Add deepseek chat (#2760)

docs/model_support.md

@@ -54,6 +54,8 @@
 - [OpenLemur/lemur-70b-chat-v1](https://huggingface.co/OpenLemur/lemur-70b-chat-v1)
 - [allenai/tulu-2-dpo-7b](https://huggingface.co/allenai/tulu-2-dpo-7b)
 - [Microsoft/Orca-2-7b](https://huggingface.co/microsoft/Orca-2-7b)
+- [deepseek-ai/deepseek-llm-67b-chat](https://huggingface.co/deepseek-ai/deepseek-llm-67b-chat)
+- [deepseek-ai/deepseek-coder-33b-instruct](https://huggingface.co/deepseek-ai/deepseek-coder-33b-instruct)
 - 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

@@ -29,6 +29,7 @@ class SeparatorStyle(IntEnum):
     ROBIN = auto()
     FALCON_CHAT = auto()
     CHATGLM3 = auto()
+    DEEPSEEK_CHAT = auto()
 
 
 @dataclasses.dataclass
@@ -224,6 +225,15 @@ def get_prompt(self) -> str:
                     ret += role + ":"
 
             return ret
+        elif self.sep_style == SeparatorStyle.DEEPSEEK_CHAT:
+            seps = [self.sep, self.sep2]
+            ret = system_prompt
+            for i, (role, message) in enumerate(self.messages):
+                if message:
+                    ret += role + ": " + message + seps[i % 2]
+                else:
+                    ret += role + ":"
+            return ret
         else:
             raise ValueError(f"Invalid style: {self.sep_style}")
 
@@ -530,7 +540,7 @@ def get_conv_template(name: str) -> Conversation:
 # Deepseek code default template
 register_conv_template(
     Conversation(
-        name="deepseek",
+        name="deepseek-coder",
         system_template="You are an AI programming assistant, utilizing the DeepSeek Coder model, developed by DeepSeek Company, and you only answer questions related to computer science. For politically sensitive questions, security and privacy issues, and other non-computer science questions, you will refuse to answer.",
         roles=("### Instruction:", "### Response:"),
         sep="\n",
@@ -1265,6 +1275,20 @@ def get_conv_template(name: str) -> Conversation:
     )
 )
 
+# Deepseek-chat template
+# reference: https://huggingface.co/deepseek-ai/deepseek-llm-67b-chat/blob/main/tokenizer_config.json
+register_conv_template(
+    Conversation(
+        name="deepseek-chat",
+        system_message="<｜begin▁of▁sentence｜>",  # must add a bos token before first message
+        roles=("User", "Assistant"),
+        sep_style=SeparatorStyle.DEEPSEEK_CHAT,
+        sep="\n\n",
+        sep2="<｜end▁of▁sentence｜>",
+        stop_str="<｜end▁of▁sentence｜>",
+    )
+)
+
 if __name__ == "__main__":
     from fastchat.conversation import get_conv_template
 

fastchat/model/model_adapter.py

@@ -1934,6 +1934,28 @@ def get_default_conv_template(self, model_path: str) -> Conversation:
         return get_conv_template("Yi-34b-chat")
 
 
+class DeepseekCoderAdapter(BaseModelAdapter):
+    """The model adapter for deepseek-ai's coder models"""
+
+    def match(self, model_path: str):
+        return "deepseek-coder" in model_path.lower()
+
+    def get_default_conv_template(self, model_path: str) -> Conversation:
+        return get_conv_template("deepseek-coder")
+
+
+class DeepseekChatAdapter(BaseModelAdapter):
+    """The model adapter for deepseek-ai's chat models"""
+
+    # Note: that this model will require tokenizer version >= 0.13.3 because the tokenizer class is LlamaTokenizerFast
+
+    def match(self, model_path: str):
+        return "deepseek-llm" in model_path.lower() and "chat" in model_path.lower()
+
+    def get_default_conv_template(self, model_path: str) -> Conversation:
+        return get_conv_template("deepseek-chat")
+
+
 # Note: the registration order matters.
 # The one registered earlier has a higher matching priority.
 register_model_adapter(PeftModelAdapter)
@@ -2005,6 +2027,8 @@ def get_default_conv_template(self, model_path: str) -> Conversation:
 register_model_adapter(PygmalionAdapter)
 register_model_adapter(MicrosoftOrcaAdapter)
 register_model_adapter(YiAdapter)
+register_model_adapter(DeepseekCoderAdapter)
+register_model_adapter(DeepseekChatAdapter)
 
 # After all adapters, try the default base adapter.
 register_model_adapter(BaseModelAdapter)