feat(model): Support Phi-3 models (#1554)

Author: fangyinc

README.md

@@ -158,6 +158,7 @@ At present, we have introduced several key features to showcase our current capa
   We offer extensive model support, including dozens of large language models (LLMs) from both open-source and API agents, such as LLaMA/LLaMA2, Baichuan, ChatGLM, Wenxin, Tongyi, Zhipu, and many more. 
 
   - News
+    - 🔥🔥🔥  [Phi-3](https://huggingface.co/collections/microsoft/phi-3-6626e15e9585a200d2d761e3)
     - 🔥🔥🔥  [Yi-1.5-34B-Chat](https://huggingface.co/01-ai/Yi-1.5-34B-Chat)
     - 🔥🔥🔥  [Yi-1.5-9B-Chat](https://huggingface.co/01-ai/Yi-1.5-9B-Chat)
     - 🔥🔥🔥  [Yi-1.5-6B-Chat](https://huggingface.co/01-ai/Yi-1.5-6B-Chat)

README.zh.md

@@ -152,6 +152,7 @@
   海量模型支持，包括开源、API代理等几十种大语言模型。如LLaMA/LLaMA2、Baichuan、ChatGLM、文心、通义、智谱等。当前已支持如下模型: 
 
   - 新增支持模型
+    - 🔥🔥🔥  [Phi-3](https://huggingface.co/collections/microsoft/phi-3-6626e15e9585a200d2d761e3)
     - 🔥🔥🔥  [Yi-1.5-34B-Chat](https://huggingface.co/01-ai/Yi-1.5-34B-Chat)
     - 🔥🔥🔥  [Yi-1.5-9B-Chat](https://huggingface.co/01-ai/Yi-1.5-9B-Chat)
     - 🔥🔥🔥  [Yi-1.5-6B-Chat](https://huggingface.co/01-ai/Yi-1.5-6B-Chat)

dbgpt/configs/model_config.py

@@ -187,6 +187,16 @@ def get_device() -> str:
     "gemma-2b-it": os.path.join(MODEL_PATH, "gemma-2b-it"),
     "starling-lm-7b-beta": os.path.join(MODEL_PATH, "Starling-LM-7B-beta"),
     "deepseek-v2-lite-chat": os.path.join(MODEL_PATH, "DeepSeek-V2-Lite-Chat"),
+    "sailor-14b-chat": os.path.join(MODEL_PATH, "Sailor-14B-Chat"),
+    # https://huggingface.co/microsoft/Phi-3-medium-128k-instruct
+    "phi-3-medium-128k-instruct": os.path.join(
+        MODEL_PATH, "Phi-3-medium-128k-instruct"
+    ),
+    "phi-3-medium-4k-instruct": os.path.join(MODEL_PATH, "Phi-3-medium-4k-instruct"),
+    "phi-3-small-128k-instruct": os.path.join(MODEL_PATH, "Phi-3-small-128k-instruct"),
+    "phi-3-small-8k-instruct": os.path.join(MODEL_PATH, "Phi-3-small-8k-instruct"),
+    "phi-3-mini-128k-instruct": os.path.join(MODEL_PATH, "Phi-3-mini-128k-instruct"),
+    "phi-3-mini-4k-instruct": os.path.join(MODEL_PATH, "Phi-3-mini-4k-instruct"),
 }
 
 EMBEDDING_MODEL_CONFIG = {

dbgpt/model/adapter/hf_adapter.py

@@ -396,6 +396,61 @@ def load(self, model_path: str, from_pretrained_kwargs: dict):
         return model, tokenizer
 
 
+class SailorAdapter(QwenAdapter):
+    """
+    https://huggingface.co/sail/Sailor-14B-Chat
+    """
+
+    def do_match(self, lower_model_name_or_path: Optional[str] = None):
+        return (
+            lower_model_name_or_path
+            and "sailor" in lower_model_name_or_path
+            and "chat" in lower_model_name_or_path
+        )
+
+
+class PhiAdapter(NewHFChatModelAdapter):
+    """
+    https://huggingface.co/microsoft/Phi-3-medium-128k-instruct
+    """
+
+    support_4bit: bool = True
+    support_8bit: bool = True
+    support_system_message: bool = False
+
+    def do_match(self, lower_model_name_or_path: Optional[str] = None):
+        return (
+            lower_model_name_or_path
+            and "phi-3" in lower_model_name_or_path
+            and "instruct" in lower_model_name_or_path
+        )
+
+    def load(self, model_path: str, from_pretrained_kwargs: dict):
+        if not from_pretrained_kwargs:
+            from_pretrained_kwargs = {}
+        if "trust_remote_code" not in from_pretrained_kwargs:
+            from_pretrained_kwargs["trust_remote_code"] = True
+        return super().load(model_path, from_pretrained_kwargs)
+
+    def get_str_prompt(
+        self,
+        params: Dict,
+        messages: List[ModelMessage],
+        tokenizer: Any,
+        prompt_template: str = None,
+        convert_to_compatible_format: bool = False,
+    ) -> Optional[str]:
+        str_prompt = super().get_str_prompt(
+            params,
+            messages,
+            tokenizer,
+            prompt_template,
+            convert_to_compatible_format,
+        )
+        params["custom_stop_words"] = ["<|end|>"]
+        return str_prompt
+
+
 # The following code is used to register the model adapter
 # The last registered model adapter is matched first
 register_model_adapter(YiAdapter)
@@ -408,3 +463,5 @@ def load(self, model_path: str, from_pretrained_kwargs: dict):
 register_model_adapter(QwenMoeAdapter)
 register_model_adapter(Llama3Adapter)
 register_model_adapter(DeepseekV2Adapter)
+register_model_adapter(SailorAdapter)
+register_model_adapter(PhiAdapter)

dbgpt/model/llm_out/hf_chat_llm.py

@@ -22,6 +22,7 @@ def huggingface_chat_generate_stream(
     max_new_tokens = int(params.get("max_new_tokens", 2048))
     stop_token_ids = params.get("stop_token_ids", [])
     do_sample = params.get("do_sample", None)
+    custom_stop_words = params.get("custom_stop_words", [])
 
     input_ids = tokenizer(prompt).input_ids
     # input_ids = input_ids.to(device)
@@ -62,4 +63,8 @@ def huggingface_chat_generate_stream(
     out = ""
     for new_text in streamer:
         out += new_text
+        if custom_stop_words:
+            for stop_word in custom_stop_words:
+                if out.endswith(stop_word):
+                    out = out[: -len(stop_word)]
         yield out