Skip to content

Commit

Permalink
llama : refactor unicode stuff (ggerganov#5992)
Browse files Browse the repository at this point in the history 
* llama : refactor unicode stuff

ggml-ci

* unicode : names

* make : fix c++ compiler

* unicode : names

* unicode : straighten tables

* zig : fix build

* unicode : put nfd normalization behind API

ggml-ci

* swift : fix build

* unicode : add BOM

* unicode : add <cstdint>

ggml-ci

* unicode : pass as cpts as const ref
  • Loading branch information
@ggerganov @NeoZhangJianyu
ggerganov authored and NeoZhangJianyu committed Mar 12, 2024
1 parent 0a8db3f commit d0b01ef
Show file tree
Hide file tree
Showing 9 changed files with 1,744 additions and 836 deletions.
2 changes: 2 additions & 0 deletions CMakeLists.txt
View file
Expand Up @@ -1141,6 +1141,8 @@ endif()
add_library(llama
llama.cpp
llama.h
unicode.h
unicode.cpp
)

target_include_directories(llama PUBLIC .)
Expand Down
7 changes: 5 additions & 2 deletions Makefile
View file
Expand Up @@ -633,9 +633,12 @@ ggml-backend.o: ggml-backend.c ggml.h ggml-backend.h
ggml-quants.o: ggml-quants.c ggml.h ggml-quants.h ggml-common.h
$(CC) $(CFLAGS) -c $< -o $@

OBJS += ggml-alloc.o ggml-backend.o ggml-quants.o
unicode.o: unicode.cpp unicode.h
$(CXX) $(CXXFLAGS) -c $< -o $@

OBJS += ggml-alloc.o ggml-backend.o ggml-quants.o unicode.o

llama.o: llama.cpp ggml.h ggml-alloc.h ggml-backend.h ggml-cuda.h ggml-metal.h llama.h
llama.o: llama.cpp unicode.h ggml.h ggml-alloc.h ggml-backend.h ggml-cuda.h ggml-metal.h llama.h
$(CXX) $(CXXFLAGS) -c $< -o $@

COMMON_H_DEPS = common/common.h common/sampling.h common/log.h
Expand Down
1 change: 1 addition & 0 deletions Package.swift
View file
Expand Up @@ -31,6 +31,7 @@ let package = Package(
sources: [
"ggml.c",
"llama.cpp",
"unicode.cpp",
"ggml-alloc.c",
"ggml-backend.c",
"ggml-quants.c",
Expand Down
15 changes: 8 additions & 7 deletions build.zig
View file
Expand Up @@ -115,6 +115,7 @@ pub fn build(b: *std.build.Builder) !void {
const ggml_alloc = make.obj("ggml-alloc", "ggml-alloc.c");
const ggml_backend = make.obj("ggml-backend", "ggml-backend.c");
const ggml_quants = make.obj("ggml-quants", "ggml-quants.c");
const unicode = make.obj("unicode", "unicode.cpp");
const llama = make.obj("llama", "llama.cpp");
const buildinfo = make.obj("common", "common/build-info.cpp");
const common = make.obj("common", "common/common.cpp");
Expand All @@ -125,14 +126,14 @@ pub fn build(b: *std.build.Builder) !void {
const clip = make.obj("clip", "examples/llava/clip.cpp");
const llava = make.obj("llava", "examples/llava/llava.cpp");

_ = make.exe("main", "examples/main/main.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, sampling, console, grammar_parser });
_ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo });
_ = make.exe("perplexity", "examples/perplexity/perplexity.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo });
_ = make.exe("embedding", "examples/embedding/embedding.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo });
_ = make.exe("finetune", "examples/finetune/finetune.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, train });
_ = make.exe("train-text-from-scratch", "examples/train-text-from-scratch/train-text-from-scratch.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, train });
_ = make.exe("main", "examples/main/main.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, common, buildinfo, sampling, console, grammar_parser });
_ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, common, buildinfo });
_ = make.exe("perplexity", "examples/perplexity/perplexity.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, common, buildinfo });
_ = make.exe("embedding", "examples/embedding/embedding.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, common, buildinfo });
_ = make.exe("finetune", "examples/finetune/finetune.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, common, buildinfo, train });
_ = make.exe("train-text-from-scratch", "examples/train-text-from-scratch/train-text-from-scratch.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, common, buildinfo, train });

const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, sampling, grammar_parser, clip, llava });
const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, common, buildinfo, sampling, grammar_parser, clip, llava });
if (server.target.isWindows()) {
server.linkSystemLibrary("ws2_32");
}
Expand Down
87 changes: 37 additions & 50 deletions llama.cpp
View file
Expand Up @@ -3703,7 +3703,7 @@ static void llm_load_vocab(

