llama.cpp LLM_CHAT_TEMPLATE_DEEPSEEK_3

不宜吹捧中国大语言模型的同时，又去贬低美国大语言模型。

水是人体的主要化学成分，约占体重的 50% 至 70%，大语言模型的含水量也不会太低。

科技发展靠的是硬实力，而不是情怀和口号。

llama.cpp
https://github.com/ggerganov/llama.cpp

1. LLAMA_VOCAB_PRE_TYPE_DEEPSEEK3_LLM

• LLAMA_VOCAB_PRE_TYPE_DEEPSEEK3_LLM, LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_LLM and LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_CODER

/home/yongqiang/llm_work/llama_cpp_25_01_05/llama.cpp/include/llama.h

    enum llama_vocab_type {LLAMA_VOCAB_TYPE_NONE = 0, // For models without vocabLLAMA_VOCAB_TYPE_SPM  = 1, // LLaMA tokenizer based on byte-level BPE with byte fallbackLLAMA_VOCAB_TYPE_BPE  = 2, // GPT-2 tokenizer based on byte-level BPELLAMA_VOCAB_TYPE_WPM  = 3, // BERT tokenizer based on WordPieceLLAMA_VOCAB_TYPE_UGM  = 4, // T5 tokenizer based on UnigramLLAMA_VOCAB_TYPE_RWKV = 5, // RWKV tokenizer based on greedy tokenization};// pre-tokenization typesenum llama_vocab_pre_type {LLAMA_VOCAB_PRE_TYPE_DEFAULT        = 0,LLAMA_VOCAB_PRE_TYPE_LLAMA3         = 1,LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_LLM   = 2,LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_CODER = 3,LLAMA_VOCAB_PRE_TYPE_FALCON         = 4,LLAMA_VOCAB_PRE_TYPE_MPT            = 5,LLAMA_VOCAB_PRE_TYPE_STARCODER      = 6,LLAMA_VOCAB_PRE_TYPE_GPT2           = 7,LLAMA_VOCAB_PRE_TYPE_REFACT         = 8,LLAMA_VOCAB_PRE_TYPE_COMMAND_R      = 9,LLAMA_VOCAB_PRE_TYPE_STABLELM2      = 10,LLAMA_VOCAB_PRE_TYPE_QWEN2          = 11,LLAMA_VOCAB_PRE_TYPE_OLMO           = 12,LLAMA_VOCAB_PRE_TYPE_DBRX           = 13,LLAMA_VOCAB_PRE_TYPE_SMAUG          = 14,LLAMA_VOCAB_PRE_TYPE_PORO           = 15,LLAMA_VOCAB_PRE_TYPE_CHATGLM3       = 16,LLAMA_VOCAB_PRE_TYPE_CHATGLM4       = 17,LLAMA_VOCAB_PRE_TYPE_VIKING         = 18,LLAMA_VOCAB_PRE_TYPE_JAIS           = 19,LLAMA_VOCAB_PRE_TYPE_TEKKEN         = 20,LLAMA_VOCAB_PRE_TYPE_SMOLLM         = 21,LLAMA_VOCAB_PRE_TYPE_CODESHELL      = 22,LLAMA_VOCAB_PRE_TYPE_BLOOM          = 23,LLAMA_VOCAB_PRE_TYPE_GPT3_FINNISH   = 24,LLAMA_VOCAB_PRE_TYPE_EXAONE         = 25,LLAMA_VOCAB_PRE_TYPE_CHAMELEON      = 26,LLAMA_VOCAB_PRE_TYPE_MINERVA        = 27,LLAMA_VOCAB_PRE_TYPE_DEEPSEEK3_LLM  = 28,};

/home/yongqiang/llm_work/llama_cpp_25_01_05/llama.cpp/src/llama-hparams.h

// bump if necessary
#define LLAMA_MAX_LAYERS  512
#define LLAMA_MAX_EXPERTS 256  // DeepSeekV3enum llama_expert_gating_func_type {LLAMA_EXPERT_GATING_FUNC_TYPE_NONE    = 0,LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX = 1,LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID = 2,
};

