-
Notifications
You must be signed in to change notification settings - Fork 1.5k
Expand file tree
/
Copy pathfused_rotary_positional_embedding_cuda.cu
More file actions
382 lines (353 loc) · 19 KB
/
fused_rotary_positional_embedding_cuda.cu
File metadata and controls
382 lines (353 loc) · 19 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
/* coding=utf-8
* Copyright (c) 2023, NVIDIA CORPORATION. All rights reserved.
*
* 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 <ATen/ATen.h>
#include "fused_rotary_positional_embedding.h"
#include "type_shim.h"
namespace fused_rope {
torch::Tensor fwd_cuda(const torch::Tensor& input, const torch::Tensor& freqs, const bool transpose_output) {
// input sizes: (s, b, h, d)
// s: sequence length
// b: batch size
// h: head num
// d: dim of each head
const int s = input.size(0);
const int b = input.size(1);
const int h = input.size(2);
const int d = input.size(3);
// input strides
const int stride_s = input.stride(0);
const int stride_b = input.stride(1);
const int stride_h = input.stride(2);
const int stride_d = input.stride(3);
// freqs' shape is always (s, 1, 1, d2), so the strides are same under
// different memory formats
const int d2 = freqs.size(3);
// output
auto act_options = input.options().requires_grad(false);
torch::Tensor output;
if (transpose_output) {
output = torch::empty({b, s, h, d}, act_options).transpose(0, 1);
} else {
output = torch::empty({s, b, h, d}, act_options);
}
// output strides
const int o_stride_s = output.stride(0);
const int o_stride_b = output.stride(1);
const int o_stride_h = output.stride(2);
const int o_stride_d = output.stride(3);
DISPATCH_FLOAT_HALF_AND_BFLOAT(
input.scalar_type(), 0, "dispatch_fused_rope_forward",
dispatch_fused_rope_forward(s, b, h, d, d2, stride_s, stride_b, stride_h, stride_d, o_stride_s, o_stride_b,
o_stride_h, o_stride_d, input.data_ptr<scalar_t_0>(), freqs.data_ptr<float>(),
output.data_ptr<scalar_t_0>()););
return output;
}
torch::Tensor bwd_cuda(const torch::Tensor& output_grads, const torch::Tensor& freqs, const bool transpose_output) {
// output_grads sizes: (s, b, h, d)
// s: sequence length
// b: batch size
// h: head num
// d: dim of each head
const int s = output_grads.size(0);
const int b = output_grads.size(1);
const int h = output_grads.size(2);
const int d = output_grads.size(3);
// output_grads strides
const int stride_s = output_grads.stride(0);
const int stride_b = output_grads.stride(1);
const int stride_h = output_grads.stride(2);
const int stride_d = output_grads.stride(3);
// freqs' shape is always (s, 1, 1, d2), so the strides are same under
// different memory formats
const int d2 = freqs.size(3);
auto act_options = output_grads.options().requires_grad(false);
torch::Tensor input_grads;
if (transpose_output) {
input_grads = torch::empty({b, s, h, d}, act_options).transpose(0, 1);
} else {
input_grads = torch::empty({s, b, h, d}, act_options);
}
const int o_stride_s = input_grads.stride(0);
const int o_stride_b = input_grads.stride(1);
const int o_stride_h = input_grads.stride(2);
const int o_stride_d = input_grads.stride(3);
DISPATCH_FLOAT_HALF_AND_BFLOAT(
output_grads.scalar_type(), 0, "dispatch_fused_rope_backward",
dispatch_fused_rope_backward(s, b, h, d, d2, stride_s, stride_b, stride_h, stride_d, o_stride_s, o_stride_b,
