1#if defined(_MSC_VER)
2#define _SILENCE_CXX17_CODECVT_HEADER_DEPRECATION_WARNING
3#endif
4
5#include "unicode.h"
6#include "unicode-data.h"
7
8#include <algorithm>
9#include <cassert>
10#include <codecvt>
11#include <cstddef>
12#include <cstdint>
13#include <locale>
14#include <map>
15#include <regex>
16#include <stdexcept>
17#include <string>
18#include <unordered_map>
19#include <utility>
20#include <vector>
21
22size_t unicode_len_utf8(char src) {
23 const size_t lookup[] = { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 3, 4 };
24 uint8_t highbits = static_cast<uint8_t>(src) >> 4;
25 return lookup[highbits];
26}
27
28static std::string unicode_cpts_to_utf8(const std::vector<uint32_t> & cps) {
29 std::string result;
30 for (size_t i = 0; i < cps.size(); ++i) {
31 result.append(unicode_cpt_to_utf8(cps[i]));
32 }
33 return result;
34}
35
36uint32_t unicode_cpt_from_utf8(const std::string & utf8, size_t & offset) {
37 assert(offset < utf8.size());
38 if (!(utf8[offset + 0] & 0x80)) {
39 auto result = utf8[offset + 0];
40 offset += 1;
41 return result;
42 }
43 if (!(utf8[offset + 0] & 0x40)) {
44 throw std::invalid_argument("invalid character");
45 }
46 if (!(utf8[offset + 0] & 0x20)) {
47 if (offset + 1 >= utf8.size() || ! ((utf8[offset + 1] & 0xc0) == 0x80)) {
48 throw std::invalid_argument("invalid character");
49 }
50 auto result = ((utf8[offset + 0] & 0x1f) << 6) | (utf8[offset + 1] & 0x3f);
51 offset += 2;
52 return result;
53 }
54 if (!(utf8[offset + 0] & 0x10)) {
55 if (offset + 2 >= utf8.size() || ! ((utf8[offset + 1] & 0xc0) == 0x80) || ! ((utf8[offset + 2] & 0xc0) == 0x80)) {
56 throw std::invalid_argument("invalid character");
57 }
58 auto result = ((utf8[offset + 0] & 0x0f) << 12) | ((utf8[offset + 1] & 0x3f) << 6) | (utf8[offset + 2] & 0x3f);
59 offset += 3;
60 return result;
61 }
62 if (!(utf8[offset + 0] & 0x08)) {
63 if (offset + 3 >= utf8.size() || ! ((utf8[offset + 1] & 0xc0) == 0x80) || ! ((utf8[offset + 2] & 0xc0) == 0x80) || !((utf8[offset + 3] & 0xc0) == 0x80)) {
64 throw std::invalid_argument("invalid character");
65 }
66 auto result = ((utf8[offset + 0] & 0x07) << 18) | ((utf8[offset + 1] & 0x3f) << 12) | ((utf8[offset + 2] & 0x3f) << 6) | (utf8[offset + 3] & 0x3f);
67 offset += 4;
68 return result;
69 }
70 throw std::invalid_argument("failed to convert utf8 to codepoint");
71}
72
73//static std::vector<uint16_t> unicode_cpt_to_utf16(uint32_t cpt) {
74// std::vector<uint16_t> result;
75// if (/* 0x0000 <= cpt && */ cpt <= 0xffff) {
76// result.emplace_back(cpt);
77// return result;
78// }
79// if (0x10000 <= cpt && cpt <= 0x10ffff) {
80// result.emplace_back(0xd800 | ((cpt - 0x10000) >> 10));
81// result.emplace_back(0xdc00 | ((cpt - 0x10000) & 0x03ff));
82// return result;
83// }
84// throw std::invalid_argument("failed to convert codepoint to utf16");
85//}
86
87//static std::vector<uint16_t> unicode_cpts_to_utf16(const std::vector<uint32_t> & cps) {
88// std::vector<uint16_t> result;
89// for (size_t i = 0; i < cps.size(); ++i) {
90// auto temp = unicode_cpt_to_utf16(cps[i]);
91// result.insert(result.end(), temp.begin(), temp.end());
92// }
93// return result;
94//}
95
96//static uint32_t unicode_cpt_from_utf16(const std::vector<uint16_t> & utf16, size_t & offset) {
97// assert(offset < utf16.size());
98// if (((utf16[0] >> 10) << 10) != 0xd800) {
99// auto result = utf16[offset + 0];
100// offset += 1;
101// return result;
102// }
103//
104// if (offset + 1 >= utf16.size() || !((utf16[1] & 0xdc00) == 0xdc00)) {
105// throw std::invalid_argument("invalid character");
106// }
107//
108// auto result = 0x10000 + (((utf16[0] & 0x03ff) << 10) | (utf16[1] & 0x03ff));
109// offset += 2;
110// return result;
111//}
112
113//static std::vector<uint32_t> unicode_cpts_from_utf16(const std::vector<uint16_t> & utf16) {
114// std::vector<uint32_t> result;
115// size_t offset = 0;
116// while (offset < utf16.size()) {
117// result.push_back(unicode_cpt_from_utf16(utf16, offset));
118// }
119// return result;
120//}
121
122static std::vector<unicode_cpt_flags> unicode_cpt_flags_array() {
123 std::vector<unicode_cpt_flags> cpt_flags(MAX_CODEPOINTS, unicode_cpt_flags::UNDEFINED);
124
125 assert (unicode_ranges_flags.begin()[0].first == 0);
126 assert (unicode_ranges_flags.begin()[unicode_ranges_flags.size()-1].first == MAX_CODEPOINTS);
127 for (size_t i = 1; i < unicode_ranges_flags.size(); ++i) {
128 const auto range_ini = unicode_ranges_flags.begin()[i-1]; // codepoint_ini, flags
129 const auto range_end = unicode_ranges_flags.begin()[i]; // codepoint_end, flags
130 for (uint32_t cpt = range_ini.first; cpt < range_end.first; ++cpt) {
131 cpt_flags[cpt] = range_ini.second;
132 }
133 }
134
135 for (auto cpt : unicode_set_whitespace) {
