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

ChatON:Cleanup: BeginEnd, Debug log

Browse source 
Update the note

Rename global-prefix|suffix to global-begin|end.

Rename chat-apply-template to chat-apply-template-single, cas it
handles only a single message.

Add some debug log messages to the helper functions
This commit is contained in:
HanishKVC 2024-04-22 23:48:24 +05:30
parent d87d27512e
commit c4cf0e9075
3 changed files with 26 additions and 17 deletions
Download patch file Download diff file Expand all files Collapse all files

29
common/chaton.hpp
View file

@ -1,14 +1,18 @@
#pragma once
/***
* Keep chatting with model and needed role tagging using special tokens simple and flexible, while building on existing interactive flow
* Keep chatting with model and needed role tagging using special tokens simple and flexible,
* while building on existing interactive flow and its in-prefix, in-suffix and antiprompt/reverse-promot
*
* 1. Use a json file to configure the needed tags for each of the supported chat-handshake-template-standard
* a. system-prefix, system-suffix,
* b. user-prefix, user-suffix,
* b. user-prefix, user-suffix, assistant-prefix
* * these override the in-prefix and in-suffix
* c. reverse-prompt
* d. global-begin-marker, global-end-marker
* e. per-msg-begin-marker, per-msg-end-marker
* f. is per-msg-begin-marker used for system+user combo
* d. Later if required look at adding
* * global-begin-marker, global-end-marker
* * per-msg-begin-marker, per-msg-end-marker
* * is system-per-msg-end-marker and user-per-msg-begin-marker used for system+user combo
* 2. Give the below option to user wrt system prompt, this should give the flexibility to either keep system prompt simple or complex in a flexible yet simple way.
* 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.
@ -51,19 +55,24 @@ inline void chaton_meta_dump() {
LOG_TEELN("\n\nINFO:%s:ChatOn Meta\n%s", __func__, conMeta.dump(4).c_str());
}
inline std::string chaton_tmpl_apply(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;
ss << conMeta[tmpl]["global"]["prefix"];
ss << conMeta[tmpl]["global"]["begin"];
ss << conMeta[tmpl][role]["prefix"] << content << conMeta[tmpl][role]["suffix"];
ss << conMeta[tmpl]["global"]["suffix"];
ss << conMeta[tmpl]["global"]["end"];
std::string taggedStr = ss.str();
LOG_TEELN("DBUG:%s:%s:%s:%s", __func__, tmpl.c_str(), role.c_str(), taggedStr.c_str());
return taggedStr;
}
inline std::string chaton_tmpl_role_part(const std::string &tmpl, const std::string &role, const std::string &part) {
return 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());
return got;
}
inline std::string chaton_tmpl_part(const std::string &tmpl, const std::string &part) {
return conMeta[tmpl][part];
std::string got = conMeta[tmpl][part];
LOG_TEELN("DBUG:%s:%s:%s:%s", __func__, tmpl.c_str(), part.c_str(), got.c_str());
return got;
}

12
examples/chaton_meta.json
View file

@ -2,8 +2,8 @@
{
"llama2": {
"global": {
"prefix": "",
"suffix": ""
"begin": "",
"end": ""
},
"system": {
"prefix": "[INST] <<SYS>>\n",
@ -21,8 +21,8 @@
},
"llama3": {
"global": {
"prefix": "",
"suffix": ""
"begin": "",
"end": ""
},
"system": {
"prefix": "<|start_header_id|>system<|end_header_id|>\n",
@ -40,8 +40,8 @@
},
"chatml": {
"global": {
"prefix": "",
"suffix": ""
"begin": "",
"end": ""
},
"system": {
"prefix": "<|im_start|>system\n",

2
examples/main/main.cpp
View file

@ -262,7 +262,7 @@ int main(int argc, char ** argv) {
params.prompt = "<|im_start|>system\n" + params.prompt + "<|im_end|>";
}
if (params.chaton) {
params.prompt = chaton_tmpl_apply(params.chaton_template_id, "system", params.prompt);
params.prompt = chaton_tmpl_apply_single(params.chaton_template_id, "system", params.prompt);
}
embd_inp = ::llama_tokenize(ctx, params.prompt, true, true);
} else {