Note

Low Risk
Small refactor in KTO loss path; behavior change only when batches lack precompute but still carry reference_logps, which should match enabled precompute mode.

Overview
Aligns KTO with DPO by using a trainer flag self.precompute_ref_logps (from args.precompute_ref_log_probs) instead of the old self.precompute_ref_log_probs name everywhere that path is checked—including Liger validation and dataset precomputation.

In _compute_loss, reference log-probs are taken from the batch only when self.precompute_ref_logps is true, replacing the previous "reference_logps" in batch check so precomputed references are tied to the config, not accidental batch keys.

^{Reviewed by Cursor Bugbot for commit 34b8124. Bugbot is set up for automated code reviews on this repo. Configure here.}