forked from ggerganov/whisper.cpp
-
Notifications
You must be signed in to change notification settings - Fork 0
/
llama-vocab.cpp
1732 lines (1482 loc) · 66.7 KB
/
llama-vocab.cpp
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 "llama-vocab.h"
#include "unicode.h"
#include <algorithm>
#include <cassert>
#include <cfloat>
#include <climits>
#include <cstdarg>
#include <cstring>
#include <forward_list>
#include <queue>
#include <sstream>
//
// helpers
//
LLAMA_ATTRIBUTE_FORMAT(1, 2)
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);
}
struct naive_trie {
naive_trie() : has_value(false), value(0) {
}
void insert(const char * key, size_t len, int32_t value = 0) {
if (len == 0) {
this->has_value = true;
this->value = value;
return;
}
char c = key[0];
auto res = children.find(c);
if (res != children.end()) {
res->second.insert(key + 1, len - 1, value);
} else {
auto res = children.insert(std::make_pair(c, naive_trie()));
res.first->second.insert(key + 1, len - 1, value);
}
}
std::pair<const char *, size_t> get_longest_prefix(const char * key, size_t len, size_t offset = 0) {
if (len == 0 || offset == len) {
return std::make_pair(key, offset);
}
char c = key[offset];
auto res = children.find(c);
if (res != children.end()) {
return res->second.get_longest_prefix(key, len, offset + 1);
} else {
return std::make_pair(key, offset);
}
}
struct naive_trie * traverse(const char c) {
auto res = children.find(c);
if (res != children.end()) {
return &res->second;
} else {
return NULL;
}
}
std::map<char, struct naive_trie> children;
bool has_value;
llama_token value;
};
//
// impl
//
int llama_vocab::find_bpe_rank(const std::string & token_left, const std::string & token_right) const {
GGML_ASSERT(token_left.find(' ') == std::string::npos);
GGML_ASSERT(token_left.find('\n') == std::string::npos);
GGML_ASSERT(token_right.find(' ') == std::string::npos);
GGML_ASSERT(token_right.find('\n') == std::string::npos);
auto it = bpe_ranks.find(std::make_pair(token_left, token_right));
if (it == bpe_ranks.end()) {
return -1;
}
return it->second;
}
static enum llama_vocab_type llama_vocab_get_type(const llama_vocab & vocab) {
return vocab.type;
}
static bool llama_is_normal_token(const llama_vocab & vocab, llama_token id) {
GGML_ASSERT(vocab.type != LLAMA_VOCAB_TYPE_NONE);
return vocab.id_to_token[id].attr & LLAMA_TOKEN_ATTR_NORMAL;
}
static bool llama_is_unknown_token(const llama_vocab & vocab, llama_token id) {
GGML_ASSERT(vocab.type != LLAMA_VOCAB_TYPE_NONE);
return vocab.id_to_token[id].attr & LLAMA_TOKEN_ATTR_UNKNOWN;
}
static bool llama_is_control_token(const llama_vocab & vocab, llama_token id) {
GGML_ASSERT(vocab.type != LLAMA_VOCAB_TYPE_NONE);
return vocab.id_to_token[id].attr & LLAMA_TOKEN_ATTR_CONTROL;
}
static bool llama_is_byte_token(const llama_vocab & vocab, llama_token id) {
GGML_ASSERT(vocab.type != LLAMA_VOCAB_TYPE_NONE);
return vocab.id_to_token[id].attr & LLAMA_TOKEN_ATTR_BYTE;
}
static bool llama_is_user_defined_token(const llama_vocab & vocab, llama_token id) {
GGML_ASSERT(vocab.type != LLAMA_VOCAB_TYPE_NONE);
return vocab.id_to_token[id].attr & LLAMA_TOKEN_ATTR_USER_DEFINED;
}
static bool llama_is_unused_token(const llama_vocab & vocab, llama_token id) {
GGML_ASSERT(vocab.type != LLAMA_VOCAB_TYPE_NONE);
return vocab.id_to_token[id].attr & LLAMA_TOKEN_ATTR_UNUSED;
}
static uint8_t llama_token_to_byte(const llama_vocab & vocab, llama_token id) {
GGML_ASSERT(llama_vocab_get_type(vocab) != LLAMA_VOCAB_TYPE_NONE);
GGML_ASSERT(llama_is_byte_token(vocab, id));
const auto & token_data = vocab.id_to_token.at(id);
switch (llama_vocab_get_type(vocab)) {
case LLAMA_VOCAB_TYPE_SPM:
case LLAMA_VOCAB_TYPE_UGM: {
auto buf = token_data.text.substr(3, 2);
return strtol(buf.c_str(), NULL, 16);
}
case LLAMA_VOCAB_TYPE_BPE: {
GGML_ABORT("fatal error");
//return unicode_utf8_to_byte(token_data.text); // TODO: why is this here after GGML_ASSERT?
