From a48c3df3df2220fc0df3a7038cdbdd2b9ed4eb3b Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Stanis=C5=82aw=20Szymczyk?= <sszymczy@gmail.com>
Date: Sat, 4 Jan 2025 14:36:03 +0100
Subject: [PATCH] llama : add DeepSeek-V3 chat template

---
 src/llama-chat.cpp | 18 ++++++++++++++++++
 src/llama-chat.h   |  1 +
 2 files changed, 19 insertions(+)

diff --git a/src/llama-chat.cpp b/src/llama-chat.cpp
index a07e9cf00..44670d3d8 100644
--- a/src/llama-chat.cpp
+++ b/src/llama-chat.cpp
@@ -45,6 +45,7 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
     { "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         },
@@ -148,6 +149,8 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
         return 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｜>' + message['content'] + '<｜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-Instruct
@@ -453,6 +456,21 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             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) {
         // ref: https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct/discussions/8#66bae61b1893d14ee8ed85bb
         // EXAONE-3.0-7.8B-Instruct
diff --git a/src/llama-chat.h b/src/llama-chat.h
index 364318c27..b8e94d9ef 100644
--- a/src/llama-chat.h
+++ b/src/llama-chat.h
@@ -25,6 +25,7 @@ enum llm_chat_template {
     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,