ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-01-27 01:29:51 +00:00
Code Issues Packages Projects Releases Wiki Activity

Enable CUDA graphs for MoE models + GPT-OSS support (#689)

Browse Source 
* gmp-oss: common

* gpt-oss: attnetion sinks, swiglu_oai

* gpt-oss: WIP llama

Model loads and runs (CPU only), but PPL is much to high
(~1500 for 1st batch vs ~200 in mainline).
Is it because of SWA, because of vocab, or did I introduce a bug somewhere?

* gpt-oss: CPU seems to be working

It was the SWA thta was missing in the previous commit.

There are issues with EOG tokens, so this still needs to be added.

* CUDA: ADD_ID

Just a copy from mainline

* gpt-oss: Seems to be working on CUDA

* gpt-oss: add sinks to the attn-vec kernels

* CUDA: add head size of 64 to new mma

Haven't turned it on yet, but observe slightly better PP and slightly
worse TG performance with that.

* gpt-oss: add ability to use -fmoe (only CUDA for now)

* Move row sums to the write place

* Add sinks to iqk flash attention

* gpt_oss: Implement -fmoe on the CPU

* Simdify swiglu_oai

Turning it off for now as performance becomes more variable,
so perhaps I'm running into thermal trottling imore often
because of making the CPU work too hard.

* llama: factor out model loader

* Builds successfully

* It runs, but mmap does not work

* Fix llama_mmap so mmap works

* Minor

* Fix CUDA after latest changes

* Attempt to use CUDA graphs with MoE models - not working

* CUDA graphs WIP - still not working

* CUDA graphs - seems to be working

Likely not all MLA variants are working.
I no longer remember why I added the q8_0 cpy that
transposes the tensor, but if really needed, this is now
missing. Also missing is q6_0.

* Make q8_0 cache work for DeepSeek models with CUDA graphs

* cuda: cpy for q6_0

* Fix llama_mmap on non-Linux platforms

* Adding forgotten file

* Iterating on Windows build failures

* cuda: re-add q8_0 -> q8_0 transpose

so mla = 2 can be used with CUDA graphs and q8_0 cache.

* Disable graphs without -fmoe

* Minor

* Turn graphs on by default

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
This commit is contained in:
Kawrakow
2025-08-15 09:18:07 +03:00
committed by GitHub
parent e082df47f2
commit fc06bc9d27
56 changed files with 8720 additions and 5115 deletions
Download Patch File Download Diff File Expand all files Collapse all files

