Flux Attention: Context-Aware Hybrid Attention for Efficient LLMs Inference

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Qiu, Quantong, Hong, Zhiyi, Yang, Yi, Wang, Haitian, Liu, Kebin, Dang, Qingqing, Li, Juntao, Zhang, Min
Format: Preprint
Published: 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866914458861830144
author Qiu, Quantong
Hong, Zhiyi
Yang, Yi
Wang, Haitian
Liu, Kebin
Dang, Qingqing
Li, Juntao
Zhang, Min
author_facet Qiu, Quantong
Hong, Zhiyi
Yang, Yi
Wang, Haitian
Liu, Kebin
Dang, Qingqing
Li, Juntao
Zhang, Min
contents The quadratic computational complexity of standard attention mechanisms presents a severe scalability bottleneck for LLMs in long-context scenarios. While hybrid attention mechanisms combining Full Attention (FA) and Sparse Attention (SA) offer a potential solution, existing methods typically rely on static allocation ratios that fail to accommodate the variable retrieval demands of different tasks. Furthermore, head-level dynamic sparsity often introduces severe computational load imbalance and synchronization long-tails, which hinder hardware acceleration during autoregressive decoding. To bridge this gap, we introduce Flux Attention, a context-aware framework that dynamically optimizes attention computation at the layer level. By integrating a lightweight Layer Router into frozen pretrained LLMs, the proposed method adaptively routes each layer to FA or SA based on the input context. This layer-wise routing preserves high-fidelity information retrieval while ensuring contiguous memory access, translating theoretical computational reductions into practical wall-clock speedups. As a parameter-efficient approach, our framework requires only 12 hours of training on 8$\times$A800 GPUs. Extensive experiments across multiple long-context and mathematical reasoning benchmarks demonstrate that Flux Attention achieves a superior trade-off between performance and inference speed compared with baseline models, with speed improvements of up to $2.8\times$ and $2.0\times$ in the prefill and decode stages.
format Preprint
id arxiv_https___arxiv_org_abs_2604_07394
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Flux Attention: Context-Aware Hybrid Attention for Efficient LLMs Inference
Qiu, Quantong
Hong, Zhiyi
Yang, Yi
Wang, Haitian
Liu, Kebin
Dang, Qingqing
Li, Juntao
Zhang, Min
Machine Learning
Computation and Language
The quadratic computational complexity of standard attention mechanisms presents a severe scalability bottleneck for LLMs in long-context scenarios. While hybrid attention mechanisms combining Full Attention (FA) and Sparse Attention (SA) offer a potential solution, existing methods typically rely on static allocation ratios that fail to accommodate the variable retrieval demands of different tasks. Furthermore, head-level dynamic sparsity often introduces severe computational load imbalance and synchronization long-tails, which hinder hardware acceleration during autoregressive decoding. To bridge this gap, we introduce Flux Attention, a context-aware framework that dynamically optimizes attention computation at the layer level. By integrating a lightweight Layer Router into frozen pretrained LLMs, the proposed method adaptively routes each layer to FA or SA based on the input context. This layer-wise routing preserves high-fidelity information retrieval while ensuring contiguous memory access, translating theoretical computational reductions into practical wall-clock speedups. As a parameter-efficient approach, our framework requires only 12 hours of training on 8$\times$A800 GPUs. Extensive experiments across multiple long-context and mathematical reasoning benchmarks demonstrate that Flux Attention achieves a superior trade-off between performance and inference speed compared with baseline models, with speed improvements of up to $2.8\times$ and $2.0\times$ in the prefill and decode stages.
title Flux Attention: Context-Aware Hybrid Attention for Efficient LLMs Inference
topic Machine Learning
Computation and Language
url https://arxiv.org/abs/2604.07394