Align KTO with DPO: Align error for Liger with precompute_ref_logps

[albertvillanova]

Align KTO with DPO: Align error for Liger with precompute_ref_logps.

Part of:

• KTO refactoring to be promoted to stable #4786

This PR makes a small update to the error message in the __init__ method of KTOTrainer to clarify the usage restrictions when using the Liger KTO loss with precompute_ref_logps, aligned with DPOTrainer.

Changes

• Improved the error message to more clearly explain that Liger KTO loss does not support precomputing reference log probabilities, and to suggest disabling precompute_ref_log_probs or setting use_liger_kernel to False.

---

Note

Low Risk
Copy-only change to an initialization error message; no training or runtime logic is modified.

Overview
Updates the ValueError raised in KTOTrainer.__init__ when use_liger_kernel=True and precompute_ref_log_probs=True are used together.

The message now states that Liger KTO loss does not support precomputing reference log probabilities, and tells users to either turn off precompute_ref_log_probs or set use_liger_kernel to False—matching DPOTrainer’s wording for the same incompatible combination. Validation behavior is unchanged.

^{Reviewed by Cursor Bugbot for commit 47da402. Bugbot is set up for automated code reviews on this repo. Configure here.}

[bot-ci-comment]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.