Custom LR schedulers give an error during training

My friend used Kohya's scripts to import custom schedulers like CosineAnnealingWarmRestarts from PyTorch directly, but they give an error. Below is a file with error text and what parameters were used.
[ValueError.txt](https://github.com/kohya-ss/sd-scripts/files/12569720/ValueError.txt)

Start with the attached ValueError.txt and the parameters shown in the issue image, then trace the training path that loads custom PyTorch schedulers such as CosineAnnealingWarmRestarts. Reproduce the reported error with those settings and identify the scheduler-loading entry point. Done means custom schedulers can be imported and used during training without the reported ValueError.

Written by the indexing model from the issue text.