691
src/unicode.cpp
View File

@@ -5,20 +5,19 @@
#include "unicode.h"
#include "unicode-data.h"
#include <algorithm>
#include <cassert>
#include <codecvt>
#include <cstddef>
#include <cstdint>
#include <locale>
#include <map>
#include <regex>
#include <stdexcept>
#include <string>
#include <unordered_map>
#include <unordered_set>
#include <utility>
#include <vector>
#include <locale>
#include <codecvt>
#include <iostream>
size_t unicode_len_utf8(char src) {
const size_t lookup[] = { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 3, 4 };
@@ -26,7 +25,7 @@ size_t unicode_len_utf8(char src) {
return lookup[highbits];
}
static std::string unicode_cpts_to_utf8(const std::vector<uint32_t>& cps) {
static std::string unicode_cpts_to_utf8(const std::vector<uint32_t> & cps) {
std::string result;
for (size_t i = 0; i < cps.size(); ++i) {
result.append(unicode_cpt_to_utf8(cps[i]));
@@ -34,7 +33,7 @@ static std::string unicode_cpts_to_utf8(const std::vector<uint32_t>& cps) {
return result;
}
uint32_t unicode_cpt_from_utf8(const std::string& utf8, size_t& offset) {
uint32_t unicode_cpt_from_utf8(const std::string & utf8, size_t & offset) {
assert(offset < utf8.size());
if (!(utf8[offset + 0] & 0x80)) {
auto result = utf8[offset + 0];
@@ -45,7 +44,7 @@ uint32_t unicode_cpt_from_utf8(const std::string& utf8, size_t& offset) {
throw std::invalid_argument("invalid character");
}
if (!(utf8[offset + 0] & 0x20)) {
if (offset + 1 >= utf8.size() || !((utf8[offset + 1] & 0xc0) == 0x80)) {
if (offset + 1 >= utf8.size() || ! ((utf8[offset + 1] & 0xc0) == 0x80)) {
throw std::invalid_argument("invalid character");
}
auto result = ((utf8[offset + 0] & 0x1f) << 6) | (utf8[offset + 1] & 0x3f);
@@ -53,7 +52,7 @@ uint32_t unicode_cpt_from_utf8(const std::string& utf8, size_t& offset) {
return result;
}
if (!(utf8[offset + 0] & 0x10)) {
if (offset + 2 >= utf8.size() || !((utf8[offset + 1] & 0xc0) == 0x80) || !((utf8[offset + 2] & 0xc0) == 0x80)) {
if (offset + 2 >= utf8.size() || ! ((utf8[offset + 1] & 0xc0) == 0x80) || ! ((utf8[offset + 2] & 0xc0) == 0x80)) {
throw std::invalid_argument("invalid character");
}
auto result = ((utf8[offset + 0] & 0x0f) << 12) | ((utf8[offset + 1] & 0x3f) << 6) | (utf8[offset + 2] & 0x3f);
@@ -61,7 +60,7 @@ uint32_t unicode_cpt_from_utf8(const std::string& utf8, size_t& offset) {
return result;
}
if (!(utf8[offset + 0] & 0x08)) {
if (offset + 3 >= utf8.size() || !((utf8[offset + 1] & 0xc0) == 0x80) || !((utf8[offset + 2] & 0xc0) == 0x80) || !((utf8[offset + 3] & 0xc0) == 0x80)) {
if (offset + 3 >= utf8.size() || ! ((utf8[offset + 1] & 0xc0) == 0x80) || ! ((utf8[offset + 2] & 0xc0) == 0x80) || !((utf8[offset + 3] & 0xc0) == 0x80)) {
throw std::invalid_argument("invalid character");
}
auto result = ((utf8[offset + 0] & 0x07) << 18) | ((utf8[offset + 1] & 0x3f) << 12) | ((utf8[offset + 2] & 0x3f) << 6) | (utf8[offset + 3] & 0x3f);
@@ -71,15 +70,15 @@ uint32_t unicode_cpt_from_utf8(const std::string& utf8, size_t& offset) {
throw std::invalid_argument("failed to convert utf8 to codepoint");
}
//static std::vector<uint16_t> unicode_cpt_to_utf16(uint32_t cp) {
//static std::vector<uint16_t> unicode_cpt_to_utf16(uint32_t cpt) {
// std::vector<uint16_t> result;
// if (/* 0x0000 <= cp && */ cp <= 0xffff) {
// result.emplace_back(cp);
// if (/* 0x0000 <= cpt && */ cpt <= 0xffff) {
// result.emplace_back(cpt);
// return result;
// }
// if (0x10000 <= cp && cp <= 0x10ffff) {
// result.emplace_back(0xd800 | ((cp - 0x10000) >> 10));
// result.emplace_back(0xdc00 | ((cp - 0x10000) & 0x03ff));
// if (0x10000 <= cpt && cpt <= 0x10ffff) {
// result.emplace_back(0xd800 | ((cpt - 0x10000) >> 10));
// result.emplace_back(0xdc00 | ((cpt - 0x10000) & 0x03ff));
// return result;
// }
// throw std::invalid_argument("failed to convert codepoint to utf16");
@@ -120,14 +119,14 @@ uint32_t unicode_cpt_from_utf8(const std::string& utf8, size_t& offset) {
// return result;
//}
static std::vector<codepoint_flags> unicode_cpt_flags_array() {
std::vector<codepoint_flags> cpt_flags(MAX_CODEPOINTS, codepoint_flags::UNDEFINED);
static std::vector<unicode_cpt_flags> unicode_cpt_flags_array() {
std::vector<unicode_cpt_flags> cpt_flags(MAX_CODEPOINTS, unicode_cpt_flags::UNDEFINED);
assert(unicode_ranges_flags.front().first == 0);
assert(unicode_ranges_flags.back().first == MAX_CODEPOINTS);
assert (unicode_ranges_flags.begin()[0].first == 0);
assert (unicode_ranges_flags.begin()[unicode_ranges_flags.size()-1].first == MAX_CODEPOINTS);
for (size_t i = 1; i < unicode_ranges_flags.size(); ++i) {
const auto range_ini = unicode_ranges_flags[i - 1]; // codepoint_ini, flags
const auto range_end = unicode_ranges_flags[i]; // codepoint_end, flags
const auto range_ini = unicode_ranges_flags.begin()[i-1]; // codepoint_ini, flags
const auto range_end = unicode_ranges_flags.begin()[i]; // codepoint_end, flags
for (uint32_t cpt = range_ini.first; cpt < range_end.first; ++cpt) {
