Add Llama2-Chinese model support

docs/model_support.md

@@ -42,6 +42,7 @@
   loading multiple peft models, you can have them share the base model weights by
   setting the environment variable `PEFT_SHARE_BASE_WEIGHTS=true` in any model
   worker.
+- [FlagAlpha/Llama2-Chinese-13b-Chat](https://huggingface.co/FlagAlpha/Llama2-Chinese-13b-Chat)
 
 ## How to support a new model
 

fastchat/conversation.py

@@ -936,6 +936,21 @@ def get_conv_template(name: str) -> Conversation:
     )
 )
 
+# Llama2-Chinese default template
+# source: https://huggingface.co/FlagAlpha
+register_conv_template(
+    Conversation(
+        name="llama2-chinese",
+        system_message="<s>{system_message}</s>",
+        roles=("Human", "Assistant", "System"),
+        messages=(),
+        offset=0,
+        sep_style=SeparatorStyle.ADD_COLON_TWO,
+        sep="\n",
+        sep2="\n</s><s>",
+        stop_str="</s>",
+    )
+)
 
 if __name__ == "__main__":
     print("Vicuna template:")

fastchat/model/model_adapter.py

@@ -1382,6 +1382,31 @@ def get_default_conv_template(self, model_path: str) -> Conversation:
         return get_conv_template("aquila-chat")
 
 
+class Lamma2ChineseAdapter(BaseModelAdapter):
+    """The model adapter for FlagAlpha/LLama2-Chinese sft"""
+
+    def match(self, model_path: str):
+        return "llama2-chinese" in model_path.lower()
+
+    def load_model(self, model_path: str, from_pretrained_kwargs: dict):
+        revision = from_pretrained_kwargs.get("revision", "main")
+        tokenizer = AutoTokenizer.from_pretrained(
+            model_path,
+            trust_remote_code=True,
+            revision=revision,
+        )
+        model = AutoModelForCausalLM.from_pretrained(
+            model_path,
+            trust_remote_code=True,
+            low_cpu_mem_usage=True,
+            **from_pretrained_kwargs,
+        )
+        return model, tokenizer
+
+    def get_default_conv_template(self, model_path: str) -> Conversation:
+        return get_conv_template("llama2-chinese")
+
+
 # Note: the registration order matters.
 # The one registered earlier has a higher matching priority.
 register_model_adapter(PeftModelAdapter)
@@ -1432,6 +1457,7 @@ def get_default_conv_template(self, model_path: str) -> Conversation:
 register_model_adapter(QwenChatAdapter)
 register_model_adapter(AquilaChatAdapter)
 register_model_adapter(BGEAdapter)
+register_model_adapter(Lamma2ChineseAdapter)
 
 # After all adapters, try the default base adapter.
 register_model_adapter(BaseModelAdapter)

fastchat/model/model_registry.py

@@ -248,3 +248,9 @@ def get_model_info(name: str) -> ModelInfo:
     "https://huggingface.co/Qwen/Qwen-7B-Chat",
     "Qwen is a multi-language large-scale language model (LLM), developed by Damo Academy.",
 )
+register_model_info(
+    ["Llama2-Chinese-13b-Chat", "LLama2-Chinese-13B"],
+    "Llama2-Chinese",
+    "https://huggingface.co/FlagAlpha/Llama2-Chinese-13b-Chat",
+    "Llama2-Chinese is a multi-language large-scale language model (LLM), developed by FlagAlpha.",
+)