Open 311-code opened 7 months ago
In my understanding, extracting DoRA may be possible. It needs the same preprocessing of DoRA (extracting magnitude vector), and SVD. However, I also think it doesn't improve the extracted DoRA quality from LoRA. Because SVD for DoRA and SVD for LoRA may have similar precision.
I'm not a math person, so please let me know if DoRA extraction has an advantage from LoRA.
I use the Lora extraction option from a dreambooth or finetuned checkpoint all the time as it seems to produce better results than just training a Lora. Would love to see a way to extract Dora.
If it's just a parameter change in Lora to implement Dora I don't think it would be too difficult?