vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

token healing : handle more special tokens

Browse source 
Infill tokens were being rolled back in certain cases.
This commit is contained in:
mare5x 2024-06-30 20:14:18 +02:00
parent 414fc13248
commit fc8773d309
1 changed files with 10 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

11
common/sampling.cpp
View file

@ -67,7 +67,16 @@ static llama_token_healing_output llama_token_healing_get_prefix(
const llama_model * model = llama_get_model(ctx_main);
auto is_special_token = [&](const llama_token token_id) {
return llama_token_is_control(model, token_id) || llama_token_is_eog(model, token_id);
return llama_token_is_control(model, token_id)
|| llama_token_bos (model) == token_id
|| llama_token_eos (model) == token_id
|| llama_token_cls (model) == token_id
|| llama_token_sep (model) == token_id
|| llama_token_pad (model) == token_id
|| llama_token_prefix (model) == token_id
|| llama_token_middle (model) == token_id
|| llama_token_suffix (model) == token_id
|| llama_token_eot (model) == token_id;
};
if (th_type == llama_token_healing_type::DYNAMIC_ONCE || th_type == llama_token_healing_type::DYNAMIC_MULTI) {