[albertvillanova]

This PR makes the MODEL_REVISIONS override apply to concrete config classes as well, so a tiny model under test is read at the pinned revision everywhere.

Motivation

apply_model_revisions patches from_pretrained on the Auto* classes and on PreTrainedModel, PreTrainedTokenizerBase and ProcessorMixin, but not on the base config class. AutoConfig is a factory and never appears in a concrete config's MRO:

Qwen2Config MRO: ['Qwen2Config', 'PreTrainedConfig', 'PushToHubMixin', 'RotaryEmbeddingConfigMixin']

so a caller that reaches for the concrete class slips past the override. peft does exactly that when deciding whether the embedding was resized:

vocab_size != model.config.__class__.from_pretrained(model_id).vocab_size

The model is loaded at the revision under test while its base config is read from the default branch, so any fixture PR that changes vocab_size makes the two disagree and every PEFT test emits:

UserWarning: Setting `save_embedding_layers` to `True` as the embedding layer has been resized during finetuning.

That is 34 spurious warnings while validating #7098, and it also means save_embedding_layers is silently forced on in those tests.

Solution

Add the base config class to the patched set. Concrete configs do not define their own from_pretrained, so they inherit the patched one.

PretrainedConfig is the spelling that works across the supported range: 4.56.2 exports only that name, while 5.x exports both it and PreTrainedConfig pointing at the same object.

Verification

Ran the suite with MODEL_REVISIONS pinned to the #7098 fixture PRs, with and without this change:

	without	with
save_embedding_layers warnings	34	0
passed / skipped / xfailed	2385 / 149 / 5	2385 / 149 / 5
distinct PAD/BOS/EOS variants	22	22

https://github.com/huggingface/trl/actions/runs/34141451081/job/101804124860

The token realignment counts are the control and are unchanged, confirming the patch does not alter which revision the model itself loads from. Total lines differ by 3 (463 against 460) purely because the earlier run had 4 reruns and this one had none.

With MODEL_REVISIONS empty, which is its state on main, the fixture returns early and nothing is patched at all.

Changes

• Patch from_pretrained on the base config class alongside the Auto* and other base classes

---

Note

Low Risk
Test-only conftest change; no production code paths, and patching is skipped when MODEL_REVISIONS is empty on main.

Overview
Extends the apply_model_revisions pytest fixture so MODEL_REVISIONS also applies when configs are loaded through concrete config classes (e.g. Qwen2Config.from_pretrained), not only AutoConfig and model/tokenizer loaders.

PretrainedConfig is added to the patched classes and imports (using the PretrainedConfig spelling for compatibility across transformers 4.x/5.x). Concrete configs inherit from_pretrained from that base, so pinned Hub revisions now apply consistently when libraries such as PEFT reload config by class.

This removes config-at-default-branch vs weights-at-pinned-revision mismatches during tiny-model PR CI (e.g. spurious save_embedding_layers warnings). When MODEL_REVISIONS is empty, the fixture still exits early and behavior is unchanged.

^{Reviewed by Cursor Bugbot for commit 1219797. 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]

Merging this PR as a trivial fix in CI.