DeepSeek-VL2 support

[fakerybakery]

Search before asking

• I have searched the Multimodal Maestro issues and found no similar feature requests.

Description

Hi,
Thanks for releasing maestro!
Are there any plans to add support for fine-tuning DeepSeek VL2?
Thanks!

Use case

No response

Additional

No response

Are you willing to submit a PR?

• Yes I'd like to help by submitting a PR!

[SkalskiP]

Hi @fakerybakery 👋🏻 absolutely! DeepSeek-VL2 and Moondream fine-tuning are next on my list of models to add. Stay tuned!

[SkalskiP]

Hi @fakerybakery I spent some time today refining the maestro contributor guide. You might find it useful when adding DeepSeek-VL2 ;)

[SkalskiP]

Hi @fakerybakery 👋🏻 I talked to @HubHop from the DeepSeek team yesterday and he told me that DeepSeek-VL2 does not have fine-tuning support. It seems that in this situation we will have to give up adding DeepSeek-VL2 to maestro. Maybe in this situation we could add DeepSeek-Janus-Pro instead?

[fakerybakery]

Makes sense. It looks like ModelScope's fine-tuning scripts support VL2, not sure if it would work though

[SkalskiP]

Interesting. I'll try to take a look and try to reverse engineer it. Let me know if you'll learn something more in the meantime.