[Bug(?)] Training takes much longer when using --cache_latents_to_disk

[Linaqruf]

Hi! Thanks for implementing --cache_latents_to_disk. It saves a lot of time as we don't need to use prepare_buckets_and_latents.py to save .npz files to disk, but instead we can save them directly when training.

However, I'm encountering a problem when using the --cache_latents_to_disk argument. The training takes longer time (it/s) than using the usual --cache_latents. Is this intended or a bug? Because, I thought that latent caching has its own process that is separate from the training.

Using --cache_latents_to_disk

Using --cache_latents

config_file.toml : https://dpaste.com/B6AVPNK2K#

Thank you!

[kohya-ss]

--cache_latents option is also necessary to enable the disk caching, not only --cache_latents_to_disk. Sorry for confusion.

I will add the feature to enable --cache_latents option automatically when --cache_latents_to_disk is specified.

[Linaqruf]

Oh! I'm sorry, I thought it was separate args 😅
Thank you for explanation.