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

Use model->gguf_kv for efficiency.

Browse source
This commit is contained in:
DAN™ 2024-12-17 14:00:45 -05:00
parent 919fe432c3
commit 52bfa235e3
1 changed files with 8 additions and 8 deletions
Download patch file Download diff file Expand all files Collapse all files

16
src/llama.cpp
View file

@ -22651,15 +22651,15 @@ int32_t llama_chat_apply_template(
std::string curr_tmpl(tmpl == nullptr ? "" : tmpl);
if (tmpl == nullptr) {
GGML_ASSERT(model != nullptr);
// load template from model
std::vector<char> model_template(16384, 0); // longest known template is about 12792 bytes
std::string template_key = "tokenizer.chat_template";
int32_t res = llama_model_meta_val_str(model, template_key.c_str(), model_template.data(), model_template.size());
if (res < 0) {
// load template from model, if available
const auto & it = model->gguf_kv.find("tokenizer.chat_template");
if (it != model->gguf_kv.end() && it->second.size() > 0) {
curr_tmpl = it->second;
}
else {
// worst case: there is no information about template, we will use chatml by default
curr_tmpl = "chatml"; // see llama_chat_apply_template_internal
} else {
curr_tmpl = std::string(model_template.data(), model_template.size());
curr_tmpl = "chatml"; // see llama_chat_apply_template_internal
}
}