cpt_flags[cpt] = range_ini.second;
}
@@ -145,7 +144,7 @@ static std::vector<codepoint_flags> unicode_cpt_flags_array() {
cpt_flags[p.second].is_uppercase = true;
}
for (auto& range : unicode_ranges_nfd) { // start, last, nfd
for (auto &range : unicode_ranges_nfd) { // start, last, nfd
cpt_flags[range.nfd].is_nfd = true;
}
@@ -200,55 +199,38 @@ static std::unordered_map<std::string, uint8_t> unicode_utf8_to_byte_map() {
return map;
}
static inline bool is_valid_utf8(const std::string& str) {
int remaining_bytes = 0; // 当前多字节字符剩余的字节数
for (unsigned char c : str) {
if (remaining_bytes == 0) {
if ((c & 0x80) == 0x00) continue; // 1字节字符
else if ((c & 0xE0) == 0xC0) remaining_bytes = 1; // 2字节
else if ((c & 0xF0) == 0xE0) remaining_bytes = 2; // 3字节
else if ((c & 0xF8) == 0xF0) remaining_bytes = 3; // 4字节
else return false; // 非法起始字节
}
else {
// 检查后续字节是否为10xxxxxx
if ((c & 0xC0) != 0x80)
{
return false;
}
remaining_bytes--;
}
}
return (remaining_bytes == 0); // 确保多字节字符完整
}
static inline std::wstring unicode_wstring_from_utf8(const std::string& s) {
static inline std::wstring unicode_wstring_from_utf8(const std::string & s) {
#if defined(__clang__)
// disable C++17 deprecation warning for std::codecvt_utf8
# pragma clang diagnostic push
# pragma clang diagnostic ignored "-Wdeprecated-declarations"
#elif defined(__GNUC__)
# pragma GCC diagnostic push
# pragma GCC diagnostic ignored "-Wdeprecated-declarations"
#endif
bool isvalid = is_valid_utf8(s);
std::wstring_convert<std::codecvt_utf8<wchar_t>> conv;
#if defined(__clang__)
# pragma clang diagnostic pop
#elif defined(__GNUC__)
# pragma GCC diagnostic pop
#endif
return conv.from_bytes(s);
}
static std::vector<std::string> unicode_byte_encoding_process(const std::vector<std::string>& bpe_words) {
static std::vector<std::string> unicode_byte_encoding_process(const std::vector<std::string> & bpe_words) {
std::vector<std::string> bpe_encoded_words;
for (const auto& word : bpe_words) {
for (const auto & word : bpe_words) {
std::string text_utf;
auto utf_word = unicode_cpts_from_utf8(word);
auto utf_word = unicode_cpts_from_utf8(word);
for (size_t i = 0; i < utf_word.size(); ++i) {
text_utf += unicode_cpt_to_utf8(utf_word[i]);
}
std::string encoded_token;
for (char& c : text_utf) {
for (char & c : text_utf) {
encoded_token += unicode_byte_to_utf8(c);
}
bpe_encoded_words.emplace_back(encoded_token);
@@ -257,7 +239,7 @@ static std::vector<std::string> unicode_byte_encoding_process(const std::vector<
}
// GPT2 system regex: 's|'t|'re|'ve|'m|'ll|'d| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+
static std::vector<size_t> unicode_regex_split_custom_gpt2(const std::string& text, const std::vector<size_t>& offsets) {
static std::vector<size_t> unicode_regex_split_custom_gpt2(const std::string & text, const std::vector<size_t> & offsets) {
std::vector<size_t> bpe_offsets; // store the offset of each word
bpe_offsets.reserve(offsets.size()); // Reserve memory for the approximate size
@@ -271,16 +253,16 @@ static std::vector<size_t> unicode_regex_split_custom_gpt2(const std::string& te
start = offset_end;
static const uint32_t OUT_OF_RANGE = 0xFFFFFFFF;
auto _get_cpt = [&](const size_t pos) -> uint32_t {
auto _get_cpt = [&] (const size_t pos) -> uint32_t {
return (offset_ini <= pos && pos < offset_end) ? cpts[pos] : OUT_OF_RANGE;
};
auto _get_flags = [&](const size_t pos) -> codepoint_flags {
return (offset_ini <= pos && pos < offset_end) ? unicode_cpt_flags(cpts[pos]) : codepoint_flags{};
auto _get_flags = [&] (const size_t pos) -> unicode_cpt_flags {
return (offset_ini <= pos && pos < offset_end) ? unicode_cpt_flags_from_cpt(cpts[pos]) : unicode_cpt_flags{};
};
size_t _prev_end = offset_ini;
auto _add_token = [&](const size_t end) -> size_t {
auto _add_token = [&] (const size_t end) -> size_t {
assert(_prev_end <= end && end <= offset_end);
size_t len = end - _prev_end;
if (len > 0) {
@@ -296,29 +278,29 @@ static std::vector<size_t> unicode_regex_split_custom_gpt2(const std::string& te
return len;
};
for (size_t pos = offset_ini; pos < offset_end; /*pos++*/) {
for (size_t pos = offset_ini; pos < offset_end; /*pos++*/ ) {
const uint32_t cpt = _get_cpt(pos);
const auto flags = _get_flags(pos);
// regex: 's|'t|'re|'ve|'m|'ll|'d
if (cpt == '\'' && pos + 1 < offset_end) {
uint32_t cpt_next = _get_cpt(pos + 1);
if (cpt == '\'' && pos+1 < offset_end) {
uint32_t cpt_next = _get_cpt(pos+1);
if (cpt_next == 's' || cpt_next == 't' || cpt_next == 'm' || cpt_next == 'd') {
pos += _add_token(pos + 2);
pos += _add_token(pos+2);
continue;
}
if (pos + 2 < offset_end) {
uint32_t cpt_next_next = _get_cpt(pos + 2);
if (pos+2 < offset_end) {
uint32_t cpt_next_next = _get_cpt(pos+2);
if ((cpt_next == 'r' && cpt_next_next == 'e') ||
(cpt_next == 'v' && cpt_next_next == 'e') ||
(cpt_next == 'l' && cpt_next_next == 'l')) {
pos += _add_token(pos + 3);
pos += _add_token(pos+3);
continue;
}
}
}
