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

llama : add DeepSeek-V3 chat template

Browse source
This commit is contained in:
Stanisław Szymczyk 2025-01-04 14:36:03 +01:00
parent ddb2dddac1
commit a48c3df3df
2 changed files with 19 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

18
src/llama-chat.cpp
View file

@ -45,6 +45,7 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
{ "vicuna-orca", LLM_CHAT_TEMPLATE_VICUNA_ORCA }, { "vicuna-orca", LLM_CHAT_TEMPLATE_VICUNA_ORCA },
{ "deepseek", LLM_CHAT_TEMPLATE_DEEPSEEK }, { "deepseek", LLM_CHAT_TEMPLATE_DEEPSEEK },
{ "deepseek2", LLM_CHAT_TEMPLATE_DEEPSEEK_2 }, { "deepseek2", LLM_CHAT_TEMPLATE_DEEPSEEK_2 },
{ "deepseek3", LLM_CHAT_TEMPLATE_DEEPSEEK_3 },
{ "command-r", LLM_CHAT_TEMPLATE_COMMAND_R }, { "command-r", LLM_CHAT_TEMPLATE_COMMAND_R },
{ "llama3", LLM_CHAT_TEMPLATE_LLAMA_3 }, { "llama3", LLM_CHAT_TEMPLATE_LLAMA_3 },
{ "chatglm3", LLM_CHAT_TEMPLATE_CHATGML_3 }, { "chatglm3", LLM_CHAT_TEMPLATE_CHATGML_3 },
@ -148,6 +149,8 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
return LLM_CHAT_TEMPLATE_MINICPM; return LLM_CHAT_TEMPLATE_MINICPM;
} else if (tmpl_contains("'Assistant: ' + message['content'] + eos_token")) { } else if (tmpl_contains("'Assistant: ' + message['content'] + eos_token")) {
return LLM_CHAT_TEMPLATE_DEEPSEEK_2; return LLM_CHAT_TEMPLATE_DEEPSEEK_2;
} else if (tmpl_contains(LU8("'<｜Assistant｜>' + message['content'] + '<｜end▁of▁sentence｜>'"))) {
return LLM_CHAT_TEMPLATE_DEEPSEEK_3;
} else if (tmpl_contains("[|system|]") && tmpl_contains("[|assistant|]") && tmpl_contains("[|endofturn|]")) { } else if (tmpl_contains("[|system|]") && tmpl_contains("[|assistant|]") && tmpl_contains("[|endofturn|]")) {
// ref: https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct/discussions/8#66bae61b1893d14ee8ed85bb // ref: https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct/discussions/8#66bae61b1893d14ee8ed85bb
// EXAONE-3.0-7.8B-Instruct // EXAONE-3.0-7.8B-Instruct
@ -453,6 +456,21 @@ int32_t llm_chat_apply_template(
if (add_ass) { if (add_ass) {
ss << "Assistant:"; ss << "Assistant:";
} }
} else if (tmpl == LLM_CHAT_TEMPLATE_DEEPSEEK_3) {
// DeepSeek-V3
for (auto message : chat) {
std::string role(message->role);
if (role == "system") {
ss << message->content << "\n\n";
} else if (role == "user") {
ss << LU8("<｜User｜>") << message->content;
} else if (role == "assistant") {
ss << LU8("<｜Assistant｜>") << message->content << LU8("<｜end▁of▁sentence｜>");
}
}
if (add_ass) {
ss << LU8("<｜Assistant｜>");
}
} else if (tmpl == LLM_CHAT_TEMPLATE_EXAONE_3) { } else if (tmpl == LLM_CHAT_TEMPLATE_EXAONE_3) {
// ref: https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct/discussions/8#66bae61b1893d14ee8ed85bb // ref: https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct/discussions/8#66bae61b1893d14ee8ed85bb
// EXAONE-3.0-7.8B-Instruct // EXAONE-3.0-7.8B-Instruct

1
src/llama-chat.h
View file

@ -25,6 +25,7 @@ enum llm_chat_template {
LLM_CHAT_TEMPLATE_VICUNA_ORCA, LLM_CHAT_TEMPLATE_VICUNA_ORCA,
LLM_CHAT_TEMPLATE_DEEPSEEK, LLM_CHAT_TEMPLATE_DEEPSEEK,
LLM_CHAT_TEMPLATE_DEEPSEEK_2, LLM_CHAT_TEMPLATE_DEEPSEEK_2,
LLM_CHAT_TEMPLATE_DEEPSEEK_3,
LLM_CHAT_TEMPLATE_COMMAND_R, LLM_CHAT_TEMPLATE_COMMAND_R,
LLM_CHAT_TEMPLATE_LLAMA_3, LLM_CHAT_TEMPLATE_LLAMA_3,
LLM_CHAT_TEMPLATE_CHATGML_3, LLM_CHAT_TEMPLATE_CHATGML_3,