}
case LLAMA_VOCAB_TYPE_WPM: {
GGML_ABORT("fatal error");
}
default:
GGML_ABORT("fatal error");
}
}
static void llama_escape_whitespace(std::string & text) {
replace_all(text, " ", "\xe2\x96\x81");
}
static void llama_unescape_whitespace(std::string & word) {
replace_all(word, "\xe2\x96\x81", " ");
}
struct llm_symbol {
using index = int;
index prev;
index next;
const char * text;
size_t n;
};
static_assert(std::is_trivially_copyable<llm_symbol>::value, "llm_symbol is not trivially copyable");
//
// SPM tokenizer
// original implementation:
// https://github.com/ggerganov/llama.cpp/commit/074bea2eb1f1349a0118239c4152914aecaa1be4
//
struct llm_bigram_spm {
struct comparator {
bool operator()(llm_bigram_spm & l, llm_bigram_spm & r) {
return (l.score < r.score) || (l.score == r.score && l.left > r.left);
}
};
using queue_storage = std::vector<llm_bigram_spm>;
using queue = std::priority_queue<llm_bigram_spm, queue_storage, comparator>;
llm_symbol::index left;
llm_symbol::index right;
float score;
size_t size;
};
struct llm_tokenizer_spm {
llm_tokenizer_spm(const llama_vocab & vocab) : vocab(vocab) {}
void tokenize(const std::string & text, std::vector<llama_vocab::id> & output) {
// split string into utf8 chars
int index = 0;
size_t offs = 0;
while (offs < text.size()) {
llm_symbol sym;
size_t len = unicode_len_utf8(text[offs]);
sym.text = text.c_str() + offs;
sym.n = std::min(len, text.size() - offs);
offs += sym.n;
sym.prev = index - 1;
sym.next = offs == text.size() ? -1 : index + 1;
index++;
symbols.emplace_back(sym);
}
// seed the work queue with all possible 2-character tokens.
for (size_t i = 1; i < symbols.size(); ++i) {
try_add_bigram(i - 1, i);
}
// keep substituting the highest frequency pairs for as long as we can.
while (!work_queue.empty()) {
auto bigram = work_queue.top();
work_queue.pop();
auto & left_sym = symbols[bigram.left];
auto & right_sym = symbols[bigram.right];
// if one of the symbols already got merged, skip it.
if (left_sym.n == 0 || right_sym.n == 0 ||
left_sym.n + right_sym.n != bigram.size) {
continue;
}
// merge the right sym into the left one
left_sym.n += right_sym.n;
right_sym.n = 0;
//LLAMA_LOG_INFO("left = '%*s' size = %zu\n", (int) left_sym.n, left_sym.text, bigram.size);
// remove the right sym from the chain
left_sym.next = right_sym.next;
if (right_sym.next >= 0) {
symbols[right_sym.next].prev = bigram.left;
}
// find more substitutions
try_add_bigram(left_sym.prev, bigram.left);
try_add_bigram(bigram.left, left_sym.next);
}
for (int i = 0; i != -1; i = symbols[i].next) {
auto & symbol = symbols[i];
resegment(symbol, output);
}
}
private:
void resegment(llm_symbol & symbol, std::vector<llama_vocab::id> & output) {
auto text = std::string(symbol.text, symbol.n);
auto token = vocab.token_to_id.find(text);
// Do we need to support is_unused?
if (token != vocab.token_to_id.end()) {
output.push_back((*token).second);
return;
}
const auto p = rev_merge.find(text);
if (p == rev_merge.end()) {
// output any symbols that did not form tokens as bytes.
