You Only Index Once: Cross-Layer Sparse Attention with Shared Routing

2026-06-06 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, extended

Summary

Cross-layer sparse attention (CLSA) is a novel architecture for long-context Large Language Models, addressing critical inference bottlenecks in decoding efficiency, KV-cache storage, and pre-filling. Built upon KV-sharing designs like YOCO, CLSA introduces shared routing indices across cross-decoder layers. Instead of recomputing expensive token-level top-k selection for each layer, a single indexer performs this operation once, reusing the resulting index across all layers. This approach amortizes routing overhead while preserving the fine-grained selectivity of token sparse attention. Experiments demonstrate CLSA's effectiveness, achieving up to 7.6× decoding speedup and 17.1× overall throughput improvement at 128K context. The method maintains model quality, performing comparably to or better than dense baselines on benchmarks such as ARC-Challenge, GSM8K, and DROP.

Key takeaway

For ML Engineers and AI Architects deploying long-context LLMs, CLSA offers a compelling solution to persistent inference bottlenecks. You should evaluate this architecture, particularly if your current systems face challenges with decoding speed, KV-cache memory, or pre-filling efficiency. CLSA provides substantial throughput improvements, achieving up to 17.1× overall speedup at 128K context, while demonstrably preserving model quality. This unified approach reconciles efficiency and accuracy, making it a strong candidate for optimizing your production LLM deployments.

Key insights

Cross-layer sparse attention shares a single routing index across decoder layers to amortize top-k cost and boost LLM inference efficiency.

Principles

• Amortize expensive operations by sharing results across layers.
• Token-level sparsity can maintain quality if routing overhead is managed.
• Jointly optimize pre-filling, KV-cache, and decoding for unified efficiency.

Method

CLSA extends YOCO by adding a single-head indexer to the self-decoder, computing a token-level top-k routing index once. Cross-decoder layers then reuse this index and the shared KV cache, trained via multi-layer distillation.

In practice

• Integrate CLSA into vLLM for substantial inference speedups.
• Consider CLSA for long-context LLMs requiring high throughput.
• Use 2048 selected tokens for a favorable quality-efficiency trade-off.

Topics

• Cross-Layer Sparse Attention
• LLM Inference
• KV Cache Optimization
• Sparse Attention Routing
• Long-Context LLMs
• Decoding Throughput

Best for: MLOps Engineer, AI Engineer, NLP Engineer, AI Scientist, Machine Learning Engineer, AI Architect

Related on AIssential

• You Only Index Once: Cross-Layer Sparse Attention with Shared Routing — Cross-layer sparse attention (CLSA) shares KV cache and routing indices across layers to improve LLM inference efficiency and quality.
• Recent Developments in LLM Architectures: KV Sharing, mHC, and Compressed Attention — LLM architectures are evolving with complex, targeted tweaks to optimize long-context efficiency and reduce computational overhead.
• FlashMemory-DeepSeek-V4: Lightning Index Ultra-Long Context via Lookahead Sparse Attention — Lookahead Sparse Attention (LSA) uses a Neural Memory Indexer to proactively manage KV cache, reducing GPU memory for ultra-long contexts.
• The Evolution of LLM Inference: Decoding algorithms — Part 1 — LLM inference speed hinges on reducing sequential decoding steps by predicting and verifying multiple tokens in parallel.
• Move the Query, Not the Cache: Characterizing Cross-Instance Latent Attention Redistribution Across GPU Fabrics — Routing small queries is often cheaper than moving large KV-cache blocks across GPUs for sparse attention.
• Unifying Sparse Attention with Hierarchical Memory for Scalable Long-Context LLM Serving — SPIN unifies sparse attention and hierarchical memory to significantly boost long-context LLM serving efficiency.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.