136 cpt_flags[cpt].is_whitespace = true;
137 }
138
139 for (auto p : unicode_map_lowercase) {
140 cpt_flags[p.second].is_lowercase = true;
141 }
142
143 for (auto p : unicode_map_uppercase) {
144 cpt_flags[p.second].is_uppercase = true;
145 }
146
147 for (auto &range : unicode_ranges_nfd) { // start, last, nfd
148 cpt_flags[range.nfd].is_nfd = true;
149 }
150
151 return cpt_flags;
152}
153
154static std::unordered_map<uint8_t, std::string> unicode_byte_to_utf8_map() {
155 std::unordered_map<uint8_t, std::string> map;
156 for (int ch = 0x21; ch <= 0x7E; ++ch) { // u'!' to u'~'
157 assert(0 <= ch && ch < 256);
158 map[ch] = unicode_cpt_to_utf8(ch);
159 }
160 for (int ch = 0xA1; ch <= 0xAC; ++ch) { // u'¡' to u'¬'
161 assert(0 <= ch && ch < 256);
162 map[ch] = unicode_cpt_to_utf8(ch);
163 }
164 for (int ch = 0xAE; ch <= 0xFF; ++ch) { // u'®' to u'ÿ'
165 assert(0 <= ch && ch < 256);
166 map[ch] = unicode_cpt_to_utf8(ch);
167 }
168 auto n = 0;
169 for (int ch = 0; ch < 256; ++ch) {
170 if (map.find(ch) == map.end()) {
171 map[ch] = unicode_cpt_to_utf8(256 + n);
172 ++n;
173 }
174 }
175 return map;
176}
177
178static std::unordered_map<std::string, uint8_t> unicode_utf8_to_byte_map() {
179 std::unordered_map<std::string, uint8_t> map;
180 for (int ch = 0x21; ch <= 0x7E; ++ch) { // u'!' to u'~'
181 assert(0 <= ch && ch < 256);
182 map[unicode_cpt_to_utf8(ch)] = ch;
183 }
184 for (int ch = 0xA1; ch <= 0xAC; ++ch) { // u'¡' to u'¬'
185 assert(0 <= ch && ch < 256);
186 map[unicode_cpt_to_utf8(ch)] = ch;
187 }
188 for (int ch = 0xAE; ch <= 0xFF; ++ch) { // u'®' to u'ÿ'
189 assert(0 <= ch && ch < 256);
190 map[unicode_cpt_to_utf8(ch)] = ch;
191 }
192 auto n = 0;
193 for (int ch = 0; ch < 256; ++ch) {
194 if (map.find(unicode_cpt_to_utf8(ch)) == map.end()) {
195 map[unicode_cpt_to_utf8(256 + n)] = ch;
196 ++n;
197 }
198 }
199 return map;
200}
201
202static inline std::wstring unicode_wstring_from_utf8(const std::string & s) {
203#if defined(__clang__)
204 // disable C++17 deprecation warning for std::codecvt_utf8
205# pragma clang diagnostic push
206# pragma clang diagnostic ignored "-Wdeprecated-declarations"
207#elif defined(__GNUC__)
208# pragma GCC diagnostic push
209# pragma GCC diagnostic ignored "-Wdeprecated-declarations"
210#endif
211
212 std::wstring_convert<std::codecvt_utf8<wchar_t>> conv;
213
214#if defined(__clang__)
215# pragma clang diagnostic pop
216#elif defined(__GNUC__)
217# pragma GCC diagnostic pop
218#endif
219
220 return conv.from_bytes(s);
221}
222
223static std::vector<std::string> unicode_byte_encoding_process(const std::vector<std::string> & bpe_words) {
224 std::vector<std::string> bpe_encoded_words;
225 for (const auto & word : bpe_words) {
226 std::string text_utf;
227 auto utf_word = unicode_cpts_from_utf8(word);
228 for (size_t i = 0; i < utf_word.size(); ++i) {
229 text_utf += unicode_cpt_to_utf8(utf_word[i]);
230 }
231
232 std::string encoded_token;
233 for (char & c : text_utf) {
234 encoded_token += unicode_byte_to_utf8(c);
235 }
236 bpe_encoded_words.emplace_back(encoded_token);
237 }
238 return bpe_encoded_words;
239}
240
241// GPT2 system regex: 's|'t|'re|'ve|'m|'ll|'d| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+
242static std::vector<size_t> unicode_regex_split_custom_gpt2(const std::string & text, const std::vector<size_t> & offsets) {
243 std::vector<size_t> bpe_offsets; // store the offset of each word
244 bpe_offsets.reserve(offsets.size()); // Reserve memory for the approximate size
245
246 const auto cpts = unicode_cpts_from_utf8(text);
247
248 size_t start = 0;
249 for (auto offset : offsets) {
250 const size_t offset_ini = start;
251 const size_t offset_end = start + offset;
252 assert(offset_end <= cpts.size());
253 start = offset_end;
254
255 static const uint32_t OUT_OF_RANGE = 0xFFFFFFFF;
256 auto _get_cpt = [&] (const size_t pos) -> uint32_t {
257 return (offset_ini <= pos && pos < offset_end) ? cpts[pos] : OUT_OF_RANGE;
258 };
259
260 auto _get_flags = [&] (const size_t pos) -> unicode_cpt_flags {
261 return (offset_ini <= pos && pos < offset_end) ? unicode_cpt_flags_from_cpt(cpts[pos]) : unicode_cpt_flags{};
262 };
263
264 size_t _prev_end = offset_ini;
265 auto _add_token = [&] (const size_t end) -> size_t {
266 assert(_prev_end <= end && end <= offset_end);
267 size_t len = end - _prev_end;
268 if (len > 0) {
269 bpe_offsets.push_back(len);
270 }
271 _prev_end = end;
272 //if (len > 0) {
273 // std::string s = "";
274 // for(size_t p = end-len; p < end; p++)
275 // s += unicode_cpt_to_utf8(cpts[p]);
276 // printf(">>> '%s'\n", s.c_str());
277 //}
278 return len;
279 };
280
281 for (size_t pos = offset_ini; pos < offset_end; /*pos++*/ ) {
282 const uint32_t cpt = _get_cpt(pos);
283 const auto flags = _get_flags(pos);
284