output.reserve(output.size() + symbol.n);
for (int j = 0; j < (int)symbol.n; ++j) {
llama_vocab::id token_id = llama_byte_to_token_impl(vocab, symbol.text[j]);
output.push_back(token_id);
}
return;
}
resegment(symbols[p->second.first], output);
resegment(symbols[p->second.second], output);
}
void try_add_bigram(int left, int right) {
if (left == -1 || right == -1) {
return;
}
const std::string text = std::string(symbols[left].text, symbols[left].n + symbols[right].n);
auto token = vocab.token_to_id.find(text);
if (token == vocab.token_to_id.end()) {
return;
}
if (static_cast<size_t>((*token).second) >= vocab.id_to_token.size()) {
return;
}
const auto & tok_data = vocab.id_to_token[(*token).second];
llm_bigram_spm bigram;
bigram.left = left;
bigram.right = right;
bigram.score = tok_data.score;
bigram.size = text.size();
work_queue.push(bigram);
// Do we need to support is_unused?
rev_merge[text] = std::make_pair(left, right);
}
const llama_vocab & vocab;
std::vector<llm_symbol> symbols;
llm_bigram_spm::queue work_queue;
std::map<std::string, std::pair<int, int>> rev_merge;
};
//
// BPE tokenizer
// adapted from https://github.com/cmp-nct/ggllm.cpp [MIT License]
// tried to simplify unicode stuff, so most likely does not work 100% correctly!
//
// TODO: there are a lot of common parts between spm and bpe tokenizers, should be refactored and reused
template<typename T, typename Container = std::vector<T>, typename Compare = std::less<typename Container::value_type>>
class llama_priority_queue : public std::priority_queue<T, Container, Compare> {
public:
using std::priority_queue<T, Container, Compare>::priority_queue;
T pop_move() {
T item = std::move(this->c.front());
std::pop_heap(this->c.begin(), this->c.end(), this->comp);
this->c.pop_back();
return item;
}
void pop() = delete;
};
struct llm_bigram_bpe {
struct comparator {
bool operator()(const llm_bigram_bpe & l, const llm_bigram_bpe & r) const {
return l.rank > r.rank || (l.rank == r.rank && l.left > r.left);
}
};
using queue_storage = std::vector<llm_bigram_bpe>;
using queue = llama_priority_queue<llm_bigram_bpe, queue_storage, comparator>;
llm_symbol::index left;
llm_symbol::index right;
std::string text;
int rank;
size_t size;
};
struct llm_tokenizer_bpe {
llm_tokenizer_bpe(const llama_vocab & vocab): vocab(vocab) {
GGML_ASSERT(vocab.type == LLAMA_VOCAB_TYPE_BPE);
switch (vocab.type_pre) {
case LLAMA_VOCAB_PRE_TYPE_LLAMA3:
regex_exprs = {
// original regex from tokenizer.json
//"(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
// adapted: https://github.com/ggerganov/llama.cpp/pull/6920#issuecomment-2080233989
"(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
};
break;
case LLAMA_VOCAB_PRE_TYPE_DBRX:
case LLAMA_VOCAB_PRE_TYPE_SMAUG:
regex_exprs = {
// same as llama3
"(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
};
break;
case LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_LLM:
regex_exprs = {
"[\r\n]",
"\\s?[A-Za-zµÀ-ÖØ-öø-ƺƼ-ƿDŽ-ʓʕ-ʯͰ-ͳͶͷͻ-ͽͿΆΈ-ΊΌΎ-ΡΣ-ϵϷ-ҁҊ-ԯԱ-ՖႠ-ჅᎠ-Ᏽᏸ-ᏽᲐ-ᲺᲽ-Ჿᴀ-ᴫᵫ-ᵷᵹ-ᶚḀ-ἕἘ-Ἕἠ-ὅὈ-Ὅὐ-ὗὙὛὝὟ-ώᾀ-ᾴᾶ-ᾼιῂ-ῄῆ-ῌῐ-ΐῖ-Ίῠ-Ῥῲ-ῴῶ-ῼℂℇℊ-ℓℕℙ-ℝℤΩℨK-ℭℯ-ℴℹℼ-ℿⅅ-ⅉⅎↃↄⰀ-ⱻⱾ-ⳤⳫ-ⳮⳲⳳꙀ-ꙭꚀ-ꚛꜢ-ꝯꝱ-ꞇꞋ-ꞎꭰ-ꮿff-stﬓ-ﬗＡ-Ｚａ-ｚ𐐀-𐑏𐒰-𐓓𐓘-𐓻𐲀-𐲲𐳀-𐳲𑢠-𑣟𞤀-𞥃]+",
"\\s?[!-/:-~！-／：-～‘-‟ -。]+",
"\\s+$",
"[一-龥ࠀ-一가-]+",
"\\p{N}+",
};
break;
case LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_CODER:
regex_exprs = {
"[\r\n]",
"\\s?\\p{L}+",
"\\s?\\p{P}+",
"[一-龥ࠀ-一가-]+",
"\\p{N}",
};
break;
case LLAMA_VOCAB_PRE_TYPE_FALCON:
