fix

README.md

@@ -24,9 +24,10 @@ hf_oauth: true
 
 ## 功能特性
 
-- 💬 实时对话：支持多轮对话，保持上下文
-- 🧠 金融专业：基于Llama 3-8B微调的金融领域模型
+- 💬 智能对话：基于金融领域微调的对话系统
+- 🧠 金融专业：使用Llama 3-8B + LoRA适配器
 - 🚀 GPU加速：使用Hugging Face Spaces的GPU支持
+- 💾 智能缓存：模型文件本地缓存，加速启动
 
 ## 使用说明
 

app.py

@@ -1,7 +1,7 @@
 import gradio as gr
 import spaces
 import torch
-from transformers import AutoTokenizer, AutoModelForCausalLM
+from transformers import LlamaTokenizerFast, LlamaForCausalLM
 from peft import PeftModel
 import os
 
@@ -25,26 +25,30 @@ model_loaded = False
 
 try:
     print("\n[1/3] 加载tokenizer...")
-    tokenizer = AutoTokenizer.from_pretrained(
+    tokenizer = LlamaTokenizerFast.from_pretrained(
         model_name,
         trust_remote_code=True,
-        token=hf_token
+        token=hf_token,
     )
     tokenizer.pad_token = tokenizer.eos_token
     print("✓ Tokenizer加载成功")
 
     print("\n[2/3] 加载基础模型...")
-    base_model = AutoModelForCausalLM.from_pretrained(
+    base_model = LlamaForCausalLM.from_pretrained(
         model_name,
         torch_dtype=torch.float16,
         device_map="auto",
         trust_remote_code=True,
-        token=hf_token
+        token=hf_token,
+        cache_dir=cache_dir
     )
-    print("✓ 基础模型加载成功")
-
     print("\n[3/3] 加载LoRA适配器...")
-    model = PeftModel.from_pretrained(base_model, adapter_name)
+    model = PeftModel.from_pretrained(base_model, adapter_name, cache_dir=cache_dir)
+    model = model.eval()
+
+    # 确保模型在正确的设备上
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    print(f"✓ LoRA适配器加载成功 (设备: {device})")_pretrained(base_model, adapter_name)
     model = model.eval()
     print("✓ LoRA适配器加载成功")