fix gemma3n case failure

[yao-matrix]

Problem

pytest -rA tests/models/gemma3n/test_modeling_gemma3n.py::Gemma3nIntegrationTest::test_generation_beyond_sliding_window_with_generation_config generate different output even do_sample=False in generation_config = GenerationConfig(max_new_tokens=20, do_sample=False), which makes UT fails

Root cause

custom generation_config will be overode by model generation_config even it explicitly set, when the set value equals global default.

Fix

Add a condition to make the overriding only happens when the custom_generation_config doesn't explicitly set the field.

@SunMarc , pls help review, thx very much.

[yao-matrix]

seems the ci failures are not related to my PR.

[yao-matrix]

Here is the case:
When user sets generation_config = GenerationConfig(max_new_tokens=20, do_sample=False), here do_sample is False, while global_default_value is also False, and model_gen_config_value is True, the current condition cannot distinguish whether the value is explicitly set by user or just the global default, so it just modify it to model default. Add another condition to make sure it happens only when user didn't explicitly specify the value.

[zucchini-nlp]

Indeed this is an issue, and afair we discussed it with @gante after merging the PR with config value overwriting. cc @gante for review

[gante]

There are probably a few more cases where this inheritance is still broken, but the fix looks sensible in general :)

[gante]

Actually, we can see in the CI that these changes break other cases :'(

@yao-matrix , let's instead add use_model_defaults=False in the model.generate call that is creating issues. We need to fully rewrite this attribute inheritance :(

[HuggingFaceDocBuilderDev]

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.

[yao-matrix]

pytest -rA tests/models/gemma3n/test_modeling_gemma3n.py::Gemma3nIntegrationTest::test_generation_beyond_sliding_window_with_generation_config

done, pls review again, and do you need me create an issue to track the generation_config issue? It seems a HF recommended way to use generate in some docs and blogs.

[yao-matrix]

@zucchini-nlp @gante , could you pls help review again? Thx

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: gemma3n