2. static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES

• LLM_CHAT_TEMPLATE_DEEPSEEK_3, LLM_CHAT_TEMPLATE_DEEPSEEK_2 and LLM_CHAT_TEMPLATE_DEEPSEEK

/home/yongqiang/llm_work/llama_cpp_25_01_05/llama.cpp/src/llama-chat.h

enum llm_chat_template {LLM_CHAT_TEMPLATE_CHATML,LLM_CHAT_TEMPLATE_LLAMA_2,LLM_CHAT_TEMPLATE_LLAMA_2_SYS,LLM_CHAT_TEMPLATE_LLAMA_2_SYS_BOS,LLM_CHAT_TEMPLATE_LLAMA_2_SYS_STRIP,LLM_CHAT_TEMPLATE_MISTRAL_V1,LLM_CHAT_TEMPLATE_MISTRAL_V3,LLM_CHAT_TEMPLATE_MISTRAL_V3_TEKKEN,LLM_CHAT_TEMPLATE_MISTRAL_V7,LLM_CHAT_TEMPLATE_PHI_3,LLM_CHAT_TEMPLATE_PHI_4,LLM_CHAT_TEMPLATE_FALCON_3,LLM_CHAT_TEMPLATE_ZEPHYR,LLM_CHAT_TEMPLATE_MONARCH,LLM_CHAT_TEMPLATE_GEMMA,LLM_CHAT_TEMPLATE_ORION,LLM_CHAT_TEMPLATE_OPENCHAT,LLM_CHAT_TEMPLATE_VICUNA,LLM_CHAT_TEMPLATE_VICUNA_ORCA,LLM_CHAT_TEMPLATE_DEEPSEEK,LLM_CHAT_TEMPLATE_DEEPSEEK_2,LLM_CHAT_TEMPLATE_DEEPSEEK_3,LLM_CHAT_TEMPLATE_COMMAND_R,LLM_CHAT_TEMPLATE_LLAMA_3,LLM_CHAT_TEMPLATE_CHATGML_3,LLM_CHAT_TEMPLATE_CHATGML_4,LLM_CHAT_TEMPLATE_MINICPM,LLM_CHAT_TEMPLATE_EXAONE_3,LLM_CHAT_TEMPLATE_RWKV_WORLD,LLM_CHAT_TEMPLATE_GRANITE,LLM_CHAT_TEMPLATE_GIGACHAT,LLM_CHAT_TEMPLATE_MEGREZ,LLM_CHAT_TEMPLATE_UNKNOWN,
};

• { "deepseek3", LLM_CHAT_TEMPLATE_DEEPSEEK_3 }, { "deepseek2", LLM_CHAT_TEMPLATE_DEEPSEEK_2 } and { "deepseek", LLM_CHAT_TEMPLATE_DEEPSEEK }