auto flags2 = (cpt == ' ' ? _get_flags(pos + 1) : flags);
auto flags2 = (cpt == ' ' ? _get_flags(pos+1) : flags);
// regex: <space>?\p{L}+
if (flags2.is_letter) {
pos += (cpt == ' ');
@@ -348,12 +330,12 @@ static std::vector<size_t> unicode_regex_split_custom_gpt2(const std::string& te
}
size_t num_whitespaces = 0;
while (_get_flags(pos + num_whitespaces).is_whitespace) {
while (_get_flags(pos+num_whitespaces).is_whitespace) {
num_whitespaces++;
}
// regex: \s+(?!\S)
if (num_whitespaces > 1 && _get_cpt(pos + num_whitespaces) != OUT_OF_RANGE) {
if (num_whitespaces > 1 && _get_cpt(pos+num_whitespaces) != OUT_OF_RANGE) {
pos += num_whitespaces - 1;
_add_token(pos);
continue;
@@ -374,6 +356,207 @@ static std::vector<size_t> unicode_regex_split_custom_gpt2(const std::string& te
return bpe_offsets;
}
// 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+"
static std::vector<size_t> unicode_regex_split_custom_llama3(const std::string & text, const std::vector<size_t> & offsets) {
std::vector<size_t> bpe_offsets; // store the offset of each word
bpe_offsets.reserve(offsets.size()); // Reserve memory for the approximate size
const auto cpts = unicode_cpts_from_utf8(text);
size_t start = 0;
for (auto offset : offsets) {
const size_t offset_ini = start;
const size_t offset_end = start + offset;
assert(offset_end <= cpts.size());
start = offset_end;
static const uint32_t OUT_OF_RANGE = 0xFFFFFFFF;
auto _get_cpt = [&] (const size_t pos) -> uint32_t {
return (offset_ini <= pos && pos < offset_end) ? cpts[pos] : OUT_OF_RANGE;
};
auto _get_flags = [&] (const size_t pos) -> unicode_cpt_flags {
return (offset_ini <= pos && pos < offset_end) ? unicode_cpt_flags_from_cpt(cpts[pos]) : unicode_cpt_flags{};
};
size_t _prev_end = offset_ini;
auto _add_token = [&] (const size_t end) -> size_t {
assert(_prev_end <= end && end <= offset_end);
size_t len = end - _prev_end;
if (len > 0) {
bpe_offsets.push_back(len);
}
_prev_end = end;
//if (len > 0) {
// std::string s = "";
// for(size_t p = end-len; p < end; p++)
// s += unicode_cpt_to_utf8(cpts[p]);
// printf(">>> '%s'\n", s.c_str());
//}
return len;
};
for (size_t pos = offset_ini; pos < offset_end; /*pos++*/ ) {
const uint32_t cpt = _get_cpt(pos);
const auto flags = _get_flags(pos);
// regex: (?i:'s|'t|'re|'ve|'m|'ll|'d) // case insensitive
if (cpt == '\'' && pos+1 < offset_end) {
uint32_t cpt_next = unicode_tolower(_get_cpt(pos+1));
if (cpt_next == 's' || cpt_next == 't' || cpt_next == 'm' || cpt_next == 'd') {
pos += _add_token(pos+2);
continue;
}
if (pos+2 < offset_end) {
uint32_t cpt_next_next = unicode_tolower(_get_cpt(pos+2));
if ((cpt_next == 'r' && cpt_next_next == 'e') ||
(cpt_next == 'v' && cpt_next_next == 'e') ||
(cpt_next == 'l' && cpt_next_next == 'l')) {
pos += _add_token(pos+3);
continue;
}
}
}
// regex: [^\r\n\p{L}\p{N}]?\p{L}+
if (!(cpt == '\r' || cpt == '\n' || flags.is_number)) {
if (flags.is_letter || _get_flags(pos+1).is_letter) { // one or more letters
pos++;
while (_get_flags(pos).is_letter) {
pos++;
}
_add_token(pos);
continue;
}
}
// regex: \p{N}{1,3}
if (flags.is_number) {
size_t ini = pos;
while (_get_flags(pos).is_number) {
if (++pos - ini >= 3 ) {
_add_token(pos);
ini = pos;
}
}
_add_token(pos);
continue;
}
// regex: <space>?[^\s\p{L}\p{N}]+[\r\n]*
auto flags2 = (cpt == ' ' ? _get_flags(pos+1) : flags);
if (!(flags2.is_whitespace | flags2.is_letter | flags2.is_number) && flags.as_uint()) {
pos += (cpt == ' ');
while (!(flags2.is_whitespace | flags2.is_letter | flags2.is_number) && flags2.as_uint()) {
flags2 = _get_flags(++pos);
}
uint32_t cpt2 = _get_cpt(pos);
while (cpt2 == '\r' || cpt2 == '\n') {
cpt2 = _get_cpt(++pos);
}
_add_token(pos);
continue;
}
size_t num_whitespaces = 0;
size_t last_end_r_or_n = 0;
while (_get_flags(pos+num_whitespaces).is_whitespace) {
uint32_t cpt2 = _get_cpt(pos+num_whitespaces);
if (cpt2 == '\r' || cpt2 == '\n') {
last_end_r_or_n = pos + num_whitespaces + 1;
}
num_whitespaces++;
}
// regex: \s*[\r\n]+
if (last_end_r_or_n > 0) {
pos = last_end_r_or_n;
_add_token(pos);
continue;
}
// regex: \s+(?!\S)
if (num_whitespaces > 1 && _get_cpt(pos+num_whitespaces) != OUT_OF_RANGE) {
pos += num_whitespaces - 1;
_add_token(pos);
continue;
}
// regex: \s+
if (num_whitespaces > 0) {
pos += num_whitespaces;
_add_token(pos);
continue;
}
// no matches
_add_token(++pos);
}
}
return bpe_offsets;
}
// use std::wregex to split the text
static std::vector<size_t> unicode_regex_split_stl(const std::wstring & wtext, const std::wstring & regex_expr, const std::vector<size_t> & offsets) {
std::wregex expr(regex_expr);
std::vector<size_t> bpe_offsets; // store the offset of each word
bpe_offsets.reserve(offsets.size()); // Reserve memory for the approximate size
size_t start = 0;
for (auto offset : offsets) {
std::wcregex_iterator it(wtext.data() + start, wtext.data() + start + offset, expr);
std::wcregex_iterator end;
int64_t start_idx = 0;
while (it != end) {
