NVlabs/VILA

Fine-tuning LongVILA

lyluh opened this issue · 2 comments

I am able to fine-tune the original models fine with the /scripts/v1_5/release/3b/3_sft.sh file using my own custom videos, but when I try to train the new LongVILA model I get this warning message and the loss is 0.0 and eval loss is NaN:
Screenshot 2024-09-17 at 4 51 24 PM

Any ideas?