Choose LoRA for Fine-Tuning and Explain Its Limitations

Constraints

Talk about standard LoRA where the base model is frozen and only low-rank weight updates are trained. Separate LoRA from quantization and from the training objective itself—LoRA can be paired with various objectives. No specific task, rank, or memory budget is given.

Clarifying Questions

• Is the primary bottleneck optimizer memory, checkpoint storage, training throughput, or the need to serve many adaptations?
• What is the magnitude of the domain shift, and which parts of the model require adaptation?
• At serving time, will the adapters stay separate or be fused into compatible base weights?

Hint — Account for both trainable and retained state: Freezing the base eliminates its optimizer updates, but the base weights and associated activations still consume resources.

What a Strong Answer Covers

• The low-rank parameterization and where the savings in parameters and optimizer state come from.
• Trade-offs involving rank and module selection, task quality, and deployment factors.
• Residual memory costs and how this differs from quantized fine-tuning.

Follow-up Questions

• How can you determine if a low rank is hurting task quality?
• What validation steps are necessary before using an adapter with a different base-model version?

Overview: Describe low-rank adaptation, the parameter and memory savings, choices of rank and modules, remaining costs, and deployment constraints relative to full fine-tuning.