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.

@kashif
Gentle ping on this. Let me know if there’s anything I can help with to get it over the line.

@codex review

I was thinking perhaps a simpler fix could be to store the logits before deleting them in the liger path, and then use them for returning the outputs instead of re-runing the model() forward?

💡 Codex Review

Replace getattr-based skip with explicit fused-loss check

Using getattr(trainer, "use_liger_gkd_loss", False) lets this test skip silently if the attribute is renamed or accidentally removed, which can mask regressions in the Liger path while CI still passes. In this repo, AGENTS.md explicitly requires avoiding getattr/hasattr; reading trainer.use_liger_gkd_loss directly (or checking an explicit runtime/version condition) makes contract breaks fail loudly instead of being hidden as skips.

Propagate the duplicated Liger return_outputs fix consistently

This adds a special return_outputs handling branch only in GKDTrainer, but the repository’s AGENTS.md requires duplicated trainer logic to stay aligned across trainers. Leaving this as a one-off divergence in the shared Liger JSD pattern makes cross-trainer behavior inconsistent and increases the chance that equivalent paths (e.g., in other distillation trainers) keep the old bug/structure and drift further over time.

I was thinking perhaps a simpler fix could be to store the logits before deleting them in the liger path, and then use them for returning the outputs instead of re-runing the model() forward?

Thanks for the suggestion! Applied it in the latest commit.