regex_exprs = {
"[\\p{P}\\$\\+<=>\\^~\\|`]+",
"'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)",
"[0-9][0-9][0-9]",
};
break;
case LLAMA_VOCAB_PRE_TYPE_STARCODER:
case LLAMA_VOCAB_PRE_TYPE_REFACT:
case LLAMA_VOCAB_PRE_TYPE_COMMAND_R:
case LLAMA_VOCAB_PRE_TYPE_SMOLLM:
case LLAMA_VOCAB_PRE_TYPE_CODESHELL:
case LLAMA_VOCAB_PRE_TYPE_EXAONE:
regex_exprs = {
"\\p{N}",
"'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)",
};
break;
case LLAMA_VOCAB_PRE_TYPE_GPT2:
case LLAMA_VOCAB_PRE_TYPE_MPT:
case LLAMA_VOCAB_PRE_TYPE_OLMO:
case LLAMA_VOCAB_PRE_TYPE_JAIS:
regex_exprs = {
"'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)",
};
break;
case LLAMA_VOCAB_PRE_TYPE_STABLELM2:
case LLAMA_VOCAB_PRE_TYPE_QWEN2:
regex_exprs = {
// original regex from tokenizer.json
// "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+"
"(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
};
break;
case LLAMA_VOCAB_PRE_TYPE_PORO:
case LLAMA_VOCAB_PRE_TYPE_BLOOM:
case LLAMA_VOCAB_PRE_TYPE_GPT3_FINNISH:
regex_exprs = {
" ?[^(\\s|.,!?…。，、।۔،)]+",
};
break;
case LLAMA_VOCAB_PRE_TYPE_CHATGLM4:
regex_exprs = {
"(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
};
break;
case LLAMA_VOCAB_PRE_TYPE_VIKING:
regex_exprs = {
" ?[^(\\s|.,!?…。，、।۔،)]+",
"\\p{N}",
};
break;
case LLAMA_VOCAB_PRE_TYPE_TEKKEN:
// original regex from tokenizer.json
// "[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]*[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]+|[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]+[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]*|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+"
regex_exprs = {
"[^\\r\\n\\p{L}\\p{N}]?((?=[\\p{L}])([^a-z]))*((?=[\\p{L}])([^A-Z]))+|[^\\r\\n\\p{L}\\p{N}]?((?=[\\p{L}])([^a-z]))+((?=[\\p{L}])([^A-Z]))*|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
};
break;
default:
// default regex for BPE tokenization pre-processing
regex_exprs = {
"[\\p{P}\\$\\+<=>\\^~\\|]+",
"'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)",
"\\p{N}+",
"[0-9][0-9][0-9]",
};
break;
}
}
void append(const llama_vocab::id token_id, std::vector<llama_vocab::id> & output) const {
output.push_back(token_id);
}
bool append_bos(std::vector<llama_vocab::id> & output) const {
if (vocab.tokenizer_add_bos) {
GGML_ASSERT(vocab.special_bos_id != -1);
output.push_back(vocab.special_bos_id);
return true;
}
return false;
}
bool append_eos(std::vector<llama_vocab::id> & output) const {
if (vocab.tokenizer_add_eos) {
GGML_ASSERT(vocab.special_eos_id != -1);
output.push_back(vocab.special_eos_id);
return true;
}
return false;
}
void check_double_bos_eos(const std::vector<llama_vocab::id> & output) const {
if (vocab.tokenizer_add_bos && output.size() >= 2 && output[1] == vocab.special_bos_id) {
LLAMA_LOG_WARN(
"%s: Added a BOS token to the prompt as specified by the model but the prompt "
"also starts with a BOS token. So now the final prompt starts with 2 BOS tokens. "
"Are you sure this is what you want?\n", __FUNCTION__);
}
if (vocab.tokenizer_add_eos && output.size() >= 2 && *(output.end()-2) == vocab.special_eos_id) {
LLAMA_LOG_WARN(
"%s: Added a EOS token to the prompt as specified by the model but the prompt "
"also ends with a EOS token. So now the final prompt ends with 2 EOS tokens. "
"Are you sure this is what you want?\n", __FUNCTION__);
}
}
void tokenize(const std::string & text, std::vector<llama_vocab::id> & output) {
int final_prev_index = -1;
