Update app.py

app.py

@@ -85,7 +85,7 @@ def predict(message, history, system_prompt, temperature, max_new_tokens, top_k,
         attention_mask = attention_mask[:, -CONTEXT_LENGTH:]
 
     generate_kwargs = dict(
-        input_ids=input_ids,
+        input_ids=input_ids.to(device),
         attention_mask=attention_mask,
         streamer=streamer,
         do_sample=True,