Update app.py

app.py

@@ -82,7 +82,9 @@ def loadModel():
                 lora_weights,
                 device_map="auto",
                 cache_dir=cache_dir,
-                torch_dtype=torch.float16
+                torch_dtype=torch.float16,
+                assign=True,
+                is_trainable=False,
             )
     model.eval()
     return model