Hugging Face's logo

Buckets:
hf-doc-build
/
doc

Files
xet
hf-doc-build/doc / diffusers /v0.43.0 /zh /optimization /cache.html
mishig's picture
mishig
download
raw
37.7 kB
<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;缓存&quot;,&quot;local&quot;:&quot;caching&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;金字塔注意力广播&quot;,&quot;local&quot;:&quot;pyramid-attention-broadcast&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;FasterCache&quot;,&quot;local&quot;:&quot;fastercache&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;SeaCache&quot;,&quot;local&quot;:&quot;seacache&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Cosmos 3&quot;,&quot;local&quot;:&quot;cosmos-3&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3}],&quot;depth&quot;:2},{&quot;title&quot;:&quot;FirstBlockCache&quot;,&quot;local&quot;:&quot;firstblockcache&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;TaylorSeer 缓存&quot;,&quot;local&quot;:&quot;taylorseer-cache&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;MagCache&quot;,&quot;local&quot;:&quot;magcache&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2}],&quot;depth&quot;:1}"/>
<link href="/docs/diffusers/v0.43.0/en/_app/immutable/entry/start.1gaXe03k.js" rel="modulepreload">
<link href="/docs/diffusers/v0.43.0/en/_app/immutable/chunks/BObs6oQO.js" rel="modulepreload">
<link href="/docs/diffusers/v0.43.0/en/_app/immutable/chunks/DPnyVmXs.js" rel="modulepreload">
<link href="/docs/diffusers/v0.43.0/en/_app/immutable/entry/app.BwdYRbCn.js" rel="modulepreload">
<link href="/docs/diffusers/v0.43.0/en/_app/immutable/chunks/DSGUT_OW.js" rel="modulepreload">
<link href="/docs/diffusers/v0.43.0/en/_app/immutable/chunks/C045MWWx.js" rel="modulepreload">
<link href="/docs/diffusers/v0.43.0/en/_app/immutable/chunks/6tuJ3nyi.js" rel="modulepreload">
<link href="/docs/diffusers/v0.43.0/en/_app/immutable/chunks/DsnmJJEf.js" rel="modulepreload">
<link href="/docs/diffusers/v0.43.0/en/_app/immutable/nodes/0.CornNeQR.js" rel="modulepreload">
<link href="/docs/diffusers/v0.43.0/en/_app/immutable/chunks/IaDvMxLQ.js" rel="modulepreload">
<link href="/docs/diffusers/v0.43.0/en/_app/immutable/nodes/308.BXNeqIJG.js" rel="modulepreload">
<link href="/docs/diffusers/v0.43.0/en/_app/immutable/chunks/bKdwVYpU.js" rel="modulepreload">
<!--fs68f1--><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;Caching&quot;,&quot;local&quot;:&quot;caching&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Pyramid Attention Broadcast&quot;,&quot;local&quot;:&quot;pyramid-attention-broadcast&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;FasterCache&quot;,&quot;local&quot;:&quot;fastercache&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;SeaCache&quot;,&quot;local&quot;:&quot;seacache&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Cosmos 3&quot;,&quot;local&quot;:&quot;cosmos-3&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3}],&quot;depth&quot;:2},{&quot;title&quot;:&quot;FirstBlockCache&quot;,&quot;local&quot;:&quot;firstblockcache&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;TaylorSeer Cache&quot;,&quot;local&quot;:&quot;taylorseer-cache&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;MagCache&quot;,&quot;local&quot;:&quot;magcache&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2}],&quot;depth&quot;:1}"/><!---->
