Long-context inference in large language models is often bottlenecked by KV cache loading during the decoding stage. While Multi-Head Latent Attention (MLA) reduces the total KV cache size, it suffers from a sharding bottleneck during distributed decoding via Tensor Parallelism (TP).
MLRA addresses this by enabling partitionable latent states for efficient 4-way TP decoding. Experimental results show that MLRA achieves state-of-the-art perplexity and downstream task performance, while delivering a 2.8$\times$ decoding speedup over MLA.
1@inproceedings{liu2026multi,
2 title = {Multi-Head Low-Rank Attention},
3 author = {Liu, Songtao and Peng, Hongwu and Zhang, Zhiwei and Chen, Zhengyu and Guo, Yue},
4 booktitle = {International Conference on Learning Representations},
5 year = {2026}
6}
