{{ message }}
-
Notifications
You must be signed in to change notification settings - Fork 6.1k
Expand file tree
/
Copy pathwhisper.cpp
More file actions
9293 lines (7483 loc) · 333 KB
/
Copy pathwhisper.cpp
File metadata and controls
9293 lines (7483 loc) · 333 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
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
#include "whisper.h"
#include "whisper-arch.h"
#include "ggml.h"
#include "ggml-cpp.h"
#include "ggml-alloc.h"
#include "ggml-backend.h"
#ifdef WHISPER_USE_COREML
#include "coreml/whisper-encoder.h"
#endif
#ifdef WHISPER_USE_OPENVINO
#include "openvino/whisper-openvino-encoder.h"
#endif
#ifdef WHISPER_USE_VITISAI
#include "vitisai/whisper-vitisai-encoder.h"
#endif
#include <atomic>
#include <algorithm>
#include <cassert>
#include <cfloat>
#define _USE_MATH_DEFINES
#include <cmath>
#include <climits>
#include <cstdarg>
#include <cstdio>
#include <cstring>
#include <fstream>
#include <functional>
#include <map>
#include <random>
#include <regex>
#include <set>
#include <string>
#include <thread>
#include <vector>
#ifdef _MSC_VER
#include <codecvt>
#endif
#if defined(WHISPER_BIG_ENDIAN)
template<typename T>
static T byteswap(T value) {
T value_swapped;
char * source = reinterpret_cast<char *>(&value);
char * target = reinterpret_cast<char *>(&value_swapped);
int size = sizeof(T);
for (int i = 0; i < size; i++) {
target[size - 1 - i] = source[i];
}
return value_swapped;
}
template<typename T>
static void byteswap_tensor_data(ggml_tensor * tensor) {
T * datum = reinterpret_cast<T *>(tensor->data);
for (int i = 0; i < ggml_nelements(tensor); i++) {
datum[i] = byteswap(datum[i]);
}
}
static void byteswap_tensor(ggml_tensor * tensor) {
switch (tensor->type) {
case GGML_TYPE_I16: {
byteswap_tensor_data<int16_t>(tensor);
break;
}
case GGML_TYPE_F16: {
byteswap_tensor_data<ggml_fp16_t>(tensor);
break;
}
case GGML_TYPE_I32: {
byteswap_tensor_data<int32_t>(tensor);
break;
}
case GGML_TYPE_F32: {
byteswap_tensor_data<float>(tensor);
break;
}
default: { // GML_TYPE_I8
break;
}
}
}
#define BYTESWAP_VALUE(d) d = byteswap(d)
#define BYTESWAP_FILTERS(f) \
do { \
for (auto & datum : f.data) { \
datum = byteswap(datum); \
} \
} while (0)
#define BYTESWAP_TENSOR(t) \
do { \
byteswap_tensor(t); \
} while (0)
#else
#define BYTESWAP_VALUE(d) do {} while (0)
#define BYTESWAP_FILTERS(f) do {} while (0)
#define BYTESWAP_TENSOR(t) do {} while (0)
#endif
#ifdef __GNUC__
#ifdef __MINGW32__
#define WHISPER_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__)))
#else
#define WHISPER_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))
#endif
#else
#define WHISPER_ATTRIBUTE_FORMAT(...)
#endif
//
// logging
//
WHISPER_ATTRIBUTE_FORMAT(2, 3)
static void whisper_log_internal (ggml_log_level level, const char * format, ...);
static void whisper_log_callback_default(ggml_log_level level, const char * text, void * user_data);
#define WHISPER_LOG_ERROR(...) whisper_log_internal(GGML_LOG_LEVEL_ERROR, __VA_ARGS__)
#define WHISPER_LOG_WARN(...) whisper_log_internal(GGML_LOG_LEVEL_WARN , __VA_ARGS__)
#define WHISPER_LOG_INFO(...) whisper_log_internal(GGML_LOG_LEVEL_INFO , __VA_ARGS__)
// define this to enable verbose trace logging - useful for debugging purposes
//#define WHISPER_DEBUG
#if defined(WHISPER_DEBUG)
#define WHISPER_LOG_DEBUG(...) whisper_log_internal(GGML_LOG_LEVEL_DEBUG, __VA_ARGS__)
#else
#define WHISPER_LOG_DEBUG(...)
