Raise a clear error when GKD student and teacher vocab sizes differ

[sergiopaniego]

What does this PR do?

White-box logit KD in GKDTrainer requires the student and teacher to share vocab_size. When they differ, training failed with a cryptic broadcast error in kl_div (e.g. Qwen2.5-0.5B is 151936 but Qwen2.5-7B is 152064). GKDTrainer now validates this at init and raises a clear error pointing to GOLD for cross-tokenizer distillation. DistillationTrainer already handles this by resizing the teacher embeddings.

Before submitting

• Did you read the contributor guideline, Pull Request section?

AI writing disclosure

• AI-assisted: some parts were suggested or improved by AI, but the PR was written and reviewed by a human.

---

Note

Low Risk
Init-time validation only; no change to training logic when vocab sizes already match.

Overview
GKDTrainer now checks that the student and teacher share the same vocab_size right after the teacher is loaded, instead of failing later with an obscure kl_div broadcast error when logits shapes disagree.

On mismatch it raises a ValueError that states both sizes, notes that GKD needs a shared vocabulary for full next-token distributions, and suggests matching teachers or GOLD for cross-tokenizer distillation (unlike DistillationTrainer, which can resize teacher embeddings).

^{Reviewed by Cursor Bugbot for commit 408920f. 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.