Are the Model Weights correct?
I see that some files were uploaded 1 month ago and others 6 days ago. I'm also having a strange output from the model even not using the 4bits. It repeats the same sentence until hit the max number of tokens.
Hi,
Yes you can verify this by running the conversion script: https://github.com/huggingface/transformers/blob/main/src/transformers/models/llava_next/convert_llava_next_weights_to_hf.py.
I ran this script with the --push_to_hub flag which pushes the weights to this repo. I noticed indeed that it only updated 2 shards out of the possible 4 (which were initially updated when working on the llava-next port), but it gave me correct logits, which means that the weights are correctly uploaded.
Great, I just wanted to double check. Thanks for the update 🤗