I want to help Keras Hub be compatible with the Qwen and Interlm models. Before that, I have some questions I'd like to ask.

[pass-lin]

Qwen and interlm is chinese llm
Their architectures are basically the same as Llama3, with just a few more biases in the attention layer. If I want to support them in Keras, should I create a new class called QwenCasualModel, or simply add a few more configuration options to the existing Llama3?
Additionally, is it possible for me not to provide a Kaggle link, but instead convert the weights directly through HF?