DIM] = num_samples;
dimA[CUDNN_SEQDATA_TIME_DIM] = qoSeqLength;
dimA[CUDNN_SEQDATA_VECT_DIM] = oProjSize;
checkCUDNN(cudnnSetSeqDataDescriptor(oDesc,
CUDNN_DATA_FLOAT,
CUDNN_SEQDATA_DIM_COUNT,
dimA,
axes,
num_samples,
qoSeqArray,
NULL));
}
// allocate memory for the seqArray and reserve space
{
size_t totalSize = reserveSpaceSize + sizeof(int) * num_samples * 2;
devQoSeqArray = (int *)allocator.allocate(totalSize);
checkCUDA(cudaMemcpy(devQoSeqArray,
qoSeqArray,
sizeof(int) * num_samples,
cudaMemcpyHostToDevice));
devKvSeqArray = devQoSeqArray + num_samples;
checkCUDA(cudaMemcpy(devKvSeqArray,
kvSeqArray,
sizeof(int) * num_samples,
cudaMemcpyHostToDevice));
reserveSpace = devKvSeqArray + num_samples;
}
// allocate memory for loWinIdx/hiWinIdx
int *loWinIdx = (int *)malloc(sizeof(int) * qoSeqLength);
int *hiWinIdx = (int *)malloc(sizeof(int) * qoSeqLength);
for (int i = 0; i < qoSeqLength; i++) {
loWinIdx[i] = 0;
hiWinIdx[i] = kvSeqLength;
}
MHAPerDeviceState per_device_state = {handle,
weightSize,
reserveSpaceSize,
attnDesc,
qDesc,
kDesc,
vDesc,
oDesc,
devQoSeqArray,
devKvSeqArray,
loWinIdx,
hiWinIdx,
reserveSpace,
allocator};
free(qoSeqArray);
free(kvSeqArray);
}
void forward_kernel(cudaStream_t stream,
MHAPerDeviceState *m,
float const *query_ptr,
float const *key_ptr,
float const *value_ptr,
float const *weight_ptr,
float *output_ptr) {
checkCUDNN(cudnnSetStream(m->handle.dnn, stream));
checkCUDNN(cudnnMultiHeadAttnForward(m->handle.dnn,
m->attnDesc,
-1,
m->loWinIdx,
m->hiWinIdx,
m->devQoSeqArray,
m->devKvSeqArray,
m->qDesc,
query_ptr,
nullptr /*residual*/,
m->kDesc,
key_ptr,
m->vDesc,
value_ptr,
m->oDesc,
output_ptr,
m->weightSize,
weight_ptr,
m->handle.workSpaceSize,
m->handle.workSpace,
m->reserveSpaceSize,
m->reserveSpace));
}
void backward_kernel(cudaStream_t stream,
MHAPerDeviceState *m,
float const *query_ptr,
float *query_grad_ptr,
float const *key_ptr,
float *key_grad_ptr,
float const *value_ptr,
float *value_grad_ptr,
float const *weight_ptr,
float *weight_grad_ptr,
float const *output_grad_ptr) {
checkCUDNN(cudnnSetStream(m->handle.dnn, stream));
checkCUDNN(cudnnMultiHeadAttnBackwardData(m->handle.dnn,
m->attnDesc,
m->loWinIdx,
m->hiWinIdx,
m->devQoSeqArray,
m->devKvSeqArray,
m->oDesc,
output_grad_ptr,
m->qDesc,
query_grad_ptr,
query_ptr,
m->kDesc,
key_grad_ptr,
key_ptr,
m->vDesc,
value_grad_ptr,
value_ptr,
m->weightSize,
weight_ptr,
m->handle.workSpaceSize,
m->handle.workSpace,
m->reserveSpaceSize,
m->reserveSpace));
checkCUDNN(cudnnMultiHeadAttnBackwardWeights(m->handle.dnn,
m->attnDesc,
CUDNN_WGRAD_MODE_ADD,
m->qDesc,
query_ptr,
m->kDesc,
key_ptr,
m->vDesc,
value_ptr,
m->oDesc,
output_grad_ptr,
m->weightSize,
weight_ptr,
weight_grad_ptr,
m->handle.workSpaceSize,
m->handle.workSpace,
m->reserveSpaceSize,
m->reserveSpace));
}
void cleanup_kernel(int *loWinIdx,
int *hiWinIdx,
ffAttnDescriptor_t attnDesc,
ffSeqDataDescriptor_t qDesc,
ffSeqDataDescriptor_t kDesc,
ffSeqDataDescriptor_t vDesc,
ffSeqDataDescriptor_t oDesc) {
free(loWinIdx);
free(hiWinIdx);
checkCUDNN(cudnnDestroyAttnDescriptor(attnDesc));
checkCUDNN(cudnnDestroySeqDataDescriptor(qDesc));
checkCUDNN(cudnnDestroySeqDataDescriptor(kDesc));
checkCUDNN(cudnnDestroySeqDataDescriptor(vDesc));
checkCUDNN(cudnnDestroySeqDataDescriptor(oDesc));
}
} // namespace MultiHeadAttention
} // namespace Kernels
} // namespace FlexFlow