---
title: "TileWeave: Heterogeneous Tiling for Sparse Prefill on Mobile SoCs"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2611.00408v1"
permalink: "https://feed.astralor.com/items/itm_0e84057e85952209f7d47a4d"
published: "2026-10-30T21:56:01.000Z"
score: 54
channel: "tech"
tags: ["LLM","推理效率","RAG","TileWeave","Heterogeneous Tiling","Sparse Prefill","Mobile SoCs Smartphone","Block-sparse"]
---

# TileWeave: Heterogeneous Tiling for Sparse Prefill on Mobile SoCs

## 速览

Smartphone applications increasingly require large language models to process long contexts, whereas high prefill latency and memory demands constrain local inference. Block-sparse attention reduces computation and memo…

## 关键点

- Smartphone applications increasingly require large language models to process long contexts, whereas high prefill latency and memory demands constrain local inference.
- Block-sparse attention reduces computation and memory traffic, but its irregular selections conflict with mobile NPUs' preference for large, regular tiles.
- Efficient execution must exploit cross-query reuse while avoiding redundant computation and preserving selections needed for accuracy.

## 全文

Smartphone applications increasingly require large language models to process long contexts, whereas high prefill latency and memory demands constrain local inference. Block-sparse attention reduces computation and memory traffic, but its irregular selections conflict with mobile NPUs' preference for large, regular tiles. Efficient execution must exploit cross-query reuse while avoiding redundant computation and preserving selections needed for accuracy. In this paper, we present TileWeave, which partitions sparse attention by this cross-query reuse: highly shared key blocks run in coarse NPU tiles, while low-reuse blocks run on the GPU only for the query blocks that selected them. In addition, we design an NPU-native pooled-key selector that keeps selection inexpensive, and a runtime that publishes GPU partials in the order the NPU consumes them and merges them in its existing output store. On RULER, HeteroSparse matches XAttention's accuracy with Qwen3-1.7B and improves it with Llama-3.2-3B-Instruct; its mean LongBench scores match XAttention as well. On Snapdragon SM8750 and SM8850 phones, it achieves 1.35--3.23$\times$ faster full-model prefill than XAttention over 1k--16k tokens and up to 1.58$\times$ faster than dense attention.

---

引用前请核对原文;本文由星际流动整理:https://feed.astralor.com/items/itm_0e84057e85952209f7d47a4d