KBaba7/llama.cpp
0
1#if defined(_MSC_VER)2#define _SILENCE_CXX17_CODECVT_HEADER_DEPRECATION_WARNING3#endif4 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, flags129 const auto range_end = unicode_ranges_flags.begin()[i]; // codepoint_end, flags130 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, nfd148 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_utf8205# pragma clang diagnostic push206# pragma clang diagnostic ignored "-Wdeprecated-declarations"207#endif208 209 std::wstring_convert<std::codecvt_utf8<wchar_t>> conv;210 211#if defined(__clang__)212# pragma clang diagnostic pop213#endif214 215 return conv.from_bytes(s);216}217 218static std::vector<std::string> unicode_byte_encoding_process(const std::vector<std::string> & bpe_words) {219 std::vector<std::string> bpe_encoded_words;220 for (const auto & word : bpe_words) {221 std::string text_utf;222 auto utf_word = unicode_cpts_from_utf8(word);223 for (size_t i = 0; i < utf_word.size(); ++i) {224 text_utf += unicode_cpt_to_utf8(utf_word[i]);225 }226 227 std::string encoded_token;228 for (char & c : text_utf) {229 encoded_token += unicode_byte_to_utf8(c);230 }231 bpe_encoded_words.emplace_back(encoded_token);232 }233 return bpe_encoded_words;234}235 236// GPT2 system regex: 's|'t|'re|'ve|'m|'ll|'d| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+237static std::vector<size_t> unicode_regex_split_custom_gpt2(const std::string & text, const std::vector<size_t> & offsets) {238 std::vector<size_t> bpe_offsets; // store the offset of each word239 bpe_offsets.reserve(offsets.size()); // Reserve memory for the approximate size240 241 const auto cpts = unicode_cpts_from_utf8(text);242 243 size_t start = 0;244 for (auto offset : offsets) {245 const size_t offset_ini = start;246 const size_t offset_end = start + offset;247 assert(offset_end <= cpts.size());248 start = offset_end;249 250 static const uint32_t OUT_OF_RANGE = 0xFFFFFFFF;251 auto _get_cpt = [&] (const size_t pos) -> uint32_t {252 return (offset_ini <= pos && pos < offset_end) ? cpts[pos] : OUT_OF_RANGE;253 };254 255 auto _get_flags = [&] (const size_t pos) -> unicode_cpt_flags {256 return (offset_ini <= pos && pos < offset_end) ? unicode_cpt_flags_from_cpt(cpts[pos]) : unicode_cpt_flags{};257 };258 259 size_t _prev_end = offset_ini;260 auto _add_token = [&] (const size_t end) -> size_t {261 assert(_prev_end <= end && end <= offset_end);262 size_t len = end - _prev_end;263 if (len > 0) {264 bpe_offsets.push_back(len);265 }266 _prev_end = end;267 //if (len > 0) {268 // std::string s = "";269 // for(size_t p = end-len; p < end; p++)270 // s += unicode_cpt_to_utf8(cpts[p]);271 // printf(">>> '%s'\n", s.c_str());272 //}273 return len;274 };275 276 for (size_t pos = offset_ini; pos < offset_end; /*pos++*/ ) {277 const uint32_t cpt = _get_cpt(pos);278 const auto flags = _get_flags(pos);279 280 // regex: 's|'t|'re|'ve|'m|'ll|'d281 if (cpt == '\'' && pos+1 < offset_end) {282 uint32_t cpt_next = _get_cpt(pos+1);283 if (cpt_next == 's' || cpt_next == 't' || cpt_next == 'm' || cpt_next == 'd') {284 pos += _add_token(pos+2);285 continue;286 }287 if (pos+2 < offset_end) {288 uint32_t cpt_next_next = _get_cpt(pos+2);289 if ((cpt_next == 'r' && cpt_next_next == 'e') ||290 (cpt_next == 'v' && cpt_next_next == 'e') ||291 (cpt_next == 'l' && cpt_next_next == 'l')) {292 pos += _add_token(pos+3);293 continue;294 }295 }296 }297 298 auto flags2 = (cpt == ' ' ? _get_flags(pos+1) : flags);299 // regex: <space>?