llama.cpp LLM_CHAT_TEMPLATE_DEEPSEEK_3

llama.cpp LLM_CHAT_TEMPLATE_DEEPSEEK_3

  • 1. `LLAMA_VOCAB_PRE_TYPE_DEEPSEEK3_LLM`
  • 2. `static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES`
  • 3. `LLM_CHAT_TEMPLATE_DEEPSEEK_3`
  • References

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

本文来自互联网用户投稿，该文观点仅代表作者本人，不代表本站立场。本站仅提供信息存储空间服务，不拥有所有权，不承担相关法律责任。如若转载，请注明出处：http://www.rhkb.cn/news/10895.html

如若内容造成侵权/违法违规/事实不符，请联系长河编程网进行投诉反馈email:809451989@qq.com，一经查实，立即删除！

相关文章

高速稳定，功能强大的免费下载工具！！

高速稳定，功能强大的免费下载工具！！

今天&#xff0c;我向你们介绍一款超强的下载器——破姐版本&#xff0c;它完全免费&#xff0c;支持高速下载&#xff0c;将彻底解决你的下载难题。 01 软件介绍 P2P下载器免费下载工具&#xff0c;它以高速稳定、功能强大、无广告的特点&#xff0c;为你提供卓越的下载体验…
阅读更多...
C++：虚函数与多态性习题2

C++：虚函数与多态性习题2

题目内容&#xff1a; 编写程序&#xff0c;声明抽象基类Shape&#xff0c;由它派生出3个派生类&#xff1a;Circle、Rectangle、Triangle&#xff0c;用虚函数分别计算图形面积&#xff0c;并求它们的和。要求用基类指针数组&#xff0c;使它每一个元素指向一个派生类对象。 …
阅读更多...
C#，入门教程(10)——常量、变量与命名规则的基础知识

C#，入门教程(10)——常量、变量与命名规则的基础知识

上一篇&#xff1a; C#&#xff0c;入门教程(09)——运算符的基础知识https://blog.csdn.net/beijinghorn/article/details/123908269 C#用于保存计算数据的元素&#xff0c;称为“变量”。 其中一般不改变初值的变量&#xff0c;称为常变量&#xff0c;简称“常量”。 无论…
阅读更多...
list的使用，及部分功能的模拟实现（C++）

list的使用，及部分功能的模拟实现（C++）

目录&#xff08;文章中"节点"和"结点"是同一个意思&#xff09; 1. list的介绍及使用 1.1 list的介绍 1.2 list的使用 1.2.1 list的构造 1.2.2 list iterator的使用 1.2.3 list capacity 1.2.4 list element access 1.2.5 list modifiers 1.2.6 list…
阅读更多...
MySQL 索引

MySQL 索引

MySQL 索引 文章目录 MySQL 索引1. 索引概念2. 索引结构3. 索引分类4. 索引使用4.1 单列索引和联合索引4.2 覆盖索引4.3 前缀索引 5. SQL提示6. 索引失效情况 1. 索引概念 索引可以理解为MySQL中用来高效检索数据的数据结构&#xff0c;它是有序的&#xff0c;因为它底层使用的…
阅读更多...
JVM方法区

JVM方法区

一、栈、堆、方法区的交互关系 二、方法区的理解: 尽管所有的方法区在逻辑上属于堆的一部分&#xff0c;但是一些简单的实现可能不会去进行垃圾收集或者进行压缩&#xff0c;方法区可以看作是一块独立于Java堆的内存空间。 方法区(Method Area)与Java堆一样&#xff0c;是各个…
阅读更多...
STM32 TIM定时器配置

STM32 TIM定时器配置

TIM简介 TIM&#xff08;Timer&#xff09;定时器 定时器可以对输入的时钟进行计数&#xff0c;并在计数值达到设定值时触发中断 16位计数器、预分频器、自动重装寄存器的时基单元&#xff0c;在72MHz计数时钟下可以实现最大59.65s的定时 不仅具备基本的定时中断功能&#xff…
阅读更多...
自定义数据集 使用pytorch框架实现逻辑回归并保存模型，然后保存模型后再加载模型进行预测，对预测结果计算精确度和召回率及F1分数

自定义数据集 使用pytorch框架实现逻辑回归并保存模型，然后保存模型后再加载模型进行预测，对预测结果计算精确度和召回率及F1分数

import numpy as np import torch import torch.nn as nn import torch.optim as optim from sklearn.metrics import precision_score, recall_score, f1_score# 数据准备 class1_points np.array([[1.9, 1.2],[1.5, 2.1],[1.9, 0.5],[1.5, 0.9],[0.9, 1.2],[1.1, 1.7],[1.4,…
阅读更多...
＜ OS 有关 ＞ 阿里云 几个小时前 使用密钥替换 SSH 密码认证后， 发现主机正在被“攻击” 分析与应对

＜ OS 有关 ＞ 阿里云 几个小时前 使用密钥替换 SSH 密码认证后， 发现主机正在被“攻击” 分析与应对

信息来源&#xff1a; 文件&#xff1a;/var/log/auth.log 因为在 sshd_config 配置文件中&#xff0c;已经定义 LogLevel INFO 部分内容&#xff1a; 2025-01-27T18:18:55.68272708:00 jpn sshd[15891]: Received disconnect from 45.194.37.171 port 58954:11: Bye Bye […
阅读更多...
[创业之路-270]：《向流程设计要效率》-2-企业流程架构模式 POS架构(规划、业务运营、支撑）、OES架构（业务运营、使能、支撑）

