vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-11 17:44:38 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 3

Merge branch 'upstream' into concedo_experimental

Browse source 
# Conflicts:
#	.gitignore
#	CMakeLists.txt
#	flake.lock
#	llama.cpp
This commit is contained in:
Concedo 2024-06-03 14:46:12 +08:00
commit 8b29d5f848
4 changed files with 35 additions and 13 deletions
Download patch file Download diff file Expand all files Collapse all files

22
llama.cpp
View file

@ -2190,8 +2190,7 @@ 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;
@ -4908,23 +4907,19 @@ 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);
}
@ -18638,9 +18633,14 @@ int32_t llama_token_to_piece(const struct llama_model * model, llama_token token
return llama_token_to_piece_old(model, token, buf, length);
}
// 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);