FlashMemory-DeepSeek-V4: Lightning Index Ultra-Long Context via Lookahead Sparse Attention
2026-06-08 • Machine Learning
Machine LearningArtificial Intelligence
AI summaryⓘ
The authors address the problem of high GPU memory use when large language models keep all past information during text generation. They introduce Lookahead Sparse Attention (LSA), which predicts important future information to keep in memory, reducing what needs to be stored. Their approach trains a smaller indexing model separately, avoiding heavy memory demands during training. Tests show their method uses much less memory while maintaining or slightly improving accuracy, especially for very long texts.
Large Language ModelsKV cacheGPU memoryAttention mechanismInferenceDual-encoder architectureRetrieval trainingLong-context evaluationMemory compressionDeepSeek-V4
Authors
Yan Wang, Qifan Zhang, Jiachen Yu, Tian Liang, Dongyang Ma, Xiang Hu, Zibo Lin, Chunyang Li, Zhichao Wang, Jia Li, Yujiu Yang, Haitao Mi, Dong Yu
Abstract
Conventional LLMs keep the full KV cache loaded during decoding, causing a severe GPU memory bottleneck for ultra-long context serving. In this report, we propose Lookahead Sparse Attention (LSA), a novel inference paradigm powered by a Neural Memory Indexer built upon the DeepSeek-V4 architecture. Rather than passively attending to all historical tokens, LSA proactively predicts future context demands and preserves only the query-critical KV chunks in the GPU memory. Crucially, we instantiate this architecture via a backbone-free decoupled training strategy. By formulating the indexer as a standard dual-encoder architecture, we train it independently using standard retrieval training frameworks without ever loading the massive backbone model into GPU memory. We demonstrate that this "less is more" paradigm significantly maximizes serving efficiency while acting as an effective attention denoiser in tasks that rely on long-term global memory. Across primary long-context evaluation suites (e.g., LongBench-v2, LongMemEval, and RULER), FM-DS-V4 compresses the average physical KV cache footprint down to merely 13.5% of the full-context baseline, while consistently preserving or slightly elevating downstream accuracy (+0.6% absolute margin on average). Crucially, at extreme 500K scales, FlashMemory suppresses the physical KV cache overhead by over 90% without destabilizing the backbone's core reasoning capacities.