---
id: 20260915-T0-19
title: "恒定大小缓存让块扩散模型在规模化时更省显存"
title_en: "Constant-Size Cache Boosts Block Diffusion at Scale"
url: https://ai.daily.yangsir.net/daily/20260915-T0-19
issue_date: 2026-09-15
publish_date: 2026-09-14T04:00:00.000Z
category: research
source_name: "arXiv cs.LG (ML)"
source_url: https://arxiv.org/abs/2609.11998
---

# 恒定大小缓存让块扩散模型在规模化时更省显存

一篇新论文研究块扩散语言模型的缓存机制。扩散模型并行解码token，但其双向去噪结构无法直接使用自回归推理中的KV缓存。块扩散通过逐块解码恢复缓存能力，论文分析了恒定大小缓存在大规模场景下带来的收益与限制。

## English Version

**Constant-Size Cache Boosts Block Diffusion at Scale**

A new paper examines caching in block diffusion language models. Diffusion models decode tokens in parallel, but their bidirectional denoiser rules out the naive key-value cache used in fast autoregressive inference. Block diffusion restores caching by decoding block-by-block. The paper analyzes what a constant-size cache buys block diffusion at scale.

---

**来源**：[arXiv cs.LG (ML)](https://arxiv.org/abs/2609.11998)

**详情页**：https://ai.daily.yangsir.net/daily/20260915-T0-19

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*