#endif
#define WHISPER_ASSERT(x) \
do { \
if (!(x)) { \
WHISPER_LOG_ERROR("WHISPER_ASSERT: %s:%d: %s\n", __FILE__, __LINE__, #x); \
abort(); \
} \
} while (0)
#define WHISPER_MAX_DECODERS 8
// temperature below which we condition on past text history
static constexpr float WHISPER_HISTORY_CONDITIONING_TEMP_CUTOFF = 0.5f;
#define WHISPER_MAX_NODES 4096
static std::string format(const char * fmt, ...) {
va_list ap;
va_list ap2;
va_start(ap, fmt);
va_copy(ap2, ap);
int size = vsnprintf(NULL, 0, fmt, ap);
GGML_ASSERT(size >= 0 && size < INT_MAX); // NOLINT
std::vector<char> buf(size + 1);
int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2);
GGML_ASSERT(size2 == size);
va_end(ap2);
va_end(ap);
return std::string(buf.data(), size);
}
//
// ggml helpers
//
static bool ggml_graph_compute_helper(
struct ggml_cgraph * graph,
int n_threads,
ggml_abort_callback abort_callback,
void * abort_callback_data) {
ggml_backend_ptr backend { ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr) };
auto * reg = ggml_backend_dev_backend_reg(ggml_backend_get_device(backend.get()));
auto * set_abort_callback_fn = (ggml_backend_set_abort_callback_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_set_abort_callback");
if (set_abort_callback_fn) {
set_abort_callback_fn(backend.get(), abort_callback, abort_callback_data);
}
auto ggml_backend_set_n_threads_fn = (ggml_backend_set_n_threads_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_set_n_threads");
if (ggml_backend_set_n_threads_fn) {
ggml_backend_set_n_threads_fn(backend.get(), n_threads);
}
return ggml_backend_graph_compute(backend.get(), graph) == GGML_STATUS_SUCCESS;
}
static bool ggml_graph_compute_helper(
ggml_backend_sched_t sched,
struct ggml_cgraph * graph,
int n_threads,
bool sched_reset = true) {
for (int i = 0; i < ggml_backend_sched_get_n_backends(sched); ++i) {
ggml_backend_t backend = ggml_backend_sched_get_backend(sched, i);
ggml_backend_dev_t dev = ggml_backend_get_device(backend);
ggml_backend_reg_t reg = dev ? ggml_backend_dev_backend_reg(dev) : nullptr;
auto * fn_set_n_threads = (ggml_backend_set_n_threads_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_set_n_threads");
if (fn_set_n_threads) {
fn_set_n_threads(backend, n_threads);
}
}
const bool t = (ggml_backend_sched_graph_compute(sched, graph) == GGML_STATUS_SUCCESS);
if (!t || sched_reset) {
ggml_backend_sched_reset(sched);
}
return t;
}
// TODO: move these functions to ggml-base with support for ggml-backend?
static ggml_tensor * whisper_set_f32(struct ggml_tensor * t, float v) {
GGML_ASSERT(t->type == GGML_TYPE_F32);
GGML_ASSERT(ggml_is_contiguous(t));
size_t nels = ggml_nelements(t);
for (size_t i = 0; i < nels; ++i) {
((float *) t->data)[i] = v;
}
return t;
}
static ggml_tensor * whisper_set_i32(struct ggml_tensor * t, int32_t v) {
GGML_ASSERT(t->type == GGML_TYPE_I32);
GGML_ASSERT(ggml_is_contiguous(t));
size_t nels = ggml_nelements(t);
for (size_t i = 0; i < nels; ++i) {
((int32_t *) t->data)[i] = v;
}
return t;
}
static float whisper_get_f32_nd(const struct ggml_tensor * t, int64_t i0, int64_t i1, int64_t i2, int64_t i3) {
GGML_ASSERT(t->type == GGML_TYPE_F32);
void * data = (char *) t->data + i0*t->nb[0] + i1*t->nb[1] + i2*t->nb[2] + i3*t->nb[3];
return *(float *) data;
}
