Vision support

#6
by raulalonsoctic - opened

Any chance you could add the vision decoder to this model? This is the only NVFP4 quantization that works reliably under vLLM

raulalonsoctic changed discussion status to closed
raulalonsoctic changed discussion status to open
Owner

The vision encoder is in the model in shard 5, it is unquantized (therefore in ignore list) and it should work.

Owner

I have uploaded a new revision that includes some configs that were missing for multi-modal purposes. Can you try with that and see if vision works? Thanks.

Working great on my first tests. Thank you very much!!!

Owner

Thanks, Im uploading the nvfp4 for the 122B version as we speak, with all configs etc.

Sign up or log in to comment