Add llama3-llava-next-8b to convert_llava_next_weights_to_hf.py

[jamt9000]

Feature request

The convert_llava_next_weights_to_hf.py script does not support converting the LLaVA-NeXT model based on Llama3-8B, llama3-llava-next-8b (announced here) meaning it is hard to load the weights of this model with LlavaNextForConditionalGeneration from the transformer library.

Motivation

Adding support would allow loading llama3-llava-next-8b, which is a strong multimodal model, with LlavaNextForConditionalGeneration included in the transformers lib, allowing full support for transformers functionality (in particular I'd like to use it with vLLM which only implements support for LlavaNextForConditionalGeneration rather than the model implementation from the LLaVA repo).

Your contribution

I have confirmed that modifying the script to add lmms-lab/llama3-llava-next-8b and setting

        text_model_id = "meta-llama/Meta-Llama-3-8B-Instruct"
        image_token_index = 128256

works and the model output seems sensible (although I'm unsure of the exact subtleties of the conversion and extra tokens). I have made a PR: #31395