LoRA training example

[JohannesGaessler]

LoRA training should in principle be achievable with the same tools as the full finetune, it's simply not implemented. I don't know the exact details of LoRA training so I will need to read up on them. Investigate the quality difference between

1. training a LoRA on top of a full-precision model, merging the LoRA, and then quantizing the model vs.
2. training a LoRA on top of a quantized model and merging it vs.
3. training a LoRA on top of a quantized model and not merging it.

[pannous]

i'm willing to assist