static void whisper_set_f32_nd(struct ggml_tensor * t, int64_t i0, int64_t i1, int64_t i2, int64_t i3, float v) {
GGML_ASSERT(t->type == GGML_TYPE_F32);
void * data = (char *) t->data + i0*t->nb[0] + i1*t->nb[1] + i2*t->nb[2] + i3*t->nb[3];
*(float *) data = v;
}
static int32_t whisper_get_i32_nd(const struct ggml_tensor * t, int64_t i0, int64_t i1, int64_t i2, int64_t i3) {
GGML_ASSERT(t->type == GGML_TYPE_I32);
void * data = (char *) t->data + i0*t->nb[0] + i1*t->nb[1] + i2*t->nb[2] + i3*t->nb[3];
return *(int32_t *) data;
}
static void whisper_set_i32_nd(struct ggml_tensor * t, int64_t i0, int64_t i1, int64_t i2, int64_t i3, int32_t v) {
GGML_ASSERT(t->type == GGML_TYPE_I32);
void * data = (char *) t->data + i0*t->nb[0] + i1*t->nb[1] + i2*t->nb[2] + i3*t->nb[3];
*(int32_t *) data = v;
}
// available whisper models
enum e_model {
MODEL_UNKNOWN,
MODEL_TINY,
MODEL_BASE,
MODEL_SMALL,
MODEL_MEDIUM,
MODEL_LARGE,
};
static const std::map<e_model, std::string> g_model_name = {
{ MODEL_UNKNOWN, "unknown" },
{ MODEL_TINY, "tiny" },
{ MODEL_BASE, "base" },
{ MODEL_SMALL, "small" },
{ MODEL_MEDIUM, "medium" },
{ MODEL_LARGE, "large" },
};
static const std::map<std::string, std::pair<int, std::string>> g_lang = {
{ "en", { 0, "english", } },
{ "zh", { 1, "chinese", } },
{ "de", { 2, "german", } },
{ "es", { 3, "spanish", } },
{ "ru", { 4, "russian", } },
{ "ko", { 5, "korean", } },
{ "fr", { 6, "french", } },
{ "ja", { 7, "japanese", } },
{ "pt", { 8, "portuguese", } },
{ "tr", { 9, "turkish", } },
{ "pl", { 10, "polish", } },
{ "ca", { 11, "catalan", } },
{ "nl", { 12, "dutch", } },
{ "ar", { 13, "arabic", } },
{ "sv", { 14, "swedish", } },
{ "it", { 15, "italian", } },
{ "id", { 16, "indonesian", } },
{ "hi", { 17, "hindi", } },
{ "fi", { 18, "finnish", } },
{ "vi", { 19, "vietnamese", } },
{ "he", { 20, "hebrew", } },
{ "uk", { 21, "ukrainian", } },
{ "el", { 22, "greek", } },
{ "ms", { 23, "malay", } },
{ "cs", { 24, "czech", } },
{ "ro", { 25, "romanian", } },
{ "da", { 26, "danish", } },
{ "hu", { 27, "hungarian", } },
{ "ta", { 28, "tamil", } },
{ "no", { 29, "norwegian", } },
{ "th", { 30, "thai", } },
{ "ur", { 31, "urdu", } },
{ "hr", { 32, "croatian", } },
{ "bg", { 33, "bulgarian", } },
{ "lt", { 34, "lithuanian", } },
{ "la", { 35, "latin", } },
{ "mi", { 36, "maori", } },
{ "ml", { 37, "malayalam", } },
{ "cy", { 38, "welsh", } },
{ "sk", { 39, "slovak", } },
{ "te", { 40, "telugu", } },
{ "fa", { 41, "persian", } },
{ "lv", { 42, "latvian", } },
{ "bn", { 43, "bengali", } },
{ "sr", { 44, "serbian", } },
{ "az", { 45, "azerbaijani", } },
{ "sl", { 46, "slovenian", } },
{ "kn", { 47, "kannada", } },
{ "et", { 48, "estonian", } },
{ "mk", { 49, "macedonian", } },
{ "br", { 50, "breton", } },
{ "eu", { 51, "basque", } },
{ "is", { 52, "icelandic", } },
{ "hy", { 53, "armenian", } },
{ "ne", { 54, "nepali", } },
{ "mn", { 55, "mongolian", } },
{ "bs", { 56, "bosnian", } },
{ "kk", { 57, "kazakh", } },
{ "sq", { 58, "albanian", } },
{ "sw", { 59, "swahili", } },
{ "gl", { 60, "galician", } },
{ "mr", { 61, "marathi", } },
{ "pa", { 62, "punjabi", } },
{ "si", { 63, "sinhala", } },
{ "km", { 64, "khmer", } },
{ "sn", { 65, "shona", } },
{ "yo", { 66, "yoruba", } },
{ "so", { 67, "somali", } },
{ "af", { 68, "afrikaans", } },
{ "oc", { 69, "occitan", } },
{ "ka", { 70, "georgian", } },
{ "be", { 71, "belarusian", } },