std::wcmatch match = *it;
if (match.position() > start_idx) {
bpe_offsets.emplace_back(match.position() - start_idx);
}
bpe_offsets.emplace_back(match.length());
start_idx = match.position() + match.length();
++it;
}
if (start_idx < (int64_t) offset) {
bpe_offsets.emplace_back(offset - start_idx);
}
start += offset;
}
return bpe_offsets;
}
// use std::regex to split the text
static std::vector<size_t> unicode_regex_split_stl(const std::string & text, const std::string & regex_expr, const std::vector<size_t> & offsets) {
std::regex expr(regex_expr);
std::vector<size_t> bpe_offsets; // store the offset of each word
bpe_offsets.reserve(offsets.size()); // Reserve memory for the approximate size
size_t start = 0;
for (auto offset : offsets) {
std::cregex_iterator it(text.data() + start, text.data() + start + offset, expr);
std::cregex_iterator end;
int64_t start_idx = 0;
while (it != end) {
std::cmatch match = *it;
if (match.position() > start_idx) {
bpe_offsets.emplace_back(match.position() - start_idx);
}
bpe_offsets.emplace_back(match.length());
start_idx = match.position() + match.length();
++it;
}
if (start_idx < (int64_t) offset) {
bpe_offsets.emplace_back(offset - start_idx);
}
start += offset;
}
return bpe_offsets;
}
// K2 system regex patterns (from tokenization_kimi.py):
// [\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+
static std::vector<size_t> unicode_regex_split_custom_kimi_k2(const std::string & text, const std::vector<size_t> & offsets) {
@@ -394,8 +577,8 @@ static std::vector<size_t> unicode_regex_split_custom_kimi_k2(const std::string
return (offset_ini <= pos && pos < offset_end) ? cpts[pos] : OUT_OF_RANGE;
};
auto _get_flags = [&] (const size_t pos) -> codepoint_flags {
return (offset_ini <= pos && pos < offset_end) ? unicode_cpt_flags(cpts[pos]) : codepoint_flags{};
auto _get_flags = [&] (const size_t pos) -> unicode_cpt_flags {
return (offset_ini <= pos && pos < offset_end) ? unicode_cpt_flags_from_cpt(cpts[pos]) : unicode_cpt_flags{};
};
size_t _prev_end = offset_ini;
@@ -546,220 +729,17 @@ static std::vector<size_t> unicode_regex_split_custom_kimi_k2(const std::string
return bpe_offsets;
}
// 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+"
static std::vector<size_t> unicode_regex_split_custom_llama3(const std::string& text, const std::vector<size_t>& offsets) {
std::vector<size_t> bpe_offsets; // store the offset of each word
bpe_offsets.reserve(offsets.size()); // Reserve memory for the approximate size
const auto cpts = unicode_cpts_from_utf8(text);
size_t start = 0;
for (auto offset : offsets) {
const size_t offset_ini = start;
const size_t offset_end = start + offset;
assert(offset_end <= cpts.size());
start = offset_end;
static const uint32_t OUT_OF_RANGE = 0xFFFFFFFF;
auto _get_cpt = [&](const size_t pos) -> uint32_t {
return (offset_ini <= pos && pos < offset_end) ? cpts[pos] : OUT_OF_RANGE;
};
auto _get_flags = [&](const size_t pos) -> codepoint_flags {
return (offset_ini <= pos && pos < offset_end) ? unicode_cpt_flags(cpts[pos]) : codepoint_flags{};
};
size_t _prev_end = offset_ini;
auto _add_token = [&](const size_t end) -> size_t {
assert(_prev_end <= end && end <= offset_end);
size_t len = end - _prev_end;
if (len > 0) {
bpe_offsets.push_back(len);
}
_prev_end = end;
//if (len > 0) {
// std::string s = "";
// for(size_t p = end-len; p < end; p++)
// s += unicode_cpt_to_utf8(cpts[p]);
// printf(">>> '%s'\n", s.c_str());
//}
return len;
};
for (size_t pos = offset_ini; pos < offset_end; /*pos++*/) {
const uint32_t cpt = _get_cpt(pos);
const auto flags = _get_flags(pos);
// regex: (?i:'s|'t|'re|'ve|'m|'ll|'d) // case insensitive
if (cpt == '\'' && pos + 1 < offset_end) {
uint32_t cpt_next = unicode_tolower(_get_cpt(pos + 1));
if (cpt_next == 's' || cpt_next == 't' || cpt_next == 'm' || cpt_next == 'd') {
pos += _add_token(pos + 2);
continue;
}
if (pos + 2 < offset_end) {
uint32_t cpt_next_next = unicode_tolower(_get_cpt(pos + 2));
if ((cpt_next == 'r' && cpt_next_next == 'e') ||
(cpt_next == 'v' && cpt_next_next == 'e') ||
(cpt_next == 'l' && cpt_next_next == 'l')) {
pos += _add_token(pos + 3);
continue;
}
}
}
// regex: [^\r\n\p{L}\p{N}]?\p{L}+
if (!(cpt == '\r' || cpt == '\n' || flags.is_number)) {
if (flags.is_letter || _get_flags(pos + 1).is_letter) { // one or more letters
pos++;
while (_get_flags(pos).is_letter) {
pos++;
}
_add_token(pos);
continue;
}
}
// regex: \p{N}{1,3}
if (flags.is_number) {
size_t ini = pos;
while (_get_flags(pos).is_number) {
if (++pos - ini >= 3) {
_add_token(pos);
ini = pos;
}
}
_add_token(pos);
continue;
}
// regex: <space>?[^\s\p{L}\p{N}]+[\r\n]*
auto flags2 = (cpt == ' ' ? _get_flags(pos + 1) : flags);
if (!(flags2.is_whitespace | flags2.is_letter | flags2.is_number) && flags.as_uint()) {
pos += (cpt == ' ');
while (!(flags2.is_whitespace | flags2.is_letter | flags2.is_number) && flags2.as_uint()) {
flags2 = _get_flags(++pos);
}