for (int i = 0; i < n_merges; i++) {
const std::string word = gguf_get_arr_str(ctx, merges_keyidx, i);
GGML_ASSERT(codepoints_from_utf8(word).size() > 0);
GGML_ASSERT(unicode_cpts_from_utf8(word).size() > 0);

std::string first;
std::string second;
Expand Down Expand Up @@ -3748,7 +3748,7 @@ static void llm_load_vocab(

for (uint32_t i = 0; i < n_vocab; i++) {
std::string word = gguf_get_arr_str(ctx, token_idx, i);
GGML_ASSERT(codepoints_from_utf8(word).size() > 0);
GGML_ASSERT(unicode_cpts_from_utf8(word).size() > 0);

vocab.token_to_id[word] = i;

Expand Down Expand Up @@ -9350,7 +9350,7 @@ static uint8_t llama_token_to_byte(const llama_vocab& vocab, llama_token id) {
}
case LLAMA_VOCAB_TYPE_BPE: {
GGML_ASSERT(false);
return unicode_to_bytes_bpe(token_data.text);
return unicode_utf8_to_byte(token_data.text);
}
case LLAMA_VOCAB_TYPE_WPM: {
GGML_ASSERT(false);
Expand All @@ -9375,7 +9375,7 @@ static llama_token llama_byte_to_token(const llama_vocab & vocab, uint8_t ch) {
}
case LLAMA_VOCAB_TYPE_WPM:
case LLAMA_VOCAB_TYPE_BPE: {
return vocab.token_to_id.at(bytes_to_unicode_bpe(ch));
return vocab.token_to_id.at(unicode_byte_to_utf8(ch));
}
default:
GGML_ASSERT(false);
Expand Down Expand Up @@ -9715,9 +9715,9 @@ struct llm_tokenizer_bpe {
bpe_words.reserve(text.size());
bpe_encoded_words.reserve(text.size());

auto cps = codepoints_from_utf8(text);
for (size_t i = 0; i < cps.size(); ++i)
text_utf.emplace_back(codepoint_to_utf8(cps[i]));
const auto cpts = unicode_cpts_from_utf8(text);
for (size_t i = 0; i < cpts.size(); ++i)
text_utf.emplace_back(unicode_cpt_to_utf8(cpts[i]));

for (int i = 0; i < (int)text_utf.size(); i++) {
const std::string & utf_char = text_utf[i];
Expand Down Expand Up @@ -9767,40 +9767,40 @@ struct llm_tokenizer_bpe {
}

