diffusers/pr_13881/zh/_app/immutable/nodes/21.D0sjm6BA.js

import"../chunks/DsnmJJEf.js";import{i as w,h as y,C as u,H as l,a as m,E as f,s as g}from"../chunks/DdZvggmf.js";import{p as Z,o as U,s as a,f as j,a as d,b as T,c as h,n as _}from"../chunks/BbekZcyp.js";const X='{"title":"缓存","local":"缓存","sections":[{"title":"金字塔注意力广播","local":"金字塔注意力广播","sections":[],"depth":2},{"title":"FasterCache","local":"fastercache","sections":[],"depth":2}],"depth":1}';var W=h('<meta name="hf:doc:metadata"/>'),k=h(`<p></p> <!> <!> <p>缓存通过存储和重用不同层的中间输出（如注意力层和前馈层）来加速推理，而不是在每个推理步骤执行整个计算。它显著提高了生成速度，但以更多内存为代价，并且不需要额外的训练。</p> <p>本指南向您展示如何在 Diffusers 中使用支持的缓存方法。</p> <!> <p><a href="https://huggingface.co/papers/2408.12588" rel="nofollow">金字塔注意力广播 (PAB)</a> 基于这样一种观察：在生成过程的连续时间步之间，注意力输出差异不大。注意力差异在交叉注意力层中最小，并且通常在一个较长的时间步范围内被缓存。其次是时间注意力和空间注意力层。</p> <blockquote class="tip"><p>并非所有视频模型都有三种类型的注意力（交叉、时间和空间）！</p></blockquote> <p>PAB 可以与其他技术（如序列并行性和无分类器引导并行性（数据并行性））结合，实现近乎实时的视频生成。</p> <p>设置并传递一个 <code>PyramidAttentionBroadcastConfig</code> 到管道的变换器以启用它。<code>spatial_attention_block_skip_range</code> 控制跳过空间注意力块中注意力计算的频率，<code>spatial_attention_timestep_skip_range</code> 是要跳过的时间步范围。注意选择一个合适的范围，因为较小的间隔可能导致推理速度变慢，而较大的间隔可能导致生成质量降低。</p> <!> <!> <p><a href="https://huggingface.co/papers/2410.19355" rel="nofollow">FasterCache</a> 缓存并重用注意力特征，类似于 <a href="#pyramid-attention-broadcast">PAB</a>，因为每个连续时间步的输出差异很小。</p> <p>此方法在使用无分类器引导进行采样时（在大多数基础模型中常见），也可能选择跳过无条件分支预测，并且
如果连续时间步之间的预测潜在输出存在显著冗余，则从条件分支预测中估计它。</p> <p>设置并将 <code>FasterCacheConfig</code> 传递给管道的 transformer 以启用它。</p> <!> <!> <p></p>`,1);function Y(b,J){Z(J,!1),U(()=>{new URLSearchParams(window.location.search).get("fw")}),w();var s=k();y("fs68f1",c=>{var r=W();g(r,"content",X),d(c,r)});var e=a(j(s),2);u(e,{containerStyle:"float: right; margin-left: 10px; display: inline-flex; position: relative; z-index: 10;"});var n=a(e,2);l(n,{title:"缓存",local:"缓存",headingTag:"h1"});var t=a(n,6);l(t,{title:"金字塔注意力广播",local:"金字塔注意力广播",headingTag:"h2"});var p=a(t,10);m(p,{code:"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",highlighted:`<span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> CogVideoXPipeline, PyramidAttentionBroadcastConfig

pipeline = CogVideoXPipeline.from_pretrained(<span class="hljs-string">&quot;THUDM/CogVideoX-5b&quot;</span>, torch_dtype=torch.bfloat16)
pipeline.to(<span class="hljs-string">&quot;cuda&quot;</span>)

config = PyramidAttentionBroadcastConfig(
    spatial_attention_block_skip_range=<span class="hljs-number">2</span>,
    spatial_attention_timestep_skip_range=(<span class="hljs-number">100</span>, <span class="hljs-number">800</span>),
    current_timestep_callback=<span class="hljs-keyword">lambda</span>: pipe.current_timestep,
)
pipeline.transformer.enable_cache(config)`,lang:"python",wrap:!1});var o=a(p,2);l(o,{title:"FasterCache",local:"fastercache",headingTag:"h2"});var i=a(o,8);m(i,{code:"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",highlighted:`<span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> CogVideoXPipeline, FasterCacheConfig

pipe line= CogVideoXPipeline.from_pretrained(<span class="hljs-string">&quot;THUDM/CogVideoX-5b&quot;</span>, torch_dtype=torch.bfloat16)
pipeline.to(<span class="hljs-string">&quot;cuda&quot;</span>)

config = FasterCacheConfig(
    spatial_attention_block_skip_range=<span class="hljs-number">2</span>,
    spatial_attention_timestep_skip_range=(-<span class="hljs-number">1</span>, <span class="hljs-number">681</span>),
    current_timestep_callback=<span class="hljs-keyword">lambda</span>: pipe.current_timestep,
    attention_weight_callback=<span class="hljs-keyword">lambda</span> _: <span class="hljs-number">0.3</span>,
    unconditional_batch_skip_range=<span class="hljs-number">5</span>,
    unconditional_batch_timestep_skip_range=(-<span class="hljs-number">1</span>, <span class="hljs-number">781</span>),
    tensor_format=<span class="hljs-string">&quot;BFCHW&quot;</span>,
)
pipeline.transformer.enable_cache(config)`,lang:"python",wrap:!1});var M=a(i,2);f(M,{source:"https://github.com/huggingface/diffusers/blob/main/docs/source/zh/optimization/cache.md"}),_(2),d(b,s),T()}export{Y as component};