\p{L}+300 if (flags2.is_letter) {301 pos += (cpt == ' ');302 while (flags2.is_letter) {303 flags2 = _get_flags(++pos);304 }305 _add_token(pos);306 continue;307 }308 // regex: <space>?\p{N}+309 if (flags2.is_number) {310 pos += (cpt == ' ');311 while (flags2.is_number) {312 flags2 = _get_flags(++pos);313 }314 _add_token(pos);315 continue;316 }317 // regex: <space>?[^\s\p{L}\p{N}]+318 if (!(flags2.is_whitespace | flags2.is_letter | flags2.is_number) && flags2.as_uint()) {319 pos += (cpt == ' ');320 while (!(flags2.is_whitespace | flags2.is_letter | flags2.is_number) && flags2.as_uint()) {321 flags2 = _get_flags(++pos);322 }323 _add_token(pos);324 continue;325 }326 327 size_t num_whitespaces = 0;328 while (_get_flags(pos+num_whitespaces).is_whitespace) {329 num_whitespaces++;330 }331 332 // regex: \s+(?!\S)333 if (num_whitespaces > 1 && _get_cpt(pos+num_whitespaces) != OUT_OF_RANGE) {334 pos += num_whitespaces - 1;335 _add_token(pos);336 continue;337 }338 339 // regex: \s+340 if (num_whitespaces > 0) {341 pos += num_whitespaces;342 _add_token(pos);343 continue;344 }345 346 // no matches347 _add_token(++pos);348 }349 }350 351 return bpe_offsets;352}353 354// 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+"355static std::vector<size_t> unicode_regex_split_custom_llama3(const std::string & text, const std::vector<size_t> & offsets) {356 std::vector<size_t> bpe_offsets; // store the offset of each word357 bpe_offsets.reserve(offsets.size()); // Reserve memory for the approximate size358 359 const auto cpts = unicode_cpts_from_utf8(text);360 361 size_t start = 0;362 for (auto offset : offsets) {363 const size_t offset_ini = start;364 const size_t offset_end = start + offset;365 assert(offset_end <= cpts.size());366 start = offset_end;367 368 static const uint32_t OUT_OF_RANGE = 0xFFFFFFFF;369 auto _get_cpt = [&] (const size_t pos) -> uint32_t {370 return (offset_ini <= pos && pos < offset_end) ? cpts[pos] : OUT_OF_RANGE;371 };372 373 auto _get_flags = [&] (const size_t pos) -> unicode_cpt_flags {374 return (offset_ini <= pos && pos < offset_end) ? unicode_cpt_flags_from_cpt(cpts[pos]) : unicode_cpt_flags{};375 };376 377 size_t _prev_end = offset_ini;378 auto _add_token = [&] (const size_t end) -> size_t {379 assert(_prev_end <= end && end <= offset_end);380 size_t len = end - _prev_end;381 if (len > 0) {382 bpe_offsets.push_back(len);383 }384 _prev_end = end;385 //if (len > 0) {386 // std::string s = "";387 // for(size_t p = end-len; p < end; p++)388 // s += unicode_cpt_to_utf8(cpts[p]);389 // printf(">>> '%s'\n", s.c_str());390 //}391 return len;392 };393 394 for (size_t pos = offset_ini; pos < offset_end; /*pos++*/ ) {395 const uint32_t cpt = _get_cpt(pos);396 const auto flags = _get_flags(pos);397 398 // regex: (?i:'s|'t|'re|'ve|'m|'ll|'d) // case insensitive399 if (cpt == '\'' && pos+1 < offset_end) {400 uint32_t cpt_next = unicode_tolower(_get_cpt(pos+1));401 if (cpt_next == 's' || cpt_next == 't' || cpt_next == 'm' || cpt_next == 'd') {402 pos += _add_token(pos+2);403 continue;404 }405 if (pos+2 < offset_end) {406 uint32_t cpt_next_next = unicode_tolower(_get_cpt(pos+2));407 if ((cpt_next == 'r' && cpt_next_next == 'e') ||408 (cpt_next == 'v' && cpt_next_next == 'e') ||409 (cpt_next == 'l' && cpt_next_next == 'l')) {410 pos += _add_token(pos+3);411 continue;412 }413 }414 }415 416 // regex: [^\r\n\p{L}\p{N}]?\p{L}+417 if (!