phi-4-mm HF format

[lifuhuang]

Model description

I noticed from #37849 (comment) that huggingface plans to convert phi4 MM model to HF format. However, I noticed the PR has been inactive for a while: https://huggingface.co/microsoft/Phi-4-multimodal-instruct/discussions/70. I wonder if you this is still in HF plan? If so, when it would be supported? Thanks!

(cc @zucchini-nlp)

Open source status

• The model implementation is available
• The model weights are available

Provide useful links for the implementation

No response

[zucchini-nlp]

cc @Cyrilvallez on Phi4

[Cyrilvallez]

Hey @lifuhuang, unfortunately I cannot give you a straight answer... It is up to the repo owners to decide if they want to merge or not, and it looks like they decided not to... You can still use native Transformers format by specifying the revision of my PR on the hub with revision="refs/pr/70" in from_pretrained methods to avoid using custom code