MISA: Mixture of Indexer Sparse Attention for Long-Context LLM Inference

Hugging Face Daily Papers Papers

Summary

The paper introduces MISA, a method that applies a mixture-of-experts approach to the indexer heads in sparse attention mechanisms, significantly reducing computational costs for long-context LLM inference while maintaining performance.

DeepSeek Sparse Attention (DSA) sets the state of the art for fine-grained inference-time sparse attention by introducing a learned token-wise indexer that scores every prefix token and selects the most relevant ones for the main attention. To remain expressive, the indexer uses many query heads (for example, 64 on DeepSeek-V3.2) that share the same selected token set; this multi-head design is precisely what makes the indexer the dominant cost on long contexts. We propose MISA (Mixture of Indexer Sparse Attention), a drop-in replacement for the DSA indexer that treats its indexer heads as a pool of mixture-of-experts. A lightweight router uses cheap block-level statistics to pick a query-dependent subset of only a few active heads, and only those heads run the heavy token-level scoring. This preserves the diversity of the original indexer pool while reducing the per-query cost from scoring every prefix token with every head to scoring it with only a handful of routed heads, plus a negligible router term computed on a small set of pooled keys. We further introduce a hierarchical variant of MISA that uses the routed pass to keep an enlarged candidate set and then re-ranks it with the original DSA indexer to recover the final selected tokens almost exactly. With only eight active heads and no additional training, MISA matches the dense DSA indexer on LongBench across DeepSeek-V3.2 and GLM-5 while running with eight and four times fewer indexer heads respectively, and outperforms HISA on average. It also preserves fully green Needle-in-a-Haystack heatmaps up to a 128K-token context and recovers more than 92% of the tokens selected by the DSA indexer per layer. Our TileLang kernel delivers roughly a 3.82 times speedup over DSA's original indexer kernel on a single NVIDIA H200 GPU.
Original Article
View Cached Full Text

Cached at: 05/11/26, 02:43 AM

Paper page - MISA: Mixture of Indexer Sparse Attention for Long-Context LLM Inference

Source: https://huggingface.co/papers/2605.07363

Abstract

MISA replaces the dense token-wise indexing in sparse attention with a routed mixture-of-experts approach that reduces computational cost while maintaining performance and handling long contexts effectively.

DeepSeekSparse Attention(DSA) sets the state of the art for fine-grained inference-timesparse attentionby introducing a learnedtoken-wise indexerthat scores every prefix token and selects the most relevant ones for the main attention. To remain expressive, the indexer uses manyquery heads(for example, 64 on DeepSeek-V3.2) that share the same selected token set; this multi-head design is precisely what makes the indexer the dominant cost on long contexts. We propose MISA (Mixture of IndexerSparse Attention), a drop-in replacement for the DSA indexer that treats its indexer heads as a pool ofmixture-of-experts. A lightweightrouteruses cheapblock-level statisticsto pick a query-dependent subset of only a few active heads, and only those heads run the heavy token-level scoring. This preserves the diversity of the original indexer pool while reducing the per-query cost from scoring every prefix token with every head to scoring it with only a handful of routed heads, plus a negligiblerouterterm computed on a small set of pooled keys. We further introduce ahierarchical variantof MISA that uses the routed pass to keep an enlarged candidate set and then re-ranks it with the original DSA indexer to recover the final selected tokens almost exactly. With only eight active heads and no additional training, MISA matches the dense DSA indexer on LongBench across DeepSeek-V3.2 and GLM-5 while running with eight and four times fewer indexer heads respectively, and outperforms HISA on average. It also preserves fully greenNeedle-in-a-Haystackheatmaps up to a 128K-token context and recovers more than 92% of the tokens selected by the DSA indexer per layer. OurTileLang kerneldelivers roughly a 3.82 times speedup over DSA’s original indexer kernel on a single NVIDIA H200 GPU.

View arXiv pageView PDFAdd to collection

Get this paper in your agent:

hf papers read 2605\.07363

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2605.07363 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2605.07363 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2605.07363 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

MiniMax Sparse Attention

Hugging Face Daily Papers

MiniMax Sparse Attention introduces a blockwise sparse attention mechanism that achieves significant speedups for ultra-long-context LLMs, reducing per-token attention compute by 28.4x at 1M context with wall-clock speedups of 14.2x for prefill and 7.6x for decoding on H800 GPUs. The method is accompanied by an open-source inference kernel and a publicly released multimodal model.

Accurate and Efficient Long-Term Memory for LLM Agents

arXiv cs.AI

MOSAIC is a structured, conflict-aware long-term memory framework for LLM agents that uses entity-typed graph storage, hash-accelerated retrieval, and active conflict detection to achieve high accuracy and efficiency on long-conversation QA and factual conflict detection tasks.

Inference Time Context Sparsity: Illusion or Opportunity?

arXiv cs.AI

This paper argues that extreme context sparsity is a principled and feasible foundation for LLM inference, showing that current models tolerate up to 100× sparsity without quality loss and that sparse decode kernels can accelerate processing by 10× on existing hardware.