(cpt == '\r' || cpt == '\n' || flags.is_number)) {418 if (flags.is_letter || _get_flags(pos+1).is_letter) { // one or more letters419 pos++;420 while (_get_flags(pos).is_letter) {421 pos++;422 }423 _add_token(pos);424 continue;425 }426 }427 428 // regex: \p{N}{1,3}429 if (flags.is_number) {430 size_t ini = pos;431 while (_get_flags(pos).is_number) {432 if (++pos - ini >= 3 ) {433 _add_token(pos);434 ini = pos;435 }436 }437 _add_token(pos);438 continue;439 }440 441 // regex: <space>?[^\s\p{L}\p{N}]+[\r\n]*442 auto flags2 = (cpt == ' ' ? _get_flags(pos+1) : flags);443 if (!(flags2.is_whitespace | flags2.is_letter | flags2.is_number) && flags.as_uint()) {444 pos += (cpt == ' ');445 while (!(flags2.is_whitespace | flags2.is_letter | flags2.is_number) && flags2.as_uint()) {446 flags2 = _get_flags(++pos);447 }448 uint32_t cpt2 = _get_cpt(pos);449 while (cpt2 == '\r' || cpt2 == '\n') {450 cpt2 = _get_cpt(++pos);451 }452 _add_token(pos);453 continue;454 }455 456 size_t num_whitespaces = 0;457 size_t last_end_r_or_n = 0;458 while (_get_flags(pos+num_whitespaces).is_whitespace) {459 uint32_t cpt2 = _get_cpt(pos+num_whitespaces);460 if (cpt2 == '\r' || cpt2 == '\n') {461 last_end_r_or_n = pos + num_whitespaces + 1;462 }463 num_whitespaces++;464 }465 466 // regex: \s*[\r\n]+467 if (last_end_r_or_n > 0) {468 pos = last_end_r_or_n;469 _add_token(pos);470 continue;471 }472 473 // regex: \s+(?!\S)474 if (num_whitespaces > 1 && _get_cpt(pos+num_whitespaces) != OUT_OF_RANGE) {475 pos += num_whitespaces - 1;476 _add_token(pos);477 continue;478 }479 480 // regex: \s+481 if (num_whitespaces > 0) {482 pos += num_whitespaces;483 _add_token(pos);484 continue;485 }486 487 // no matches488 _add_token(++pos);489 }490 }491 492 return bpe_offsets;493}494 495// use std::wregex to split the text496static std::vector<size_t> unicode_regex_split_stl(const std::wstring & wtext, const std::wstring & regex_expr, const std::vector<size_t> & offsets) {497 std::wregex expr(regex_expr);498 std::vector<size_t> bpe_offsets; // store the offset of each word499 bpe_offsets.reserve(offsets.size()); // Reserve memory for the approximate size500 size_t start = 0;501 for (auto offset : offsets) {502 std::wcregex_iterator it(wtext.data() + start, wtext.data() + start + offset, expr);503 std::wcregex_iterator end;504 505 int64_t start_idx = 0;506 while (it != end) {507 std::wcmatch match = *it;508 if (match.position() > start_idx) {509 bpe_offsets.emplace_back(match.position() - start_idx);510 }511 bpe_offsets.emplace_back(match.length());512 start_idx = match.position() + match.length();513 ++it;514 }515 516 if (start_idx < (int64_t) offset) {517 bpe_offsets.emplace_back(offset - start_idx);518 }519 start += offset;520 }521 522 return bpe_offsets;523}524 525// use std::regex to split the text526static std::vector<size_t> unicode_regex_split_stl(const std::string & text, const std::string & regex_expr, const std::vector<size_t> & offsets) {527 std::regex expr(regex_expr);528 std::vector<size_t> bpe_offsets; // store the offset of each word529 bpe_offsets.reserve(offsets.size()); // Reserve memory for the approximate size530 size_t start = 0;531 for (auto offset : offsets) {532 std::cregex_iterator it(text.data() + start, text.data() + start + offset, expr);533 std::cregex_iterator end;534 535 int64_t start_idx = 0;536 while (it != end) {537 std::cmatch match = *it;538 if (match.position() > start_idx) {539 bpe_offsets.emplace_back(match.position() - start_idx);540 }541 bpe_offsets.emplace_back(match.length());542 start_idx = match.position() + match.length();543 ++it;544 }545 546 if (start_idx < (int64_t) offset) {547 bpe_offsets.emplace_back(offset - start_idx);548 }549 start += offset;550 }551 552 return bpe_offsets;553}554 555static std::vector<size_t> unicode_regex_split_custom(const std::string & text, const std::string & regex_expr, const std::vector<size_t> & offsets) {556 std::vector<size_t> bpe_offsets;557 558 if (regex_expr == "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)") {559 bpe_offsets = unicode_regex_split_custom_gpt2(text, offsets);560 } else if (561 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+" ||562 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+") {563 564 bpe_offsets = unicode_regex_split_custom_llama3(text, offsets);565 }566 567 return bpe_offsets;568}569 570//571// interface572//573 574std::string unicode_cpt_to_utf8(uint32_t cpt) {575 std::string result;576 577 if (/* 0x00 <= cpt && */ cpt <= 0x7f) {578 result.push_back(cpt);579 return result;580 }581 if (0x80 <= cpt && cpt <= 0x7ff) {582 result.push_back(0xc0 | ((cpt >> 6) & 0x1f));583 result.push_back(0x80 | (cpt & 0x3f));584 return result;585 }586 if (0x800 <= cpt && cpt <= 0xffff) {587 result.push_back(0xe0 | ((cpt >> 12) & 0x0f));588 result.push_back(0x80 | ((cpt >> 6) & 0x3f));589 result.push_back(0x80 | (cpt & 0x3f));590 return result;591 }592 if (0x10000 <= cpt && cpt <= 0x10ffff) {593 result.push_back(0xf0 | ((cpt >> 18) & 0x07));594 result.push_back(0x80 | ((cpt >> 12) & 0x3f));595 result.push_back(0x80 | ((cpt >> 6) & 0x3f));596 result.push_back(0x80 | (cpt & 0x3f));597 return result;598 }599 600 throw std::invalid_argument("invalid codepoint");601}602 603std::vector<uint32_t> unicode_cpts_normalize_nfd(const std::vector<uint32_t> & cpts) {604 auto comp = [] (const uint32_t cpt, const range_nfd & range) {605 return cpt < range.first;606 };607 std::vector<uint32_t> result(cpts.size());608 for (size_t i = 0; i < cpts.size(); ++i) {609 const uint32_t cpt = cpts[i];610 auto it = std::upper_bound(unicode_ranges_nfd.begin(), unicode_ranges_nfd.end(), cpt, comp) - 1;611 result[i] = (it->first <= cpt && cpt <= it->last) ? it->nfd : cpt;612 }613 return result;614}615 616std::vector<uint32_t> unicode_cpts_from_utf8(const std::string & utf8) {617 std::vector<uint32_t> result;618 result.reserve(utf8.size());619 size_t offset = 0;620 while (offset < utf8.size()) {621 try {622 result.push_back(unicode_cpt_from_utf8(utf8, offset));623 }624 catch (const std::invalid_argument & /*ex*/) {625 // Silently ignore invalid UTF-8 input to avoid leaking the exception beyond llama_tokenize626 ++offset;627 result.emplace_back(0xFFFD); // replacement character628 }629 }630 return result;631}632 633unicode_cpt_flags unicode_cpt_flags_from_cpt(const uint32_t cpt) {634 static const unicode_cpt_flags undef(unicode_cpt_flags::UNDEFINED);635 static const auto cpt_flags = unicode_cpt_flags_array();636 return cpt < cpt_flags.size() ? cpt_flags[cpt] : undef;637}638 639unicode_cpt_flags unicode_cpt_flags_from_utf8(const std::string & utf8) {640 static const unicode_cpt_flags undef(unicode_cpt_flags::UNDEFINED);641 if (utf8.empty()) {642 return undef; // undefined643 }644 size_t offset = 0;645 return unicode_cpt_flags_from_cpt(unicode_cpt_from_utf8(utf8, offset));646}647 648std::string unicode_byte_to_utf8(uint8_t byte) {649 static std::unordered_map<uint8_t, std::string> map = unicode_byte_to_utf8_map();650 return map.at(byte);651}652 653uint8_t unicode_utf8_to_byte(const