uint32_t cpt2 = _get_cpt(pos);
while (cpt2 == '\r' || cpt2 == '\n') {
cpt2 = _get_cpt(++pos);
}
_add_token(pos);
continue;
}
size_t num_whitespaces = 0;
size_t last_end_r_or_n = 0;
while (_get_flags(pos + num_whitespaces).is_whitespace) {
uint32_t cpt2 = _get_cpt(pos + num_whitespaces);
if (cpt2 == '\r' || cpt2 == '\n') {
last_end_r_or_n = pos + num_whitespaces + 1;
}
num_whitespaces++;
}
// regex: \s*[\r\n]+
if (last_end_r_or_n > 0) {
pos = last_end_r_or_n;
_add_token(pos);
continue;
}
// regex: \s+(?!\S)
if (num_whitespaces > 1 && _get_cpt(pos + num_whitespaces) != OUT_OF_RANGE) {
pos += num_whitespaces - 1;
_add_token(pos);
continue;
}
// regex: \s+
if (num_whitespaces > 0) {
pos += num_whitespaces;
_add_token(pos);
continue;
}
// no matches
_add_token(++pos);
}
}
return bpe_offsets;
}
// use std::wregex to split the text
static std::vector<size_t> unicode_regex_split_stl(const std::wstring& wtext, const std::wstring& regex_expr, const std::vector<size_t>& offsets) {
std::wregex expr(regex_expr);
std::vector<size_t> bpe_offsets; // store the offset of each word
bpe_offsets.reserve(offsets.size()); // Reserve memory for the approximate size
size_t start = 0;
for (auto offset : offsets) {
std::wcregex_iterator it(wtext.data() + start, wtext.data() + start + offset, expr);
std::wcregex_iterator end;
int64_t start_idx = 0;
while (it != end) {
std::wcmatch match = *it;
if (match.position() > start_idx) {
bpe_offsets.emplace_back(match.position() - start_idx);
}
bpe_offsets.emplace_back(match.length());
start_idx = match.position() + match.length();
++it;
}
if (start_idx < (int64_t)offset) {
bpe_offsets.emplace_back(offset - start_idx);
}
start += offset;
}
return bpe_offsets;
}
// use std::regex to split the text
static std::vector<size_t> unicode_regex_split_stl(const std::string& text, const std::string& regex_expr, const std::vector<size_t>& offsets) {
std::regex expr(regex_expr);
std::vector<size_t> bpe_offsets; // store the offset of each word
bpe_offsets.reserve(offsets.size()); // Reserve memory for the approximate size
size_t start = 0;
for (auto offset : offsets) {
std::cregex_iterator it(text.data() + start, text.data() + start + offset, expr);
std::cregex_iterator end;
int64_t start_idx = 0;
while (it != end) {
std::cmatch match = *it;
if (match.position() > start_idx) {
bpe_offsets.emplace_back(match.position() - start_idx);
}
bpe_offsets.emplace_back(match.length());
start_idx = match.position() + match.length();
++it;
}
if (start_idx < (int64_t)offset) {
bpe_offsets.emplace_back(offset - start_idx);
}
start += offset;
}
return bpe_offsets;
}
static std::vector<size_t> unicode_regex_split_custom(const std::string& text, const std::string& regex_expr, const std::vector<size_t>& offsets) {
static std::vector<size_t> unicode_regex_split_custom(const std::string & text, const std::string & regex_expr, const std::vector<size_t> & offsets) {
std::vector<size_t> bpe_offsets;
if (regex_expr == "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)") {
bpe_offsets = unicode_regex_split_custom_gpt2(text, offsets);
}
else if (
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+" ||
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+") {
} else if (
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+" ||
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+") {
bpe_offsets = unicode_regex_split_custom_llama3(text, offsets);
}
else if (regex_expr == "\\p{Han}+") {
} else if (regex_expr == "\\p{Han}+") {
// K2's first pattern - handle all K2 patterns together
bpe_offsets = unicode_regex_split_custom_kimi_k2(text, offsets);
}
@@ -771,71 +751,100 @@ static std::vector<size_t> unicode_regex_split_custom(const std::string& text, c
// interface
//
std::string unicode_cpt_to_utf8(uint32_t cp) {
std::string unicode_cpt_to_utf8(uint32_t cpt) {
std::string result;
if (/* 0x00 <= cp && */ cp <= 0x7f) {
result.push_back(cp);
if (/* 0x00 <= cpt && */ cpt <= 0x7f) {
result.push_back(cpt);
return result;
}
if (0x80 <= cp && cp <= 0x7ff) {
result.push_back(0xc0 | ((cp >> 6) & 0x1f));
result.push_back(0x80 | (cp & 0x3f));
if (0x80 <= cpt && cpt <= 0x7ff) {
result.push_back(0xc0 | ((cpt >> 6) & 0x1f));
result.push_back(0x80 | (cpt & 0x3f));
return result;
}
if (0x800 <= cp && cp <= 0xffff) {
result.push_back(0xe0 | ((cp >> 12) & 0x0f));
result.push_back(0x80 | ((cp >> 6) & 0x3f));
result.push_back(0x80 | (cp & 0x3f));
if (0x800 <= cpt && cpt <= 0xffff) {
result.push_back(0xe0 | ((cpt >> 12) & 0x0f));
result.push_back(0x80 | ((cpt >> 6) & 0x3f));
result.push_back(0x80 | (cpt & 0x3f));
return result;
}
if (0x10000 <= cp && cp <= 0x10ffff) {
result.push_back(0xf0 | ((cp >> 18) & 0x07));
result.push_back(0x80 | ((cp >> 12) & 0x3f));
result.push_back(0x80 | ((cp >> 6) & 0x3f));
result.push_back(0x80 | (cp & 0x3f));
if (0x10000 <= cpt && cpt <= 0x10ffff) {
result.push_back(0xf0 | ((cpt >> 18) & 0x07));
result.push_back(0x80 | ((cpt >> 12) & 0x3f));