285 // regex: 's|'t|'re|'ve|'m|'ll|'d
286 if (cpt == '\'' && pos+1 < offset_end) {
287 uint32_t cpt_next = _get_cpt(pos+1);
288 if (cpt_next == 's' || cpt_next == 't' || cpt_next == 'm' || cpt_next == 'd') {
289 pos += _add_token(pos+2);
290 continue;
291 }
292 if (pos+2 < offset_end) {
293 uint32_t cpt_next_next = _get_cpt(pos+2);
294 if ((cpt_next == 'r' && cpt_next_next == 'e') ||
295 (cpt_next == 'v' && cpt_next_next == 'e') ||
296 (cpt_next == 'l' && cpt_next_next == 'l')) {
297 pos += _add_token(pos+3);
298 continue;
299 }
300 }
301 }
302
303 auto flags2 = (cpt == ' ' ? _get_flags(pos+1) : flags);
304 // regex: <space>?\p{L}+
305 if (flags2.is_letter) {
306 pos += (cpt == ' ');
307 while (flags2.is_letter) {
308 flags2 = _get_flags(++pos);
309 }
310 _add_token(pos);
311 continue;
312 }
313 // regex: <space>?\p{N}+
314 if (flags2.is_number) {
315 pos += (cpt == ' ');
316 while (flags2.is_number) {
317 flags2 = _get_flags(++pos);
318 }
319 _add_token(pos);
320 continue;
321 }
322 // regex: <space>?[^\s\p{L}\p{N}]+
323 if (!(flags2.is_whitespace | flags2.is_letter | flags2.is_number) && flags2.as_uint()) {
324 pos += (cpt == ' ');
325 while (!(flags2.is_whitespace | flags2.is_letter | flags2.is_number) && flags2.as_uint()) {
326 flags2 = _get_flags(++pos);
327 }
328 _add_token(pos);
329 continue;
330 }
331
332 size_t num_whitespaces = 0;
333 while (_get_flags(pos+num_whitespaces).is_whitespace) {
334 num_whitespaces++;
335 }
336
337 // regex: \s+(?!\S)
338 if (num_whitespaces > 1 && _get_cpt(pos+num_whitespaces) != OUT_OF_RANGE) {
339 pos += num_whitespaces - 1;
340 _add_token(pos);
341 continue;
342 }
343
344 // regex: \s+
345 if (num_whitespaces > 0) {
346 pos += num_whitespaces;
347 _add_token(pos);
348 continue;
349 }
350
351 // no matches
352 _add_token(++pos);
353 }
354 }
355
356 return bpe_offsets;
357}
358
359// LLAMA3 system regex: "(?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+"
360static std::vector<size_t> unicode_regex_split_custom_llama3(const std::string & text, const std::vector<size_t> & offsets) {
361 std::vector<size_t> bpe_offsets; // store the offset of each word
362 bpe_offsets.reserve(offsets.size()); // Reserve memory for the approximate size
363
364 const auto cpts = unicode_cpts_from_utf8(text);
365
366 size_t start = 0;
367 for (auto offset : offsets) {
368 const size_t offset_ini = start;
369 const size_t offset_end = start + offset;
370 assert(offset_end <= cpts.size());
371 start = offset_end;
372
373 static const uint32_t OUT_OF_RANGE = 0xFFFFFFFF;
374 auto _get_cpt = [&] (const size_t pos) -> uint32_t {
375 return (offset_ini <= pos && pos < offset_end) ? cpts[pos] : OUT_OF_RANGE;
376 };
377
378 auto _get_flags = [&] (const size_t pos) -> unicode_cpt_flags {
379 return (offset_ini <= pos && pos < offset_end) ? unicode_cpt_flags_from_cpt(cpts[pos]) : unicode_cpt_flags{};
380 };
381
382 size_t _prev_end = offset_ini;
383 auto _add_token = [&] (const size_t end) -> size_t {
384 assert(_prev_end <= end && end <= offset_end);
385 size_t len = end - _prev_end;
386 if (len > 0) {
387 bpe_offsets.push_back(len);
388 }
389 _prev_end = end;
390 //if (len > 0) {
391 // std::string s = "";
392 // for(size_t p = end-len; p < end; p++)
393 // s += unicode_cpt_to_utf8(cpts[p]);
394 // printf(">>> '%s'\n", s.c_str());
395 //}
396 return len;
397 };
398
399 for (size_t pos = offset_ini; pos < offset_end; /*pos++*/ ) {
400 const uint32_t cpt = _get_cpt(pos);
401 const auto flags = _get_flags(pos);
402
403 // regex: (?i:'s|'t|'re|'ve|'m|'ll|'d) // case insensitive
404 if (cpt == '\'' && pos+1 < offset_end) {
405 uint32_t cpt_next = unicode_tolower(_get_cpt(pos+1));
406 if (cpt_next == 's' || cpt_next == 't' || cpt_next == 'm' || cpt_next == 'd') {
407 pos += _add_token(pos+2);
408 continue;
409 }
410 if (pos+2 < offset_end) {
411 uint32_t cpt_next_next = unicode_tolower(_get_cpt(pos+2));
412 if ((cpt_next == 'r' && cpt_next_next == 'e') ||
413 (cpt_next == 'v' && cpt_next_next == 'e') ||
414 (cpt_next == 'l' && cpt_next_next == 'l')) {
415 pos += _add_token(pos+3);
416 continue;
417 }
418 }
419 }
420
421 // regex: [^\r\n\p{L}\p{N}]?\p{L}+
422 if (!(cpt == '\r' || cpt == '\n' || flags.is_number)) {
423 if (flags.is_letter || _get_flags(pos+1).is_letter) { // one or more letters
424 pos++;
425 while (_get_flags(pos).is_letter) {
426 pos++;
427 }
428 _add_token(pos);
429 continue;
430 }
431 }
432
433 // regex: \p{N}{1,3}
434 if (flags.is_number) {
435 size_t ini = pos;
436 while (_get_flags(pos).is_number) {
437 if (++pos - ini >= 3 ) {
438 _add_token(pos);
439 ini = pos;
440 }
441 }
442 _add_token(pos);
443 continue;
444 }
445
446 // regex: <space>?[^\s\p{L}\p{N}]+[\r\n]*
447 auto flags2 = (cpt == ' ' ? _get_flags(pos+1) : flags);