{ "tg", { 72, "tajik", } },
{ "sd", { 73, "sindhi", } },
{ "gu", { 74, "gujarati", } },
{ "am", { 75, "amharic", } },
{ "yi", { 76, "yiddish", } },
{ "lo", { 77, "lao", } },
{ "uz", { 78, "uzbek", } },
{ "fo", { 79, "faroese", } },
{ "ht", { 80, "haitian creole", } },
{ "ps", { 81, "pashto", } },
{ "tk", { 82, "turkmen", } },
{ "nn", { 83, "nynorsk", } },
{ "mt", { 84, "maltese", } },
{ "sa", { 85, "sanskrit", } },
{ "lb", { 86, "luxembourgish", } },
{ "my", { 87, "myanmar", } },
{ "bo", { 88, "tibetan", } },
{ "tl", { 89, "tagalog", } },
{ "mg", { 90, "malagasy", } },
{ "as", { 91, "assamese", } },
{ "tt", { 92, "tatar", } },
{ "haw", { 93, "hawaiian", } },
{ "ln", { 94, "lingala", } },
{ "ha", { 95, "hausa", } },
{ "ba", { 96, "bashkir", } },
{ "jw", { 97, "javanese", } },
{ "su", { 98, "sundanese", } },
{ "yue", { 99, "cantonese", } },
};
// [EXPERIMENTAL] Token-level timestamps with DTW
static const whisper_ahead g_aheads_tiny_en[] = { {1, 0}, {2, 0}, {2, 5}, {3, 0}, {3, 1}, {3, 2}, {3, 3}, {3, 4} };
static const whisper_ahead g_aheads_tiny[] = { {2, 2}, {3, 0}, {3, 2}, {3, 3}, {3, 4}, {3, 5} };
static const whisper_ahead g_aheads_base_en[] = { {3, 3}, {4, 7}, {5, 1}, {5, 5}, {5, 7} };
static const whisper_ahead g_aheads_base[] = { {3, 1}, {4, 2}, {4, 3}, {4, 7}, {5, 1}, {5, 2}, {5, 4}, {5, 6} };
static const whisper_ahead g_aheads_small_en[] = { {6, 6}, {7, 0}, {7, 3}, {7, 8}, {8, 2}, {8, 5}, {8, 7}, {9, 0}, {9, 4}, {9, 8}, {9, 10}, {10, 0}, {10, 1}, {10, 2}, {10, 3}, {10, 6}, {10, 11}, {11, 2}, {11, 4} };
static const whisper_ahead g_aheads_small[] = { {5, 3}, {5, 9}, {8, 0}, {8, 4}, {8, 7}, {8, 8}, {9, 0}, {9, 7}, {9, 9}, {10, 5} };
static const whisper_ahead g_aheads_medium_en[] = { {11, 4}, {14, 1}, {14, 12}, {14, 14}, {15, 4}, {16, 0}, {16, 4}, {16, 9}, {17, 12}, {17, 14}, {18, 7}, {18, 10}, {18, 15}, {20, 0}, {20, 3}, {20, 9}, {20, 14}, {21, 12} };
static const whisper_ahead g_aheads_medium[] = { {13, 15}, {15, 4}, {15, 15}, {16, 1}, {20, 0}, {23, 4} };
static const whisper_ahead g_aheads_large_v1[] = { {9, 19}, {11, 2}, {11, 4}, {11, 17}, {22, 7}, {22, 11}, {22, 17}, {23, 2}, {23, 15} };
static const whisper_ahead g_aheads_large_v2[] = { {10, 12}, {13, 17}, {16, 11}, {16, 12}, {16, 13}, {17, 15}, {17, 16}, {18, 4}, {18, 11}, {18, 19}, {19, 11}, {21, 2}, {21, 3}, {22, 3}, {22, 9}, {22, 12}, {23, 5}, {23, 7}, {23, 13}, {25, 5}, {26, 1}, {26, 12}, {27, 15} };
static const whisper_ahead g_aheads_large_v3[] = { {7, 0}, {10, 17}, {12, 18}, {13, 12}, {16, 1}, {17, 14}, {19, 11}, {21, 4}, {24, 1}, {25, 6} };
static const whisper_ahead g_aheads_large_v3_turbo[] = { {2, 4}, {2, 11}, {3, 3}, {3, 6}, {3, 11}, {3, 14} };
static const std::map<whisper_alignment_heads_preset, whisper_aheads> g_aheads {
{ WHISPER_AHEADS_TINY_EN, { 8, g_aheads_tiny_en } },
{ WHISPER_AHEADS_TINY, { 6, g_aheads_tiny } },
{ WHISPER_AHEADS_BASE_EN, { 5, g_aheads_base_en } },
{ WHISPER_AHEADS_BASE, { 8, g_aheads_base } },
{ WHISPER_AHEADS_SMALL_EN, { 19, g_aheads_small_en } },
{ WHISPER_AHEADS_SMALL, { 10, g_aheads_small } },
{ WHISPER_AHEADS_MEDIUM_EN, { 18, g_aheads_medium_en } },
{ WHISPER_AHEADS_MEDIUM, { 6, g_aheads_medium } },
{ WHISPER_AHEADS_LARGE_V1, { 9, g_aheads_large_v1 } },
{ WHISPER_AHEADS_LARGE_V2, { 23, g_aheads_large_v2 } },
{ WHISPER_AHEADS_LARGE_V3, { 10, g_aheads_large_v3 } },
