vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-10 00:54:41 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

llama : keep track of all EOG tokens in the vocab (#9609)

Browse source 
ggml-ci
This commit is contained in:
Georgi Gerganov 2024-09-24 10:16:06 +03:00 • committed by GitHub
parent cea1486ecf
commit 31ac5834fe
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
3 changed files with 61 additions and 18 deletions
Download patch file Download diff file Expand all files Collapse all files

59
src/llama.cpp
View file

@ -6509,21 +6509,21 @@ static void llm_load_vocab(
// for now, we apply this workaround to find the EOT token based on its text
if (vocab.special_eot_id == -1) {
for (const auto & t : vocab.token_to_id) {
if (
if (false
// TODO: gemma "<end_of_turn>" is exported as a normal token, so the following check does not work
// need to fix convert script
//vocab.id_to_token[t.second].type == LLAMA_TOKEN_TYPE_CONTROL &&
(t.first == "<|eot_id|>" ||
t.first == "<|im_end|>" ||
t.first == "<|end|>" ||
t.first == "<end_of_turn>" ||
t.first == "<|endoftext|>"
)
|| t.first == "<|eot_id|>"
|| t.first == "<|im_end|>"
|| t.first == "<|end|>"
|| t.first == "<end_of_turn>"
|| t.first == "<|endoftext|>"
|| t.first == "<EOT>"
) {
vocab.special_eot_id = t.second;
if ((vocab.id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
LLAMA_LOG_WARN("%s: control-looking token: '%s' was not control-type; this is probably a bug in the model. its type will be overridden\n",
__func__, t.first.c_str());
__func__, t.first.c_str());
vocab.id_to_token[t.second].attr = LLAMA_TOKEN_ATTR_CONTROL;
}
break;
@ -6546,6 +6546,44 @@ static void llm_load_vocab(
}
}
}
// maintain a list of tokens that cause end-of-generation
// this is currently determined based on the token text, which is obviously not ideal
// ref: https://github.com/ggerganov/llama.cpp/issues/9606
vocab.special_eog_ids.clear();
for (const auto & t : vocab.token_to_id) {
if (false
|| t.first == "<|eot_id|>"
|| t.first == "<|im_end|>"
|| t.first == "<|end|>"
|| t.first == "<end_of_turn>"
|| t.first == "<|endoftext|>"
|| t.first == "<|eom_id|>"
|| t.first == "<EOT>"
) {
vocab.special_eog_ids.insert(t.second);
if ((vocab.id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
LLAMA_LOG_WARN("%s: control-looking token: '%s' was not control-type; this is probably a bug in the model. its type will be overridden\n",
__func__, t.first.c_str());
vocab.id_to_token[t.second].attr = LLAMA_TOKEN_ATTR_CONTROL;
}
}
}
if (vocab.special_eos_id != -1 && vocab.special_eog_ids.count(vocab.special_eos_id) == 0) {
vocab.special_eog_ids.insert(vocab.special_eos_id);
LLAMA_LOG_WARN("%s: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect\n", __func__);
}
if (vocab.special_eot_id != -1 && vocab.special_eog_ids.count(vocab.special_eot_id) == 0) {
vocab.special_eog_ids.insert(vocab.special_eot_id);
LLAMA_LOG_WARN("%s: special_eot_id is not in special_eog_ids - the tokenizer config may be incorrect\n", __func__);
}
if (vocab.special_eom_id != -1 && vocab.special_eog_ids.count(vocab.special_eom_id) == 0) {
vocab.special_eog_ids.insert(vocab.special_eom_id);
LLAMA_LOG_WARN("%s: special_eom_id is not in special_eog_ids - the tokenizer config may be incorrect\n", __func__);
}
}
// build special tokens cache
@ -6749,6 +6787,11 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) {
if (vocab.special_suffix_id != -1) { LLAMA_LOG_INFO( "%s: SUF token = %d '%s'\n", __func__, vocab.special_suffix_id, vocab.id_to_token[vocab.special_suffix_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_eom_id != -1) { LLAMA_LOG_INFO( "%s: EOM token = %d '%s'\n", __func__, vocab.special_eom_id, vocab.id_to_token[vocab.special_eom_id].text.c_str() ); }
for (const auto & id : vocab.special_eog_ids) {
LLAMA_LOG_INFO( "%s: EOG token = %d '%s'\n", __func__, id, vocab.id_to_token[id].text.c_str() );
}
LLAMA_LOG_INFO("%s: max token length = %d\n", __func__, vocab.max_token_len);