it's trained with a lora adaptor, so it's either an error or they also count the adaptor. they use a 16 param inner lora dimension however, so it's unlikely tha

it's trained with a lora adaptor, so it's either an error or they also count the adaptor. they use a 16 param inner lora dimension however, so it's unlikely that that's the reason (too small) an important point to keep in mind is that at inference, the lora adaptors are made to be merged into the base model so they don't affect inference speed. (you need to explicitly do it though, if you train your own adaptor)