FYI something I noticed is that the rkllm code is using the wrong prompt format for some models. For Phi3, for example, it uses different prompt tags. Probably it should give different prompts for different models, or at least give a warning about it.