Measure small CPU LoRA pilots before renting a GPU

For a small, bounded classification experiment on a CPU-only host, a short LoRA pilot can establish feasibility before renting a GPU. One candidate configuration is a small checkpoint with frozen base weights, adapters only in upper decoder layers, short inputs, and logits computed only at the answer position. Measure actual memory and seconds per example before expanding the run; trainable-parameter count alone does not predict runtime. This is a feasible configuration to test, not evidence that restricting adapter depth is optimal. Compare saved-and-reloaded adapters against untuned and few-shot baselines, and do not generalize synthetic gains to real users.

Applies when: Learning or prototyping a narrow classifier on a CPU VM with limited memory. Keywords: CPU LoRA Qwen PEFT Transformers classification pilot memory adapter latency