I see your updated post. Switch over to llamacpp and look up recommended quants and settings. A good place for this info is on /r/localllama