When deploying LLMs in production, what is a common strategy for parameter-efficient fine-tuning?

Parameter-efficient fine-tuning (PEFT) strategies, like LoRA or adapters, freeze most pretrained parameters and train only lightweight modules, reducing computational costs while adapting to new tasks. This preserves general knowledge, prevents catastrophic forgetting, and enables quick deployments in resource-constrained settings. For LLMs, it's crucial for efficiency in production, allowing specialization without retraining billions of parameters. Security-wise, it minimizes exposure to new data risks. Exact extract: "A common strategy is freezing the majority of model parameters and updating only a small task-relevant subset, ensuring efficiency in fine-tuning for production deployment." (Reference: Cyber Security for AI by SISA Study Guide, Section on Efficient Fine-Tuning in SDLC, Page 90-92).