Llama3 video fix

[sywangyi]

• vision models: @yonigozlan @molbap
• CIs: @ydshieh

when running tests/models/video_llama_3/test_modeling_video_llama_3.py::VideoLlama3IntegrationTest all fail cause by lm_head.weight is missing.

[github-actions]

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

run-slow: video_llama_3

[ydshieh]

@zucchini-nlp

The current failure (on main) is from

Untangle config inheritance (#41541)

For this PR, it fixed the CI, but I don't know if it's the best / correct way.

Could you check, taking the above context.

[zucchini-nlp]

so the model ties weights by default always or some ckpt leave untied? We need to change the default if it's always tied, that should work for most cases

[ydshieh]

I don't know if we should change code defaults according "if there are checkpoints on the hub" untied.

Does the PR (on Jan) change some thing that won't load the tie_weight attribute from the text model correctly like before?

[zucchini-nlp]

I added the config attr in Jan, it didn't exist before (but the base class had default True) and I might have added it with wrong default ig