首页 > AI前沿 > LoopPrune: Recurrent Pruning for Looped Diffusion Transformers

LoopPrune: Recurrent Pruning for Looped Diffusion Transformers

arXiv机器学习 2026-11-17 21:25 7 阅读 查看原文
用户现在需要把给定的纯文本新闻正文排版成语义化HTML片段,要遵守规则。首先,先理清楚内容,开头是背景部分,然后提出方法,接着是实验结果。首先,整个内容都是自然段落吗?不对,先看,首先第一部分是背景:Looped diffusion Transformers...deployment burdens. 然后下一句是Yet repeated...costly. 不对,原文的段落怎么分?哦,原文的正文是连续的,需要把每个自然段落用p包裹。不对,先拆分:首先第一段应该是背景介绍,从开头到"reducing storage and deployment burdens. Yet repeated attention and MLP computations remain costly."?不对,不对,再看,然后接下来是"Pruning reduces...final outputs." 这部分是现有方法的问题,然后"We propose LoopPrune...requirements." 是提出方法,然后后面是方法的三个部分,然后实验部分。哦,不对,用户要求每个自然段落用p包裹,长文用h2/h3划分小节。那首先,先划分小节:比如第一个小节可以是背景与现有问题,用h2,然后段落用p。然后下一个小节是提出的框架,用h3?然后实验验证部分用h2?不对,得保留所有信息,不能增删。等下,原文的内容: 1. 背景:Looped diffusion Transformers reuse shared blocks to deepen computation and match larger models with fewer parameters, reducing storage and deployment burdens. Yet repeated attention and MLP computations remain costly. 2. 现有剪枝方法的不足:Pruning reduces storage and computation by removing redundant parameters or units. Training-free methods struggle to resolve recurrent contributions, while training-based methods lack explicit constraints on position-specific errors affecting later inputs and final outputs. 3. 提出的方法:We propose LoopPrune, a pruning and recovery framework addressing these requirements. 然后方法的三个部分:Recurrent importance estimation combines shared-removal effects and position variation to select a shared structure; shared reconstruction pools recurrent input statistics to compensate weights. Loop-position recovery compares complete denoising predictions at identical student states and matches text guidance, updating all student parameters against a fixed dense teacher. 4. 实验结果:Extensive experiments across four open-source models, six benchmarks, and three pruning paradigms demonstrate effectiveness. At 30% structured parameter pruning, Looped-DiT B/16 retains 93.14% of dense six-benchmark average accuracy, exceeding the best baseline by 9.49%. Across PyTorch, vLLM-Omni, and SGLang Diffusion, pruned SDXSv3 reduces peak allocated GPU memory by 34.36-44.64% and accelerates denoising 1.46-1.74x over dense models. 现在排版,首先,可能用h2划分小节: 首先

研究背景与现有问题

?不对,不对,不能改内容,哦不对,h2是划分小节,但是小节标题不能自己加吗?不对,等下用户说"长文用

/

划分小节",但是不能增删内容,那是不是不能自己加标题?哦,不对,等下看用户的规则4是"保留原文所有信息,不得增删改内容",那如果我自己加小节标题就属于新增内容了,不行。那怎么办?哦,不对,原文有没有可以作为小节的内容?哦,不对,原文的结构:首先讲背景和现有剪枝的问题,然后提出LoopPrune框架,然后讲实验结果。那是不是可以把每一部分的开头用h3?不对,不对,不能加内容,那是不是所有的自然段落用p,然后如果有需要划分的话,比如把方法部分的三个点用列表?哦,等下原文里的方法部分:"Recurrent importance estimation combines shared-removal effects and position variation to select a shared structure; shared reconstruction pools recurrent input statistics to compensate weights. Loop-position recovery compares complete denoising predictions at identical student states and matches text guidance, updating all student parameters against a fixed dense teacher." 这里有三个方法组件,对吗?可以拆成列表?等下,原文是两个句子,第一个句子有两个组件(Recurrent importance estimation和shared reconstruction),第二个是Loop-position recovery。那可以拆成三个li吗?要保留原内容,不能改。 等下重新理,先逐部分来: 首先,第一个自然段(背景):

Looped diffusion Transformers reuse shared blocks to deepen computation and match larger models with fewer parameters, reducing storage and deployment burdens. Yet repeated attention and MLP computations remain costly.

