Don't create TE layers that won't be trained or used.

Nerogar / OneTrainer

OneTrainer is a one-stop solution for all your stable diffusion training needs.

GNU Affero General Public License v3.0

1.34k stars 110 forks source link

Don't create TE layers that won't be trained or used. #359

Open mx opened 1 week ago

mx commented 1 week ago

Should work for any combination of:

Train text encoder or not.
Do not load a lora.
Load a lora with a trained text encoder.
Load a lora with an untrained text encoder.

Saves a (possibly significant) amount of output disk space.

hameerabbasi commented 1 week ago

Question: Would it be possible to generalise this to all model components? Unet as well, etc.

mx commented 6 days ago

Question: Would it be possible to generalise this to all model components? Unet as well, etc.

Better addressed in another pull request. Possibly yes, for unet/transformer will need to look through the rest of the codebase for any assumptions that they are present though.