Update app.py

app.py

@@ -12,8 +12,9 @@ CACHE_DIR = "/home/user/cache"
 os.makedirs(CACHE_DIR, exist_ok=True)
 
 # 减少内存占用
-os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "max_split_size_mb:32"
-torch.set_num_threads(1)
+# os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "max_split_size_mb:32"
+# torch.set_num_threads(1)
+os.environ["OMP_NUM_THREADS"] = "8"
 
 # 全局变量 - 移除了clip_model和clip_processor
 index = None
@@ -71,6 +72,8 @@ def load_resources():
                 print(f"❌ 元数据加载失败: {str(e)}")
                 raise
 
+from functools import lru_cache
+@lru_cache(maxsize=1000)
 def predict(vector):
     """处理768维向量输入并返回答案"""
     print(f"输入向量维度: {np.array(vector).shape}")