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

Avoid double bos w/ jinja

Browse source
This commit is contained in:
ochafik 2025-02-03 01:20:11 +00:00
parent 130ca222c9
commit 04d511b5b5
1 changed files with 8 additions and 3 deletions
Download patch file Download diff file Expand all files Collapse all files

11
common/common.cpp
View file

@ -1871,7 +1871,6 @@ std::string common_chat_format_example(const common_chat_template & tmpl, bool u
common_chat_templates common_chat_templates_from_model(const struct llama_model * model, const std::string & chat_template_override) common_chat_templates common_chat_templates_from_model(const struct llama_model * model, const std::string & chat_template_override)
{ {
auto vocab = llama_model_get_vocab(model);
std::string default_template_src = chat_template_override; std::string default_template_src = chat_template_override;
std::string template_tool_use_src = chat_template_override; std::string template_tool_use_src = chat_template_override;
bool has_explicit_template = !chat_template_override.empty(); bool has_explicit_template = !chat_template_override.empty();
@ -1901,6 +1900,11 @@ common_chat_templates common_chat_templates_from_model(const struct llama_model
)"; )";
} }
} }
std::string token_bos;
std::string token_eos;
// TODO: update logic that adds BOS and EOS tokens to the tokenized prompt, in favour of the template.
#if 0
auto vocab = llama_model_get_vocab(model);
const auto get_token = [&](llama_token token, const char * name, const char * jinja_variable_name) { const auto get_token = [&](llama_token token, const char * name, const char * jinja_variable_name) {
if (token == LLAMA_TOKEN_NULL) { if (token == LLAMA_TOKEN_NULL) {
if (default_template_src.find(jinja_variable_name) != std::string::npos if (default_template_src.find(jinja_variable_name) != std::string::npos
@ -1912,8 +1916,9 @@ common_chat_templates common_chat_templates_from_model(const struct llama_model
return common_token_to_piece(vocab, token, true); return common_token_to_piece(vocab, token, true);
} }
}; };
auto token_bos = get_token(llama_vocab_bos(vocab), "BOS", "bos_token"); token_bos = get_token(llama_vocab_bos(vocab), "BOS", "bos_token");
auto token_eos = get_token(llama_vocab_eos(vocab), "EOS", "eos_token"); token_eos = get_token(llama_vocab_eos(vocab), "EOS", "eos_token");
#endif
return { return {
has_explicit_template, has_explicit_template,
std::make_unique<minja::chat_template>(default_template_src, token_bos, token_eos), std::make_unique<minja::chat_template>(default_template_src, token_bos, token_eos),