Seeking help: How to improve the efficiency of the training process?

Hello!

I've want to train the PreDiff on my own dataset. I'm trying to predict data for the next 24 times using data from the previous 12 times. During the training of the model, I found that the validation on the validation set should have completed all 1000 noise reduction steps (the default value), this would take a lot of time. Are there any methods to improve model efficiency in this situation?

Additionally, since I'm training the model on a single GPU, the default model settings result in a large model size, making training on a single GPU quite difficult. Therefore, I'd like to reduce the model size. Could you advise on a suitable method to reduce the model's complexity?

Looking forward to your reply! Thanks!