feat: 加入通义千问支持

modules/models/Qwen.py

@@ -0,0 +1,57 @@
+from transformers import AutoModelForCausalLM, AutoTokenizer
+from transformers.generation import GenerationConfig
+import logging
+import colorama
+from .base_model import BaseLLMModel
+from ..presets import MODEL_METADATA
+
+
+class Qwen_Client(BaseLLMModel):
+    def __init__(self, model_name, user_name="") -> None:
+        super().__init__(model_name=model_name, user=user_name)
+        self.tokenizer = AutoTokenizer.from_pretrained(MODEL_METADATA[model_name]["repo_id"], trust_remote_code=True, resume_download=True)
+        self.model = AutoModelForCausalLM.from_pretrained(MODEL_METADATA[model_name]["repo_id"], device_map="auto", trust_remote_code=True, resume_download=True).eval()
+
+    def generation_config(self):
+        return GenerationConfig.from_dict({
+            "chat_format": "chatml",
+            "do_sample": True,
+            "eos_token_id": 151643,
+            "max_length": self.token_upper_limit,
+            "max_new_tokens": 512,
+            "max_window_size": 6144,
+            "pad_token_id": 151643,
+            "top_k": 0,
+            "top_p": self.top_p,
+            "transformers_version": "4.33.2",
+            "trust_remote_code": True,
+            "temperature": self.temperature,
+            })
+
+    def _get_glm_style_input(self):
+        history = [x["content"] for x in self.history]
+        query = history.pop()
+        logging.debug(colorama.Fore.YELLOW +
+                      f"{history}" + colorama.Fore.RESET)
+        assert (
+            len(history) % 2 == 0
+        ), f"History should be even length. current history is: {history}"
+        history = [[history[i], history[i + 1]]
+                   for i in range(0, len(history), 2)]
+        return history, query
+
+    def get_answer_at_once(self):
+        history, query = self._get_glm_style_input()
+        self.model.generation_config = self.generation_config()
+        response, history = self.model.chat(self.tokenizer, query, history=history)
+        return response, len(response)
+
+    def get_answer_stream_iter(self):
+        history, query = self._get_glm_style_input()
+        self.model.generation_config = self.generation_config()
+        for response in self.model.chat_stream(
+                self.tokenizer,
+                query,
+                history,
+            ):
+                yield response

modules/models/base_model.py

@@ -146,6 +146,7 @@ class ModelType(Enum):
     Spark = 12
     OpenAIInstruct = 13
     Claude = 14
+    Qwen = 15
 
     @classmethod
     def get_type(cls, model_name: str):
@@ -181,7 +182,9 @@ class ModelType(Enum):
         elif "星火大模型" in model_name_lower:
             model_type = ModelType.Spark
         elif "claude" in model_name_lower:
-            model_type = ModelType.Claude        
+            model_type = ModelType.Claude
+        elif "qwen" in model_name_lower:
+            model_type = ModelType.Qwen
         else:
             model_type = ModelType.LLaMA
         return model_type
@@ -656,14 +659,13 @@ class BaseLLMModel:
     def delete_last_conversation(self, chatbot):
         if len(chatbot) > 0 and STANDARD_ERROR_MSG in chatbot[-1][1]:
             msg = "由于包含报错信息，只删除chatbot记录"
-            chatbot.pop()
+            chatbot = chatbot[:-1]
             return chatbot, self.history
         if len(self.history) > 0:
-            self.history.pop()
-            self.history.pop()
+            self.history = self.history[:-2]
         if len(chatbot) > 0:
             msg = "删除了一组chatbot对话"
-            chatbot.pop()
+            chatbot = chatbot[:-1]
         if len(self.all_token_counts) > 0:
             msg = "删除了一组对话的token计数记录"
             self.all_token_counts.pop()

modules/models/models.py

@@ -116,9 +116,12 @@ def get_model(
             from .spark import Spark_Client
             model = Spark_Client(model_name, os.getenv("SPARK_APPID"), os.getenv(
                 "SPARK_API_KEY"), os.getenv("SPARK_API_SECRET"), user_name=user_name)
-         elif model_type == ModelType.Claude:
+        elif model_type == ModelType.Claude:
             from .Claude import Claude_Client
             model = Claude_Client(model_name="claude-2", api_secret=os.getenv("CLAUDE_API_SECRET"))
+        elif model_type == ModelType.Qwen:
+            from .Qwen import Qwen_Client
+            model = Qwen_Client(model_name, user_name=user_name)
         elif model_type == ModelType.Unknown:
             raise ValueError(f"未知模型: {model_name}")
         logging.info(msg)

modules/presets.py

@@ -87,6 +87,8 @@ LOCAL_MODELS = [
     "StableLM",
     "MOSS",
     "Llama-2-7B-Chat",
+    "Qwen 7B",
+    "Qwen 14B"
 ]
 
 # Additional metadate for local models
@@ -98,6 +100,12 @@ MODEL_METADATA = {
     "Llama-2-7B-Chat":{
         "repo_id": "TheBloke/Llama-2-7b-Chat-GGUF",
         "filelist": ["llama-2-7b-chat.Q6_K.gguf"],
+    },
+    "Qwen 7B": {
+        "repo_id": "Qwen/Qwen-7B-Chat-Int4",
+    },
+    "Qwen 14B": {
+        "repo_id": "Qwen/Qwen-14B-Chat-Int4",
     }
 }
 

requirements_advanced.txt

@@ -6,3 +6,7 @@ sentence_transformers
 accelerate
 sentencepiece
 llama-cpp-python
+transformers_stream_generator
+einops
+optimum
+auto-gptq