Mirrors/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2025-01-12 21:37:19 +01:00
Code Issues Packages Projects Releases Wiki Activity

vocab : minor tokenization optimizations (#11160)

Browse Source 
ggml-ci

Co-authored-by: Diego Devesa <slarengh@gmail.com>
This commit is contained in:
Georgi Gerganov 2025-01-09 17:12:54 +02:00
parent a857dc50af
commit aeeb9420a3
1 changed files with 19 additions and 17 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

36
src/llama-vocab.cpp
View File

@ -2345,19 +2345,21 @@ std::vector<llama_token> llama_vocab::impl::tokenize(
for (const auto & fragment : fragment_buffer) {
if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) {
auto raw_text = fragment.raw_text.substr(fragment.offset, fragment.length);
std::string text;
// prefix with space if previous is special
if (tokenizer_add_space_prefix && is_prev_special) {
raw_text = " " + raw_text;
text = ' ';
}
text += fragment.raw_text.substr(fragment.offset, fragment.length);
#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", text.length(), fragment.offset, fragment.length, text.c_str());
#endif
llama_escape_whitespace(raw_text);
llama_escape_whitespace(text);
llm_tokenizer_spm_session session(vocab);
session.tokenize(raw_text, output);
session.tokenize(text, output);
is_prev_special = false;
} else { // if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN)
output.push_back(fragment.token);
@ -2387,12 +2389,12 @@ std::vector<llama_token> llama_vocab::impl::tokenize(
}
for (const auto & fragment : fragment_buffer) {
if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) {
auto raw_text = fragment.raw_text.substr(fragment.offset, fragment.length);
std::string text = fragment.raw_text.substr(fragment.offset, fragment.length);
#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", text.length(), fragment.offset, fragment.length, text.c_str());
#endif
session.tokenize(raw_text, output);
session.tokenize(text, output);
} else { // if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN)
session.append(fragment.token, output);
}
@ -2414,12 +2416,12 @@ std::vector<llama_token> llama_vocab::impl::tokenize(
for (const auto & fragment : fragment_buffer) {
if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) {
auto raw_text = fragment.raw_text.substr(fragment.offset, fragment.length);
std::string text = fragment.raw_text.substr(fragment.offset, fragment.length);
#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", text.length(), fragment.offset, fragment.length, text.c_str());
#endif
session.tokenize(raw_text, output);
session.tokenize(text, output);
} else { // if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN)
output.push_back(fragment.token);
}
@ -2440,11 +2442,11 @@ std::vector<llama_token> llama_vocab::impl::tokenize(
for (const auto & fragment : fragment_buffer) {
if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) {
auto raw_text = fragment.raw_text.substr(fragment.offset, fragment.length);
std::string text = fragment.raw_text.substr(fragment.offset, fragment.length);
#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", text.length(), fragment.offset, fragment.length, text.c_str());
#endif
session.tokenize(raw_text, output);
session.tokenize(text, output);
} else { // if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN)
output.push_back(fragment.token);
}
@ -2467,13 +2469,13 @@ std::vector<llama_token> llama_vocab::impl::tokenize(
llm_tokenizer_rwkv_session session(vocab, *static_cast<const llm_tokenizer_rwkv *>(tokenizer.get()));
for (const auto & fragment : fragment_buffer) {
if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) {
auto raw_text = fragment.raw_text.substr(fragment.offset, fragment.length);
std::string text = fragment.raw_text.substr(fragment.offset, fragment.length);
#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", text.length(), fragment.offset, fragment.length, text.c_str());
#endif
session.tokenize(raw_text, output);
session.tokenize(text, output);
} else { // if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN)
output.push_back(fragment.token);
}