448 if (!(flags2.is_whitespace | flags2.is_letter | flags2.is_number) && flags.as_uint()) {
449 pos += (cpt == ' ');
450 while (!(flags2.is_whitespace | flags2.is_letter | flags2.is_number) && flags2.as_uint()) {
451 flags2 = _get_flags(++pos);
452 }
453 uint32_t cpt2 = _get_cpt(pos);
454 while (cpt2 == '\r' || cpt2 == '\n') {
455 cpt2 = _get_cpt(++pos);
456 }
457 _add_token(pos);
458 continue;
459 }
460
461 size_t num_whitespaces = 0;
462 size_t last_end_r_or_n = 0;
463 while (_get_flags(pos+num_whitespaces).is_whitespace) {
464 uint32_t cpt2 = _get_cpt(pos+num_whitespaces);
465 if (cpt2 == '\r' || cpt2 == '\n') {
466 last_end_r_or_n = pos + num_whitespaces + 1;
467 }
468 num_whitespaces++;
469 }
470
471 // regex: \s*[\r\n]+
472 if (last_end_r_or_n > 0) {
473 pos = last_end_r_or_n;
474 _add_token(pos);
475 continue;
476 }
477
478 // regex: \s+(?!\S)
479 if (num_whitespaces > 1 && _get_cpt(pos+num_whitespaces) != OUT_OF_RANGE) {
480 pos += num_whitespaces - 1;
481 _add_token(pos);
482 continue;
483 }
484
485 // regex: \s+
486 if (num_whitespaces > 0) {
487 pos += num_whitespaces;
488 _add_token(pos);
489 continue;
490 }
491
492 // no matches
493 _add_token(++pos);
494 }
495 }
496
497 return bpe_offsets;
498}
499
500template <typename CharT>
501static std::vector<size_t> unicode_regex_split_stl(const std::basic_string<CharT> & text, const std::basic_string<CharT> & regex, const std::vector<size_t> & offsets) {
502 using BidirIt = typename std::basic_string<CharT>::const_iterator;
503#ifdef _MSC_VER
504 // Bypass bug in MSVC: https://github.com/ggml-org/llama.cpp/issues/17830
505 constexpr auto regex_flags = std::regex_constants::ECMAScript;
506#else
507 constexpr auto regex_flags = std::regex_constants::optimize | std::regex_constants::nosubs;
508#endif
509 std::basic_regex<CharT> expr(regex, regex_flags);
510 std::vector<size_t> bpe_offsets; // store the offset of each word
511 bpe_offsets.reserve(offsets.size()); // Reserve memory for the approximate size
512 size_t start = 0;
513 for (auto offset : offsets) {
514 std::regex_iterator<BidirIt> it(text.begin() + start, text.begin() + start + offset, expr);
515 std::regex_iterator<BidirIt> end;
516
517 int64_t start_idx = 0;
518 while (it != end) {
519 std::match_results<BidirIt> match = *it;
520 if (match.position() > start_idx) {
521 bpe_offsets.emplace_back(match.position() - start_idx);
522 }
523 bpe_offsets.emplace_back(match.length());
524 start_idx = match.position() + match.length();
525 ++it;
526 }
527
528 if (start_idx < (int64_t) offset) {
529 bpe_offsets.emplace_back(offset - start_idx);
530 }
531 start += offset;
532 }
533
534 return bpe_offsets;
535}
536
537// K2 system regex patterns (from tokenization_kimi.py):
538// [\p{Han}]+|[^\r\n\p{L}\p{N}]?[\p{Lu}\p{Lt}\p{Lm}\p{Lo}\p{M}&&[^\p{Han}]]*[\p{Ll}\p{Lm}\p{Lo}\p{M}&&[^\p{Han}]]+(?i:'s|'t|'re|'ve|'m|'ll|'d)?|[^\r\n\p{L}\p{N}]?[\p{Lu}\p{Lt}\p{Lm}\p{Lo}\p{M}&&[^\p{Han}]]+[\p{Ll}\p{Lm}\p{Lo}\p{M}&&[^\p{Han}]]*(?i:'s|'t|'re|'ve|'m|'ll|'d)?|\p{N}{1,3}| ?[^\s\p{L}\p{N}]+[\r\n]*|\s*[\r\n]+|\s+(?!\S)|\s+
539static std::vector<size_t> unicode_regex_split_custom_kimi_k2(const std::string & text, const std::vector<size_t> & offsets) {
540 std::vector<size_t> bpe_offsets;
541 bpe_offsets.reserve(offsets.size());
542
543 const auto cpts = unicode_cpts_from_utf8(text);
544
545 size_t start = 0;
546 for (auto offset : offsets) {
547 const size_t offset_ini = start;
548 const size_t offset_end = start + offset;
549 assert(offset_end <= cpts.size());
550 start = offset_end;
551
552 static const uint32_t OUT_OF_RANGE = 0xFFFFFFFF;
553 auto _get_cpt = [&] (const size_t pos) -> uint32_t {
554 return (offset_ini <= pos && pos < offset_end) ? cpts[pos] : OUT_OF_RANGE;
555 };
556
557 auto _get_flags = [&] (const size_t pos) -> unicode_cpt_flags {
558 return (offset_ini <= pos && pos < offset_end) ? unicode_cpt_flags_from_cpt(cpts[pos]) : unicode_cpt_flags{};
559 };
560
561 size_t _prev_end = offset_ini;
562 auto _add_token = [&] (const size_t end) -> size_t {
563 assert(_prev_end <= end && end <= offset_end);
564 size_t len = end - _prev_end;
565 if (len > 0) {
566 bpe_offsets.push_back(len);
567 }
568 _prev_end = end;
569 return len;
570 };
571
572 for (size_t pos = offset_ini; pos < offset_end; /*pos++*/ ) {
573 const uint32_t cpt = _get_cpt(pos);
574 const auto flags = _get_flags(pos);
575
576 // Pattern 1: [\p{Han}]+ (Chinese characters)
577 if (unicode_cpt_is_han(cpt)) {
578 while (unicode_cpt_is_han(_get_cpt(pos))) {
579 pos++;
580 }
581 _add_token(pos);
582 continue;
583 }
584
585 // Pattern 2 & 3: Letter words excluding Han characters with optional contractions
586 // [^\r\n\p{L}\p{N}]?[\p{Lu}\p{Lt}\p{Lm}\p{Lo}\p{M}&&[^\p{Han}]]*[\p{Ll}\p{Lm}\p{Lo}\p{M}&&[^\p{Han}]]+(?:'s|'t|'re|'ve|'m|'ll|'d)?