o_stride_h, o_stride_d, output_grads.data_ptr<scalar_t_0>(), freqs.data_ptr<float>(),
input_grads.data_ptr<scalar_t_0>()););
return input_grads;
}
#define DISPATCH_FUSED_ROPE_TYPES(TYPE1, TYPE2, NAME, ...) \
switch (TYPE1) { \
case at::ScalarType::Float: { \
using scalar_t_0 = float; \
switch (TYPE2) { \
case at::ScalarType::Float: { \
using scalar_t_1 = float; \
__VA_ARGS__; \
break; \
} \
default: \
TORCH_CHECK(false, #NAME, " not supported for '", toString(TYPE1), "' with '", toString(TYPE2), "'"); \
} \
break; \
} \
case at::ScalarType::Half: { \
using scalar_t_0 = at::Half; \
switch (TYPE2) { \
case at::ScalarType::Float: { \
using scalar_t_1 = float; \
__VA_ARGS__; \
break; \
} \
case at::ScalarType::Half: { \
using scalar_t_1 = at::Half; \
__VA_ARGS__; \
break; \
} \
default: \
TORCH_CHECK(false, #NAME, " not supported for '", toString(TYPE1), "' with '", toString(TYPE2), "'"); \
} \
break; \
} \
case at::ScalarType::BFloat16: { \
using scalar_t_0 = at::BFloat16; \
switch (TYPE2) { \
case at::ScalarType::Float: { \
using scalar_t_1 = float; \
__VA_ARGS__; \
break; \
} \
case at::ScalarType::BFloat16: { \
using scalar_t_1 = at::BFloat16; \
__VA_ARGS__; \
break; \
} \
default: \
TORCH_CHECK(false, #NAME, " not supported for '", toString(TYPE1), "' with '", toString(TYPE2), "'"); \
} \
break; \
} \
default: \
TORCH_CHECK(false, #NAME, " not supported for '", toString(TYPE1), "' with '", toString(TYPE2), "'"); \
}
torch::Tensor fwd_cached_cuda(const torch::Tensor& input, const torch::Tensor& cos, const torch::Tensor& sin,
const bool transpose_output) {
// input sizes: (s, b, h, d)
// s: sequence length
// b: batch size
// h: head num
// d: dim of each head
const int s = input.size(0);
const int b = input.size(1);
const int h = input.size(2);
const int d = input.size(3);
// input strides
const int stride_s = input.stride(0);
const int stride_b = input.stride(1);
const int stride_h = input.stride(2);
const int stride_d = input.stride(3);
// cos/sin's shape is always (s, 1, 1, d2), so the strides are same under
// different memory formats
const int d2 = cos.size(3);
// output
auto act_options = input.options().requires_grad(false);
torch::Tensor output;
if (transpose_output) {
output = torch::empty({b, s, h, d}, act_options).transpose(0, 1);
} else {
output = torch::empty({s, b, h, d}, act_options);
}
// output strides
const int o_stride_s = output.stride(0);
const int o_stride_b = output.stride(1);
const int o_stride_h = output.stride(2);
const int o_stride_d = output.stride(3);
DISPATCH_FUSED_ROPE_TYPES(input.scalar_type(), cos.scalar_type(), "dispatch_fused_rope_cached_forward",
dispatch_fused_rope_cached_forward(
s, b, h, d, d2, stride_s, stride_b, stride_h, stride_d, o_stride_s, o_stride_b,
o_stride_h, o_stride_d, input.data_ptr<scalar_t_0>(), cos.data_ptr<scalar_t_1>(),
sin.data_ptr<scalar_t_1>(), output.data_ptr<scalar_t_0>()););
return output;
}
torch::Tensor bwd_cached_cuda(const torch::Tensor& output_grads, const torch::Tensor& cos, const torch::Tensor& sin,
const bool transpose_output) {
// output_grads sizes: (s, b, h, d)
// s: sequence length
// b: batch size
// h: head num
// d: dim of each head
const int s = output_grads.size(0);
const int b = output_grads.size(1);
const int h = output_grads.size(2);
const int d = output_grads.size(3);
