Choose LoRA for Fine-Tuning and Explain Its Limitations

When would you use a low-rank adaptation method such as LoRA for fine-tuning a language model, and what limitations would you evaluate before choosing it over full fine-tuning? ### Constraints Discuss standard LoRA with a frozen base model and trainable low-rank weight updates. Distinguish LoRA from quantization and from the training objective: LoRA can be used with different objectives. No task, rank, or memory budget is prescribed. ### Clarifying Questions - Is the main constraint optimizer memory, checkpoint storage, training throughput, or serving multiple adaptations? - How large is the domain shift, and which model modules need adaptation? - Will adapters remain separate at serving time or be merged into compatible base weights? ```hint Count both trainable and retained state Freezing the base removes its optimizer updates, but the base weights and relevant activations still require resources. ``` ### What a Strong Answer Covers - The low-rank parameterization and the source of parameter and optimizer-state savings. - Rank and module-selection tradeoffs, task quality, and deployment considerations. - Remaining memory costs and the distinction from quantized fine-tuning. ### Follow-up Questions - How would you tell whether a low rank is limiting task quality? - What checks are needed before applying an adapter to a different base-model version?