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

Add Gemma chat template (#5665)

Browse source 
* add gemma chat template

* gemma: only apply system_prompt on non-model message
This commit is contained in:
Xuan Son Nguyen 2024-02-22 19:10:21 +01:00 • committed by GitHub
parent 4cb4d8b22d
commit 373ee3fbba
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
2 changed files with 26 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

22
llama.cpp
View file

@ -12782,6 +12782,28 @@ static int32_t llama_chat_apply_template_internal(
if (add_ass) {
ss << "<s>assistant\n";
}
} else if (tmpl.find("<start_of_turn>") != std::string::npos) {
// google/gemma-7b-it
std::string system_prompt = "";
for (auto message : chat) {
std::string role(message->role);
if (role == "system") {
// there is no system message for gemma, but we will merge it with user prompt, so nothing is broken
system_prompt = trim(message->content);
continue;
}
// in gemma, "assistant" is "model"
role = role == "assistant" ? "model" : message->role;
ss << "<start_of_turn>" << role << "\n";
if (!system_prompt.empty() && role != "model") {
ss << system_prompt << "\n\n";
system_prompt = "";
}
ss << trim(message->content) << "<end_of_turn>\n";
}
if (add_ass) {
ss << "<start_of_turn>model\n";
}
} else {
// template not supported
return -1;