---
title: "CLiS: Plug-and-Play Attention Acceleration for Video Diffusion on TPUs via Certified Stabilization"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2611.00444v1"
permalink: "https://feed.astralor.com/items/itm_20baba65552de16d16261da2"
published: "2026-10-30T22:57:09.000Z"
score: 53
channel: "tech"
tags: ["LLM","推理效率","安全","多模态","CLiS","Plug-and-Play Attention Acceleration","Video Diffusion","TPUs"]
---

# CLiS: Plug-and-Play Attention Acceleration for Video Diffusion on TPUs via Certified Stabilization

## 速览

High-resolution video diffusion relies on attention over long token sequences, making attention efficiency critical to generation latency. Our analysis identifies repeated maximum reductions and accumulator rescaling in…

## 关键点

- High-resolution video diffusion relies on attention over long token sequences, making attention efficiency critical to generation latency.
- Our analysis identifies repeated maximum reductions and accumulator rescaling in online softmax as a bottleneck in long-context TPU attention.
- We propose CLiS (Certified Loop-invariant Stabilization), a training-free, plug-and-play optimization for video diffusion inference.

## 全文

High-resolution video diffusion relies on attention over long token sequences, making attention efficiency critical to generation latency. Our analysis identifies repeated maximum reductions and accumulator rescaling in online softmax as a bottleneck in long-context TPU attention. We propose CLiS (Certified Loop-invariant Stabilization), a training-free, plug-and-play optimization for video diffusion inference. CLiS fixes per-row stabilizers before the KV loop and certifies numerical safety per head, falling back to online softmax when certification fails. It integrates into existing fused-attention kernels without changing model weights or architecture. On Wan 2.2 and LTX-2 across TPU v6e and TPU7x, CLiS reduces diffusion transformer (DiT) step latency by up to 19.6% relative to tuned online softmax without measurable quality loss. Controlled TPU adaptations of prior GPU stabilization methods show performance or numerical-safety limitations. Preliminary causal LLM experiments also show up to 13.5% higher training throughput in continued pretraining and 7.7% in supervised fine-tuning.

---

引用前请核对原文;本文由星际流动整理:https://feed.astralor.com/items/itm_20baba65552de16d16261da2