@article{xecomlaupcyclingpretrainedattentioninto, title = {X-EcoMLA: Upcycling Pre-Trained Attention into MLA for Efficient and Extreme KV Compression}, author = {Guihong Li and Mehdi Rezagholizadeh and Mingyu Yang and Vikram Appia and Emad Barsoum}, year = {2025}, eprint = {2503.11132}, archivePrefix = {arXiv}, url = {https://arxiv.org/abs/2503.11132v2}, }