/home/yongqiang/llm_work/llama_cpp_25_01_05/llama.cpp/src/llama-chat.cpp

static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {{ "chatml",            LLM_CHAT_TEMPLATE_CHATML            },{ "llama2",            LLM_CHAT_TEMPLATE_LLAMA_2           },{ "llama2-sys",        LLM_CHAT_TEMPLATE_LLAMA_2_SYS       },{ "llama2-sys-bos",    LLM_CHAT_TEMPLATE_LLAMA_2_SYS_BOS   },{ "llama2-sys-strip",  LLM_CHAT_TEMPLATE_LLAMA_2_SYS_STRIP },{ "mistral-v1",        LLM_CHAT_TEMPLATE_MISTRAL_V1        },{ "mistral-v3",        LLM_CHAT_TEMPLATE_MISTRAL_V3        },{ "mistral-v3-tekken", LLM_CHAT_TEMPLATE_MISTRAL_V3_TEKKEN },{ "mistral-v7",        LLM_CHAT_TEMPLATE_MISTRAL_V7        },{ "phi3",              LLM_CHAT_TEMPLATE_PHI_3             },{ "phi4",              LLM_CHAT_TEMPLATE_PHI_4             },{ "falcon3",           LLM_CHAT_TEMPLATE_FALCON_3          },{ "zephyr",            LLM_CHAT_TEMPLATE_ZEPHYR            },{ "monarch",           LLM_CHAT_TEMPLATE_MONARCH           },{ "gemma",             LLM_CHAT_TEMPLATE_GEMMA             },{ "orion",             LLM_CHAT_TEMPLATE_ORION             },{ "openchat",          LLM_CHAT_TEMPLATE_OPENCHAT          },{ "vicuna",            LLM_CHAT_TEMPLATE_VICUNA            },{ "vicuna-orca",       LLM_CHAT_TEMPLATE_VICUNA_ORCA       },{ "deepseek",          LLM_CHAT_TEMPLATE_DEEPSEEK          },{ "deepseek2",         LLM_CHAT_TEMPLATE_DEEPSEEK_2        },{ "deepseek3",         LLM_CHAT_TEMPLATE_DEEPSEEK_3        },{ "command-r",         LLM_CHAT_TEMPLATE_COMMAND_R         },{ "llama3",            LLM_CHAT_TEMPLATE_LLAMA_3           },{ "chatglm3",          LLM_CHAT_TEMPLATE_CHATGML_3         },{ "chatglm4",          LLM_CHAT_TEMPLATE_CHATGML_4         },{ "minicpm",           LLM_CHAT_TEMPLATE_MINICPM           },{ "exaone3",           LLM_CHAT_TEMPLATE_EXAONE_3          },{ "rwkv-world",        LLM_CHAT_TEMPLATE_RWKV_WORLD        },{ "granite",           LLM_CHAT_TEMPLATE_GRANITE           },{ "gigachat",          LLM_CHAT_TEMPLATE_GIGACHAT          },{ "megrez",            LLM_CHAT_TEMPLATE_MEGREZ            },
};