{ WHISPER_AHEADS_LARGE_V3_TURBO, { 6, g_aheads_large_v3_turbo } },
};
static std::vector<uint32_t> get_alignment_heads_by_layer(const whisper_context_params & cparams, int il, int32_t n_text_layer, int32_t n_head);
struct whisper_mel {
int n_len = 0;
int n_len_org = 0;
int n_mel = 0;
std::vector<float> data;
};
struct whisper_filters {
int32_t n_mel;
int32_t n_fft;
std::vector<float> data;
};
struct whisper_vocab {
using id = int32_t;
using token = std::string;
int n_vocab = 51864;
std::map<token, id> token_to_id;
std::map<id, token> id_to_token;
// reference: https://github.com/openai/whisper/blob/248b6cb124225dd263bb9bd32d060b6517e067f8/whisper/tokenizer.py#L334-L349
id token_eot = 50256;
id token_sot = 50257;
// task tokens (used only for multilingual models)
id token_translate = 50357;
id token_transcribe = 50358;
// other special tokens
id token_solm = 50359; // [TDRZ] used by tinydiarize models to indicate speaker turn
id token_prev = 50360;
id token_nosp = 50361;
id token_not = 50362; // no timestamps
id token_beg = 50363; // begin timestamps
bool is_multilingual() const {
return n_vocab >= 51865;
}
int num_languages() const {
return n_vocab - 51765 - (is_multilingual() ? 1 : 0);
}
};
struct whisper_segment {
int64_t t0;
int64_t t1;
std::string text;
float no_speech_prob;
std::vector<whisper_token_data> tokens;
bool speaker_turn_next;
};
struct whisper_batch {
int32_t n_tokens;
whisper_token * token;
whisper_pos * pos;
int32_t * n_seq_id; // always 1, here for consistency with llama.cpp
whisper_seq_id ** seq_id; // null terminated
int8_t * logits;
};
static struct whisper_batch whisper_batch_init(int32_t n_tokens, int32_t n_seq_max) {
whisper_batch batch = { 0, nullptr, nullptr, nullptr, nullptr, nullptr, };
batch.token = (whisper_token * ) malloc(sizeof(whisper_token) * (n_tokens));
batch.pos = (whisper_pos *) malloc(sizeof(whisper_pos) * (n_tokens));
batch.n_seq_id = (int32_t *) malloc(sizeof(int32_t) * (n_tokens));
batch.seq_id = (whisper_seq_id **) malloc(sizeof(whisper_seq_id *) * (n_tokens + 1));
for (int i = 0; i < n_tokens; ++i) {
batch.seq_id[i] = (whisper_seq_id *) malloc(sizeof(whisper_seq_id) * n_seq_max);
}
batch.seq_id[n_tokens] = nullptr;
batch.logits = (int8_t *) malloc(sizeof(int8_t) * n_tokens);
return batch;
}
static void whisper_batch_free(struct whisper_batch batch) {
if (batch.token) free(batch.token);
if (batch.pos) free(batch.pos);
if (batch.n_seq_id) free(batch.n_seq_id);
if (batch.seq_id) {
for (int i = 0; batch.seq_id[i]; ++i) {
free(batch.seq_id[i]);
}
free(batch.seq_id);
}
if (batch.logits) free(batch.logits);
}
static void whisper_batch_prep_legacy(whisper_batch & batch, const whisper_token * tokens, int n_tokens, int n_past, int seq_id) {
batch.n_tokens = n_tokens;
for (int i = 0; i < n_tokens; ++i) {
if (tokens) {
batch.token[i] = tokens[i];
}
batch.pos [i] = n_past + i;
batch.n_seq_id[i] = 1;
batch.seq_id [i][0] = seq_id;
batch.logits [i] = 0;
}
batch.logits[n_tokens - 1] = 1;
}
// replace std::pair by using customized pair struct (reason: std::pair is very slow)
template<typename A, typename B>
struct whisper_pair {
A first;
B second;
// Define a constructor that takes two arguments.
whisper_pair(const A& a, const B& b) : first(a), second(b) {}
// Define a constructor that takes no argument.