if (!split_condition && !collecting) {
if (codepoint_type(utf_char) == CODEPOINT_TYPE_LETTER || (!token.size() && utf_char == " " && codepoint_type(utf_char_next) == CODEPOINT_TYPE_LETTER)) {
if (unicode_cpt_type(utf_char) == CODEPOINT_TYPE_LETTER || (!token.size() && utf_char == " " && unicode_cpt_type(utf_char_next) == CODEPOINT_TYPE_LETTER)) {
collecting_letter = true;
collecting = true;
}
else if (codepoint_type(utf_char) == CODEPOINT_TYPE_DIGIT || (!token.size() && utf_char == " " && codepoint_type(utf_char_next) == CODEPOINT_TYPE_DIGIT)) {
else if (unicode_cpt_type(utf_char) == CODEPOINT_TYPE_DIGIT || (!token.size() && utf_char == " " && unicode_cpt_type(utf_char_next) == CODEPOINT_TYPE_DIGIT)) {
collecting_numeric = true;
collecting = true;
}
else if (
((codepoint_type(utf_char) != CODEPOINT_TYPE_LETTER && codepoint_type(utf_char) != CODEPOINT_TYPE_DIGIT) && (codepoint_type(utf_char) != CODEPOINT_TYPE_WHITESPACE)) ||
(!token.size() && utf_char == " " && codepoint_type(utf_char_next) != CODEPOINT_TYPE_LETTER && codepoint_type(utf_char_next) != CODEPOINT_TYPE_DIGIT && codepoint_type(utf_char_next) != CODEPOINT_TYPE_WHITESPACE)
((unicode_cpt_type(utf_char) != CODEPOINT_TYPE_LETTER && unicode_cpt_type(utf_char) != CODEPOINT_TYPE_DIGIT) && (unicode_cpt_type(utf_char) != CODEPOINT_TYPE_WHITESPACE)) ||
(!token.size() && utf_char == " " && unicode_cpt_type(utf_char_next) != CODEPOINT_TYPE_LETTER && unicode_cpt_type(utf_char_next) != CODEPOINT_TYPE_DIGIT && unicode_cpt_type(utf_char_next) != CODEPOINT_TYPE_WHITESPACE)
) {
collecting_special = true;
collecting = true;
}
else if (codepoint_type(utf_char) == CODEPOINT_TYPE_WHITESPACE && codepoint_type(utf_char_next) == CODEPOINT_TYPE_WHITESPACE) {
else if (unicode_cpt_type(utf_char) == CODEPOINT_TYPE_WHITESPACE && unicode_cpt_type(utf_char_next) == CODEPOINT_TYPE_WHITESPACE) {
collecting_whitespace_lookahead = true;
collecting = true;
}
else if (codepoint_type(utf_char) == CODEPOINT_TYPE_WHITESPACE) {
else if (unicode_cpt_type(utf_char) == CODEPOINT_TYPE_WHITESPACE) {
split_condition = true;
}
}
else if (!split_condition && collecting) {
if (collecting_letter && codepoint_type(utf_char) != CODEPOINT_TYPE_LETTER) {
if (collecting_letter && unicode_cpt_type(utf_char) != CODEPOINT_TYPE_LETTER) {
split_condition = true;
}
else if (collecting_numeric && codepoint_type(utf_char) != CODEPOINT_TYPE_DIGIT) {
else if (collecting_numeric && unicode_cpt_type(utf_char) != CODEPOINT_TYPE_DIGIT) {
split_condition = true;
}
else if (collecting_special && (codepoint_type(utf_char) == CODEPOINT_TYPE_LETTER || codepoint_type(utf_char) == CODEPOINT_TYPE_DIGIT || codepoint_type(utf_char) == CODEPOINT_TYPE_WHITESPACE)) {
else if (collecting_special && (unicode_cpt_type(utf_char) == CODEPOINT_TYPE_LETTER || unicode_cpt_type(utf_char) == CODEPOINT_TYPE_DIGIT || unicode_cpt_type(utf_char) == CODEPOINT_TYPE_WHITESPACE)) {
split_condition = true;
}
else if (collecting_whitespace_lookahead && (codepoint_type(utf_char_next) == CODEPOINT_TYPE_LETTER || codepoint_type(utf_char_next) == CODEPOINT_TYPE_DIGIT)) {
else if (collecting_whitespace_lookahead && (unicode_cpt_type(utf_char_next) == CODEPOINT_TYPE_LETTER || unicode_cpt_type(utf_char_next) == CODEPOINT_TYPE_DIGIT)) {
split_condition = true;
}
}
Expand Down Expand Up @@ -9829,7 +9829,7 @@ struct llm_tokenizer_bpe {
for (std::string & word : bpe_words) {
std::string encoded_token = "";
for (char & c : word) {
encoded_token += bytes_to_unicode_bpe(c);
encoded_token += unicode_byte_to_utf8(c);
}
bpe_encoded_words.emplace_back(encoded_token);
}
Expand Down Expand Up @@ -9903,33 +9903,21 @@ struct llm_tokenizer_wpm {
}

std::vector<std::string> preprocess(const std::string & text) {
// normalalization form D
std::vector<uint32_t> codepoints = codepoints_from_utf8(text);
std::vector<uint32_t> nfd_codepoints;
for (uint32_t code : codepoints) {
auto it = nfd_map.equal_range(code);
if (it.first != it.second) {
for (auto jt = it.first; jt != it.second; jt++) {
nfd_codepoints.push_back(jt->second);
}
} else {
nfd_codepoints.push_back(code);
}
}
std::vector<uint32_t> cpts_nfd = unicode_cpts_normalize_nfd(unicode_cpts_from_utf8(text));

// strip accents, strip control, uniformize whitespace,
// to lowercase, pad chinese characters, pad punctuation
std::string new_str = "";
for (uint32_t code : nfd_codepoints) {
int type = codepoint_type(code);
for (uint32_t code : cpts_nfd) {
int type = unicode_cpt_type(code);
if (type == CODEPOINT_TYPE_ACCENT_MARK || type == CODEPOINT_TYPE_CONTROL) {
continue;
}
code = to_lower(code);
if (type == CODEPOINT_TYPE_WHITESPACE) {
code = ' ';
}
std::string s = codepoint_to_utf8(code);
std::string s = unicode_cpt_to_utf8(code);
if (type == CODEPOINT_TYPE_PUNCTUATION || is_ascii_punct(code) || is_chinese_char(code)) {
new_str += " ";
new_str += s;
Expand All @@ -9949,8 +9937,7 @@ struct llm_tokenizer_wpm {
if (r > l) words.push_back(new_str.substr(l, (r - l)));
l = r + 1;
r = l;
}
else {
} else {
r += 1;
}
}
Expand All @@ -9974,17 +9961,17 @@ struct llm_tokenizer_wpm {
return code < 256 && ispunct(code);
}