3. LLM_CHAT_TEMPLATE_DEEPSEEK_3

• LLM_CHAT_TEMPLATE_DEEPSEEK_3, LLM_CHAT_TEMPLATE_DEEPSEEK_2 and LLM_CHAT_TEMPLATE_DEEPSEEK

/home/yongqiang/llm_work/llama_cpp_25_01_05/llama.cpp/src/llama-chat.cpp

// Simple version of "llama_apply_chat_template" that only works with strings
// This function uses heuristic checks to determine commonly used template. It is not a jinja parser.
int32_t llm_chat_apply_template(llm_chat_template tmpl,const std::vector<const llama_chat_message *> & chat,std::string & dest, bool add_ass) {// Taken from the research: https://github.com/ggerganov/llama.cpp/issues/5527std::stringstream ss;if (tmpl == LLM_CHAT_TEMPLATE_CHATML) {// chatml templatefor (auto message : chat) {ss << "<|im_start|>" << message->role << "\n" << message->content << "<|im_end|>\n";}if (add_ass) {ss << "<|im_start|>assistant\n";}} else if (tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V7) {// Official mistral 'v7' template// See: https://huggingface.co/mistralai/Mistral-Large-Instruct-2411#basic-instruct-template-v7for (auto message : chat) {std::string role(message->role);std::string content(message->content);if (role == "system") {ss << "[SYSTEM_PROMPT] " << content << "[/SYSTEM_PROMPT]";} else if (role == "user") {ss << "[INST] " << content << "[/INST]";}else {ss << " " << content << "</s>";}}} else if (tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V1|| tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V3|| tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V3_TEKKEN) {// See: https://github.com/mistralai/cookbook/blob/main/concept-deep-dive/tokenization/chat_templates.md// See: https://github.com/mistralai/cookbook/blob/main/concept-deep-dive/tokenization/templates.mdstd::string leading_space = tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V1 ? " " : "";std::string trailing_space = tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V3_TEKKEN ? "" : " ";bool trim_assistant_message = tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V3;bool is_inside_turn = false;for (auto message : chat) {if (!is_inside_turn) {ss << leading_space << "[INST]" << trailing_space;is_inside_turn = true;}std::string role(message->role);std::string content(message->content);if (role == "system") {ss << content << "\n\n";} else if (role == "user") {ss << content << leading_space << "[/INST]";} else {ss << trailing_space << (trim_assistant_message ? trim(content) : content) << "</s>";is_inside_turn = false;}}} else if (tmpl == LLM_CHAT_TEMPLATE_LLAMA_2|| tmpl == LLM_CHAT_TEMPLATE_LLAMA_2_SYS|| tmpl == LLM_CHAT_TEMPLATE_LLAMA_2_SYS_BOS|| tmpl == LLM_CHAT_TEMPLATE_LLAMA_2_SYS_STRIP) {// llama2 template and its variants// [variant] support system message// See: https://huggingface.co/blog/llama2#how-to-prompt-llama-2bool support_system_message = tmpl != LLM_CHAT_TEMPLATE_LLAMA_2;// [variant] add BOS inside historybool add_bos_inside_history = tmpl == LLM_CHAT_TEMPLATE_LLAMA_2_SYS_BOS;// [variant] trim spaces from the input messagebool strip_message = tmpl == LLM_CHAT_TEMPLATE_LLAMA_2_SYS_STRIP;// construct the promptbool is_inside_turn = true; // skip BOS at the beginningss << "[INST] ";for (auto message : chat) {std::string content = strip_message ? trim(message->content) : message->content;std::string role(message->role);if (!is_inside_turn) {is_inside_turn = true;ss << (add_bos_inside_history ? "<s>[INST] " : "[INST] ");}if (role == "system") {if (support_system_message) {ss << "<<SYS>>\n" << content << "\n<</SYS>>\n\n";} else {// if the model does not support system message, we still include it in the first message, but without <<SYS>>ss << content << "\n";}} else if (role == "user") {ss << content << " [/INST]";} else {ss << content << "</s>";is_inside_turn = false;}}} else if (tmpl == LLM_CHAT_TEMPLATE_PHI_3) {// Phi 3for (auto