Add support for Vigogne models (#2236)

docs/model_support.md

@@ -36,6 +36,8 @@
 - [Qwen/Qwen-7B-Chat](https://huggingface.co/Qwen/Qwen-7B-Chat)
 - [BAAI/AquilaChat-7B](https://huggingface.co/BAAI/AquilaChat-7B)
 - [HuggingFaceH4/starchat-beta](https://huggingface.co/HuggingFaceH4/starchat-beta)
+- [bofenghuang/vigogne-2-7b-instruct](https://huggingface.co/bofenghuang/vigogne-2-7b-instruct)
+- [bofenghuang/vigogne-2-7b-chat](https://huggingface.co/bofenghuang/vigogne-2-7b-chat)
 - Any [EleutherAI](https://huggingface.co/EleutherAI) pythia model such as [pythia-6.9b](https://huggingface.co/EleutherAI/pythia-6.9b)
 - Any [Peft](https://github.com/huggingface/peft) adapter trained on top of a
   model above.  To activate, must have `peft` in the model path.  Note: If

fastchat/conversation.py

@@ -952,6 +952,24 @@ def get_conv_template(name: str) -> Conversation:
     )
 )
 
+# Vigogne Chat default template
+# source: https://github.com/bofenghuang/vigogne
+register_conv_template(
+    Conversation(
+        name="vigogne-chat",
+        system_template="<|system|>: {system_message}",
+        system_message="Vous êtes l'assistant IA nommé Vigogne, créé par Zaion Lab (https://zaion.ai). "
+        "Vous suivez extrêmement bien les instructions. Aidez autant que vous le pouvez.",
+        roles=("<|user|>", "<|assistant|>"),
+        messages=(),
+        offset=0,
+        sep_style=SeparatorStyle.ADD_COLON_TWO,
+        sep="\n",
+        sep2="</s>\n",
+        stop_str="<|user|>",
+    )
+)
+
 if __name__ == "__main__":
     print("Vicuna template:")
     conv = get_conv_template("vicuna_v1.1")

fastchat/model/model_adapter.py

@@ -1407,6 +1407,62 @@ def get_default_conv_template(self, model_path: str) -> Conversation:
         return get_conv_template("llama2-chinese")
 
 
+class VigogneInstructAdapter(BaseModelAdapter):
+    """The model adapter for Vigogne-Instruct"""
+
+    use_fast_tokenizer = False
+
+    def match(self, model_path: str):
+        return "vigogne" in model_path.lower() and "instruct" 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,
+            use_fast=self.use_fast_tokenizer,
+            trust_remote_code=True,
+            revision=revision,
+        )
+        model = AutoModelForCausalLM.from_pretrained(
+            model_path,
+            trust_remote_code=True,
+            low_cpu_mem_usage=True,
+            **from_pretrained_kwargs,
+        ).eval()
+        return model, tokenizer
+
+    def get_default_conv_template(self, model_path: str) -> Conversation:
+        return get_conv_template("alpaca")
+
+
+class VigogneChatAdapter(BaseModelAdapter):
+    """The model adapter for Vigogne-Chat"""
+
+    use_fast_tokenizer = False
+
+    def match(self, model_path: str):
+        return "vigogne" in model_path.lower() and "chat" 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,
+            use_fast=self.use_fast_tokenizer,
+            trust_remote_code=True,
+            revision=revision,
+        )
+        model = AutoModelForCausalLM.from_pretrained(
+            model_path,
+            trust_remote_code=True,
+            low_cpu_mem_usage=True,
+            **from_pretrained_kwargs,
+        ).eval()
+        return model, tokenizer
+
+    def get_default_conv_template(self, model_path: str) -> Conversation:
+        return get_conv_template("vigogne-chat")
+
+
 # Note: the registration order matters.
 # The one registered earlier has a higher matching priority.
 register_model_adapter(PeftModelAdapter)
@@ -1458,6 +1514,8 @@ def get_default_conv_template(self, model_path: str) -> Conversation:
 register_model_adapter(AquilaChatAdapter)
 register_model_adapter(BGEAdapter)
 register_model_adapter(Lamma2ChineseAdapter)
+register_model_adapter(VigogneInstructAdapter)
+register_model_adapter(VigogneChatAdapter)
 
 # After all adapters, try the default base adapter.
 register_model_adapter(BaseModelAdapter)

fastchat/model/model_registry.py

@@ -254,3 +254,15 @@ def get_model_info(name: str) -> ModelInfo:
     "https://huggingface.co/FlagAlpha/Llama2-Chinese-13b-Chat",
     "Llama2-Chinese is a multi-language large-scale language model (LLM), developed by FlagAlpha.",
 )
+register_model_info(
+    ["Vigogne-2-7B-Instruct", "Vigogne-2-13B-Instruct"],
+    "Vigogne-Instruct",
+    "https://huggingface.co/bofenghuang/vigogne-2-7b-instruct",
+    "Vigogne-Instruct is a French large language model (LLM) optimized for instruction-following, developed by Bofeng Huang",
+)
+register_model_info(
+    ["Vigogne-2-7B-Chat", "Vigogne-2-13B-Chat"],
+    "Vigogne-Chat",
+    "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",
+)