code question

[lclszsdnr]

作者您好，今天仔细读了一下您的代码，有个困惑在于训练时每次都需要额外加载一次clip的单独权重，我看您代码中有注释
TODO: In the new HiLoRA stage, the CLIP model has changed, and the CLIP parameters from the previous stage
# can no longer be loaded. They need to be loaded separately
但是我没有很明白，之前的全局权重中不就包含了clip的权重吗，而且为什么说CLIP会改变呢，不是在不同阶段仅有训练参数的不同，但是结构是相同的吗。

[linhuixiao]

@lclszsdnr Hi, you can perform some experiments and print out the model structure. Then you will find that using LoRA will add a structure layer outside the original network structure, and directly loading the structure will not match.