Update app.py

app.py

@@ -146,7 +146,7 @@ def gradient_cuff_reject(message,with_defense, sample_times,perturb_times,thresh
         original_embedding,shift_embeddings,prefix_embedding,suffix_embedding
     )
     with torch.no_grad():
-        responses = engine(input_embeds)
+        responses = engine(input_embeds.to(torch.bfloat16))
     labels=get_labels(responses)
     results=[None for _ in range(perturb_times+1)]
     results[0]=sum(labels)/len(labels)