root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-25 10:54:36 +00:00
Code Issues Actions 9 Packages Projects Releases Wiki Activity

llama : optimize long word tokenization with WPM (#8034)

Browse Source 
ggml-ci
This commit is contained in:
Georgi Gerganov 2024-06-21 08:51:28 +03:00 committed by GitHub
parent 80ea089d77
commit a927b0f3dd
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 13 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

17
llama.cpp
View File

@ -2293,6 +2293,8 @@ struct llama_vocab {
enum llama_vocab_type type = LLAMA_VOCAB_TYPE_SPM; enum llama_vocab_type type = LLAMA_VOCAB_TYPE_SPM;
enum llama_vocab_pre_type type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT; enum llama_vocab_pre_type type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
int max_token_len = 0; // used for optimizing longest token search
std::unordered_map<token, id> token_to_id; std::unordered_map<token, id> token_to_id;
std::vector<token_data> id_to_token; std::vector<token_data> id_to_token;
@ -4939,6 +4941,7 @@ static void llm_load_vocab(
GGML_ASSERT(unicode_cpts_from_utf8(word).size() > 0); GGML_ASSERT(unicode_cpts_from_utf8(word).size() > 0);
vocab.token_to_id[word] = i; vocab.token_to_id[word] = i;
vocab.max_token_len = std::max(vocab.max_token_len, (int) word.size());
auto & token_data = vocab.id_to_token[i]; auto & token_data = vocab.id_to_token[i];
token_data.text = std::move(word); token_data.text = std::move(word);
@ -5249,6 +5252,8 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) {
if (vocab.special_middle_id != -1) { LLAMA_LOG_INFO( "%s: MID token = %d '%s'\n", __func__, vocab.special_middle_id, vocab.id_to_token[vocab.special_middle_id].text.c_str() ); } if (vocab.special_middle_id != -1) { LLAMA_LOG_INFO( "%s: MID token = %d '%s'\n", __func__, vocab.special_middle_id, vocab.id_to_token[vocab.special_middle_id].text.c_str() ); }
if (vocab.special_eot_id != -1) { LLAMA_LOG_INFO( "%s: EOT token = %d '%s'\n", __func__, vocab.special_eot_id, vocab.id_to_token[vocab.special_eot_id].text.c_str() ); } if (vocab.special_eot_id != -1) { LLAMA_LOG_INFO( "%s: EOT token = %d '%s'\n", __func__, vocab.special_eot_id, vocab.id_to_token[vocab.special_eot_id].text.c_str() ); }
LLAMA_LOG_INFO("%s: max token length = %d\n", __func__, vocab.max_token_len);
if (model.arch == LLM_ARCH_DEEPSEEK2) { if (model.arch == LLM_ARCH_DEEPSEEK2) {
LLAMA_LOG_INFO("%s: n_layer_dense_lead = %d\n", __func__, hparams.n_layer_dense_lead); LLAMA_LOG_INFO("%s: n_layer_dense_lead = %d\n", __func__, hparams.n_layer_dense_lead);
LLAMA_LOG_INFO("%s: n_lora_q = %d\n", __func__, hparams.n_lora_q); LLAMA_LOG_INFO("%s: n_lora_q = %d\n", __func__, hparams.n_lora_q);
@ -13488,7 +13493,7 @@ private:
struct llm_tokenizer_wpm { struct llm_tokenizer_wpm {
llm_tokenizer_wpm(const llama_vocab & vocab): vocab(vocab) {} llm_tokenizer_wpm(const llama_vocab & vocab): vocab(vocab) {}
void tokenize(const std::string & text, std::vector<llama_vocab::id> & output) { void tokenize(const std::string & text, std::vector<llama_vocab::id> & output) const {
const auto & token_map = vocab.token_to_id; const auto & token_map = vocab.token_to_id;
// normalize and split by whitespace // normalize and split by whitespace
@ -13497,7 +13502,7 @@ struct llm_tokenizer_wpm {
// bos token prepended already // bos token prepended already
// find the longest tokens that form the words // find the longest tokens that form the words
for (const std::string &word : words) { for (const std::string & word : words) {
// skip empty words // skip empty words
if (word.size() == 0) { if (word.size() == 0) {
continue; continue;
@ -13514,7 +13519,7 @@ struct llm_tokenizer_wpm {
for (int i = 0; i < n; ++i) { for (int i = 0; i < n; ++i) {
// loop through possible match length // loop through possible match length
bool match = false; bool match = false;
for (int j = n; j > i; j--) { for (int j = std::min(n, i + vocab.max_token_len + 1); j > i; j--) {
auto it = token_map.find(word1.substr(i, j - i)); auto it = token_map.find(word1.substr(i, j - i));
if (it != token_map.end()) { if (it != token_map.end()) {
output.push_back(it->second); output.push_back(it->second);
@ -13537,7 +13542,8 @@ struct llm_tokenizer_wpm {
} }
} }
std::vector<std::string> preprocess(const std::string & text) { // TODO: reduce string copies by using cpts_offs array
std::vector<std::string> preprocess(const std::string & text) const {
const std::vector<uint32_t> cpts_nfd = unicode_cpts_normalize_nfd(unicode_cpts_from_utf8(text)); const std::vector<uint32_t> cpts_nfd = unicode_cpts_normalize_nfd(unicode_cpts_from_utf8(text));
std::vector<std::string> words(1, ""); std::vector<std::string> words(1, "");
@ -13832,6 +13838,8 @@ static std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab &
output.push_back(vocab.special_cls_id); output.push_back(vocab.special_cls_id);
} }
llm_tokenizer_wpm tokenizer(vocab);
for (const auto & fragment : fragment_buffer) { for (const auto & fragment : fragment_buffer) {
if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) { if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) {
auto raw_text = fragment.raw_text.substr(fragment.offset, fragment.length); auto raw_text = fragment.raw_text.substr(fragment.offset, fragment.length);
@ -13839,7 +13847,6 @@ static std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab &
#ifdef PRETOKENIZERDEBUG #ifdef PRETOKENIZERDEBUG
LLAMA_LOG_WARN("TT: (%ld %ld %ld) '%s'\n", raw_text.length(), fragment.offset, fragment.length, raw_text.c_str()); LLAMA_LOG_WARN("TT: (%ld %ld %ld) '%s'\n", raw_text.length(), fragment.offset, fragment.length, raw_text.c_str());
#endif #endif
llm_tokenizer_wpm tokenizer(vocab);
tokenizer.tokenize(raw_text, output); tokenizer.tokenize(raw_text, output);
} else { // if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN) } else { // if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN)
output.push_back(fragment.token); output.push_back(fragment.token);

1
unicode.cpp
View File

@ -596,6 +596,7 @@ std::vector<uint32_t> unicode_cpts_normalize_nfd(const std::vector<uint32_t> & c
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; std::vector<uint32_t> result;
result.reserve(utf8.size());
size_t offset = 0; size_t offset = 0;
while (offset < utf8.size()) { while (offset < utf8.size()) {
result.push_back(unicode_cpt_from_utf8(utf8, offset)); result.push_back(unicode_cpt_from_utf8(utf8, offset));