ausboss

/

llygma-13b-lora

like 0

Alright, I change the name now that I finally got around to trying to test this. I get a LoRA loading error:

peft\peft_model.py", line 163, in from_pretrained
    config = PEFT_TYPE_TO_CONFIG_MAPPING[
KeyError: None

Doesn't happen with VicUnLocked LoRA or SuperCOT 13b LoRA.

I'm marking as private until I get this worked out.

I think I got it worked out

Same error still. I'd say have someone else test it just in case it's my ooba or PEFT version or whatever. Entirely possible since I've been messing with some stuff lately.

Honestly same. Everything has been kind of wonky since I started trying to load the MPT models. I can't even really merge properly right now but I'm still fiddling with it.

4bit LoRA loading has always been a little inconsistent for me. I'll try to remember to test it with koboldcpp later. I think they added support. Testing LoRAs is always such a pain. It'd help if base llama wasn't competent.

I'm fairly certain there was an issue with the config I just noticed and fixed. I am trying to merge it with another model right now so I was seeing the same issue but its fixed now.

Awesome, I'll wait for the merge then since the 4bit LoRA loading hurts me inside.