message : chat) {std::string role(message->role);ss << "<|" << role << "|>\n" << message->content << "<|end|>\n";}if (add_ass) {ss << "<|assistant|>\n";}} else if (tmpl == LLM_CHAT_TEMPLATE_PHI_4) {// chatml templatefor (auto message : chat) {ss << "<|im_start|>" << message->role << "<|im_sep|>" << message->content << "<|im_end|>";}if (add_ass) {ss << "<|im_start|>assistant<|im_sep|>";}} else if (tmpl == LLM_CHAT_TEMPLATE_FALCON_3) {// Falcon 3for (auto message : chat) {std::string role(message->role);ss << "<|" << role << "|>\n" << message->content << "\n";}if (add_ass) {ss << "<|assistant|>\n";}} else if (tmpl == LLM_CHAT_TEMPLATE_ZEPHYR) {// zephyr templatefor (auto message : chat) {ss << "<|" << message->role << "|>" << "\n" << message->content << "<|endoftext|>\n";}if (add_ass) {ss << "<|assistant|>\n";}} else if (tmpl == LLM_CHAT_TEMPLATE_MONARCH) {// mlabonne/AlphaMonarch-7B template (the <s> is included inside history)for (auto message : chat) {std::string bos = (message == chat.front()) ? "" : "<s>"; // skip BOS for first messagess << bos << message->role << "\n" << message->content << "</s>\n";}if (add_ass) {ss << "<s>assistant\n";}} else if (tmpl == LLM_CHAT_TEMPLATE_GEMMA) {// google/gemma-7b-itstd::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 brokensystem_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 if (tmpl == LLM_CHAT_TEMPLATE_ORION) {// OrionStarAI/Orion-14B-Chatstd::string system_prompt = "";for (auto message : chat) {std::string role(message->role);if (role == "system") {// there is no system message support, we will merge it with user promptsystem_prompt = message->content;continue;} else if (role == "user") {ss << "Human: ";if (!system_prompt.empty()) {ss << system_prompt << "\n\n";system_prompt = "";}ss << message->content << "\n\nAssistant: </s>";} else {ss << message->content << "</s>";}}} else if (tmpl == LLM_CHAT_TEMPLATE_OPENCHAT) {// openchat/openchat-3.5-0106,for (auto message : chat) {std::string role(message->role);if (role == "system") {ss << message->content << "<|end_of_turn|>";} else {role[0] = toupper(role[0]);ss << "GPT4 Correct " << role << ": " << message->content << "<|end_of_turn|>";}}if (add_ass) {ss << "GPT4 Correct Assistant:";}} else if (tmpl == LLM_CHAT_TEMPLATE_VICUNA || tmpl == LLM_CHAT_TEMPLATE_VICUNA_ORCA) {// eachadea/vicuna-13b-1.1 (and Orca variant)for (auto message : chat) {std::string role(message->role);if (role == "system") {// Orca-Vicuna variant uses a system prefixif (tmpl == LLM_CHAT_TEMPLATE_VICUNA_ORCA) {ss << "SYSTEM: " << message->content << "\n";} else {ss << message->content << "\n\n";}} else if (role == "user") {ss << "USER: " << message->content << "\n";} else if (role == "assistant") {ss << "ASSISTANT: " << message->content << "</s>\n";}}if (add_ass) {ss << "ASSISTANT:";}} else if (tmpl == LLM_CHAT_TEMPLATE_DEEPSEEK) {// deepseek-ai/deepseek-coder-33b-instructfor (auto message : chat) {std::string role(message->role);if (role == "system") {ss << message->content;} else if (role == "user") {ss << "### Instruction:\n" << message->content << "\n";} else if (role == "assistant") {ss << "### Response:\n" << message->content << "\n<|EOT|>\n";}}if (add_ass) {ss << "### Response:\n";}} else if (tmpl == LLM_CHAT_TEMPLATE_COMMAND_R) {// CohereForAI/c4ai-command-r-plusfor (auto message : chat) {std::string role(message->role);if (role == "system") {ss << "<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>" << trim(message->content) << "<|END_OF_TURN_TOKEN|>";} else if (role == "user") {ss << "<|START_OF_TURN_TOKEN|><|USER_TOKEN|>" << trim(message->content) << "<|END_OF_TURN_TOKEN|>";} else if (role == "assistant") {ss << "<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>" << trim(message->content) << "<|END_OF_TURN_TOKEN|>";}}if (add_ass) {ss << "<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>";}} else if (tmpl == LLM_CHAT_TEMPLATE_LLAMA_3) {// Llama 3for (auto message : chat) {std::string role(message->role);ss << "<|start_header_id|>" << role << "<|end_header_id|>\n\n" << trim(message->content) << "<|eot_id|>";}if (add_ass) {ss << "<|start_header_id|>assistant<|end_header_id|>\n\n";}} else if (tmpl == LLM_CHAT_TEMPLATE_CHATGML_3) {// chatglm3-6bss << "[gMASK]" << "sop";for (auto message : chat) {std::string role(message->role);ss << "<|" << role << "|>" << "\n " << message->content;}if (add_ass) {ss << "<|assistant|>";}} else if (tmpl == LLM_CHAT_TEMPLATE_CHATGML_4) {ss << "[gMASK]" << "<sop>";for (auto message : chat) {std::string role(message->role);ss << "<|" << role << "|>" << "\n" << message->content;}if (add_ass) {ss << "<|assistant|>";}} else if (tmpl == LLM_CHAT_TEMPLATE_MINICPM) {// MiniCPM-3B-OpenHermes-2.5-v2-GGUFfor (auto message : chat) {std::string role(message->role);if (role == "user") {ss << LU8("<用户>");ss << trim(message->content);ss << "<AI>";} else {ss << trim(message->content);}}} else if (tmpl == LLM_CHAT_TEMPLATE_DEEPSEEK_2) {// DeepSeek-V2for (auto message : chat) {std::string role(message->role);if (role == "system") {ss << message->content << "\n\n";} else if (role == "user") {ss << "User: " << message->content << "\n\n";} else if (role == "assistant") {ss << "Assistant: " << message->content << LU8("<｜end▁of▁sentence｜>");}}if (add_ass) {ss << "Assistant:";}} else if (tmpl == LLM_CHAT_TEMPLATE_DEEPSEEK_3) {// DeepSeek-V3for (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) {// ref: https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct/discussions/8#66bae61b1893d14ee8ed85bb// EXAONE-3.0-7.8B-Instructfor (auto message : chat) {std::string role(message->role);if (role == "system") {ss << "[|system|]" << trim(message->content) << "[|endofturn|]\n";} else if (role == "user") {ss << "[|user|]" << trim(message->content) << "\n";} else if (role == "assistant") {ss << "[|assistant|]" << trim(message->content) << "[|endofturn|]\n";}}if (add_ass) {ss << "[|assistant|]";}} else if (tmpl == LLM_CHAT_TEMPLATE_RWKV_WORLD) {// this template requires the model to have "\n\n" as EOT tokenfor (auto message : chat) {std::string role(message->role);if (role == "user") {ss << "User: " << message->content << "\n\nAssistant:";} else {ss << message->content << "\n\n";}}} else if (tmpl == LLM_CHAT_TEMPLATE_GRANITE) {// IBM Granite templatefor (const auto & message : chat) {std::string role(message->role);ss << "<|start_of_role|>" << role << "<|end_of_role|>";if (role == "assistant_tool_call") {ss << "<|tool_call|>";}ss << message->content << "<|end_of_text|>\n";}if (add_ass) {ss << "<|start_of_role|>assistant<|end_of_role|>\n";}} else if (tmpl == LLM_CHAT_TEMPLATE_GIGACHAT) {// GigaChat templatebool has_system = !chat.empty() && std::string(chat[0]->role) == "system";// Handle system message if presentif (has_system) {ss << "<s>" << chat[0]->content << "<|message_sep|>";} else {ss << "<s>";}// Process remaining messagesfor (size_t i = has_system ? 1 : 0; i < chat.size(); i++) {std::string role(chat[i]->role);if (role == "user") {ss << "user<|role_sep|>" << chat[i]->content << "<|message_sep|>"<< "available functions<|role_sep|>[]<|message_sep|>";} else if (role == "assistant") {ss << "assistant<|role_sep|>" << chat[i]->content << "<|message_sep|>";}}// Add generation prompt if neededif (add_ass) {ss << "assistant<|role_sep|>";}}  else if (tmpl == LLM_CHAT_TEMPLATE_MEGREZ) {// Megrez templatefor (auto message : chat) {std::string role(message->role);ss << "<|role_start|>" << role << "<|role_end|>" << message->content << "<|turn_end|>";}if (add_ass) {ss << "<|role_start|>assistant<|role_end|>";}} else {// template not supportedreturn -1;}dest = ss.str();return dest.size();
}