587 // [^\r\n\p{L}\p{N}]?[\p{Lu}\p{Lt}\p{Lm}\p{Lo}\p{M}&&[^\p{Han}]]+[\p{Ll}\p{Lm}\p{Lo}\p{M}&&[^\p{Han}]]*(?:'s|'t|'re|'ve|'m|'ll|'d)?
588 // Check if current char is a letter OR if current char could be a leading char and next char is a letter
589 bool is_letter_pattern = (flags.is_letter && !unicode_cpt_is_han(cpt)) ||
590 (!(cpt == '\r' || cpt == '\n' || flags.is_letter || flags.is_number) &&
591 _get_flags(pos + 1).is_letter && !unicode_cpt_is_han(_get_cpt(pos + 1)));
592
593 if (is_letter_pattern) {
594 // Handle optional leading non-letter/non-number character
595 bool has_leading_char = false;
596 if (!(cpt == '\r' || cpt == '\n' || flags.is_letter || flags.is_number)) {
597 has_leading_char = true;
598 pos++;
599 }
600
601 // Match letter sequence (excluding Han characters)
602 bool has_letters = false;
603 while (_get_flags(pos).is_letter && !unicode_cpt_is_han(_get_cpt(pos))) {
604 has_letters = true;
605 pos++;
606 }
607
608 // Only proceed if we found letters (after potentially skipping leading char)
609 if (has_letters || (!has_leading_char && _get_flags(pos).is_letter && !unicode_cpt_is_han(_get_cpt(pos)))) {
610 if (!has_letters) pos++; // consume the first letter if we didn't already
611
612 // Continue consuming letters
613 while (_get_flags(pos).is_letter && !unicode_cpt_is_han(_get_cpt(pos))) {
614 pos++;
615 }
616
617 // Check for optional contractions (?:'s|'t|'re|'ve|'m|'ll|'d)
618 if (_get_cpt(pos) == '\'' && pos + 1 < offset_end) {
619 uint32_t cpt_next = unicode_tolower(_get_cpt(pos + 1));
620 if (cpt_next == 's' || cpt_next == 't' || cpt_next == 'm' || cpt_next == 'd') {
621 pos += 2;
622 } else if (pos + 2 < offset_end) {
623 uint32_t cpt_next_next = unicode_tolower(_get_cpt(pos + 2));
624 if ((cpt_next == 'r' && cpt_next_next == 'e') ||
625 (cpt_next == 'v' && cpt_next_next == 'e') ||
626 (cpt_next == 'l' && cpt_next_next == 'l')) {
627 pos += 3;
628 }
629 }
630 }
631
632 _add_token(pos);
633 continue;
634 } else if (has_leading_char) {
635 // We consumed a leading char but found no letters, backtrack
636 pos--;
637 }
638 }
639
640 // Pattern 4: \p{N}{1,3} (numbers 1-3 digits)
641 if (flags.is_number) {
642 size_t ini = pos;
643 while (_get_flags(pos).is_number) {
644 if (++pos - ini >= 3) {
645 _add_token(pos);
646 ini = pos;
647 }
648 }
649 _add_token(pos);
650 continue;
651 }
652
653 // Pattern 5: ?[^\s\p{L}\p{N}]+[\r\n]* (optional space + non-word chars + optional newlines)
654 auto flags2 = (cpt == ' ' ? _get_flags(pos + 1) : flags);
655 if (!(flags2.is_whitespace || flags2.is_letter || flags2.is_number) && flags2.as_uint()) {
656 pos += (cpt == ' ');
657 while (!(flags2.is_whitespace || flags2.is_letter || flags2.is_number) && flags2.as_uint()) {
658 flags2 = _get_flags(++pos);
659 }
660 // Match optional [\r\n]*
661 uint32_t cpt2 = _get_cpt(pos);
662 while (cpt2 == '\r' || cpt2 == '\n') {
663 cpt2 = _get_cpt(++pos);
664 }
665 _add_token(pos);
666 continue;
667 }
668
669 // Count whitespace characters
670 size_t num_whitespaces = 0;
671 size_t last_end_r_or_n = 0;
672 while (_get_flags(pos + num_whitespaces).is_whitespace) {
673 uint32_t cpt2 = _get_cpt(pos + num_whitespaces);
674 if (cpt2 == '\r' || cpt2 == '\n') {
675 last_end_r_or_n = pos + num_whitespaces + 1;
676 }
677 num_whitespaces++;
678 }
679
680 // Pattern 6: \s*[\r\n]+ (whitespace with newlines)
681 if (last_end_r_or_n > 0) {
682 pos = last_end_r_or_n;
683 _add_token(pos);
684 continue;
685 }
686
687 // Pattern 7: \s+(?!\S) (trailing whitespace)
688 if (num_whitespaces > 1 && _get_cpt(pos + num_whitespaces) != OUT_OF_RANGE) {
689 pos += num_whitespaces - 1;
690 _add_token(pos);
691 continue;
692 }
693
694 // Pattern 8: \s+ (general whitespace)
695 if (num_whitespaces > 0) {
696 pos += num_whitespaces;
697 _add_token(pos);
698 continue;
699 }
700
701 // No matches - consume single character
702 _add_token(++pos);
703 }
704 }
705
706 return bpe_offsets;
707}
708
709// AFMOE digit handling: splits digits with leading 1-2 based on total length modulo 3
710static std::vector<size_t> unicode_regex_split_custom_afmoe(const std::string & text, const std::vector<size_t> & offsets) {
711 std::vector<size_t> bpe_offsets;
712 bpe_offsets.reserve(offsets.size());
713
714 const auto cpts = unicode_cpts_from_utf8(text);
715
716 size_t start = 0;
717 for (auto offset : offsets) {
718 const size_t offset_ini = start;
719 const size_t offset_end = start + offset;
720 assert(offset_end <= cpts.size());
721 start = offset_end;
722
723 auto _get_flags = [&] (const size_t pos) -> unicode_cpt_flags {
724 return (offset_ini <= pos && pos < offset_end) ? unicode_cpt_flags_from_cpt(cpts[pos]) : unicode_cpt_flags{};
