rhymes-ai/Aria-sequential_mlp · Converting another Aria-based model

Hi. I would like to convert another Aria-based model to use the sequential MLP instead of grouped GEMM so that it can be quantized. Aria-UI

I understand that this code replaces the GroupedMLP with a SequentialMLP in the MoELayer. However, it is not clear to me how I can take the Aria-UI model and convert it to using this SequentialMLP for the MoELayer. Do I need to reshape the weights in the safetensors for the MoELayers? Is there a script or pseudocode that you used for this that can be shared?

Thanks for your help in advance.