Align Liger loss naming across core trainers

[albertvillanova]

Align Liger loss naming across core trainers.

This PR standardizes the naming of the Liger loss attribute in both the DPO and GRPO trainers, changing it from liger_loss_fn or liger_grpo_loss to a consistent liger_loss. It also updates all references in the code and associated tests to use the new attribute name. This change improves code readability and maintainability by ensuring a single, clear attribute name is used throughout.

Changes

Attribute renaming and code consistency

• Renamed the Liger loss attribute from liger_grpo_loss and liger_loss_fn to liger_loss in both trl/trainer/grpo_trainer.py and trl/trainer/dpo_trainer.py, and updated all internal references accordingly.
• Updated method calls to use self.liger_loss instead of the previous attribute names in both trainers.

Test updates

• Updated all test assertions and mocks in tests/test_grpo_trainer.py to use liger_loss instead of liger_grpo_loss.

---

Note

Low Risk
Rename-only refactor across trainers and tests with no logic changes; any external code that referenced the old attribute names would break at import/runtime.

Overview
Standardizes how DPO and GRPO trainers expose the fused Liger loss module by using a single liger_loss attribute instead of liger_loss_fn (DPO) and liger_grpo_loss (GRPO).

Initialization and forward paths in dpo_trainer.py and grpo_trainer.py now assign and call self.liger_loss with the same LigerFusedLinear* instances as before. GRPO tests that mock or assert on the Liger module were updated to use trainer.liger_loss.

This is a naming-only change; loss computation and trainer behavior are unchanged.

^{Reviewed by Cursor Bugbot for commit 16d0d51. 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.

[qgallouedec]

better. cc @kashif fyi