// output_grads strides
const int stride_s = output_grads.stride(0);
const int stride_b = output_grads.stride(1);
const int stride_h = output_grads.stride(2);
const int stride_d = output_grads.stride(3);
// cos/sin's shape is always (s, 1, 1, d2), so the strides are same under
// different memory formats
const int d2 = cos.size(3);
auto act_options = output_grads.options().requires_grad(false);
torch::Tensor input_grads;
if (transpose_output) {
input_grads = torch::empty({b, s, h, d}, act_options).transpose(0, 1);
} else {
input_grads = torch::empty({s, b, h, d}, act_options);
}
const int o_stride_s = input_grads.stride(0);
const int o_stride_b = input_grads.stride(1);
const int o_stride_h = input_grads.stride(2);
const int o_stride_d = input_grads.stride(3);
DISPATCH_FUSED_ROPE_TYPES(output_grads.scalar_type(), cos.scalar_type(), "dispatch_fused_rope_cached_backward",
dispatch_fused_rope_cached_backward(
s, b, h, d, d2, stride_s, stride_b, stride_h, stride_d, o_stride_s, o_stride_b,
o_stride_h, o_stride_d, output_grads.data_ptr<scalar_t_0>(), cos.data_ptr<scalar_t_1>(),
sin.data_ptr<scalar_t_1>(), input_grads.data_ptr<scalar_t_0>()););
return input_grads;
}
torch::Tensor fwd_thd_cuda(const torch::Tensor& input, const torch::Tensor& cu_seqlens, const torch::Tensor& freqs) {
// input sizes: (t, h, d)
// t: cumulative sum of sequence lengths
// h: head num
// d: dim of each head
const int t = input.size(0);
const int h = input.size(1);
const int d = input.size(2);
// input strides
const int stride_t = input.stride(0);
const int stride_h = input.stride(1);
const int stride_d = input.stride(2);
// batch size
const int b = cu_seqlens.size(0) - 1;
// freqs' shape is (max_s, 1, 1, d2)
const int max_s = freqs.size(0);
const int d2 = freqs.size(3);
// output
auto act_options = input.options().requires_grad(false);
auto output = torch::empty({t, h, d}, act_options);
// output strides
const int o_stride_t = output.stride(0);
const int o_stride_h = output.stride(1);
const int o_stride_d = output.stride(2);
DISPATCH_FLOAT_HALF_AND_BFLOAT(
input.scalar_type(), 0, "dispatch_fused_rope_thd_forward",
dispatch_fused_rope_thd_forward(max_s, b, h, d, d2, stride_t, stride_h, stride_d, o_stride_t, o_stride_h,
o_stride_d, input.data_ptr<scalar_t_0>(), cu_seqlens.data_ptr<int>(),
freqs.data_ptr<float>(), output.data_ptr<scalar_t_0>()););
return output;
}
torch::Tensor bwd_thd_cuda(const torch::Tensor& output_grads, const torch::Tensor& cu_seqlens,
const torch::Tensor& freqs) {
// output_grads sizes: (t, h, d)
// t: cumulative sum of sequence lengths
// h: head num
// d: dim of each head
const int t = output_grads.size(0);
const int h = output_grads.size(1);
const int d = output_grads.size(2);
// output_grads strides
const int stride_t = output_grads.stride(0);
const int stride_h = output_grads.stride(1);
const int stride_d = output_grads.stride(2);
// batch size
const int b = cu_seqlens.size(0) - 1;
// freqs' shape is (max_s, 1, 1, d2)
const int max_s = freqs.size(0);
const int d2 = freqs.size(3);
auto act_options = output_grads.options().requires_grad(false);
auto input_grads = torch::empty({t, h, d}, act_options);
const int o_stride_t = input_grads.stride(0);
const int o_stride_h = input_grads.stride(1);
const int o_stride_d = input_grads.stride(2);
DISPATCH_FLOAT_HALF_AND_BFLOAT(
