[Bug]: There is a bug when saving diffusers backups

[FurkanGozukara]

the error obviously happening due to / \ mismatch in folder formatting on Windows 10 @Nerogar

caching resolutions:   0%|                                                                                        | 0/30 [00:00<?, ?it/s]
Creating Backup F:/onetrainer_workspace_diffusers_test\backup\2024-04-14_20-45-11-backup-870-29-0                  | 0/30 [00:00<?, ?it/s]
Traceback (most recent call last):
  File "F:\OneTrainer\modules\trainer\GenericTrainer.py", line 330, in backup
    self.model_saver.save(
  File "F:\OneTrainer\modules\modelSaver\StableDiffusionXLModelSaver.py", line 139, in save
    self.__save_internal(model, output_model_destination)
  File "F:\OneTrainer\modules\modelSaver\StableDiffusionXLModelSaver.py", line 101, in __save_internal
    torch.save(model.optimizer.state_dict(), os.path.join(destination, "optimizer", "optimizer.pt"))
  File "F:\OneTrainer\venv\lib\site-packages\torch\serialization.py", line 628, in save
    with _open_zipfile_writer(f) as opened_zipfile:
  File "F:\OneTrainer\venv\lib\site-packages\torch\serialization.py", line 502, in _open_zipfile_writer
    return container(name_or_buffer)
  File "F:\OneTrainer\venv\lib\site-packages\torch\serialization.py", line 473, in __init__
    super().__init__(torch._C.PyTorchFileWriter(self.name))
RuntimeError: Parent directory F: does not exist.
Could not save backup. Check your disk space!

[FurkanGozukara]

I changed folder path to F:\onetrainer_workspace_diffusers_test and testing again now

edit: this also fails

[FurkanGozukara]

I recorded the error video so weird @Nerogar

I have over 600 GB disk space

onetrainer_error.mp4

[FurkanGozukara]

to fix this issue you can't give a custom workspace folder sadly

so you have to use wherever the OneTrainer installed

this is it currently

@mx if you could reply this i think pytorch team could fix

pytorch/pytorch#105488 (comment)

[Nerogar]

I don't think the pytorch bug will be fixed anytime soon. We might need to add a better workaround to the model saving code.

There are two possible solutions:

1. Pass _use_new_zipfile_serialization=False to torch.save. This will use an older format. I don't know if there are any problems with this, or if the flag will be removed in a future release.
2. Always sanitize file names before saving checkpoint files (by replacing / with \). This won't always work. The pytorch bug still prevents saving to the root directory of any drive. But it might be a good compromise, I don't think anyone will save directly to the root directory. And if they do, we can at least print out a more useful error message.