<link href="/docs/diffusers/v0.43.0/en/_app/immutable/assets/0.tn0RQdqM.css" rel="modulepreload"> <!--[--><!--[0--><!--[--><!--[0--><!--[--><p></p> <div class="items-center shrink-0 min-w-[100px] max-sm:min-w-[50px] justify-end ml-auto flex" style="float: inline-end; margin-inline-start: 10px; display: inline-flex; position: relative; z-index: 10;"><div dir="ltr" class="inline-flex rounded-md max-sm:rounded-sm"><button class="inline-flex items-center gap-1 h-7 max-sm:h-7 px-2 max-sm:px-1.5 text-sm font-medium text-gray-800 border border-r-0 rounded-l-md max-sm:rounded-l-sm border-gray-200 bg-white hover:shadow-inner dark:border-gray-850 dark:bg-gray-950 dark:text-gray-200 dark:hover:bg-gray-800" aria-live="polite"><span class="inline-flex items-center justify-center rounded-md p-0.5 max-sm:p-0 hover:text-gray-800 dark:hover:text-gray-200"><svg class="sm:size-3.5 size-3" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----></span> <span dir="auto">Copy page</span></button> <button class="inline-flex items-center justify-center w-6 max-sm:w-5 h-7 max-sm:h-7 disabled:pointer-events-none text-sm text-gray-500 hover:text-gray-700 dark:hover:text-white rounded-r-md max-sm:rounded-r-sm border border-l transition border-gray-200 bg-white hover:shadow-inner dark:border-gray-850 dark:bg-gray-950 dark:text-gray-200 dark:hover:bg-gray-800" aria-haspopup="menu" aria-expanded="false" aria-label="Open copy menu"><svg class="transition-transform text-gray-400 overflow-visible sm:size-3.5 size-3 rotate-0" width="1em" height="1em" viewBox="0 0 12 7" fill="none" xmlns="http://www.w3.org/2000/svg"><path d="M1 1L6 6L11 1" stroke="currentColor"></path></svg><!----></button></div> <!--[-1--><!--]--></div><!----> <!--[0--><h1 class="relative group"><a id="caching" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#caching"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span><!---->缓存<!----></span></h1><!--]--><!----> <p>缓存通过存储并复用不同层（如注意力层和前馈层）的中间输出，而非在每个推理步骤中执行完整计算，从而加速推理。它以占用更多内存为代价，显著提升生成速度，且无需额外训练。</p> <p>本指南将向您展示如何使用 Diffusers 支持的缓存方法。</p> <!--[1--><h2 class="relative group"><a id="pyramid-attention-broadcast" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#pyramid-attention-broadcast"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span><!---->金字塔注意力广播<!----></span></h2><!--]--><!----> <p><a href="https://huggingface.co/papers/2408.12588" rel="nofollow">Pyramid Attention Broadcast (PAB)</a>基于以下观察：在生成过程的连续时间步之间，注意力输出之间的差异并不大。交叉注意力层中的注意力差异最小，通常会在更长的时间步范围内进行缓存。随后是时间注意力层和空间注意力层。</p> <blockquote class="tip"><p>并非所有视频模型都包含三种注意力机制（交叉注意力、时间注意力和空间注意力）！</p></blockquote> <p>PAB 可以与其他技术（如序列并行和 classifier-free guidance 并行（数据并行））结合使用，以实现近乎实时的视频生成。</p> <p>设置并传递 <a href="/docs/diffusers/v0.43.0/zh/api/cache#diffusers.PyramidAttentionBroadcastConfig">PyramidAttentionBroadcastConfig</a> 到流水线的 transformer 以启用该功能。<code>spatial_attention_block_skip_range</code> 用于控制空间注意力块中跳过注意力计算的频率，而 <code>spatial_attention_timestep_skip_range</code> 则是需要跳过的时间步范围。请注意选择合适的范围，因为较小的间隔可能导致推理速度变慢，而较大的间隔则可能降低生成质量。</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> 已复制</div><!----></button><!----></div> <pre class="language-python "><!----><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>, dtype=torch.bfloat16)