whisper_pair() : first(A()), second(B()) {}
};
// ggml_backend_sched wrapper for whisper usage
struct whisper_sched {
ggml_backend_sched_t sched = nullptr;
std::vector<uint8_t> meta;
};
static size_t whisper_sched_size(struct whisper_sched & allocr) {
size_t size = allocr.meta.size();
for (int i = 0; i < ggml_backend_sched_get_n_backends(allocr.sched); ++i) {
ggml_backend_t backend = ggml_backend_sched_get_backend(allocr.sched, i);
size += ggml_backend_sched_get_buffer_size(allocr.sched, backend);
}
return size;
}
// measure the memory usage of a graph and prepare the allocr's internal data buffer
static bool whisper_sched_graph_init(struct whisper_sched & allocr, std::vector<ggml_backend_t> backends, std::function<struct ggml_cgraph *()> && get_graph) {
auto & sched = allocr.sched;
auto & meta = allocr.meta;
sched = ggml_backend_sched_new(backends.data(), nullptr, backends.size(), WHISPER_MAX_NODES, false, true);
meta.resize(ggml_tensor_overhead()*WHISPER_MAX_NODES + ggml_graph_overhead());
// since there are dependencies between the different graphs,
// we need to allocate them instead of only reserving to get the correct compute buffer size
if (!ggml_backend_sched_alloc_graph(sched, get_graph())) {
// failed to allocate the compute buffer
WHISPER_LOG_ERROR("%s: failed to allocate the compute buffer\n", __func__);
return false;
}
ggml_backend_sched_reset(sched);
return true;
}
// medium
// hparams: {
// 'n_mels': 80,
// 'n_vocab': 51864,
// 'n_audio_ctx': 1500,
// 'n_audio_state': 1024,
// 'n_audio_head': 16,
// 'n_audio_layer': 24,
// 'n_text_ctx': 448,
// 'n_text_state': 1024,
// 'n_text_head': 16,
// 'n_text_layer': 24
// }
//
// default hparams (Whisper tiny)
struct whisper_hparams {
int32_t n_vocab = 51864;
int32_t n_audio_ctx = 1500;
int32_t n_audio_state = 384;
int32_t n_audio_head = 6;
int32_t n_audio_layer = 4;
int32_t n_text_ctx = 448;
int32_t n_text_state = 384;
int32_t n_text_head = 6;
int32_t n_text_layer = 4;
int32_t n_mels = 80;
int32_t ftype = 1;
float eps = 1e-5f;
};
// audio encoding layer
struct whisper_layer_encoder {
// encoder.blocks.*.attn_ln
struct ggml_tensor * attn_ln_0_w;
struct ggml_tensor * attn_ln_0_b;
// encoder.blocks.*.attn.out
struct ggml_tensor * attn_ln_1_w;
struct ggml_tensor * attn_ln_1_b;
// encoder.blocks.*.attn.query
struct ggml_tensor * attn_q_w;
struct ggml_tensor * attn_q_b;
// encoder.blocks.*.attn.key
struct ggml_tensor * attn_k_w;
// encoder.blocks.*.attn.value
struct ggml_tensor * attn_v_w;
struct ggml_tensor * attn_v_b;
// encoder.blocks.*.mlp_ln
struct ggml_tensor * mlp_ln_w;
struct ggml_tensor * mlp_ln_b;
// encoder.blocks.*.mlp.0
struct ggml_tensor * mlp_0_w;
struct ggml_tensor * mlp_0_b;
// encoder.blocks.*.mlp.2
struct ggml_tensor * mlp_1_w;
struct ggml_tensor * mlp_1_b;
};
// token decoding layer
struct whisper_layer_decoder {
// decoder.blocks.*.attn_ln
struct ggml_tensor * attn_ln_0_w;
struct ggml_tensor * attn_ln_0_b;
// decoder.blocks.*.attn.out
struct ggml_tensor * attn_ln_1_w;
struct ggml_tensor * attn_ln_1_b;
// decoder.blocks.*.attn.query
struct ggml_tensor * attn_q_w;
struct ggml_tensor * attn_q_b;
// decoder.blocks.*.attn.key
struct ggml_tensor * attn_k_w;
// decoder.blocks.*.attn.value
struct ggml_tensor * attn_v_w;
struct ggml_tensor * attn_v_b;
// decoder.blocks.*.cross_attn_ln
struct ggml_tensor * cross_attn_ln_0_w;
struct ggml_tensor * cross_attn_ln_0_b;
// decoder.blocks.*.cross_attn.out
struct ggml_tensor * cross_attn_ln_1_w;
struct ggml_tensor * cross_attn_ln_1_b;
// decoder.blocks.*.cross_attn.query
struct ggml_tensor * cross_attn_q_w;
struct ggml_tensor * cross_attn_q_b;
// decoder.blocks.*.cross_attn.key
struct ggml_tensor * cross_attn_k_w;
// decoder.blocks.*.cross_attn.value
struct ggml_tensor * cross_attn_v_w;
struct ggml_tensor * cross_attn_v_b;
// decoder.blocks.*.mlp_ln
