[sergiopaniego]

What does this PR do?

Since transformers v5, unset GenerationConfig fields default to None, which means "inherit the value from the model's generation_config" (huggingface/transformers#42702). GKDTrainer builds its on-policy generation config with an explicit top_k=0 but no top_p, so models that ship sampling defaults (e.g. Qwen3 or Gemma with top_p=0.95) silently truncate the on-policy sampling distribution during training.

Repro (CPU):

from transformers import AutoModelForCausalLM, GenerationConfig

model = AutoModelForCausalLM.from_pretrained("trl-internal-testing/tiny-Qwen3ForCausalLM")
model.generation_config.top_p = 0.33
generation_config = GenerationConfig(max_new_tokens=4, temperature=0.9, do_sample=True, top_k=0)
prepared, _ = model._prepare_generation_config(generation_config)
print(prepared.top_p)  # 0.33 (inherited), expected 1.0

This sets an explicit neutral top_p=1.0, completing the intent of the explicit top_k=0 and matching the sibling trainers (PPO hardcodes top_p: 1.0; GOLD, Distillation, SDFT, SSD and SDPO set top_p explicitly from their configs).

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline, Pull Request section?
• Was this discussed/approved via a GitHub issue? Please add a link to it if that's the case.
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?

AI writing disclosure

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

Who can review?

@kashif

---

Note

Low Risk
Small, targeted change to GKD training-time generation kwargs; improves consistency with intended temperature-only sampling and aligns with other TRL trainers, with no auth or data-path impact.

Overview
GKD on-policy sampling no longer inherits the student model’s default top_p (e.g. 0.95 on Qwen/Gemma). GKDTrainer now sets top_p=1.0 alongside the existing top_k=0, so transformers v5’s “unset field inherits from model.generation_config” behavior cannot narrow the training-time sampling distribution.

test_generation_config_init asserts trainer.generation_config.top_p == 1.0.

^{Reviewed by Cursor Bugbot for commit ddf75ee. 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.

[albertvillanova]

Thanks for for the detailed report and the fix!

The fix is minimal, the reasoning is sound, and it closes a real silent-truncation bug for models that ship sampling defaults (Qwen3, Gemma, etc.).