model : avoid hardcoded chat template constant

ggml-ci

Author: ggerganov

src/llama-arch.cpp

@@ -176,6 +176,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP, "tokenizer.ggml.precompiled_charsmap"     },
     { LLM_KV_TOKENIZER_HF_JSON,              "tokenizer.huggingface.json"              },
     { LLM_KV_TOKENIZER_RWKV,                 "tokenizer.rwkv.world"                    },
+    { LLM_KV_TOKENIZER_CHAT_TEMPLATE,        "tokenizer.chat.template"                 },
     { LLM_KV_TOKENIZER_FIM_PRE_ID,           "tokenizer.ggml.fim_pre_token_id"         },
     { LLM_KV_TOKENIZER_FIM_SUF_ID,           "tokenizer.ggml.fim_suf_token_id"         },
     { LLM_KV_TOKENIZER_FIM_MID_ID,           "tokenizer.ggml.fim_mid_token_id"         },

src/llama-arch.h

@@ -174,6 +174,7 @@ enum llm_kv {
     LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP,
     LLM_KV_TOKENIZER_HF_JSON,
     LLM_KV_TOKENIZER_RWKV,
+    LLM_KV_TOKENIZER_CHAT_TEMPLATE,
     LLM_KV_TOKENIZER_FIM_PRE_ID,
     LLM_KV_TOKENIZER_FIM_SUF_ID,
     LLM_KV_TOKENIZER_FIM_MID_ID,

src/llama-model.cpp

@@ -3836,7 +3836,7 @@ uint64_t llama_model_size(const struct llama_model * model) {
 }
 
 const char * llama_model_chat_template(const struct llama_model * model) {
-    const auto & it = model->gguf_kv.find("tokenizer.chat_template");
+    const auto & it = model->gguf_kv.find(LLM_KV(model->arch)(LLM_KV_TOKENIZER_CHAT_TEMPLATE));
     if (it == model->gguf_kv.end()) {
         return nullptr;
     }