725 };
726
727 size_t _prev_end = offset_ini;
728 auto _add_token = [&] (const size_t end) -> size_t {
729 assert(_prev_end <= end && end <= offset_end);
730 size_t len = end - _prev_end;
731 if (len > 0) {
732 bpe_offsets.push_back(len);
733 }
734 _prev_end = end;
735 return len;
736 };
737
738 for (size_t pos = offset_ini; pos < offset_end; ) {
739 const auto flags = _get_flags(pos);
740
741 // Handle digit sequences with special splitting logic
742 if (flags.is_number) {
743 size_t digit_start = pos;
744 size_t digit_count = 0;
745
746 // Count consecutive digits
747 while (_get_flags(pos).is_number && pos < offset_end) {
748 digit_count++;
749 pos++;
750 }
751
752 // Split based on total length modulo 3
753 size_t remainder = digit_count % 3;
754 size_t current = digit_start;
755
756 // Emit leading 1-2 digits if needed
757 if (remainder > 0) {
758 _add_token(current + remainder);
759 current += remainder;
760 }
761
762 // Emit groups of 3
763 while (current < digit_start + digit_count) {
764 _add_token(current + 3);
765 current += 3;
766 }
767 continue;
768 }
769
770 // For non-digits, just move forward
771 pos++;
772 }
773
774 // Add any remaining content
775 if (_prev_end < offset_end) {
776 _add_token(offset_end);
777 }
778 }
779
780 return bpe_offsets;
781}
782
783static std::vector<size_t> unicode_regex_split_custom(const std::string & text, const std::string & regex_expr, const std::vector<size_t> & offsets) {
784 std::vector<size_t> bpe_offsets;
785
786 if (regex_expr == "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)") {
787 bpe_offsets = unicode_regex_split_custom_gpt2(text, offsets);
788 } else if (
789 regex_expr == "(?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+" ||
790 regex_expr == "(?:'[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+") {
791
792 bpe_offsets = unicode_regex_split_custom_llama3(text, offsets);
793 } else if (regex_expr == "\\p{Han}+") {
794 // K2's first pattern - handle all K2 patterns together
795 bpe_offsets = unicode_regex_split_custom_kimi_k2(text, offsets);
796 } else if (regex_expr == "\\p{AFMoE_digits}") {
797 // AFMOE digit pattern - use custom implementation for proper splitting
798 bpe_offsets = unicode_regex_split_custom_afmoe(text, offsets);
799 }
800
801 return bpe_offsets;
802}
803
804//
805// interface
806//
807
808std::string unicode_cpt_to_utf8(uint32_t cpt) {
809 std::string result;
810
811 if (/* 0x00 <= cpt && */ cpt <= 0x7f) {
812 result.push_back(cpt);
813 return result;
814 }
815 if (0x80 <= cpt && cpt <= 0x7ff) {
816 result.push_back(0xc0 | ((cpt >> 6) & 0x1f));
817 result.push_back(0x80 | (cpt & 0x3f));
818 return result;
819 }
820 if (0x800 <= cpt && cpt <= 0xffff) {
821 result.push_back(0xe0 | ((cpt >> 12) & 0x0f));
822 result.push_back(0x80 | ((cpt >> 6) & 0x3f));
823 result.push_back(0x80 | (cpt & 0x3f));
824 return result;
825 }
826 if (0x10000 <= cpt && cpt <= 0x10ffff) {
827 result.push_back(0xf0 | ((cpt >> 18) & 0x07));
828 result.push_back(0x80 | ((cpt >> 12) & 0x3f));
829 result.push_back(0x80 | ((cpt >> 6) & 0x3f));
830 result.push_back(0x80 | (cpt & 0x3f));
831 return result;
832 }
833
834 throw std::invalid_argument("invalid codepoint");
835}
836
837std::vector<uint32_t> unicode_cpts_normalize_nfd(const std::vector<uint32_t> & cpts) {
838 auto comp = [] (const uint32_t cpt, const range_nfd & range) {
839 return cpt < range.first;
840 };
841 std::vector<uint32_t> result(cpts.size());
842 for (size_t i = 0; i < cpts.size(); ++i) {
843 const uint32_t cpt = cpts[i];
844 auto it = std::upper_bound(unicode_ranges_nfd.begin(), unicode_ranges_nfd.end(), cpt, comp) - 1;
845 result[i] = (it->first <= cpt && cpt <= it->last) ? it->nfd : cpt;
846 }
847 return result;
848}
849
850std::vector<uint32_t> unicode_cpts_from_utf8(const std::string & utf8) {
851 std::vector<uint32_t> result;
852 result.reserve(utf8.size());
853 size_t offset = 0;
854 while (offset < utf8.size()) {
855 try {
856 result.push_back(unicode_cpt_from_utf8(utf8, offset));
857 }
858 catch (const std::invalid_argument & /*ex*/) {
859 // Silently ignore invalid UTF-8 input to avoid leaking the exception beyond llama_tokenize
860 ++offset;
861 result.emplace_back(0xFFFD); // replacement character
862 }
863 }
864 return result;
865}
866
867unicode_cpt_flags unicode_cpt_flags_from_cpt(const uint32_t cpt) {
868 static const unicode_cpt_flags undef(unicode_cpt_flags::UNDEFINED);
869 static const auto cpt_flags = unicode_cpt_flags_array();
870 return cpt < cpt_flags.size() ? cpt_flags[cpt] : undef;
871}
872
873unicode_cpt_flags unicode_cpt_flags_from_utf8(const std::string & utf8) {
