ROMA: a Read-Only-Memory-based Accelerator for QLoRA-based On-Device LLM