use "turboquant" various version in git actually - only one got success is llama.cpp one it can load even higher models i heard