Why your fine-tuning job takes longer than the FLOPs say
Comments
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Summary
A preview paper reports that a CatBoost regressor, using only pre-dispatch job configuration, predicts end-to-end fine-tuning time on shared H200 clusters far better than the standard FLOP-based estimate.