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

ChatOn:No global-begin/end in ChatApplyTmplSingle, ChatApplyTmpl

Browse source 
Avoid adding global begin/end markers wrt ChatApplyTmplSingle.

Add ChatApplyTmpl which goes through a vector of messages.
This commit is contained in:
HanishKVC 2024-04-22 23:58:19 +05:30
parent c4cf0e9075
commit f03dd2439f
1 changed files with 18 additions and 4 deletions
Download patch file Download diff file Expand all files Collapse all files

22
common/chaton.hpp
View file

@ -17,8 +17,8 @@
* a. the system prompt they specify using -f, is used as is with parse_special when tokenising or * a. the system prompt they specify using -f, is used as is with parse_special when tokenising or
* b. whether the system prefix and suffix is added, but without parse_special tokenisation of system-prompt provided by user. * b. whether the system prefix and suffix is added, but without parse_special tokenisation of system-prompt provided by user.
* 3. chat-apply-template uses the json file, which was loaded, to decide on how to generate the tagged messages for tokenisation * 3. chat-apply-template uses the json file, which was loaded, to decide on how to generate the tagged messages for tokenisation
* a. input: [ { role: message }, { role: message}, ....] * a. input: [ { role, message }, { role, message}, ....]
* b. output: [ {flag: data}, { flag: data}, {flag: data}, ....] * b. output: [ {flag, data}, { flag, data}, {flag, data}, ....]
* * flag is whether to do parse_special for this data, during tokenization or not * * flag is whether to do parse_special for this data, during tokenization or not
* *
*/ */
@ -55,16 +55,30 @@ inline void chaton_meta_dump() {
LOG_TEELN("\n\nINFO:%s:ChatOn Meta\n%s", __func__, conMeta.dump(4).c_str()); LOG_TEELN("\n\nINFO:%s:ChatOn Meta\n%s", __func__, conMeta.dump(4).c_str());
} }
// NOTE: This currently doesnt return about which parts of the tagged message contain tags and which parts the user message
inline std::string chaton_tmpl_apply_single(const std::string &tmpl, const std::string &role, const std::string &content) { inline std::string chaton_tmpl_apply_single(const std::string &tmpl, const std::string &role, const std::string &content) {
std::stringstream ss; std::stringstream ss;
ss << conMeta[tmpl]["global"]["begin"];
ss << conMeta[tmpl][role]["prefix"] << content << conMeta[tmpl][role]["suffix"]; ss << conMeta[tmpl][role]["prefix"] << content << conMeta[tmpl][role]["suffix"];
ss << conMeta[tmpl]["global"]["end"];
std::string taggedStr = ss.str(); std::string taggedStr = ss.str();
LOG_TEELN("DBUG:%s:%s:%s:%s", __func__, tmpl.c_str(), role.c_str(), taggedStr.c_str()); LOG_TEELN("DBUG:%s:%s:%s:%s", __func__, tmpl.c_str(), role.c_str(), taggedStr.c_str());
return taggedStr; return taggedStr;
} }
// NOTE: This currently doesnt return about which parts of the tagged message contain tags and which parts the user message
inline std::string chaton_tmpl_apply(const std::string &tmpl, const std::vector<llama_chat_message> &msgs) {
std::stringstream ss;
ss << conMeta[tmpl]["global"]["begin"];
for(auto msg: msgs) {
auto role = msg.role;
auto content = msg.content;
ss << conMeta[tmpl][role]["prefix"] << content << conMeta[tmpl][role]["suffix"];
}
ss << conMeta[tmpl]["global"]["end"];
std::string taggedMsgs = ss.str();
LOG_TEELN("DBUG:%s:%s:%s", __func__, tmpl.c_str(), taggedMsgs.c_str());
return taggedMsgs;
}
inline std::string chaton_tmpl_role_part(const std::string &tmpl, const std::string &role, const std::string &part) { inline std::string chaton_tmpl_role_part(const std::string &tmpl, const std::string &role, const std::string &part) {
std::string got = conMeta[tmpl][role][part]; std::string got = conMeta[tmpl][role][part];
LOG_TEELN("DBUG:%s:%s:%s:%s:%s", __func__, tmpl.c_str(), role.c_str(), part.c_str(), got.c_str()); LOG_TEELN("DBUG:%s:%s:%s:%s:%s", __func__, tmpl.c_str(), role.c_str(), part.c_str(), got.c_str());