请问作者大大，encoder如何训练？

看了知乎链接的教程，尝试训练encoder
练了半天，这结果似乎没什么变化
数据自建的，有2个多G
问题1：这种情况是正常的吗？如果不正常是什么原因造成的？
问题2：根据知乎上的说法“**实测了一次 训练synthesizer时，4000左右step就能attention收敛，22k step的时候loss就到0.35了，可以很快进行finetune，算是超越预期。**”，训练synthesizer时，如何把encoder加入？

Start by reviewing the referenced Zhihu tutorial and reproducing encoder training with the reported self-built dataset, comparing the two UMAP screenshots and training behavior. Then trace the repository's encoder and synthesizer training workflows to determine whether encoder outputs are already included; done means documenting the expected result, likely causes of no change, and the supported integration steps.

Written by the indexing model from the issue text.