Views
No views yet
raw_r6/: final trained PEFT LoRA and its Q8_0 llama.cpp LoRA (r=6, lora_alpha=9, rsLoRA).smoothed_r6/: rank-6 PEFT approximation and its Q8_0 llama.cpp LoRA (r=6, lora_alpha=6, standard LoRA).9/sqrt(6)). The raw GGUF stores the equivalent converted alpha, so do not multiply either form again. The smoothed adapters have the intended delta baked into their factors. For PEFT merging, load normally and call merge_and_unload(). For llama.cpp, attach the selected GGUF LoRA at scale 1.0.