Fix teacher quantization kwargs and guard eval callback in GKD example

[sergiopaniego]

What does this PR do?

Fixes two bugs in examples/scripts/gkd.py:

• Under --load_in_4bit, the teacher's quantization block set model_kwargs (the student) instead of teacher_model_kwargs, so the teacher loaded unquantized. Aligns with distillation.py.
• LogCompletionsCallback requires a prompt column and crashed eval on conversational (messages) datasets (ValueError: Column 'prompt' doesn't exist.); it is now attached only when that column exists.

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
Example-script-only fixes with no changes to core TRL training logic or APIs.

Overview
Fixes two bugs in the GKD example script (gkd.py).

When --load_in_4bit (or other quantization) is enabled, quantization config is now applied to teacher_model_kwargs instead of mistakenly updating the student’s model_kwargs, so the teacher actually loads quantized—consistent with distillation.py.

LogCompletionsCallback is only registered when the eval split has a prompt column, avoiding a crash on conversational datasets that only expose messages.

^{Reviewed by Cursor Bugbot for commit 758035b. 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.