support stable-vicuna model (lm-sys#2696)

fastchat/conversation.py

@@ -1139,6 +1139,20 @@ def get_conv_template(name: str) -> Conversation:
     )
 )
 
+# Stable Vicuna default template
+# source: https://huggingface.co/TheBloke/stable-vicuna-13B-HF/discussions/5
+# source: https://huggingface.co/spaces/CarperAI/StableVicuna/blob/main/app.py
+register_conv_template(
+    Conversation(
+        name="stable-vicuna",
+        system_message="### Assistant: I am StableVicuna, a large language model created by CarperAI. I am here to chat!\n",
+        roles=("### Human", "### Assistant"),
+        sep_style=SeparatorStyle.ADD_COLON_TWO,
+        sep="\n",
+        sep2="\n\n",
+    )
+)
+
 register_conv_template(
     Conversation(
         name="vigogne_chat_v3",

fastchat/model/model_adapter.py

@@ -1829,6 +1829,22 @@ def get_default_conv_template(self, model_path: str) -> Conversation:
         return get_conv_template("llama-2")
 
 
+class StableVicunaAdapter(BaseModelAdapter):
+    """The model adapter for StableVicuna"""
+
+    def match(self, model_path: str):
+        return "stable-vicuna" in model_path.lower()
+
+    def load_model(self, model_path: str, from_pretrained_kwargs: dict):
+        model, tokenizer = super().load_model(model_path, from_pretrained_kwargs)
+        model.config.eos_token_id = tokenizer.eos_token_id
+        model.config.pad_token_id = tokenizer.pad_token_id
+        return model, tokenizer
+
+    def get_default_conv_template(self, model_path: str) -> Conversation:
+        return get_conv_template("stable-vicuna")
+
+
 class PhindCodeLlamaAdapter(CodeLlamaAdapter):
     """The model adapter for Phind-CodeLlama (e.g., Phind/Phind-CodeLlama-34B-v2)"""
 
@@ -1913,6 +1929,7 @@ def get_default_conv_template(self, model_path: str) -> Conversation:
 # The one registered earlier has a higher matching priority.
 register_model_adapter(RerankAdapter)
 register_model_adapter(PeftModelAdapter)
+register_model_adapter(StableVicunaAdapter)
 register_model_adapter(VicunaAdapter)
 register_model_adapter(AiroborosAdapter)
 register_model_adapter(LongChatAdapter)

fastchat/model/model_registry.py

@@ -348,6 +348,12 @@ def get_model_info(name: str) -> ModelInfo:
     "https://huggingface.co/bofenghuang/vigogne-2-7b-chat",
     "Vigogne-Chat is a French large language model (LLM) optimized for instruction-following and multi-turn dialogues, developed by Bofeng Huang",
 )
+register_model_info(
+    ["stable-vicuna-13B-HF"],
+    "stable-vicuna",
+    "https://huggingface.co/TheBloke/stable-vicuna-13B-HF",
+    "StableVicuna is a Vicuna model fine-tuned using RLHF via PPO on various conversational and instructional datasets.",
+)
 register_model_info(
     ["deluxe-chat-v1", "deluxe-chat-v1.1"],
     "DeluxeChat",