May be a bug?

[xiao10ma]

https://github.com/THUDM/CogVideo/blob/main/finetune/models/cogvideox_t2v/lora_trainer.py#L111

Why first_frame.repeat(1, 1, ncopy, 1, 1)?

I think that it should be repeating patch_size_t - ncopy times so that the assert latent.shape[2] % patch_size_t == 0 can be realized. Here the patch_size_t is set to 2, so it may not raise an error. But if it is 4, and the latent't frame is 13, then it causes error.