[Bug]: Out of VRAM if training layer filter is used for full finetuning

[dxqb]

What happened?

requires_grad is set up before the parameters are created:

OneTrainer/modules/modelSetup/FluxFineTuneSetup.py

Line 83 in 3250a32

self.__setup_requires_grad(model, config)

This means requires_grad is set to True even for frozen (filtered) parameters for the first training step. They aren't trained, but gradients are created and as much vram is used as would be necessary without "fused backpass"

https://github.com/Nerogar/OneTrainer/blob/3250a32898acc8608a21ac91f7f68f8a1ef542d2/modules/modelSetup/BaseModelSetup.py#L233C13-L234C74

What did you expect would happen?

Relevant log output

Generate and upload debug_report.log

No response