bool is_chinese_char(uint32_t codepoint) {
if ((codepoint >= 0x4E00 && codepoint <= 0x9FFF) ||
(codepoint >= 0x3400 && codepoint <= 0x4DBF) ||
(codepoint >= 0x20000 && codepoint <= 0x2A6DF) ||
(codepoint >= 0x2A700 && codepoint <= 0x2B73F) ||
(codepoint >= 0x2B740 && codepoint <= 0x2B81F) ||
(codepoint >= 0x2B920 && codepoint <= 0x2CEAF) || // this should be 0x2B820 but in hf rust code it is 0x2B920
(codepoint >= 0xF900 && codepoint <= 0xFAFF) ||
(codepoint >= 0x2F800 && codepoint <= 0x2FA1F) ||
(codepoint >= 0x3000 && codepoint <= 0x303F) ||
(codepoint >= 0xFF00 && codepoint <= 0xFFEF)) {
bool is_chinese_char(uint32_t cpt) {
if ((cpt >= 0x4E00 && cpt <= 0x9FFF) ||
(cpt >= 0x3400 && cpt <= 0x4DBF) ||
(cpt >= 0x20000 && cpt <= 0x2A6DF) ||
(cpt >= 0x2A700 && cpt <= 0x2B73F) ||
(cpt >= 0x2B740 && cpt <= 0x2B81F) ||
(cpt >= 0x2B920 && cpt <= 0x2CEAF) || // this should be 0x2B820 but in hf rust code it is 0x2B920
(cpt >= 0xF900 && cpt <= 0xFAFF) ||
(cpt >= 0x2F800 && cpt <= 0x2FA1F) ||
(cpt >= 0x3000 && cpt <= 0x303F) ||
(cpt >= 0xFF00 && cpt <= 0xFFEF)) {
return true; // NOLINT
}
return false;
Expand Down Expand Up @@ -13966,9 +13953,9 @@ int32_t llama_tokenize(

static std::string llama_decode_text(const std::string & text) {
std::string decoded_text;
auto unicode_sequences = codepoints_from_utf8(text);
for (auto& unicode_sequence : unicode_sequences) {
decoded_text += unicode_to_bytes_bpe(codepoint_to_utf8(unicode_sequence));
auto unicode_sequences = unicode_cpts_from_utf8(text);
for (auto & unicode_sequence : unicode_sequences) {
decoded_text += unicode_utf8_to_byte(unicode_cpt_to_utf8(unicode_sequence));
}

return decoded_text;
Expand Down
4 changes: 2 additions & 2 deletions tests/test-tokenizer-1-bpe.cpp
View file
Expand Up @@ -64,7 +64,7 @@ int main(int argc, char **argv) {
for (int i = 0; i < n_vocab; ++i) {
std::string str = llama_detokenize_bpe(ctx, std::vector<int>(1, i));
try {
auto cps = codepoints_from_utf8(str);
auto cps = unicode_cpts_from_utf8(str);
std::vector<llama_token> tokens = llama_tokenize(ctx, str, false);
std::string check = llama_detokenize_bpe(ctx, tokens);
if (check != str) {
Expand Down Expand Up @@ -97,7 +97,7 @@ int main(int argc, char **argv) {
continue;
}

std::string str = codepoint_to_utf8(cp);
std::string str = unicode_cpt_to_utf8(cp);
std::vector<llama_token> tokens = llama_tokenize(ctx, str, false);
std::string check = llama_detokenize_bpe(ctx, tokens);
if (cp != 9601 && str != check) {
Expand Down
2 changes: 1 addition & 1 deletion tests/test-tokenizer-1-llama.cpp
View file
Expand Up @@ -85,7 +85,7 @@ int main(int argc, char **argv) {
continue;
}

std::string str = codepoint_to_utf8(cp);
std::string str = unicode_cpt_to_utf8(cp);
std::vector<llama_token> tokens = llama_tokenize(ctx, str, false);
std::string check = llama_detokenize_spm(ctx, tokens);
if (cp != 9601 && str != check) {
Expand Down

0 comments on commit d0b01ef

Please sign in to comment.