Cross-Architecture Distillation Recipe for Transformer to Mamba Migration

This method enables distilling pretrained Transformer models into Mamba-like State Space Models using a principled initialization recipe. This approach aims to preserve teacher performance while achieving the higher throughput and lower memory consumption characteristic of SSMs.