const auto word_collection = unicode_regex_split(text, regex_exprs);
symbols_final.clear();
for (auto & word : word_collection) {
work_queue = llm_bigram_bpe::queue();
symbols.clear();
int index = 0;
size_t offset = 0;
if (vocab.tokenizer_ignore_merges && vocab.token_to_id.find(word) != vocab.token_to_id.end()) {
symbols.emplace_back(llm_symbol{-1, -1, word.c_str(), word.size()});
offset = word.size();
}
while (offset < word.size()) {
llm_symbol sym;
size_t char_len = std::min(word.size() - offset, (size_t) unicode_len_utf8(word[offset]));
sym.text = word.c_str() + offset;
sym.n = char_len;
offset += sym.n;
sym.prev = index - 1;
sym.next = offset == word.size() ? -1 : index + 1;
index++;
symbols.emplace_back(sym);
}
for (size_t i = 1; i < symbols.size(); ++i) {
add_new_bigram(i - 1, i);
}
// build token(s)
while (!work_queue.empty()) {
auto bigram = work_queue.pop_move();
auto & left_symbol = symbols[bigram.left];
auto & right_symbol = symbols[bigram.right];
if (left_symbol.n == 0 || right_symbol.n == 0) {
continue;
}
std::string left_token = std::string(left_symbol.text, left_symbol.n);
std::string right_token = std::string(right_symbol.text, right_symbol.n);
if (left_token + right_token != bigram.text) {
continue; // Skip this bigram if it's outdated
}
// merge the right sym into the left one
left_symbol.n += right_symbol.n;
right_symbol.n = 0;
// remove the right sym from the chain
left_symbol.next = right_symbol.next;
if (right_symbol.next >= 0) {
symbols[right_symbol.next].prev = bigram.left;
}
add_new_bigram(left_symbol.prev, bigram.left); // left side of current symbol
add_new_bigram(bigram.left, left_symbol.next); // right side of current symbol
}
// add the finished tokens to the final list keeping correct order for next and prev
for (auto & sym : symbols) {
if (sym.n > 0) {
sym.prev = final_prev_index;
sym.next = -1;
if (final_prev_index != -1) {
symbols_final[final_prev_index].next = symbols_final.size();
}
symbols_final.emplace_back(sym);
final_prev_index = symbols_final.size() - 1;
}
}
}
symbols = symbols_final;
if (!symbols.empty()) {
for (int i = 0; i != -1; i = symbols[i].next) {
auto & symbol = symbols[i];
if (symbol.n == 0) {
continue;
}
const std::string str = std::string(symbol.text, symbol.n);
const auto token = vocab.token_to_id.find(str);
if (token == vocab.token_to_id.end()) {
for (auto j = str.begin(); j != str.end(); ++j) {
std::string byte_str(1, *j);
auto token_multibyte = vocab.token_to_id.find(byte_str);
if (token_multibyte != vocab.token_to_id.end()) {
output.push_back(token_multibyte->second);
}
}
} else {
output.push_back((*token).second);
}
}
}
}
private:
void add_new_bigram(int left, int right) {
if (left == -1 || right == -1) {
return;
}
std::string left_token = std::string(symbols[left].text, symbols[left].n);
std::string right_token = std::string(symbols[right].text, symbols[right].n);
int rank_found = -1;
rank_found = vocab.find_bpe_rank(left_token, right_token);
if (rank_found < 0) {
return;
}
llm_bigram_bpe bigram;
bigram.left = left;
bigram.right = right;
bigram.text = left_token + right_token;
bigram.size = left_token.size() + right_token.size();
bigram.rank = rank_found;
work_queue.push(bigram);
}
const llama_vocab & vocab;
std::vector<std::string> regex_exprs;
std::vector<llm_symbol> symbols;
std::vector<llm_symbol> symbols_final;
llm_bigram_bpe::queue work_queue;
};
//
// WPM tokenizer
//
struct llm_tokenizer_wpm {
