A town clerk has one GPU evening and a wish list: full-weight training of a huge model, a small PEFT run, and a prompt-only trial. How should the evening be designed?
Select an answer to reveal the explanation.
Short Explanation
One GPU evening cannot finish a huge full-weight job plus PEFT plus a prompt trial. Fit the comparisons to time and memory so the affordable runs actually return a score. A design that cannot complete is not a design.
Full Explanation
Compute budget is part of experimental design, not an afterthought. With one GPU evening, the clerk should schedule comparisons that can finish — typically the prompt-only trial and a small PEFT run — rather than a full-weight job that never returns a score. A plan that cannot complete produces no comparable results. Refusing PEFT on principle, or spending the night on collectives, also wastes the only evening.