result.push_back(0x80 | ((cpt >> 6) & 0x3f));
result.push_back(0x80 | (cpt & 0x3f));
return result;
}
throw std::invalid_argument("invalid codepoint");
}
std::vector<uint32_t> unicode_cpts_normalize_nfd(const std::vector<uint32_t>& cpts) {
auto comp = [](const uint32_t cpt, const range_nfd& range) {
std::vector<uint32_t> unicode_cpts_normalize_nfd(const std::vector<uint32_t> & cpts) {
auto comp = [] (const uint32_t cpt, const range_nfd & range) {
return cpt < range.first;
};
std::vector<uint32_t> result(cpts.size());
for (size_t i = 0; i < cpts.size(); ++i) {
const uint32_t cpt = cpts[i];
auto it = std::upper_bound(unicode_ranges_nfd.cbegin(), unicode_ranges_nfd.cend(), cpt, comp) - 1;
auto it = std::upper_bound(unicode_ranges_nfd.begin(), unicode_ranges_nfd.end(), cpt, comp) - 1;
result[i] = (it->first <= cpt && cpt <= it->last) ? it->nfd : cpt;
}
return result;
}
std::vector<uint32_t> unicode_cpts_from_utf8(const std::string& utf8) {
std::vector<uint32_t> unicode_cpts_from_utf8(const std::string & utf8) {
std::vector<uint32_t> result;
result.reserve(utf8.size());
size_t offset = 0;
while (offset < utf8.size()) {
result.push_back(unicode_cpt_from_utf8(utf8, offset));
try {
result.push_back(unicode_cpt_from_utf8(utf8, offset));
}
catch (const std::invalid_argument & /*ex*/) {
// Silently ignore invalid UTF-8 input to avoid leaking the exception beyond llama_tokenize
++offset;
result.emplace_back(0xFFFD); // replacement character
}
}
return result;
}
codepoint_flags unicode_cpt_flags(const uint32_t cp) {
static const codepoint_flags undef(codepoint_flags::UNDEFINED);
unicode_cpt_flags unicode_cpt_flags_from_cpt(const uint32_t cpt) {
static const unicode_cpt_flags undef(unicode_cpt_flags::UNDEFINED);
static const auto cpt_flags = unicode_cpt_flags_array();
return cp < cpt_flags.size() ? cpt_flags[cp] : undef;
return cpt < cpt_flags.size() ? cpt_flags[cpt] : undef;
}
codepoint_flags unicode_cpt_flags(const std::string& utf8) {
static const codepoint_flags undef(codepoint_flags::UNDEFINED);
unicode_cpt_flags unicode_cpt_flags_from_utf8(const std::string & utf8) {
static const unicode_cpt_flags undef(unicode_cpt_flags::UNDEFINED);
if (utf8.empty()) {
return undef; // undefined
}
size_t offset = 0;
return unicode_cpt_flags(unicode_cpt_from_utf8(utf8, offset));
return unicode_cpt_flags_from_cpt(unicode_cpt_from_utf8(utf8, offset));
}
std::string unicode_byte_to_utf8(uint8_t byte) {
static std::unordered_map<uint8_t, std::string> map = unicode_byte_to_utf8_map();
return map.at(byte);
}
uint8_t unicode_utf8_to_byte(const std::string & utf8) {
static std::unordered_map<std::string, uint8_t> map = unicode_utf8_to_byte_map();
return map.at(utf8);
}
uint32_t unicode_tolower(uint32_t cpt) {
// binary search
auto it = std::lower_bound(unicode_map_lowercase.begin(), unicode_map_lowercase.end(), cpt,
[](const std::pair<uint32_t, uint32_t> & pair, uint32_t value) {
return pair.first < value;
});
if (it != unicode_map_lowercase.end() && it->first == cpt) {
return it->second;
}
return cpt; // Return the original code point if no lowercase mapping is found
}
bool unicode_cpt_is_han(uint32_t cpt) {
@@ -870,53 +879,37 @@ bool unicode_cpt_is_han(uint32_t cpt) {
return false;
}
std::string unicode_byte_to_utf8(uint8_t byte) {
static std::unordered_map<uint8_t, std::string> map = unicode_byte_to_utf8_map();
return map.at(byte);
}
uint8_t unicode_utf8_to_byte(const std::string& utf8) {
static std::unordered_map<std::string, uint8_t> map = unicode_utf8_to_byte_map();
return map.at(utf8);
}
uint32_t unicode_tolower(uint32_t cp) {
auto it = unicode_map_lowercase.find(cp);
return it == unicode_map_lowercase.end() ? cp : it->second;
}
std::vector<std::string> unicode_regex_split(const std::string& text, const std::vector<std::string>& regex_exprs) {
std::vector<std::string> unicode_regex_split(const std::string & text, const std::vector<std::string> & regex_exprs) {
// unicode categories
static const std::map<std::string, int> k_ucat_enum = {
{ "\\p{N}", codepoint_flags::NUMBER },
{ "\\p{L}", codepoint_flags::LETTER },
{ "\\p{P}", codepoint_flags::PUNCTUATION },
{ "\\p{M}", codepoint_flags::ACCENT_MARK },
{ "\\p{S}", codepoint_flags::SYMBOL },
{ "\\p{N}", unicode_cpt_flags::NUMBER },
{ "\\p{L}", unicode_cpt_flags::LETTER },
{ "\\p{P}", unicode_cpt_flags::PUNCTUATION },
{ "\\p{M}", unicode_cpt_flags::ACCENT_MARK },
{ "\\p{S}", unicode_cpt_flags::SYMBOL },
};
static const std::map<int, int> k_ucat_cpt = {
{ codepoint_flags::NUMBER, 0xD1 },
{ codepoint_flags::LETTER, 0xD2 },
{ codepoint_flags::PUNCTUATION, 0xD3 },
{ codepoint_flags::ACCENT_MARK, 0xD4 },
{ codepoint_flags::SYMBOL, 0xD5 },
{ unicode_cpt_flags::NUMBER, 0xD1 },
{ unicode_cpt_flags::LETTER, 0xD2 },
{ unicode_cpt_flags::PUNCTUATION, 0xD3 },
{ unicode_cpt_flags::ACCENT_MARK, 0xD4 },
{ unicode_cpt_flags::SYMBOL, 0xD5 },
};
static const std::map<int, std::string> k_ucat_map = {
