English

TransMLA: Multi-Head Latent Attention Is All You Need

Machine Learning 2025-06-13 v5 Artificial Intelligence

Abstract

In this paper, we present TransMLA, a framework that seamlessly converts any GQA-based pre-trained model into an MLA-based model. Our approach enables direct compatibility with DeepSeek's codebase, allowing these models to fully leverage DeepSeek-specific optimizations such as vLLM and SGlang. By compressing 93% of the KV cache in LLaMA-2-7B, TransMLA achieves a 10.6x inference speedup at an 8K context length while preserving meaningful output quality. Additionally, the model requires only 6 billion tokens for fine-tuning to regain performance on par with the original across multiple benchmarks. TransMLA offers a practical solution for migrating GQA-based models to the MLA structure. When combined with DeepSeek's advanced features, such as FP8 quantization and Multi-Token Prediction, even greater inference acceleration can be realized.

Keywords

Cite

@article{arxiv.2502.07864,
  title  = {TransMLA: Multi-Head Latent Attention Is All You Need},
  author = {Fanxu Meng and Pingzhi Tang and Xiaojuan Tang and Zengwei Yao and Xing Sun and Muhan Zhang},
  journal= {arXiv preprint arXiv:2502.07864},
  year   = {2025}
}

Comments

https://github.com/fxmeng/TransMLA

R2 v1 2026-06-28T21:40:44.505Z