Skip to content

llama : avoid double token-to-piece cache #7654

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Jun 3, 2024
Merged

llama : avoid double token-to-piece cache #7654

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
22 changes: 11 additions & 11 deletions llama.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2164,8 +2164,7 @@ struct llama_vocab {
std::vector<token_data> id_to_token;

std::vector<id> cache_special_tokens;
std::vector<token> cache_token_to_piece; // llama_token_to_piece(special = false);
std::vector<token> cache_token_to_piece_special; // llama_token_to_piece(special = true);
std::vector<token> cache_token_to_piece; // llama_token_to_piece(special = true);

std::map<std::pair<std::string, std::string>, int> bpe_ranks;

Expand Down Expand Up @@ -4845,23 +4844,19 @@ static void llm_load_vocab(
LLAMA_LOG_INFO("%s: special tokens cache size = %u\n", __func__, (uint32_t)vocab.cache_special_tokens.size());
}

// build token to piece caches
// build token to piece cache
{
size_t size_cache = 0;

std::vector<llama_vocab::token> cache_token_to_piece (n_vocab);
std::vector<llama_vocab::token> cache_token_to_piece_special(n_vocab);
std::vector<llama_vocab::token> cache_token_to_piece(n_vocab);

for (uint32_t id = 0; id < n_vocab; ++id) {
cache_token_to_piece[id] = llama_token_to_piece(&model, id, false);
cache_token_to_piece_special[id] = llama_token_to_piece(&model, id, true);
cache_token_to_piece[id] = llama_token_to_piece(&model, id, true);

size_cache += cache_token_to_piece[id].size();
size_cache += cache_token_to_piece_special[id].size();
}

std::swap(vocab.cache_token_to_piece, cache_token_to_piece);
std::swap(vocab.cache_token_to_piece_special, cache_token_to_piece_special);
std::swap(vocab.cache_token_to_piece, cache_token_to_piece);

LLAMA_LOG_INFO("%s: token to piece cache size = %.4f MB\n", __func__, size_cache / 1024.0 / 1024.0);
}
Expand Down Expand Up @@ -18313,9 +18308,14 @@ static std::string llama_decode_text(const std::string & text) {

// does not write null-terminator to buf
int32_t llama_token_to_piece(const struct llama_model * model, llama_token token, char * buf, int32_t length, bool special) {
// ref: https://github.com/ggerganov/llama.cpp/pull/7587#discussion_r1620983843
if (!special && llama_is_control_token(model->vocab, token)) {
return 0;
}

// if we have a cache - use it
{
const auto & cache = special ? model->vocab.cache_token_to_piece_special : model->vocab.cache_token_to_piece;
const auto & cache = model->vocab.cache_token_to_piece;

if (!cache.empty()) {
const auto & res = cache.at(token);
Expand Down