llava surgery script for new llava-arch model from Intel
KohakuBlueleaf opened this issue · 0 comments
KohakuBlueleaf commented
Intel recently have uploaded a llava-1.5 arch model which use llama3-8B as backbone:
https://huggingface.co/Intel/llava-llama-3-8b/tree/main
And looks like their keynames in the state_dict are totally different from what it was. I think we may need some updates for llava surgery/convertion script?
Or anything did I missed?
my llama.cpp version is 9afdffe