[fix] regression introduced by #45534

[eustlb]

What does this PR do?

A continuation of #46400

Verified by running (on this branch) this script

model                 lm_head   lm==embed   weights tied  config tie  match
---------------------------------------------------------------------------
qwen2_audio           yes       False       False         False       OK
voxtral               yes       False       False         False       OK
voxtral_realtime      no        n/a         True          True        OK
glmasr                yes       True        False         True        MISMATCH <<<
granite_speech        no        n/a         True          True        OK
granite_speech_plus   no        n/a         True          True        OK
audioflamingo3        yes       False       False         False       OK
musicflamingo         yes       False       False         False       OK
vibevoice_asr         yes       False       False         False       OK

• lm_head — yes/no: whether the checkpoint has a separate lm_head.* tensor in its safetensors (has_lm_head).
• lm==embed — True/False/n/a: when an lm_head exists, whether lm_head.weight is bitwise identical to embed_tokens.weight. n/a when there's no separate lm_head.
• weights tied — True/False: what the actual checkpoint implies (not has_lm_head → no separate head means weights are tied).
• config tie — True/False: what the config class resolves tie_word_embeddings to (fallback False).
• match — OK if config tie == weights tied, else MISMATCH <<< (the regression check).

For glmasr, hub weights have a lm head but it's bitwise to embed tokens so we keep weight tying

[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.

[hmellor]

For the models with no LM head, why set tie_word_embeddings=True?

[hmellor]

Should this be propagated to modelling_glmasr.py?

And tie_word_embeddings updated?

[eustlb]

It's already propagated, but since _tied_weights_keys has been removed from AudioFlamingo3ForConditionalGeneration, this needs to be readded here (so modeling stays the same)

[eustlb]

no lm_heads in the above tables means no lm_head in hub weigths, so tie_word_embeddings must be set tot True

[github-actions]

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

run-slow: audioflamingo3, glmasr, musicflamingo, qwen2_audio, vibevoice_asr, voxtral, voxtral_realtime

[eustlb]

run-slow: audioflamingo3, glmasr, musicflamingo, qwen2_audio, vibevoice_asr, voxtral, voxtral_realtime

[github-actions]

CI Dashboard: View test results in Grafana

[github-actions]

Workflow Run ⚙️💔 This comment contains run-slow, but unknown error occurred and the workflow run aborted!