然后第二个自然段(现有剪枝方法的情况和不足):

Pruning reduces storage and computation by removing redundant parameters or units. Training-free methods struggle to resolve recurrent contributions, while training-based methods lack explicit constraints on position-specific errors affecting later inputs and final outputs.

然后第三个部分,提出框架,然后方法组件:

We propose LoopPrune, a pruning and recovery framework addressing these requirements.

然后后面的方法三个部分,可以用h3?不对,不能加标题,那可以用ul?哦,原文的三个方法点是: 1. Recurrent importance estimation combines shared-removal effects and position variation to select a shared structure 2. shared reconstruction pools recurrent input statistics to compensate weights 3. Loop-position recovery compares complete denoising predictions at identical student states and matches text guidance, updating all student parameters against a fixed dense teacher. 那这三个可以作为列表,所以:
  • Recurrent importance estimation combines shared-removal effects and position variation to select a shared structure
  • shared reconstruction pools recurrent input statistics to compensate weights
  • Loop-position recovery compares complete denoising predictions at identical student states and matches text guidance, updating all student parameters against a fixed dense teacher.
不对,原文的这两个句子是用分号连接的,所以拆分的时候要保留内容,没问题。 然后接下来的实验部分,作为

?不对,实验部分可以分吗?不,原文的实验部分是:

Extensive experiments across four open-source models, six benchmarks, and three pruning paradigms demonstrate effectiveness. At 30% structured parameter pruning, Looped-DiT B/16 retains 93.14% of dense six-benchmark average accuracy, exceeding the best baseline by 9.49%. Across PyTorch, vLLM-Omni, and SGLang Diffusion, pruned SDXSv3 reduces peak allocated GPU memory by 34.36-44.64% and accelerates denoising 1.46-1.74x over dense models.

等下,那有没有必要用h2/h3?如果不用的话,会不会不符合要求?用户说"长文用

/

划分小节",那这个内容不算特别长,但是如果要划分的话,不能加自己的标题,那是不是可以用h2把各个部分的开头?不对,不对,原文没有标题,所以不能新增,那是不是就把所有自然段落用p,方法部分用列表,重要的句子可以加strong?比如LoopPrune可以加strong?还有关键数据? 等下,再检查规则:1. 只输出HTML片段,无markdown围栏,无解释。2. 每个自然段落用p包裹,长文用h2/h3划分小节。3. 列表用ul/li,引用blockquote,代码pre/code,重要句子strong。4. 保留所有信息,不增删改。5. 无文档级标签。 那现在,正确的排版应该是: 首先,背景部分作为

,然后现有方法不足作为

,然后提出框架作为

,然后方法的三个组件用

    ,然后实验部分作为

    。不对,但是有没有需要h2的?哦,不对,如果我不添加小节标题的话,用h2没有内容,不行。那是不是可以不用h2?不对,用户说"长文用

    /

    划分小节",如果内容不需要的话?不对,再想,是不是可以把"LoopPrune框架"部分用h3?但是h3需要有文本内容,不能空,那如果用原文的句子作为h3?比如

    We propose LoopPrune, a pruning and recovery framework addressing these requirements.

    ?不对,那这句话本来是正文,改成标题会不会改了?不对,用户说不得增删改内容,所以不能把正文句子改成标题。那哦,对了,可能我理解错了,h2/h3是用来划分小节,但是如果没有明确的小节标题,是不是可以不用?不对,再看用户的需求,是排版纯文本新闻正文,那原文的结构是