struct ggml_tensor * mlp_ln_w;
struct ggml_tensor * mlp_ln_b;
// decoder.blocks.*.mlp.0
struct ggml_tensor * mlp_0_w;
struct ggml_tensor * mlp_0_b;
// decoder.blocks.*.mlp.2
struct ggml_tensor * mlp_1_w;
struct ggml_tensor * mlp_1_b;
};
struct whisper_kv_cell {
whisper_pos pos = -1;
std::set<whisper_seq_id> seq_id;
bool has_seq_id(const whisper_seq_id & id) const {
return seq_id.find(id) != seq_id.end();
}
};
struct whisper_kv_cache {
uint32_t head = 0;
uint32_t size = 0;
// computed before each graph build
uint32_t n = 0;
std::vector<whisper_kv_cell> cells;
struct ggml_tensor * k;
struct ggml_tensor * v;
ggml_backend_buffer_t buffer = nullptr;
std::vector<uint8_t> ctx_buf;
};
struct whisper_model {
e_model type = MODEL_UNKNOWN;
whisper_hparams hparams;
whisper_filters filters;
// encoder.positional_embedding
struct ggml_tensor * e_pe;
// encoder.conv1
struct ggml_tensor * e_conv_1_w;
struct ggml_tensor * e_conv_1_b;
// encoder.conv2
struct ggml_tensor * e_conv_2_w;
struct ggml_tensor * e_conv_2_b;
// encoder.ln_post
struct ggml_tensor * e_ln_w;
struct ggml_tensor * e_ln_b;
// decoder.positional_embedding
struct ggml_tensor * d_pe;
// decoder.token_embedding
struct ggml_tensor * d_te;
// decoder.ln
struct ggml_tensor * d_ln_w;
struct ggml_tensor * d_ln_b;
std::vector<whisper_layer_encoder> layers_encoder;
std::vector<whisper_layer_decoder> layers_decoder;
// ggml context that contains all the meta information about the model tensors
std::vector<ggml_context *> ctxs;
// the model backend data is read-only and can be shared between processors
std::vector<ggml_backend_buffer_t> buffers;
// tensors
int n_loaded;
std::map<std::string, struct ggml_tensor *> tensors;
};
struct whisper_partial_utf8 {
uint32_t value; // bit value so far (unshifted)
int n_remain; // num bytes remaining; -1 indicates invalid sequence
};
struct whisper_grammar {
/*const*/ std::vector<std::vector<whisper_grammar_element>> rules;
std::vector<std::vector<const whisper_grammar_element *>> stacks;
// buffer for partially generated UTF-8 sequence from accepted tokens
whisper_partial_utf8 partial_utf8;
};
struct whisper_grammar_candidate {
whisper_token id;
const uint32_t * code_points;
whisper_partial_utf8 partial_utf8;
};
struct whisper_sequence {
std::vector<whisper_token_data> tokens;
// the accumulated transcription in the current iteration (used to truncate the tokens array)
int result_len;
double sum_logprobs_all; // the sum of the log probabilities of the tokens
double sum_logprobs; // the sum of the log probabilities of the tokens (first result_len tokens)
double avg_logprobs; // the average log probability of the tokens
double entropy; // the entropy of the tokens
double score; // likelihood rank score
};
// TAGS: WHISPER_DECODER_INIT
struct whisper_decoder {
// the currently generated sequence of tokens
whisper_sequence sequence;
// grammar parse state of generated sequence of tokens
whisper_grammar grammar;
int i_batch; // the index of the token in the current batch
int seek_delta; // the window shift found so far based on the decoded timestamp tokens
bool failed; // has the current segment failed to decode?
bool completed; // has the decoder completed the current segment?
bool has_ts; // have we already sampled a non-beg timestamp token for the current segment?
// new token probs, logits and logprobs after the last whisper_decode (1-dimensional array: [n_vocab])
std::vector<float> probs;
std::vector<float> logits;
std::vector<float> logprobs;
// work container used to avoid memory allocations
std::vector<whisper_pair<double, whisper_vocab::id>> logits_id;
mutable std::mt19937 rng; // used for sampling at t > 0.0
};
// [EXPERIMENTAL] Token-level timestamps with DTW
struct whisper_aheads_masks {
std::vector<struct ggml_tensor *> m; // One mask per text layer.
