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

ChatOn+Main: Initial go at chaton in main interactive flow

Browse source
This commit is contained in:
HanishKVC 2024-04-22 22:20:16 +05:30
parent 1374a64200
commit 050d329e7e
2 changed files with 38 additions and 3 deletions
Download patch file Download diff file Expand all files Collapse all files

17
common/chaton.hpp
View file

@ -50,3 +50,20 @@ 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());
} }
inline std::string chaton_tmpl_apply(const std::string &tmpl, const std::string &role, const std::string &content) {
std::stringstream ss;
ss << conMeta[tmpl]["global"]["prefix"];
ss << conMeta[tmpl][role]["prefix"] << content << conMeta[tmpl][role]["suffix"];
ss << conMeta[tmpl]["global"]["suffix"];
std::string taggedStr = ss.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];
}
inline std::string chaton_tmpl_part(const std::string &tmpl, const std::string &part) {
return conMeta[tmpl][part];
}

24
examples/main/main.cpp
View file

@ -256,11 +256,14 @@ int main(int argc, char ** argv) {
std::vector<llama_token> embd_inp; std::vector<llama_token> embd_inp;
if (params.interactive_first || params.instruct || params.chatml || !params.prompt.empty() || session_tokens.empty()) { if (params.interactive_first || params.instruct || params.chatml || params.chaton || !params.prompt.empty() || session_tokens.empty()) {
LOG("tokenize the prompt\n"); LOG("tokenize the prompt\n");
if (params.chatml) { if (params.chatml) {
params.prompt = "<|im_start|>system\n" + params.prompt + "<|im_end|>"; 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);
}
embd_inp = ::llama_tokenize(ctx, params.prompt, true, true); embd_inp = ::llama_tokenize(ctx, params.prompt, true, true);
} else { } else {
LOG("use session tokens\n"); LOG("use session tokens\n");
@ -338,7 +341,7 @@ int main(int argc, char ** argv) {
} }
// number of tokens to keep when resetting context // number of tokens to keep when resetting context
if (params.n_keep < 0 || params.n_keep > (int) embd_inp.size() || params.instruct || params.chatml) { if (params.n_keep < 0 || params.n_keep > (int) embd_inp.size() || params.instruct || params.chatml || params.chaton) {
params.n_keep = (int)embd_inp.size(); params.n_keep = (int)embd_inp.size();
} else { } else {
params.n_keep += add_bos; // always keep the BOS token params.n_keep += add_bos; // always keep the BOS token
@ -369,6 +372,16 @@ int main(int argc, char ** argv) {
params.antiprompt.emplace_back("<|im_start|>user\n"); params.antiprompt.emplace_back("<|im_start|>user\n");
} }
// chaton mode
const auto chaton_assitant_prefix = ::llama_tokenize(ctx, chaton_tmpl_role_part(params.chaton_template_id, "assistant", "prefix"), false, true);
if (params.chaton) {
params.interactive = true; // may remove later, by requiring user to explicitly request interactive mode
params.interactive_first = true;
params.input_prefix = chaton_tmpl_role_part(params.chaton_template_id, "user", "prefix");
params.input_suffix = chaton_tmpl_role_part(params.chaton_template_id, "user", "suffix");
params.antiprompt.emplace_back(chaton_tmpl_part(params.chaton_template_id, "reverse-prompt"));
}
// enable interactive mode if interactive start is specified // enable interactive mode if interactive start is specified
if (params.interactive_first) { if (params.interactive_first) {
params.interactive = true; params.interactive = true;
@ -822,7 +835,7 @@ int main(int argc, char ** argv) {
if (n_past > 0 && is_interacting) { if (n_past > 0 && is_interacting) {
LOG("waiting for user input\n"); LOG("waiting for user input\n");
if (params.instruct || params.chatml) { if (params.instruct || params.chatml || params.chaton) {
printf("\n> "); printf("\n> ");
} }
@ -901,6 +914,11 @@ int main(int argc, char ** argv) {
LOG("inserting chatml suffix\n"); LOG("inserting chatml suffix\n");
embd_inp.insert(embd_inp.end(), cml_sfx.begin(), cml_sfx.end()); embd_inp.insert(embd_inp.end(), cml_sfx.begin(), cml_sfx.end());
} }
// chaton mode: insert assistant prefix
if (params.chaton) {
LOG("inserting chaton assistant prefix\n");
embd_inp.insert(embd_inp.end(), chaton_assitant_prefix.begin(), chaton_assitant_prefix.end());
}
for (size_t i = original_size; i < embd_inp.size(); ++i) { for (size_t i = original_size; i < embd_inp.size(); ++i) {
const llama_token token = embd_inp[i]; const llama_token token = embd_inp[i];