feat(translator): 添加transformers作为备选模型加载方式

支持使用transformers库作为llama-cpp-python的备选方案加载模型
新增模型加载失败时的自动回退机制
更新requirements.txt添加transformers和torch依赖

models/Modelfile

@@ -0,0 +1,7 @@
+FROM ./HY-MT1.5-1.8B_bf16_Q4_K_M.gguf
+
+PARAMETER temperature 0.7
+PARAMETER top_p 0.95
+PARAMETER num_ctx 2048
+
+SYSTEM 你是一个专业的翻译助手，根据以下要求将中文翻译成英文。

requirements.txt

@@ -4,3 +4,5 @@ python-docx
 loguru
 psutil
 GPUtil
+transformers
+torch

translator.py

@@ -6,10 +6,18 @@ try:
     from llama_cpp import Llama
     llama_cpp_available = True
 except ImportError:
-    logger.warning("llama-cpp-python库未找到，将禁用翻译功能")
+    logger.warning("llama-cpp-python库未找到，将尝试使用transformers库")
     Llama = None
     llama_cpp_available = False
 
+# 尝试导入transformers库
+try:
+    from transformers import AutoModelForCausalLM, AutoTokenizer
+    transformers_available = True
+except ImportError:
+    logger.warning("transformers库未找到")
+    transformers_available = False
+
 class Translator:
     def __init__(self, model_path=None):
         self.model = None
@@ -17,13 +25,11 @@ class Translator:
         self.is_ready = False
         self.model_name = ""
         self.llama_cpp_available = llama_cpp_available
+        self.transformers_available = transformers_available
+        self.use_transformers = False
         
     def load_model(self, model_path=None):
         """加载模型"""
-        if not self.llama_cpp_available:
-            logger.error("llama-cpp-python库未找到，无法加载模型")
-            return False
-            
         if model_path:
             self.model_path = model_path
         
@@ -31,22 +37,73 @@ class Translator:
             logger.error("未提供模型路径")
             return False
         
-        if not os.path.exists(self.model_path):
-            logger.error(f"模型文件不存在: {self.model_path}")
-            return False
-        
         try:
             logger.info(f"开始加载模型: {self.model_path}")
+            
+            if self.llama_cpp_available:
+                if os.path.exists(self.model_path):
+                    try:
                         self.model = Llama(
                             model_path=self.model_path,
                             n_ctx=2048,
                             n_threads=4,
-                n_gpu_layers=100  # 尽可能使用GPU加速
+                            n_gpu_layers=100
                         )
+                        self.use_transformers = False
                         self.is_ready = True
                         self.model_name = os.path.basename(self.model_path)
                         logger.info(f"模型加载成功: {self.model_name}")
                         return True
+                    except Exception as e:
+                        logger.warning(f"使用llama-cpp-python加载模型失败: {e}")
+                else:
+                    logger.warning(f"模型文件不存在: {self.model_path}")
+            
+            if self.transformers_available:
+                try:
+                    from transformers import AutoConfig
+                    if os.path.exists(self.model_path):
+                        config = AutoConfig.from_pretrained(
+                            self.model_path,
+                            local_files_only=True,
+                            trust_remote_code=True
+                        )
+                        self.model = AutoModelForCausalLM.from_pretrained(
+                            self.model_path,
+                            local_files_only=True,
+                            trust_remote_code=True,
+                            torch_dtype="auto"
+                        )
+                        self.tokenizer = AutoTokenizer.from_pretrained(
+                            self.model_path,
+                            local_files_only=True,
+                            trust_remote_code=True
+                        )
+                    else:
+                        config = AutoConfig.from_pretrained(
+                            self.model_path,
+                            trust_remote_code=True
+                        )
+                        self.model = AutoModelForCausalLM.from_pretrained(
+                            self.model_path,
+                            trust_remote_code=True,
+                            torch_dtype="auto"
+                        )
+                        self.tokenizer = AutoTokenizer.from_pretrained(
+                            self.model_path,
+                            trust_remote_code=True
+                        )
+                    self.use_transformers = True
+                    self.is_ready = True
+                    self.model_name = self.model_path
+                    logger.info(f"使用transformers加载模型成功: {self.model_name}")
+                    return True
+                except Exception as e:
+                    logger.error(f"使用transformers加载模型失败: {e}")
+                    return False
+            
+            logger.error("没有可用的模型加载方式")
+            return False
         except Exception as e:
             logger.error(f"模型加载失败: {e}")
             self.is_ready = False
@@ -54,10 +111,6 @@ class Translator:
     
     def translate(self, text, context="", terms=None):
         """执行翻译"""
-        if not self.llama_cpp_available:
-            logger.error("llama-cpp-python库未找到，无法执行翻译")
-            return ""
-            
         if not self.is_ready or not self.model:
             logger.error("模型未就绪，无法执行翻译")
             return ""
@@ -68,7 +121,19 @@ class Translator:
             
             logger.info(f"开始翻译，输入长度: {len(text)} 字符")
             
-            # 调用模型进行翻译
+            if self.use_transformers:
+                import torch
+                inputs = self.tokenizer(prompt, return_tensors="pt")
+                with torch.no_grad():
+                    outputs = self.model.generate(
+                        **inputs,
+                        max_new_tokens=2048,
+                        temperature=0.7,
+                        top_p=0.95,
+                        do_sample=True
+                    )
+                translated_text = self.tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
+            else:
                 output = self.model(
                     prompt,
                     max_tokens=2048,
@@ -76,8 +141,8 @@ class Translator:
                     top_p=0.95,
                     stop=["\n原文:", "\n译文:", "\n###"]
                 )
-            
                 translated_text = output["choices"][0]["text"].strip()
+            
             logger.info(f"翻译完成，输出长度: {len(translated_text)} 字符")
             
             return translated_text