Improve consistency of Liger Kernel initialization in DPO

[albertvillanova]

Improve consistency of Liger Kernel initialization in DPO.

This PR refactors how the Liger kernel is integrated into the DPOTrainer. The main changes improve consistency and clarity around enabling and initializing the Liger kernel and its associated loss function. It aligns the corresponding comments in the and KTOTrainer class.

Changes

Integration and initialization improvements for Liger kernel:

• Standardized the assignment and checking of use_liger_kernel to use self.use_liger_kernel instead of directly referencing args.use_liger_kernel in both DPOTrainer and KTOTrainer constructors.
• Moved the initialization of self.liger_loss_fn for the Liger kernel in DPOTrainer to after PEFT model checks, ensuring it is only set when appropriate.

---

Note

Low Risk
Small constructor reordering and comment/consistency changes in trainer init; no change to loss computation paths for valid configurations.

Overview
Aligns Liger kernel setup in DPOTrainer and KTOTrainer with a shared # Liger loss section and consistent use of self.use_liger_kernel in conditionals (DPO previously gated on args.use_liger_kernel).

In DPOTrainer, self.liger_loss_fn (LigerFusedLinearDPOLoss) is now created after all Liger validation, including PEFT checks for lm_head adapters and prompt-learning methods, instead of immediately after the single-loss-type check. Invalid configs fail before the fused loss object is constructed; supported setups are unchanged.

KTOTrainer only gets the matching comment tweak; behavior is unchanged.

^{Reviewed by Cursor Bugbot for commit 090c74e. 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.