struct ggml_context * ctx = nullptr;
ggml_backend_buffer_t buffer = nullptr;
};
struct vad_time_mapping {
int64_t processed_time; // Time in processed (VAD) audio
int64_t original_time; // Corresponding time in original audio
};
struct whisper_state {
int64_t t_sample_us = 0;
int64_t t_encode_us = 0;
int64_t t_decode_us = 0;
int64_t t_batchd_us = 0;
int64_t t_prompt_us = 0;
int64_t t_mel_us = 0;
int32_t n_sample = 0; // number of tokens sampled
int32_t n_encode = 0; // number of encoder calls
int32_t n_decode = 0; // number of decoder calls with n_tokens == 1 (text-generation)
int32_t n_batchd = 0; // number of decoder calls with n_tokens < 16 (batch decoding)
int32_t n_prompt = 0; // number of decoder calls with n_tokens > 1 (prompt encoding)
int32_t n_fail_p = 0; // number of logprob threshold failures
int32_t n_fail_h = 0; // number of entropy threshold failures
// number of decoders for which we have constructed the KV cache
int32_t kv_self_n_dec = 0;
// unified self-attention KV cache for all decoders
whisper_kv_cache kv_self;
// cross-attention KV cache for the decoders
// shared between all decoders
whisper_kv_cache kv_cross;
// padded buffer for flash-attention
whisper_kv_cache kv_pad;
whisper_mel mel;
whisper_batch batch;
whisper_decoder decoders[WHISPER_MAX_DECODERS];
std::vector<ggml_backend_t> backends;
// - stores meta info about the intermediate tensors into the `meta` buffers
whisper_sched sched_conv;
whisper_sched sched_encode;
whisper_sched sched_cross;
whisper_sched sched_decode;
// result of the encoder
struct ggml_tensor * embd_conv = nullptr;
struct ggml_tensor * embd_enc = nullptr;
// helpers for GPU offloading
std::vector<float> inp_mel;
std::vector<float> inp_mask;
// decode output (2-dimensional array: [n_tokens][n_vocab])
std::vector<float> logits;
std::vector<whisper_segment> result_all;
// prompt history split into static prefix (prompt_past0) and dynamic rolling context (prompt_past1)
std::vector<whisper_token> prompt_past0; // static carried initial prompt (if enabled)
std::vector<whisper_token> prompt_past1; // dynamic context from decoded output
int lang_id = 0; // english by default
std::string path_model; // populated by whisper_init_from_file_with_params()
#ifdef WHISPER_USE_COREML
whisper_coreml_context * ctx_coreml = nullptr;
#endif
#ifdef WHISPER_USE_OPENVINO
whisper_openvino_context * ctx_openvino = nullptr;
#endif
#ifdef WHISPER_USE_VITISAI
whisper_vitisai_context * ctx_vitisai = nullptr;
#endif
// [EXPERIMENTAL] token-level timestamps data
int64_t t_beg = 0;
int64_t t_last = 0;
whisper_token tid_last;
std::vector<float> energy; // PCM signal energy
float no_speech_prob = 0.0f;
// [EXPERIMENTAL] Token-level timestamps with DTW
whisper_aheads_masks aheads_masks;
ggml_tensor * aheads_cross_QKs = nullptr;
std::vector<float> aheads_cross_QKs_data;
// [EXPERIMENTAL] speed-up techniques
int32_t exp_n_audio_ctx = 0; // 0 - use default
whisper_vad_context * vad_context = nullptr;
struct vad_segment_info {
int64_t orig_start;
int64_t orig_end;
int64_t vad_start;
int64_t vad_end;
};
std::vector<vad_segment_info> vad_segments;
bool has_vad_segments = false;
std::vector<vad_time_mapping> vad_mapping_table;
};
struct whisper_context {
int64_t t_load_us = 0;
int64_t t_start_us = 0;
ggml_type wtype = ggml_type::GGML_TYPE_F16; // weight type (FP32 / FP16 / QX)
ggml_type itype = ggml_type::GGML_TYPE_F16; // intermediate type (FP32 or FP16)
whisper_context_params params;
whisper_model model;
whisper_vocab vocab;
whisper_state * state = nullptr;
std::string path_model; // populated by whisper_init_from_file_with_params()
};
struct whisper_global {
// We save the log callback globally
ggml_log_callback log_callback = whisper_log_callback_default;
void * log_callback_user_data = nullptr;
};
static whisper_global g_state;
template<typename T>
static void read_safe(whisper_model_loader * loader, T & dest) {
loader->read(loader->context, &dest, sizeof(T));
BYTESWAP_VALUE(dest);
}
static bool whisper_kv_cache_init(
struct whisper_kv_cache & cache,
ggml_backend_t backend,
ggml_type wtype,
int64_t n_text_state,
int64_t n_text_layer,
int n_ctx) {
const int64_t n_mem = n_text_layer*n_ctx;
const int64_t n_elements = n_text_state*n_mem;
cache.ctx_buf.resize(2*ggml_tensor_overhead());
struct ggml_init_params params = {
/*.mem_size =*/ cache.ctx_buf.size(),
/*.mem_buffer =*/ cache.ctx_buf.data(),
/*.no_alloc =*/ true,
};
cache.head = 0;
cache.size = n_ctx;
cache.cells.clear();
cache.cells.resize(n_ctx);
struct ggml_context * ctx = ggml_init(params);
You can’t perform that action at this time.