I think if we can change the llama output layer, maybe we can get logs.
j 下一段next speechk 上一段previous speech