Fix tie_word_embeddings not lifted from text_config for some VLM configs (BC regression)

[qgallouedec]

For Qwen2_5_VLConfig, Qwen2VLConfig, Glm4vConfig, and Glm4vMoeConfig, pre-v5 saves placed tie_word_embeddings inside text_config (and dropped it from the root) because the inner text class declared the field. v5 moved the field to the outer config (#42420) but __post_init__ doesn't lift the value from text_config. The outer attribute therefore falls through to its dataclass default (False), get_expanded_tied_weights_keys returns {}, no tying happens, lm_head.weight is reported MISSING and randomly initialized — the model loads silently broken.

Regression introduced in #41541, which removed the forwarding line kwargs["tie_word_embeddings"] = self.text_config.tie_word_embeddings originally added in #42420 with a FIXME: tying has to be used from the text config comment. The vestigial tie_word_embeddings: bool = False field was later removed from the inner text classes in #44976.

This patch mirrors the existing fix in LlavaConfig.__post_init__: when text_config arrives as a dict (the deserialization path), forward tie_word_embeddings to the outer when the root is the default and the dict has a truthy value.

Reproducer

# transformers v5 (e.g. main / 5.8.0.dev0)
from transformers import Qwen2_5_VLForConditionalGeneration

m = Qwen2_5_VLForConditionalGeneration.from_pretrained("trl-internal-testing/tiny-Qwen2_5_VLForConditionalGeneration")
print("outer tie_word_embeddings:", m.config.tie_word_embeddings)
print("lm_head tied:", m.lm_head.weight.data_ptr() == m.model.language_model.embed_tokens.weight.data_ptr())

Before this PR (loud LOAD REPORT, broken model):

Qwen2_5_VLForConditionalGeneration LOAD REPORT from: trl-internal-testing/tiny-Qwen2_5_VLForConditionalGeneration
Key            | Status  |
---------------+---------+
lm_head.weight | MISSING |
- MISSING: those params were newly initialized because missing from the checkpoint.

outer tie_word_embeddings: False
lm_head tied: False

After this PR:

outer tie_word_embeddings: True
lm_head tied: True

(no LOAD REPORT printed)

The Hub artifact above was saved by transformers 4.57.5; under 4.x its __post_init__ still propagated tie_word_embeddings from text → root, so the bug is v5-only and only triggers on configs saved by 4.x (the vast majority of pre-existing Hub repos for these architectures).

Affected configs

Empirically determined by saving each composite VLM config under transformers 4.56.2 with tie_word_embeddings=True and inspecting the resulting config.json:

config	root	text_config	needs lift in v5?
qwen2_5_vl	<absent>	True	yes
qwen2_vl	<absent>	True	yes
glm4v	<absent>	True	yes
glm4v_moe	<absent>	True	yes
got_ocr2	<absent>	True	no — outer class default is True already
llava{,_next,_next_video,_onevision}	True	<absent>	no — already saved at root or already has the lift
idefics3, smolvlm, llama4, pix2struct, fuyu	True	—	no

Files touched: qwen2_5_vl, qwen2_vl, glm4v (configuration + modular), glm4v_moe.

[github-actions]

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

run-slow: glm4v, glm4v_moe, qwen2_5_vl, qwen2_vl

[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]

Huh interesting, I suppose testing models tie weights unlike official releases

We can fix it in case there are more models like that on the hub (which i am not really sure), but we have to check text config after it is fully initialized imo

[zucchini-nlp]

wont work if text config is already a PreTrainedConfig