Fix spurious liger token_accuracy CI warnings in SFT tests

[albertvillanova]

This PR fixes two test mocks in test_sft_trainer.py that were emitting spurious UserWarnings about liger-kernel not returning token_accuracy.

Motivation

Both test_compute_loss_skip_logits_on_eval_without_metrics_with_liger and test_predict_does_not_skip_logits_with_liger are tests about skip_logits behaviour. Their mocks were accidentally simulating a broken liger call (returning None or no token_accuracy), which caused the SFT trainer's liger branch to emit a warning intended for genuine liger failures. The warning is correct by design; the mocks were not.

Real liger always returns token_accuracy when return_token_accuracy=True is set, even when skip_logits=True, because token accuracy is computed inside the fused kernel without materialising the full logits tensor.

Changes

• In test_compute_loss_skip_logits_on_eval_without_metrics_with_liger: set dummy_outputs.token_accuracy = torch.tensor(0.5) instead of None
• In test_predict_does_not_skip_logits_with_liger: replace the bare tuple with a namedtuple that exposes a token_accuracy field while remaining tuple-indexable (required by transformers' prediction_step which slices outputs as outputs[1:])

---

Note

Low Risk
Test-only mock fixes; no changes to SFT trainer or runtime behavior.

Overview
Updates two Liger + skip_logits tests in test_sft_trainer.py so their patched compute_loss mocks behave like real liger-kernel (which always supplies token_accuracy when return_token_accuracy=True), instead of returning None or outputs without that field.

In test_compute_loss_skip_logits_on_eval_without_metrics_with_liger, the mock MagicMock now sets token_accuracy to a tensor instead of None. In test_predict_does_not_skip_logits_with_liger, the mock return value is a collections.namedtuple with loss, logits, and token_accuracy so it stays tuple-indexable for prediction_step while exposing the attribute the trainer reads. import collections is added for the namedtuple.

This removes spurious CI **UserWarning**s from the intentional liger failure path in SFTTrainer.compute_loss; production trainer logic is unchanged.

^{Reviewed by Cursor Bugbot for commit 9aeb466. 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.