llama : chat

ggml-ci

Author: ggerganov

src/CMakeLists.txt

@@ -11,6 +11,7 @@ add_library(llama
             llama.cpp
             llama-arch.cpp
             llama-batch.cpp
+            llama-chat.cpp
             llama-context.cpp
             llama-adapter.cpp
             llama-grammar.cpp

src/llama-arch.cpp

@@ -1385,50 +1385,3 @@ llm_arch llm_arch_from_string(const std::string & name) {
 const llm_tensor_info & llm_tensor_info_for(llm_tensor tensor) {
     return LLM_TENSOR_INFOS.at(tensor);
 }
-
-// chat templates
-
-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             },
-    { "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        },
-    { "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          },
-};
-
-llm_chat_template llm_chat_template_from_str(const std::string & name) {
-    return LLM_CHAT_TEMPLATES.at(name);
-}
-
-int32_t llama_chat_builtin_templates(const char ** output, size_t len) {
-    auto it = LLM_CHAT_TEMPLATES.begin();
-    for (size_t i = 0; i < std::min(len, LLM_CHAT_TEMPLATES.size()); i++) {
-        output[i] = it->first.c_str();
-        std::advance(it, 1);
-    }
-    return (int32_t) LLM_CHAT_TEMPLATES.size();
-}
-

src/llama-arch.h

@@ -388,42 +388,3 @@ const char * llm_arch_name(llm_arch arch);
 llm_arch llm_arch_from_string(const std::string & name);
 
 const llm_tensor_info & llm_tensor_info_for(llm_tensor tensor);
-
-//
-// Chat templates
-// TODO: maybe move these to a separate module
-//
-
-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_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_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_UNKNOWN,
-};
-
-llm_chat_template llm_chat_template_from_str(const std::string & name);