llm_chat_template llm_chat_detect_template(const std::string & tmpl) {try {return llm_chat_template_from_str(tmpl);} catch (const std::out_of_range &) {// ignore}auto tmpl_contains = [&tmpl](const char * haystack) -> bool {return tmpl.find(haystack) != std::string::npos;};if (tmpl_contains("<|im_start|>")) {return tmpl_contains("<|im_sep|>")? LLM_CHAT_TEMPLATE_PHI_4: LLM_CHAT_TEMPLATE_CHATML;} else if (tmpl.find("mistral") == 0 || tmpl_contains("[INST]")) {if (tmpl_contains("[SYSTEM_PROMPT]")) {return LLM_CHAT_TEMPLATE_MISTRAL_V7;} else if (// catches official 'v1' templatetmpl_contains("' [INST] ' + system_message")// catches official 'v3' and 'v3-tekken' templates|| tmpl_contains("[AVAILABLE_TOOLS]")) {// Official mistral 'v1', 'v3' and 'v3-tekken' templates// See: https://github.com/mistralai/cookbook/blob/main/concept-deep-dive/tokenization/chat_templates.md// See: https://github.com/mistralai/cookbook/blob/main/concept-deep-dive/tokenization/templates.mdif (tmpl_contains(" [INST]")) {return LLM_CHAT_TEMPLATE_MISTRAL_V1;} else if (tmpl_contains("\"[INST]\"")) {return LLM_CHAT_TEMPLATE_MISTRAL_V3_TEKKEN;}return LLM_CHAT_TEMPLATE_MISTRAL_V3;} else {// llama2 template and its variants// [variant] support system message// See: https://huggingface.co/blog/llama2#how-to-prompt-llama-2bool support_system_message = tmpl_contains("<<SYS>>");bool add_bos_inside_history = tmpl_contains("bos_token + '[INST]");bool strip_message = tmpl_contains("content.strip()");if (strip_message) {return LLM_CHAT_TEMPLATE_LLAMA_2_SYS_STRIP;} else if (add_bos_inside_history) {return LLM_CHAT_TEMPLATE_LLAMA_2_SYS_BOS;} else if (support_system_message) {return LLM_CHAT_TEMPLATE_LLAMA_2_SYS;} else {return LLM_CHAT_TEMPLATE_LLAMA_2;}}} else if (tmpl_contains("<|assistant|>") && tmpl_contains("<|end|>")) {return LLM_CHAT_TEMPLATE_PHI_3;} else if (tmpl_contains("<|assistant|>") && tmpl_contains("<|user|>")) {return LLM_CHAT_TEMPLATE_FALCON_3;} else if (tmpl_contains("<|user|>") && tmpl_contains("<|endoftext|>")) {return LLM_CHAT_TEMPLATE_ZEPHYR;} else if (tmpl_contains("bos_token + message['role']")) {return LLM_CHAT_TEMPLATE_MONARCH;} else if (tmpl_contains("<start_of_turn>")) {return LLM_CHAT_TEMPLATE_GEMMA;} else if (tmpl_contains("'\\n\\nAssistant: ' + eos_token")) {// OrionStarAI/Orion-14B-Chatreturn LLM_CHAT_TEMPLATE_ORION;} else if (tmpl_contains("GPT4 Correct ")) {// openchat/openchat-3.5-0106return LLM_CHAT_TEMPLATE_OPENCHAT;} else if (tmpl_contains("USER: ") && tmpl_contains("ASSISTANT: ")) {// eachadea/vicuna-13b-1.1 (and Orca variant)if (tmpl_contains("SYSTEM: ")) {return LLM_CHAT_TEMPLATE_VICUNA_ORCA;}return LLM_CHAT_TEMPLATE_VICUNA;} else if (tmpl_contains("### Instruction:") && tmpl_contains("<|EOT|>")) {// deepseek-ai/deepseek-coder-33b-instructreturn LLM_CHAT_TEMPLATE_DEEPSEEK;} else if (tmpl_contains("<|START_OF_TURN_TOKEN|>") && tmpl_contains("<|USER_TOKEN|>")) {// CohereForAI/c4ai-command-r-plusreturn LLM_CHAT_TEMPLATE_COMMAND_R;} else if (tmpl_contains("<|start_header_id|>") && tmpl_contains("<|end_header_id|>")) {return LLM_CHAT_TEMPLATE_LLAMA_3;} else if (tmpl_contains("[gMASK]sop")) {// chatglm3-6breturn LLM_CHAT_TEMPLATE_CHATGML_3;} else if (tmpl_contains("[gMASK]<sop>")) {return LLM_CHAT_TEMPLATE_CHATGML_4;} else if (tmpl_contains(LU8("<用户>"))) {// MiniCPM-3B-OpenHermes-2.5-v2-GGUFreturn LLM_CHAT_TEMPLATE_MINICPM;} else if (tmpl_contains("'Assistant: ' + message['content'] + eos_token")) {return LLM_CHAT_TEMPLATE_DEEPSEEK_2;} else if (tmpl_contains(LU8("<｜Assistant｜>")) && tmpl_contains(LU8("<｜User｜>")) && tmpl_contains(LU8("<｜end▁of▁sentence｜>"))) {return LLM_CHAT_TEMPLATE_DEEPSEEK_3;} 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// EXAONE-3.0-7.8B-Instructreturn LLM_CHAT_TEMPLATE_EXAONE_3;} else if (tmpl_contains("rwkv-world")) {return LLM_CHAT_TEMPLATE_RWKV_WORLD;} else if (tmpl_contains("<|start_of_role|>")) {return LLM_CHAT_TEMPLATE_GRANITE;} else if (tmpl_contains("message['role'] + additional_special_tokens[0] + message['content'] + additional_special_tokens[1]")) {return LLM_CHAT_TEMPLATE_GIGACHAT;} else if (tmpl_contains("<|role_start|>")) {return LLM_CHAT_TEMPLATE_MEGREZ;}return LLM_CHAT_TEMPLATE_UNKNOWN;
}

References

[1] Yongqiang Cheng, https://yongqiang.blog.csdn.net/
[2] huggingface/gguf, https://github.com/huggingface/huggingface.js/tree/main/packages/gguf