GQLA: Group-Query Latent Attention for Hardware-Adaptive Large Language Model Decoding
The article discusses a new approach called Group-Query Latent Attention (GQLA) for improving large language model decoding. GQLA offers two decoding paths, allowing for efficient inference on different hardware without the need for retraining. This method enhances performance on both high-end and commodity GPUs while maintaining tensor parallelism.
- ▪GQLA is a modification of Multi-head Latent Attention (MLA) that provides two decoding paths.
- ▪The runtime selects the optimal path based on the target hardware, eliminating the need for custom kernels.
- ▪GQLA supports up to 8-way zero-redundancy tensor parallelism on the GQA path.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Machine Learning arXiv:2605.15250 (cs) [Submitted on 14 May 2026] Title:GQLA: Group-Query Latent Attention for Hardware-Adaptive Large Language Model Decoding Authors:Fanxu Meng View a PDF of the paper titled GQLA: Group-Query Latent Attention for Hardware-Adaptive Large Language Model Decoding, by Fanxu Meng View PDF HTML (experimental) Abstract:Multi-head Latent Attention (MLA), the attention used in DeepSeek-V2/V3, jointly compresses keys and values into a low-rank latent and matches the H100 roofline almost perfectly.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