{ codepoint_flags::NUMBER, "\x30-\x39" }, // 0-9
{ codepoint_flags::LETTER, "\x41-\x5A\x61-\x7A" }, // A-Za-z
{ codepoint_flags::PUNCTUATION, "\x21-\x23\x25-\x2A\x2C-\x2F\x3A-\x3B\x3F-\x40\\\x5B-\\\x5D\x5F\\\x7B\\\x7D" }, // !-#%-*,-/:-;?-@\[-\]_\{\}i
{ codepoint_flags::ACCENT_MARK, "" }, // no sub-128 codepoints
{ codepoint_flags::SYMBOL, "\\\x24\\\x2B\x3C-\x3E\x5E\x60\\\x7C" }, // $+<=>^`|
{ unicode_cpt_flags::NUMBER, "\x30-\x39" }, // 0-9
{ unicode_cpt_flags::LETTER, "\x41-\x5A\x61-\x7A" }, // A-Za-z
{ unicode_cpt_flags::PUNCTUATION, "\x21-\x23\x25-\x2A\x2C-\x2F\x3A-\x3B\x3F-\x40\\\x5B-\\\x5D\x5F\\\x7B\\\x7D" }, // !-#%-*,-/:-;?-@\[-\]_\{\}
{ unicode_cpt_flags::ACCENT_MARK, "" }, // no sub-128 codepoints
{ unicode_cpt_flags::SYMBOL, "\\\x24\\\x2B\x3C-\x3E\x5E\x60\\\x7C" }, // $+<=>^`|
};
// compute collapsed codepoints only if needed by at least one regex
bool need_collapse = false;
for (auto& regex_expr : regex_exprs) {
for (const auto & regex_expr : regex_exprs) {
// search for unicode categories
for (const auto& ucat : k_ucat_enum) {
for (const auto & ucat : k_ucat_enum) {
if (std::string::npos != regex_expr.find(ucat.first)) {
need_collapse = true;
break;
@@ -927,7 +920,7 @@ std::vector<std::string> unicode_regex_split(const std::string& text, const std:
const auto cpts = unicode_cpts_from_utf8(text);
// generate a "collapsed" representation of the text, where all codepoints are replaced by a single byte
// ref: https://github.com/ggerganov/llama.cpp/pull/6920#issuecomment-2081479935
// ref: https://github.com/ggml-org/llama.cpp/pull/6920#issuecomment-2081479935
std::string text_collapsed;
if (need_collapse) {
// collapse all unicode categories
@@ -940,25 +933,23 @@ std::vector<std::string> unicode_regex_split(const std::string& text, const std:
continue;
}
const auto flags = unicode_cpt_flags(cpts[i]);
const auto flags = unicode_cpt_flags_from_cpt(cpts[i]);
if (flags.is_whitespace) {
//NOTE: C++ std::regex \s does not mach 0x85, Rust and Python regex does.
//text_collapsed[i] = (char) 0x85; // <Next Line> as whitespace fallback
text_collapsed[i] = (char)0x0B; // <vertical tab> as whitespace fallback
}
else if (k_ucat_cpt.find(flags.category_flag()) != k_ucat_cpt.end()) {
text_collapsed[i] = (char) 0x0B; // <vertical tab> as whitespace fallback
} else if (k_ucat_cpt.find(flags.category_flag()) != k_ucat_cpt.end()) {
text_collapsed[i] = k_ucat_cpt.at(flags.category_flag());
}
else {
text_collapsed[i] = (char)0xD0; // fallback
} else {
text_collapsed[i] = (char) 0xD0; // fallback
}
}
}
std::vector<size_t> bpe_offsets = { cpts.size() };
for (auto& regex_expr : regex_exprs) {
for (const auto & regex_expr : regex_exprs) {
// first, see if we have an efficient custom regex implementation
auto tmp = unicode_regex_split_custom(text, regex_expr, bpe_offsets);
@@ -972,7 +963,7 @@ std::vector<std::string> unicode_regex_split(const std::string& text, const std:
// if a unicode category is used in the regex, we use the collapsed text and replace the unicode category
// with the corresponding collapsed representation
bool use_collapsed = false;
for (auto& ucat : k_ucat_enum) {
for (const auto & ucat : k_ucat_enum) {
if (std::string::npos != regex_expr.find(ucat.first)) {
use_collapsed = true;
break;
@@ -1031,15 +1022,14 @@ std::vector<std::string> unicode_regex_split(const std::string& text, const std:
//printf("text_collapsed: %s\n", text_collapsed.c_str());
//printf("regex_expr_collapsed: %s\n", regex_expr_collapsed.c_str());
bpe_offsets = unicode_regex_split_stl(text_collapsed, regex_expr_collapsed, bpe_offsets);
}
else {
} else {
// no unicode category used, we can use std::wregex directly
const std::wstring wregex_expr = unicode_wstring_from_utf8(regex_expr);
// std::wregex \s does not mach non-ASCII whitespaces, using 0x0B as fallback
std::wstring wtext(cpts.begin(), cpts.end());
for (size_t i = 0; i < wtext.size(); ++i) {
if (wtext[i] > 0x7F && unicode_cpt_flags(wtext[i]).is_whitespace) {
if (wtext[i] > 0x7F && unicode_cpt_flags_from_cpt(wtext[i]).is_whitespace) {
wtext[i] = 0x0B;
}
}
@@ -1048,8 +1038,7 @@ std::vector<std::string> unicode_regex_split(const std::string& text, const std:
//printf("regex_expr: %s\n", regex_expr.c_str());
bpe_offsets = unicode_regex_split_stl(wtext, wregex_expr, bpe_offsets);
}
}
catch (std::regex_error& e) {
} catch (std::regex_error & e) {
fprintf(stderr, "Failed to process regex: '%s'\n", regex_expr.c_str());
fprintf(stderr, "Regex error: %s\n", e.what());
throw std::runtime_error("Failed to process regex");
@@ -1060,7 +1049,7 @@ std::vector<std::string> unicode_regex_split(const std::string& text, const std:
bpe_words.reserve(bpe_offsets.size()); // reserve memory for the approximate size
size_t start = 0;
for (size_t& offset : bpe_offsets) {
for (size_t & offset : bpe_offsets) {
bpe_words.emplace_back();
for (size_t i = start; i < start + offset; ++i) {
bpe_words.back() += unicode_cpt_to_utf8(cpts[i]);