llm_tokenizer_wpm(const llama_vocab & vocab): vocab(vocab) {}
void tokenize(const std::string & text, std::vector<llama_vocab::id> & output) const {
const auto & token_map = vocab.token_to_id;
// normalize and split by whitespace
std::vector<std::string> words = preprocess(text);
// bos token prepended already
// find the longest tokens that form the words
for (const std::string & word : words) {
// skip empty words
if (word.size() == 0) {
continue;
}
// prepend phantom space
const std::string word1 = "\xe2\x96\x81" + word;
const int n = word1.size();
const size_t current_tokens = output.size();
// we're at the start of a new word
// move through character position in word
for (int i = 0; i < n; ++i) {
// loop through possible match length
bool match = false;
for (int j = std::min(n, i + vocab.max_token_len + 1); j > i; j--) {
auto it = token_map.find(word1.substr(i, j - i));
if (it != token_map.end()) {
output.push_back(it->second);
match = true;
i = j - 1;
break;
}
}
if (!match) { // discard all
output.resize(current_tokens);
break; // and discard next tokens
}
}
// we didn't find any matches for this word
if (current_tokens == output.size()) {
output.push_back(vocab.special_unk_id);
}
}
}
// TODO: reduce string copies by using cpts_offs array
std::vector<std::string> preprocess(const std::string & text) const {
const std::vector<uint32_t> cpts_nfd = unicode_cpts_normalize_nfd(unicode_cpts_from_utf8(text));
std::vector<std::string> words(1, "");
for (const uint32_t cpt : cpts_nfd) {
const auto flags = unicode_cpt_flags(cpt);
if (flags.is_whitespace) {
if (words.back().size()) { // finish previous word if any
words.emplace_back();
}
continue;
}
assert (!flags.is_separator);
if (cpt == 0 || cpt == 0xFFFD || flags.is_control) {
continue;
}
const std::string s = unicode_cpt_to_utf8(unicode_tolower(cpt));
if (flags.is_punctuation || ( cpt < 0x7F && flags.is_symbol ) || is_chinese_char(cpt)) {
if (words.back().size()) { // finish previous word if any
words.emplace_back();
}
words.back() = s; // single char word
words.emplace_back(); // start a new word
} else {
words.back() += s; // append char to word
}
}
if (!words.back().size()) {
words.pop_back();
}
return words;
}
static bool is_chinese_char(uint32_t cpt) {
return
(cpt >= 0x04E00 && cpt <= 0x09FFF) ||
(cpt >= 0x03400 && cpt <= 0x04DBF) ||
(cpt >= 0x20000 && cpt <= 0x2A6DF) ||
(cpt >= 0x2A700 && cpt <= 0x2B73F) ||
(cpt >= 0x2B740 && cpt <= 0x2B81F) ||
(cpt >= 0x2B920 && cpt <= 0x2CEAF) || // this should be 0x2B820 but in hf rust code it is 0x2B920
(cpt >= 0x0F900 && cpt <= 0x0FAFF) ||
(cpt >= 0x2F800 && cpt <= 0x2FA1F);
//(cpt >= 0x3000 && cpt <= 0x303F) ||
//(cpt >= 0xFF00 && cpt <= 0xFFEF);
}
const llama_vocab & vocab;
};
//
// UGM tokenizer
//
struct llm_tokenizer_ugm {
llm_tokenizer_ugm(const llama_vocab & vocab) : vocab(vocab) {
if (vocab.precompiled_charsmap.size() > 0) {
size_t charsmap_offset = 0;
// First four bytes of precompiled_charsmap contains length of binary
// blob containing XOR-compressed compact double array (XCDA) entries
uint32_t xcda_blob_size = *(const uint32_t *) &vocab.precompiled_charsmap[0];
charsmap_offset += sizeof(xcda_blob_size);
if (xcda_blob_size + charsmap_offset >= vocab.precompiled_charsmap.size()) {
throw std::runtime_error("Index out of array bounds in precompiled charsmap!");
}
// Next xcda_blob_size bytes contain entries of XOR-compressed compact
// double array (XCDA). Each entry is bit-packed into a 32-bit integer.
