fix(camembert): add tie_word_embeddings=True to CamembertConfig

[r266-tech]

What does this PR do?

Fixes a v5 regression where CamembertForMaskedLM (and all CamemBERT masked-LM tasks) produces near-zero, near-uniform logits, making the model completely non-functional.

Root cause

In v5, modeling_utils.get_expanded_tied_weights_keys() gates all weight tying on config.tie_word_embeddings:

tie_word_embeddings = getattr(self.config, "tie_word_embeddings", False)
if not tie_word_embeddings:
    return {}   # <-- skips ALL tying when attribute is missing

CamembertConfig was missing tie_word_embeddings: bool = True, so the method returned {} and lm_head.decoder.weight was randomly initialized instead of being tied to roberta.embeddings.word_embeddings.weight.

Sibling configs RobertaConfig and BertConfig already declare tie_word_embeddings: bool = True. CamemBERT is RoBERTa-based and its modeling code already defines _tied_weights_keys mapping lm_head.decoder.weight → roberta.embeddings.word_embeddings.weight — but that mapping was silently ignored.

Before fix (transformers v5.3.0)

>>> pipeline("fill-mask", model="camembert-base")("Le camembert est un délicieux fromage <mask>.")
# score=0.000108  totalité    ← near-uniform logits (broken)
# score=0.000106  Mat
# score=0.000104  Populaire

The LOAD REPORT also showed lm_head.decoder.weight: MISSING (randomly initialized).

After fix

>>> pipeline("fill-mask", model="camembert-base")("Le camembert est un délicieux fromage <mask>.")
# score=0.1819  suisse    ← matches v4 exactly ✅
# score=0.0937  français
# score=0.0495  italien

Change

One line added to CamembertConfig:

tie_word_embeddings: bool = True

Fixes #44671

Before submitting

• Did you read the contributor guideline?
• Was this discussed/approved via a Github issue? CamemBERT produces incorrect masked LM predictions in v5 #44671
• Did you write any new necessary tests?

Who can review?

@ArthurZucker @Cyrilvallez

[github-actions]

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

run-slow: camembert

[github-actions]

View the CircleCI Test Summary for this PR:

https://huggingface.co/spaces/transformers-community/circle-ci-viz?pr=44931&sha=c35409

[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.

[zucchini-nlp]

run-slow: camembert

[github-actions]

Workflow Run ⚙️

This comment contains run-slow, running the specified jobs:

models: ["models/camembert"]
quantizations: []

[github-actions]

CI Results

Workflow Run ⚙️

Commit Info

Context	Commit	Description
RUN	7f72b3d2	workflow commit (merge commit)
PR	c354098a	branch commit (from PR)
main	55cc1a7f	base commit (on main)

✅ No failing test specific to this PR 🎉 👏 !

[Cyrilvallez]

cc @tarekziade!! This has been forgotten in a precedent refactor for sure. COuld we add the following rule for our modeling format linter: if _tied_weights_keys is present and non-empty in modeling -> Config MUST contain the tie_word_embeddings field?

[Cyrilvallez]

Indeed we need the flag! Thanks!

[Cyrilvallez]

This was forgotten in #41541 @zucchini-nlp! With the linter rule, @tarekziade will be able to determine if we missed it at other locations as well!

[tarekziade]

cc @tarekziade!! This has been forgotten in a precedent refactor for sure. COuld we add the following rule for our modeling format linter: if _tied_weights_keys is present and non-empty in modeling -> Config MUST contain the tie_word_embeddings field?

#44988