Model Request: Addition of new Lava-1.6 models

[Enyakk]

As seen here:
https://llava-vl.github.io/blog/2024-01-30-llava-1-6/
https://huggingface.co/liuhaotian/llava-v1.6-vicuna-13b/tree/main

[BenDes21]

agreeee

[jhc13]

I see there's a lot of demand for new models, including this one and #49, but I am currently very busy and do not have the time to work on adding them. I will have more free time in a few days, so I will take a look then.

[BenDes21]

I see there's a lot of demand for new models, including this one and #49, but I am currently very busy and do not have the time to work on adding them. I will have more free time in a few days, so I will take a look then.

Thanks

[Goldenkoron]

I would like to see support for moondream added, unsure if it's already compatible or not if I dragged the model somewhere.

[jhc13]

I would like to see support for moondream added, unsure if it's already compatible or not if I dragged the model somewhere.

Please create a separate issue for this. Thank you.

[jhc13]

It seems like LLaVA-1.6 might be officially supported in Transformers soon (there's a pull request in progress for it), so I will wait for now.

[BenDes21]

It seems like LLaVA-1.6 might be officially supported in Transformers soon (there's a pull request in progress for it), so I will wait for now.

nice cannot wait

[Vigilence]

Been doing some testing and this model can give impressive results. It can sometimes not detail items that are always detailed in cog va, but I think the 34b model might fix that.

[BenDes21]

Been doing some testing and this model can give impressive results. It can sometimes not detail items that are always detailed in cog va, but I think the 34b model might fix that.

cannot wait to test, is it less vram consuming than cog va ?

[Vigilence]

The ram used depends on the model. For:

• the 34gb model my gpu is using nearly all 24gb vram sometimes a little past it.
• the 13gb model uses around 20gb of vram
• the 7gb model uses around 11gb vram.

I am sure these models could use less vram, but the gui i'm using to access these models may not be fully optimized. All models are using 4-bit mode. This model is good for art identification, at least for me, as it points out many little details the other models won't point out. Not even the top rated ones like Qwen-vl-max.

[LoFiApostasy]

This would be great, v1.6 is pretty good. It looks like the pr has changed to here: huggingface/transformers#29586 and is in the final stages of approval.

[jhc13]

The models have been added in v1.19.0.

[Vigilence]

The models have been added in v1.19.0.

Thank you!