xcda_array = (const uint32_t *) &vocab.precompiled_charsmap[charsmap_offset];
xcda_array_size = xcda_blob_size / sizeof(uint32_t);
charsmap_offset += xcda_blob_size;
// Remaining bytes of precompiled charsmap contain null-terminated
// replacement strings for prefixes matched by the XCDA.
prefix_replacements = &vocab.precompiled_charsmap[charsmap_offset];
prefix_replacements_size = vocab.precompiled_charsmap.size() - charsmap_offset;
}
for (unsigned int id = 0; id < vocab.id_to_token.size(); ++id) {
const auto &token_data = vocab.id_to_token[id];
if (llama_is_normal_token(vocab, id)) {
min_score = std::min<float>(min_score, token_data.score);
max_score = std::max<float>(max_score, token_data.score);
}
if (llama_is_normal_token(vocab, id) ||
llama_is_user_defined_token(vocab, id) ||
llama_is_unused_token(vocab, id)) {
token_matcher.insert(token_data.text.data(), token_data.text.size(), id);
}
if (llama_is_user_defined_token(vocab, id)) {
user_defined_token_matcher.insert(token_data.text.data(), token_data.text.size());
}
}
unknown_token_score = min_score - unknown_token_score_penalty;
}
/* This implementation is based on SentencePiece optimized Viterbi algorithm for
* unigram language models. The general idea is to:
* - move along the input sequence in steps of one UTF code point,
* - at each step find all possible tokenizations of the prefix by
* traversing the tokens trie,
* - for each tokenization store the best one so far (by higher score)
* - use the position in sequence after given token as an index to store
* results
* - if there was no valid tokenization of the current UTF code point
* then use unknown token with additional score penalty
* After processing the whole sequence we backtrack from the end to get
* the best tokenization.
*/
void tokenize(const std::string & text, std::vector<llama_vocab::id> & output) {
// get current size of output (for reversal later)
size_t output_size = output.size();
// normalize the input first
std::string normalized;
normalize(text, &normalized);
size_t input_len = normalized.size();
if (input_len == 0) {
return;
}
// initialize score_sum to -FLT_MAX so it will be always lower than sums of token scores
std::vector<struct best_tokenization> tokenization_results(input_len + 1, {vocab.special_unk_id, 0, -FLT_MAX});
// at the beginning tokenization score is zero
tokenization_results[0] = { vocab.special_unk_id, 0, 0 };
for (size_t input_offset = 0; input_offset < input_len;) {
size_t prefix_offset = input_offset;
// calculate how many code units are in the currently processed UTF code point
size_t n_utf8_code_units = std::min<size_t>(unicode_len_utf8(normalized[input_offset]), input_len - input_offset);
// traverse the token matcher trie to find a matching token
bool single_codepoint_token_found = false;
const struct best_tokenization & current_best = tokenization_results[input_offset];
struct naive_trie * node = token_matcher.traverse(normalized[prefix_offset++]);
while (prefix_offset <= input_len && node != NULL) {
// check if we found valid token in prefix
if (node->has_value) {
// check if it corresponds to the whole UTF code point
if (prefix_offset - input_offset == n_utf8_code_units) {
single_codepoint_token_found = true;
}
llama_token token_id = node->value;
const auto & token_data = vocab.id_to_token[token_id];
// we set the user-defined token scores to 0 to make them more likely to be selected
// (normal token scores are log probabilities, so they are negative)
// score type is double here to make tokenization results exactly
// the same as in the HF tokenizer using SentencePiece
const double token_score = llama_is_user_defined_token(vocab, token_id) ? 0.0 : token_data.score;
const double challenger_score = current_best.score_sum + token_score;
struct best_tokenization & current_champ = tokenization_results[prefix_offset];
if (challenger_score > current_champ.score_sum) {
struct best_tokenization challenger = { token_id, input_offset, (float) challenger_score };
current_champ = challenger;
}
}
node = node->traverse(normalized[prefix_offset++]);