output_grads.scalar_type(), 0, "dispatch_fused_rope_thd_backward",
dispatch_fused_rope_thd_backward(max_s, b, h, d, d2, stride_t, stride_h, stride_d, o_stride_t, o_stride_h,
o_stride_d, output_grads.data_ptr<scalar_t_0>(), cu_seqlens.data_ptr<int>(),
freqs.data_ptr<float>(), input_grads.data_ptr<scalar_t_0>()););
return input_grads;
}
torch::Tensor fwd_2d_cuda(const torch::Tensor& input, const torch::Tensor& cos_h, const torch::Tensor& sin_h,
const torch::Tensor& cos_w, const torch::Tensor& sin_w) {
// input sizes: (b, ih, iw, h, d)
// b: batch size
// ih: image height
// iw: image width
// h: head num
// d: dim of each head
const int b = input.size(0);
const int ih = input.size(1);
const int iw = input.size(2);
const int h = input.size(3);
const int d = input.size(4);
// input strides
const int stride_b = input.stride(0);
const int stride_ih = input.stride(1);
const int stride_iw = input.stride(2);
const int stride_h = input.stride(3);
const int stride_d = input.stride(4);
// output
auto act_options = input.options().requires_grad(false);
auto output = torch::empty({b, ih * iw, h, d}, act_options);
// output strides
const int o_stride_b = output.stride(0);
const int o_stride_s = output.stride(1);
const int o_stride_h = output.stride(2);
const int o_stride_d = output.stride(3);
DISPATCH_FUSED_ROPE_TYPES(
input.scalar_type(), cos_h.scalar_type(), "dispatch_fused_rope_2d_forward",
dispatch_fused_rope_2d_forward(
b, ih, iw, h, d, stride_b, stride_ih, stride_iw, stride_h, stride_d, o_stride_b, o_stride_s, o_stride_h,
o_stride_d, input.data_ptr<scalar_t_0>(), cos_h.data_ptr<scalar_t_1>(), sin_h.data_ptr<scalar_t_1>(),
cos_w.data_ptr<scalar_t_1>(), sin_w.data_ptr<scalar_t_1>(), output.data_ptr<scalar_t_0>()););
return output;
}
torch::Tensor bwd_2d_cuda(const torch::Tensor& output_grads, const torch::Tensor& cos_h, const torch::Tensor& sin_h,
const torch::Tensor& cos_w, const torch::Tensor& sin_w) {
// output_grads sizes: (b, ih, iw, h, d)
// b: batch size
// ih: image height
// iw: image width
// h: head num
// d: dim of each head
const int b = output_grads.size(0);
const int ih = output_grads.size(1);
const int iw = output_grads.size(2);
const int h = output_grads.size(3);
const int d = output_grads.size(4);
// output_grads strides
const int stride_b = output_grads.stride(0);
const int stride_ih = output_grads.stride(1);
const int stride_iw = output_grads.stride(2);
const int stride_h = output_grads.stride(3);
const int stride_d = output_grads.stride(4);
auto act_options = output_grads.options().requires_grad(false);
auto input_grads = torch::empty({b, ih * iw, h, d}, act_options);
const int o_stride_b = input_grads.stride(0);
const int o_stride_s = input_grads.stride(1);
const int o_stride_h = input_grads.stride(2);
const int o_stride_d = input_grads.stride(3);
DISPATCH_FUSED_ROPE_TYPES(
output_grads.scalar_type(), cos_h.scalar_type(), "dispatch_fused_rope_2d_backward",
dispatch_fused_rope_2d_backward(
b, ih, iw, h, d, stride_b, stride_ih, stride_iw, stride_h, stride_d, o_stride_b, o_stride_s, o_stride_h,
o_stride_d, output_grads.data_ptr<scalar_t_0>(), cos_h.data_ptr<scalar_t_1>(), sin_h.data_ptr<scalar_t_1>(),
cos_w.data_ptr<scalar_t_1>(), sin_w.data_ptr<scalar_t_1>(), input_grads.data_ptr<scalar_t_0>()););
return input_grads;
}
} // end namespace fused_rope