std::string & utf8) {654 static std::unordered_map<std::string, uint8_t> map = unicode_utf8_to_byte_map();655 return map.at(utf8);656}657 658uint32_t unicode_tolower(uint32_t cpt) {659 // binary search660 auto it = std::lower_bound(unicode_map_lowercase.begin(), unicode_map_lowercase.end(), cpt,661 [](const std::pair<uint32_t, uint32_t> & pair, uint32_t value) {662 return pair.first < value;663 });664 if (it != unicode_map_lowercase.end() && it->first == cpt) {665 return it->second;666 }667 return cpt; // Return the original code point if no lowercase mapping is found668}669 670std::vector<std::string> unicode_regex_split(const std::string & text, const std::vector<std::string> & regex_exprs) {671 // unicode categories672 static const std::map<std::string, int> k_ucat_enum = {673 { "\\p{N}", unicode_cpt_flags::NUMBER },674 { "\\p{L}", unicode_cpt_flags::LETTER },675 { "\\p{P}", unicode_cpt_flags::PUNCTUATION },676 { "\\p{M}", unicode_cpt_flags::ACCENT_MARK },677 { "\\p{S}", unicode_cpt_flags::SYMBOL },678 };679 680 static const std::map<int, int> k_ucat_cpt = {681 { unicode_cpt_flags::NUMBER, 0xD1 },682 { unicode_cpt_flags::LETTER, 0xD2 },683 { unicode_cpt_flags::PUNCTUATION, 0xD3 },684 { unicode_cpt_flags::ACCENT_MARK, 0xD4 },685 { unicode_cpt_flags::SYMBOL, 0xD5 },686 };687 688 static const std::map<int, std::string> k_ucat_map = {689 { unicode_cpt_flags::NUMBER, "\x30-\x39" }, // 0-9690 { unicode_cpt_flags::LETTER, "\x41-\x5A\x61-\x7A" }, // A-Za-z691 { unicode_cpt_flags::PUNCTUATION, "\x21-\x23\x25-\x2A\x2C-\x2F\x3A-\x3B\x3F-\x40\\\x5B-\\\x5D\x5F\\\x7B\\\x7D" }, // !-#%-*,-/:-;?-@\[-\]_\{\}692 { unicode_cpt_flags::ACCENT_MARK, "" }, // no sub-128 codepoints693 { unicode_cpt_flags::SYMBOL, "\\\x24\\\x2B\x3C-\x3E\x5E\x60\\\x7C" }, // $+<=>^`|694 };695 696 // compute collapsed codepoints only if needed by at least one regex697 bool need_collapse = false;698 for (const auto & regex_expr : regex_exprs) {699 // search for unicode categories700 for (const auto & ucat : k_ucat_enum) {701 if (std::string::npos != regex_expr.find(ucat.first)) {702 need_collapse = true;703 break;704 }705 }706 }707 708 const auto cpts = unicode_cpts_from_utf8(text);709 710 // generate a "collapsed" representation of the text, where all codepoints are replaced by a single byte711 // ref: https://github.com/ggerganov/llama.cpp/pull/6920#issuecomment-2081479935712 std::string text_collapsed;713 if (need_collapse) {714 // collapse all unicode categories715 text_collapsed.resize(cpts.size());716 717 for (size_t i = 0; i < cpts.size(); ++i) {718 // keep single-byte codepoints as is719 if (cpts[i] < 128) {720 text_collapsed[i] = cpts[i];721 continue;722 }723 724 const auto flags = unicode_cpt_flags_from_cpt(cpts[i]);725 726 if (flags.is_whitespace) {727 //NOTE: C++ std::regex \s does not mach 0x85, Rust and Python regex does.728 //text_collapsed[i] = (char) 0x85; // <Next Line> as whitespace fallback729 text_collapsed[i] = (char) 0x0B; // <vertical tab> as whitespace fallback730 } else if (k_ucat_cpt.find(flags.category_flag()) != k_ucat_cpt.end()) {731 text_collapsed[i] = k_ucat_cpt.at(flags.category_flag());732 } else {733 text_collapsed[i] = (char) 0xD0; // fallback734 }735 }736 }737 738 std::vector<size_t> bpe_offsets = { cpts.size() };739 740 for (const auto & regex_expr : regex_exprs) {741 // first, see if we have an efficient custom regex implementation742 auto tmp = unicode_regex_split_custom(text, regex_expr, bpe_offsets);743 744 if (!tmp.empty()) {745 bpe_offsets = std::move(tmp);746 continue;747 }748 749 // fallback to general-purpose std::regex / std::wregex750 try {751 // if a unicode category is used in the regex, we use the collapsed text and replace the unicode category752 // with the corresponding collapsed representation753 bool use_collapsed = false;754 for (const auto & ucat : k_ucat_enum) {755 if (std::string::npos != regex_expr.find(ucat.first)) {756 use_collapsed = true;757 break;758 }759 }760 761 if (use_collapsed) {762 // sanity-check that the original regex does not contain any non-ASCII characters763 const auto cpts_regex = unicode_cpts_from_utf8(regex_expr);764 for (size_t i = 0; i < cpts_regex.size(); ++i) {765 if (cpts_regex[i] >= 128) {766 throw std::runtime_error("Regex includes both unicode categories and non-ASCII characters - not supported");767 }768 }769 770 // generate a collapsed representation of the regex771 std::string regex_expr_collapsed;772 773 // track if we are inside [], because nested [] are not allowed774 bool inside = false;775 for (size_t i = 0; i < regex_expr.size(); ++i) {776 if (regex_expr[i] == '[' && (i == 0 || regex_expr[i - 1] != '\\')) {777 regex_expr_collapsed += '[';778 inside = true;779 continue;780 }781 782 if (inside && regex_expr[i] == ']' && regex_expr[i - 1] != '\\') {783 regex_expr_collapsed += ']';784 inside = false;785 continue;786 }787 788 if (regex_expr[i + 0] == '\\' && i + 4 < regex_expr.size() &&789 regex_expr[i + 1] == 'p' &&790 regex_expr[i + 2] == '{' &&791 regex_expr[i + 4] == '}') {792 const std::string pat = regex_expr.substr(i, 5);793 if (k_ucat_enum.find(pat) != k_ucat_enum.end()) {794 if (!inside) {795 regex_expr_collapsed += '[';796 }797 regex_expr_collapsed += k_ucat_cpt.at(k_ucat_enum.at(pat));798 regex_expr_collapsed += k_ucat_map.at(k_ucat_enum.at(pat));799 if (!inside) {800 regex_expr_collapsed += ']';801 }802 i += 4;803 continue;804 }805 }806 807 regex_expr_collapsed += regex_expr[i];808 }809 810 //printf("text_collapsed: %s\n", text_collapsed.c_str());811 //printf("regex_expr_collapsed: %s\n", regex_expr_collapsed.c_str());812 bpe_offsets = unicode_regex_split_stl(text_collapsed, regex_expr_collapsed, bpe_offsets);813 } else {814 // no unicode category used, we can use std::wregex directly815 const std::wstring wregex_expr = unicode_wstring_from_utf8(regex_expr);816 817 // std::wregex \s does not mach non-ASCII whitespaces, using 0x0B as fallback818 std::wstring wtext(cpts.begin(), cpts.end());819 for (size_t i = 0; i < wtext.size(); ++i) {820 if (wtext[i] > 0x7F && unicode_cpt_flags_from_cpt(wtext[i]).is_whitespace) {821 wtext[i] = 0x0B;822 }823 }824 825 //printf("text: %s\n", text.c_str());826 //printf("regex_expr: %s\n", regex_expr.c_str());827 bpe_offsets = unicode_regex_split_stl(wtext, wregex_expr, bpe_offsets);828 }829 } catch (std::regex_error & e) {830 fprintf(stderr, "Failed to process regex: '%s'\n", regex_expr.c_str());831 fprintf(stderr, "Regex error: %s\n", e.what());832 throw std::runtime_error("Failed to process regex");833 }834 }835 836 std::vector<std::string> bpe_words;837 bpe_words.reserve(bpe_offsets.size()); // reserve memory for the approximate size838 839 size_t start = 0;840 for (size_t & offset : bpe_offsets) {841 bpe_words.emplace_back();842 for (size_t i = start; i < start + offset; ++i) {843 bpe_words.back() += unicode_cpt_to_utf8(cpts[i]);844 }845 start += offset;846 }847 848 return unicode_byte_encoding_process(bpe_words);849}850