}
// if we didn't find a valid token corresponding to the whole UTF code point
// then use unknown token as the tokenization of this UTF code point
if (!single_codepoint_token_found) {
const double challenger_score = current_best.score_sum + unknown_token_score;
prefix_offset = input_offset + n_utf8_code_units;
struct best_tokenization & current_champ = tokenization_results[prefix_offset];
if (challenger_score > current_champ.score_sum) {
struct best_tokenization challenger = { vocab.special_unk_id, input_offset, (float) challenger_score };
current_champ = challenger;
}
}
// move to the next UTF code point
input_offset += n_utf8_code_units;
}
// now backtrack from the end to gather token ids of the best tokenization
// merge sequences of consecutive unknown tokens into single unknown tokens
bool is_prev_unknown = false;
for (struct best_tokenization & tokenization = tokenization_results[input_len]; ; tokenization = tokenization_results[tokenization.input_offset]) {
bool is_unknown = tokenization.token_id == vocab.special_unk_id;
if (!(is_prev_unknown && is_unknown)) {
output.push_back(tokenization.token_id);
}
if (tokenization.input_offset == 0) {
break;
}
is_prev_unknown = is_unknown;
}
// reverse the output since we added tokens starting from the end of the input
std::reverse(output.begin() + output_size, output.end());
}
private:
const llama_vocab & vocab;
// helper structure for returning normalization results
struct normalization_result {
const char * normalized;
size_t normalized_len;
size_t consumed_input;
};
void normalize(const std::string& input, std::string * normalized) {
normalized->clear();
normalized->reserve(input.size() * 3);
const std::string space = vocab.tokenizer_escape_whitespaces ? escaped_space : " ";
bool shall_prepend_space = !vocab.tokenizer_treat_whitespace_as_suffix && vocab.tokenizer_add_space_prefix;
bool shall_append_space = vocab.tokenizer_treat_whitespace_as_suffix && vocab.tokenizer_add_space_prefix;
bool shall_merge_spaces = vocab.tokenizer_remove_extra_whitespaces;
bool is_space_prepended = false;
bool processing_non_ws = false;
size_t input_len = input.size();
for (size_t input_offset = 0; input_offset < input_len; ) {
auto norm_res = normalize_prefix(input, input_offset);
for (size_t i = 0; i < norm_res.normalized_len; i++) {
char c = norm_res.normalized[i];
if (c != ' ') {
if (!processing_non_ws) {
processing_non_ws = true;
if ((shall_prepend_space && !is_space_prepended) || shall_merge_spaces) {
normalized->append(space);
is_space_prepended = true;
}
}
normalized->push_back(c);
} else {
if (processing_non_ws) {
processing_non_ws = false;
}
if (!shall_merge_spaces) {
normalized->append(space);
}
}
}
input_offset += norm_res.consumed_input;
}
if (shall_append_space) {
normalized->append(space);
}
}
/*
* This structure is a view wrapper for XOR-compressed double array (XCDA)
* See Shunsuke Kanda (2018). Space- and Time-Efficient String Dictionaries.
* Eeach bit-packed entry contains:
* - BASE array value in bits 10-30
* - LCHECK array value in bits 0-7
* - LEAF array value in bit 9
* Entries containing indexes of replacement sequences have set bit 31
*/
struct xcda_array_view {
public:
xcda_array_view(const uint32_t * xcda_array, size_t xcda_array_size) : xcda_array(xcda_array), xcda_array_size(xcda_array_size) {
}
uint32_t get_base(size_t index) {
uint32_t packed_node = get_node(index);
return (packed_node >> 10) << ((packed_node & (1U << 9)) >> 6);
}
uint32_t get_lcheck(size_t index) {
uint32_t packed_node = get_node(index);
return packed_node & ((1U << 31) | 0xff);
}
bool get_leaf(size_t index) {
uint32_t packed_node = get_node(index);
return (packed_node >> 8) & 1;
}
uint32_t get_value(size_t index) {
uint32_t packed_node = get_node(index);
return packed_node & ((1U << 31) - 1);
}
private:
uint32_t get_node(size_t index) {
if (index > xcda_array_size) {
throw std::runtime_error("Index out of array bounds in XCDA array!");
}
return xcda_array[index];
}
const uint32_t * xcda_array;
size_t xcda_array_size;