Update app.py

app.py

@@ -156,8 +156,8 @@ def main():
 
     if model_args.pre_seq_len is not None:
         # P-tuning v2
-        model = model.half().cuda()
-        model.transformer.prefix_encoder.float().cuda()
+        model = model.half()
+        model.transformer.prefix_encoder.float()
     
     model = model.eval()
     demo.queue().launch(share=False, inbrowser=True)