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

ChatOn+Main: Cleanup the Requested ChatOn ReversePrompt handling

Browse source 
Avoid the use of the seperate vector, which inturn is copied to
the main vector on return. Now directly pass the main reverse
prompt vector and inturn directly add to passed vector.

Also keep data and return status seperate. Explicitly identify
a unknown template_id situation and return failure status.
This commit is contained in:
HanishKVC 2024-04-20 23:26:16 +05:30
parent ca55da2b6f
commit e23b5c8689
2 changed files with 11 additions and 14 deletions
Download patch file Download diff file Expand all files Collapse all files

18
common/chaton.hpp
View file

@ -36,16 +36,18 @@ inline bool llama_chat_apply_template_simple(
}
// return what should be the reverse prompt for the given template id
// ie possible end text tag(s) of specified model type's chat query response
inline std::vector<std::string> llama_chat_reverse_prompt(std::string &template_id) {
std::vector<std::string> rends;
// ie possible end text tag(s) of specified model type's chat query response.
// Note that It adds these reverse prompts to any that may already exist in the passed vector.
inline bool llama_chat_reverse_prompt(std::string &template_id, std::vector<std::string> &rprompts) {
if (template_id == "chatml") {
rends.push_back("<|im_start|>user\n");
rprompts.push_back("<|im_start|>user\n");
} else if (template_id == "llama2") {
rends.push_back("</s>");
rprompts.push_back("</s>");
} else if (template_id == "llama3") {
rends.push_back("<|eot_id|>");
rprompts.push_back("<|eot_id|>");
} else {
LOG_TEELN("WARN:%s:Unknown template [%s] requested", __func__, template_id.c_str());
return false;
}
return rends;
return true;
}

7
examples/main/main.cpp
View file

@ -373,15 +373,10 @@ int main(int argc, char ** argv) {
// handle chaton mode, it adds on to any reverse prompt specified explicitly by the user
if (params.chaton) {
params.interactive_first = true;
std::vector<std::string> resp_ends = llama_chat_reverse_prompt(params.chaton_template_id);
if (resp_ends.size() == 0) {
if (!llama_chat_reverse_prompt(params.chaton_template_id, params.antiprompt)) {
LOG_TEELN("ERRR:%s:ChatOn:Unsupported ChatTemplateType:%s", __func__, params.chaton_template_id.c_str());
exit(1);
}
for (size_t i = 0; i < resp_ends.size(); i++)
{
params.antiprompt.emplace_back(resp_ends[i]);
}
}
// enable interactive mode if interactive start is specified