874 static const unicode_cpt_flags undef(unicode_cpt_flags::UNDEFINED);
875 if (utf8.empty()) {
876 return undef; // undefined
877 }
878 size_t offset = 0;
879 return unicode_cpt_flags_from_cpt(unicode_cpt_from_utf8(utf8, offset));
880}
881
882std::string unicode_byte_to_utf8(uint8_t byte) {
883 static std::unordered_map<uint8_t, std::string> map = unicode_byte_to_utf8_map();
884 return map.at(byte);
885}
886
887uint8_t unicode_utf8_to_byte(const std::string & utf8) {
888 static std::unordered_map<std::string, uint8_t> map = unicode_utf8_to_byte_map();
889 return map.at(utf8);
890}
891
892uint32_t unicode_tolower(uint32_t cpt) {
893 // binary search
894 auto it = std::lower_bound(unicode_map_lowercase.begin(), unicode_map_lowercase.end(), cpt,
895 [](const std::pair<uint32_t, uint32_t> & pair, uint32_t value) {
896 return pair.first < value;
897 });
898 if (it != unicode_map_lowercase.end() && it->first == cpt) {
899 return it->second;
900 }
901 return cpt; // Return the original code point if no lowercase mapping is found
902}
903
904bool unicode_cpt_is_han(uint32_t cpt) {
905 // Han character ranges (Chinese/CJK characters)
906 // CJK Unified Ideographs (most common)
907 if (cpt >= 0x4E00 && cpt <= 0x9FFF) return true;
908
909 // CJK Extension A
910 if (cpt >= 0x3400 && cpt <= 0x4DBF) return true;
911
912 // CJK Extension B
913 if (cpt >= 0x20000 && cpt <= 0x2A6DF) return true;
914
915 // CJK Extension C
916 if (cpt >= 0x2A700 && cpt <= 0x2B73F) return true;
917
918 // CJK Extension D
919 if (cpt >= 0x2B740 && cpt <= 0x2B81F) return true;
920
921 // CJK Extension E
922 if (cpt >= 0x2B820 && cpt <= 0x2CEAF) return true;
923
924 // CJK Extension F
925 if (cpt >= 0x2CEB0 && cpt <= 0x2EBEF) return true;
926
927 // CJK Compatibility Ideographs
928 if (cpt >= 0xF900 && cpt <= 0xFAFF) return true;
929
930 // CJK Compatibility Ideographs Supplement
931 if (cpt >= 0x2F800 && cpt <= 0x2FA1F) return true;
932
933 return false;
934}
935
936std::vector<std::string> unicode_regex_split(const std::string & text, const std::vector<std::string> & regex_exprs) {
937 // unicode categories
938 static const std::map<std::string, int> k_ucat_enum = {
939 { "\\p{N}", unicode_cpt_flags::NUMBER },
940 { "\\p{L}", unicode_cpt_flags::LETTER },
941 { "\\p{P}", unicode_cpt_flags::PUNCTUATION },
942 { "\\p{M}", unicode_cpt_flags::ACCENT_MARK },
943 { "\\p{S}", unicode_cpt_flags::SYMBOL },
944 { "\\p{Lu}", unicode_cpt_flags::LETTER }, // Uppercase letter
945 { "\\p{Ll}", unicode_cpt_flags::LETTER }, // Lowercase letter
946 { "\\p{Lt}", unicode_cpt_flags::LETTER }, // Titlecase letter
947 { "\\p{Lm}", unicode_cpt_flags::LETTER }, // Modifier letter
948 { "\\p{Lo}", unicode_cpt_flags::LETTER }, // Other letter
949 };
950
951 static const std::map<int, int> k_ucat_cpt = {
952 { unicode_cpt_flags::NUMBER, 0xD1 },
953 { unicode_cpt_flags::LETTER, 0xD2 },
954 { unicode_cpt_flags::PUNCTUATION, 0xD3 },
955 { unicode_cpt_flags::ACCENT_MARK, 0xD4 },
956 { unicode_cpt_flags::SYMBOL, 0xD5 },
957 };
958
959 static const std::map<int, std::string> k_ucat_map = {
960 { unicode_cpt_flags::NUMBER, "\x30-\x39" }, // 0-9
961 { unicode_cpt_flags::LETTER, "\x41-\x5A\x61-\x7A" }, // A-Za-z
962 { unicode_cpt_flags::PUNCTUATION, "\x21-\x23\x25-\x2A\x2C-\x2F\x3A-\x3B\x3F-\x40\\\x5B-\\\x5D\x5F\\\x7B\\\x7D" }, // !-#%-*,-/:-;?-@\[-\]_\{\}
963 { unicode_cpt_flags::ACCENT_MARK, "" }, // no sub-128 codepoints
964 { unicode_cpt_flags::SYMBOL, "\\\x24\\\x2B\x3C-\x3E\x5E\x60\\\x7C" }, // $+<=>^`|
965 };
966
967 // compute collapsed codepoints only if needed by at least one regex
968 bool need_collapse = false;
969 for (const auto & regex_expr : regex_exprs) {
970 // search for unicode categories
971 for (const auto & ucat : k_ucat_enum) {
972 if (std::string::npos != regex_expr.find(ucat.first)) {
973 need_collapse = true;
974 break;
975 }
976 }
977 }
978
979 const auto cpts = unicode_cpts_from_utf8(text);
980
981 // generate a "collapsed" representation of the text, where all codepoints are replaced by a single byte
982 // ref: https://github.com/ggml-org/llama.cpp/pull/6920#issuecomment-2081479935
983 std::string text_collapsed;
984 if (need_collapse) {
985 // collapse all unicode categories
986 text_collapsed.resize(cpts.size());
987
988 for (size_t i = 0; i < cpts.size(); ++i) {
989 // keep single-byte codepoints as is
990 if (cpts[i] < 128) {
991 text_collapsed[i] = cpts[i];
992 continue;
993 }
994
995 const auto flags = unicode_cpt_flags_from_cpt(cpts[i]);
996
997 if (flags.is_whitespace) {
998 //NOTE: C++ std::regex \s does not mach 0x85, Rust and Python regex does.
