Can you do the optimization/re-rank magic on the new experimental lora?
This lora just dropped and it's way more high quality than sulphur/eros base models. I don't recall the specifics but it was a special merge of training checkpoints of the original sulphur.
However just like the high-rank lora released by the original sulphur creators, this experimental lora is also at an insane rank of 1024 for most layers with logits layers at 32.
Can you do your magic and shrink this bad boy down?
This lora just dropped and it's way more high quality than sulphur/eros base models. I don't recall the specifics but it was a special merge of training checkpoints of the original sulphur.
However just like the high-rank lora released by the original sulphur creators, this experimental lora is also at an insane rank of 1024 for most layers with logits layers at 32.
Can you do your magic and shrink this bad boy down?
I optimized this LoRA just like my previous ones, but I noticed its internal structure is quite unique. The weight distribution is very dense, and almost all layers carry a lot of influence. This means layer pruning is pretty much ineffective here. I chose not to force-prune it. Instead, to give you a better balance between size and quality, I added an intermediate version between fro 99 and fro 95 — LTX_SulphurEXP_LoRA_fro97
https://huggingface.co/maximsobolev275/LTX-SulphurExperimental-LoRA-Optimized/tree/main