[创业之路-270]：《向流程设计要效率》-2-企业流程架构模式 POS架构(规划、业务运营、支撑）、OES架构（业务运营、使能、支撑）

目录 一、POS架构 二、OES架构 三、POS架构与OES架构的差异 四、各自的典型示例 POS架构典型示例 OES架构典型示例 示例分析 五、各自的典型企业 POS架构典型企业 OES架构典型企业 分析 六、各自典型的流程 POS架构的典型流程 OES架构的典型流程 企业流程架构模式…
阅读更多...
【贪心算法篇】：“贪心”之旅--算法练习题中的智慧与策略（一）

【贪心算法篇】：“贪心”之旅--算法练习题中的智慧与策略（一）

✨感谢您阅读本篇文章&#xff0c;文章内容是个人学习笔记的整理&#xff0c;如果哪里有误的话还请您指正噢✨ ✨ 个人主页&#xff1a;余辉zmh–CSDN博客 ✨ 文章所属专栏&#xff1a;贪心算法篇–CSDN博客 文章目录 一.贪心算法1.什么是贪心算法2.贪心算法的特点 二.例题1.柠…
阅读更多...
Python 梯度下降法（二）：RMSProp Optimize

Python 梯度下降法（二）：RMSProp Optimize

文章目录 Python 梯度下降法&#xff08;二&#xff09;&#xff1a;RMSProp Optimize一、数学原理1.1 介绍1.2 公式 二、代码实现2.1 函数代码2.2 总代码 三、代码优化3.1 存在问题3.2 收敛判断3.3 函数代码3.4 总代码 四、优缺点4.1 优点4.2 缺点 五、相关链接 Python 梯度下…
阅读更多...
【2025年更新】1000个大数据/人工智能毕设选题推荐

【2025年更新】1000个大数据/人工智能毕设选题推荐

文章目录 前言大数据/人工智能毕设选题&#xff1a;后记 前言 正值毕业季我看到很多同学都在为自己的毕业设计发愁 Maynor在网上搜集了1000个大数据的毕设选题&#xff0c;希望对大家有帮助&#xff5e; 适合大数据毕业设计的项目&#xff0c;完全可以作为本科生当前较新的毕…
阅读更多...
three.js+WebGL踩坑经验合集(6.2):负缩放，负定矩阵和行列式的关系（3D版本）

three.js+WebGL踩坑经验合集(6.2):负缩放，负定矩阵和行列式的关系（3D版本）

本篇将紧接上篇的2D版本对3D版的负缩放矩阵进行解读。 (6.1):负缩放&#xff0c;负定矩阵和行列式的关系&#xff08;2D版本&#xff09; 既然three.js对3D版的负缩放也使用行列式进行判断&#xff0c;那么&#xff0c;2D版的结论用到3D上其实是没毛病的&#xff0c;THREE.Li…
阅读更多...
反向代理模块jmh

反向代理模块jmh

1 概念 1.1 反向代理概念 反向代理是指以代理服务器来接收客户端的请求&#xff0c;然后将请求转发给内部网络上的服务器&#xff0c;将从服务器上得到的结果返回给客户端&#xff0c;此时代理服务器对外表现为一个反向代理服务器。 对于客户端来说&#xff0c;反向代理就相当…
阅读更多...
软件工程经济学-日常作业+大作业

软件工程经济学-日常作业+大作业

目录 一、作业1 作业内容 解答 二、作业2 作业内容 解答 三、作业3 作业内容 解答 四、大作业 作业内容 解答 1.建立层次结构模型 (1)目标层 (2)准则层 (3)方案层 2.构造判断矩阵 (1)准则层判断矩阵 (2)方案层判断矩阵 3.层次单排序及其一致性检验 代码 …
阅读更多...
【回溯】目标和 字母大小全排列

【回溯】目标和 字母大小全排列

文章目录 494. 目标和解题思路&#xff1a;回溯784. 字母大小写全排列解题思路&#xff1a;回溯 494. 目标和 494. 目标和 给你一个非负整数数组 nums 和一个整数 target 。 向数组中的每个整数前添加 或 - &#xff0c;然后串联起所有整数&#xff0c;可以构造一个 表达式…
阅读更多...
告别复杂，拥抱简洁：用plusDays(7)代替plus(7, ChronoUnit.DAYS)

告别复杂，拥抱简洁：用plusDays(7)代替plus(7, ChronoUnit.DAYS)

前言 你知道吗？有时候代码里的一些小细节看起来很简单，却可能成为你调试时的大麻烦。在 Java 中，我们用 LocalDateTime 进行日期和时间的操作时，发现一个小小的替代方法可以让代码更简洁，功能更强大。这不，今天我们就来探讨如何用 LocalDateTime.now().plusDays(7) 替代…
阅读更多...
《苍穹外卖》项目学习记录-Day10订单状态定时处理

《苍穹外卖》项目学习记录-Day10订单状态定时处理

利用Cron表达式生成器生成Cron表达式 1.处理超时订单 查询订单表把超时的订单查询出来&#xff0c;也就是订单的状态为待付款&#xff0c;下单的时间已经超过了15分钟。 //select * from orders where status ? and order_time < (当前时间 - 15分钟) 遍历集合把数据库…
阅读更多...
【深度分析】微软全球裁员计划不影响印度地区，将继续增加当地就业机会

【深度分析】微软全球裁员计划不影响印度地区，将继续增加当地就业机会

当微软的裁员刀锋掠过全球办公室时&#xff0c;班加罗尔的键盘声却愈发密集——这场资本迁徙背后&#xff0c;藏着数字殖民时代最锋利的生存法则。 表面是跨国公司的区域战略调整&#xff0c;实则是全球人才市场的地壳运动。微软一边在硅谷裁撤年薪20万美金的高级工程师&#x…
阅读更多...
最新文章
推荐文章

Copyright @ 2022~2023