999 //text_collapsed[i] = (char) 0x85; // <Next Line> as whitespace fallback
1000 text_collapsed[i] = (char) 0x0B; // <vertical tab> as whitespace fallback
1001 } else if (k_ucat_cpt.find(flags.category_flag()) != k_ucat_cpt.end()) {
1002 text_collapsed[i] = k_ucat_cpt.at(flags.category_flag());
1003 } else {
1004 text_collapsed[i] = (char) 0xD0; // fallback
1005 }
1006 }
1007 }
1008
1009 std::vector<size_t> bpe_offsets = { cpts.size() };
1010
1011 for (const auto & regex_expr : regex_exprs) {
1012 // first, see if we have an efficient custom regex implementation
1013 auto tmp = unicode_regex_split_custom(text, regex_expr, bpe_offsets);
1014
1015 if (!tmp.empty()) {
1016 bpe_offsets = std::move(tmp);
1017 continue;
1018 }
1019
1020 // fallback to general-purpose std::regex / std::wregex
1021 try {
1022 // if a unicode category is used in the regex, we use the collapsed text and replace the unicode category
1023 // with the corresponding collapsed representation
1024 bool use_collapsed = false;
1025 for (const auto & ucat : k_ucat_enum) {
1026 if (std::string::npos != regex_expr.find(ucat.first)) {
1027 use_collapsed = true;
1028 break;
1029 }
1030 }
1031
1032 if (use_collapsed) {
1033 // sanity-check that the original regex does not contain any non-ASCII characters
1034 const auto cpts_regex = unicode_cpts_from_utf8(regex_expr);
1035 for (size_t i = 0; i < cpts_regex.size(); ++i) {
1036 if (cpts_regex[i] >= 128) {
1037 throw std::runtime_error("Regex includes both unicode categories and non-ASCII characters - not supported");
1038 }
1039 }
1040
1041 // generate a collapsed representation of the regex
1042 std::string regex_expr_collapsed;
1043
1044 // track if we are inside [], because nested [] are not allowed
1045 bool inside = false;
1046 for (size_t i = 0; i < regex_expr.size(); ++i) {
1047 if (regex_expr[i] == '[' && (i == 0 || regex_expr[i - 1] != '\\')) {
1048 regex_expr_collapsed += '[';
1049 inside = true;
1050 continue;
1051 }
1052
1053 if (inside && regex_expr[i] == ']' && regex_expr[i - 1] != '\\') {
1054 regex_expr_collapsed += ']';
1055 inside = false;
1056 continue;
1057 }
1058
1059 // Match \p{...} Unicode properties of varying lengths
1060 if (regex_expr[i + 0] == '\\' && i + 3 < regex_expr.size() &&
1061 regex_expr[i + 1] == 'p' &&
1062 regex_expr[i + 2] == '{') {
1063 // Find the closing brace
1064 size_t closing_brace = regex_expr.find('}', i + 3);
1065 if (closing_brace != std::string::npos && closing_brace <= i + 10) { // reasonable limit
1066 const std::string pat = regex_expr.substr(i, closing_brace - i + 1);
1067 if (k_ucat_enum.find(pat) != k_ucat_enum.end()) {
1068 if (!inside) {
1069 regex_expr_collapsed += '[';
1070 }
1071 regex_expr_collapsed += k_ucat_cpt.at(k_ucat_enum.at(pat));
1072 regex_expr_collapsed += k_ucat_map.at(k_ucat_enum.at(pat));
1073 if (!inside) {
1074 regex_expr_collapsed += ']';
1075 }
1076 i = closing_brace;
1077 continue;
1078 }
1079 }
1080 }
1081
1082 regex_expr_collapsed += regex_expr[i];
1083 }
1084
1085 //printf("text_collapsed: %s\n", text_collapsed.c_str());
1086 //printf("regex_expr_collapsed: %s\n", regex_expr_collapsed.c_str());
1087 bpe_offsets = unicode_regex_split_stl(text_collapsed, regex_expr_collapsed, bpe_offsets);
1088 } else {
1089 // no unicode category used, we can use std::wregex directly
1090 const std::wstring wregex_expr = unicode_wstring_from_utf8(regex_expr);
1091
1092 // std::wregex \s does not mach non-ASCII whitespaces, using 0x0B as fallback
1093 std::wstring wtext(cpts.begin(), cpts.end());
1094 for (size_t i = 0; i < wtext.size(); ++i) {
1095 if (wtext[i] > 0x7F && unicode_cpt_flags_from_cpt(wtext[i]).is_whitespace) {
1096 wtext[i] = 0x0B;
1097 }
1098 }
1099
1100 //printf("text: %s\n", text.c_str());
1101 //printf("regex_expr: %s\n", regex_expr.c_str());
1102 bpe_offsets = unicode_regex_split_stl(wtext, wregex_expr, bpe_offsets);
1103 }
1104 } catch (std::regex_error & e) {
1105 fprintf(stderr, "Failed to process regex: '%s'\n", regex_expr.c_str());
1106 fprintf(stderr, "Regex error: %s\n", e.what());
1107 throw std::runtime_error("Failed to process regex");
1108 }
1109 }
1110
1111 std::vector<std::string> bpe_words;
1112 bpe_words.reserve(bpe_offsets.size()); // reserve memory for the approximate size
1113
1114 size_t start = 0;
1115 for (size_t & offset : bpe_offsets) {
1116 bpe_words.emplace_back();
1117 for (size_t i = start; i < start + offset; ++i) {
1118 bpe_words.back() += unicode_cpt_to_utf8(cpts[i]);
1119 }
1120 start += offset;
1121 }
1122
1123 return unicode_byte_encoding_process(bpe_words);
1124}