Nerogar
Nerogar
> Now that OneTrainer has upgraded PyTorch and is compatible with Python 3.10, 3.11 and 3.12, it would be easy to switch to 3.11 as the minimum requirement, It would...
> In the repo you have shared above, I don't see configuration files for the different format. Where do you put the configuration attributes so that the load happens correctly?...
According to comfy this should be fixed now. Can you confirm, or do you still see the problem?
Great, thanks. That confirms it's not a OneTrainer issue
Adding a bit more context from the original discussion: The overhead is mostly caused by this issue: https://github.com/huggingface/safetensors/issues/542, but there are other problems that make it harder reduce RAM usage...
Models aren't always saved with a single dtype. Especially if you use something low precision like fp8. Some of the weights will still be saved in higher precision like fp32....
@hlky did you try this on Windows or only Linux? I remember trying something similar before, but immediately got a blue screen.
I think adding a user configurable placeholder name makes sense. But I don't really understand how this would help cloud training in any way. Can you explain that a bit...
> Yeah, the cloud training can then use the workspace name property for its own server directory That would change the behavior between cloud training and local training. Not something...
Include should not be changed for sdxl. It's only useful for models where one of the text encoders is optional like SD3. I've never seen anything suggesting disabled text encoder...