pipeline.to(<span class="hljs-string">&quot;cuda&quot;</span>) <span class="hljs-comment"># or &quot;mps&quot;, &quot;xpu&quot;, &quot;cpu&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)<!----></pre></div><!----> <!--[1--><h2 class="relative group"><a id="fastercache" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#fastercache"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span><!---->FasterCache<!----></span></h2><!--]--><!----> <p><a href="https://huggingface.co/papers/2410.19355" rel="nofollow">FasterCache</a> 会缓存并复用注意力特征，其原理与 <a href="#pyramid-attention-broadcast">PAB</a> 类似，因为相邻时间步之间的输出差异很小。</p> <p>当使用无分类器引导进行采样时（这在大多数基础模型中很常见），该方法也可以选择跳过无条件分支预测，并在连续时间步的预测潜在输出之间存在显著冗余时，从条件分支预测中对其进行估计。</p> <p>设置一个 <a href="/docs/diffusers/v0.43.0/zh/api/cache#diffusers.FasterCacheConfig">FasterCacheConfig</a> 并将其传递给 pipeline 的 transformer，以启用该功能。</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> 已复制</div><!----></button><!----></div> <pre class="language-python "><!----><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>, dtype=torch.bfloat16)
pipeline.to(<span class="hljs-string">&quot;cuda&quot;</span>) <span class="hljs-comment"># or &quot;mps&quot;, &quot;xpu&quot;, &quot;cpu&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)<!----></pre></div><!----> <!--[1--><h2 class="relative group"><a id="seacache" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#seacache"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span><!---->SeaCache<!----></span></h2><!--]--><!----> <p><a href="https://huggingface.co/papers/2602.18993" rel="nofollow">SeaCache</a> 在连续的降噪步骤之间比较 Spectral-Evolution-Aware (SEA) 指标。当累积的指标变化低于某个阈值时，它会跳过计算开销较大的 transformer block 堆栈，并利用缓存的残差来预测其输出。该指标由构成生成输出的视觉潜变量（visual latents）构建。如果干净的条件帧属于该输出轨迹的一部分（例如在图生视频和视生视频生成中），则应将其包含在内。应排除那些用于条件化生成但不属于输出的一部分的独立视觉提示。文本条件被排除在外，因为它不是视觉潜变量。</p> <p>Cosmos 3 Transfer 将控制提示作为独立的视觉序列进行打包，因此其适配器会将它们从指示器中排除。
Control-CFG 分支在比较相同的输出轨迹时，会保留各自缓存的残差。控制提示仍然会对 transformer 进行条件约束。</p> <p>该实现为以下模型提供了内置适配器：</p> <ul><li><strong>Cosmos 3</strong> 是主要经过优化和基准测试的集成方案。它通过 post-normalization 边界缓存完整的解码器堆栈。</li> <li><strong>Wan T2V</strong> 在 eager 模式下使用通用的重复块路径。此集成展示了另一种单流视频 Transformer 如何向 SeaCache 提供原始视觉潜在变量；这并不意味着相同的缓存参数对 Wan 也是最优的，也不表示支持其他 Wan 变体。</li></ul> <p>其他视频 Transformer 在满足以下条件时，可以集成到通用路径中：使用 <code>CacheMixin</code>、暴露一个可识别的重复块列表，并在 <code>TransformerBlockRegistry</code> 中注册块的输入/输出布局。流水线必须在每次 Transformer 调用时进入 <code>cache_context</code>，附加 <code>step_index</code>、<code>sigma</code> 和 <code>num_inference_steps</code>，并为独立轨迹（例如条件引导和无条件引导）使用不同的上下文名称。当没有内置适配器时，请传入一个 <code>raw_vision_callback</code>，该回调返回构成生成输出的视觉潜在变量。请针对每个模型和调度器验证输出质量并调整缓存参数；对 Cosmos 3 的支持和基准测试结果不会自动迁移。</p> <!--[2--><h3 class="relative group"><a id="cosmos-3" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#cosmos-3"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span><!---->Cosmos 3<!----></span></h3><!--]--><!----> <p>SeaCache 默认处于禁用状态。请在 transformer 上启用它；Cosmos 3 的去噪循环会将当前活跃的调度器步长、sigma 以及步数附加到每个 <code>cache_context</code> 调用中，因此无需进行额外的接线：</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> 已复制</div><!----></button><!----></div> <pre class="language-python "><!----><span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> Cosmos3OmniPipeline, SeaCacheConfig
pipe = Cosmos3OmniPipeline.from_pretrained(<span class="hljs-string">&quot;nvidia/Cosmos3-Nano&quot;</span>)
pipe.transformer.enable_cache(SeaCacheConfig(threshold=<span class="hljs-number">0.2</span>, max_consecutive_cached=<span class="hljs-number">2</span>))<!----></pre></div><!----> <p>此模型级 API 适用于 <a href="/docs/diffusers/v0.43.0/zh/api/pipelines/cosmos3#diffusers.Cosmos3OmniPipeline">Cosmos3OmniPipeline</a>、<a href="/docs/diffusers/v0.43.0/zh/api/pipelines/cosmos3#diffusers.Cosmos3OmniModularPipeline">Cosmos3OmniModularPipeline</a> 和 <a href="/docs/diffusers/v0.43.0/zh/api/pipelines/cosmos3#diffusers.Cosmos3DistilledModularPipeline">Cosmos3DistilledModularPipeline</a>。SeaCache 是一种近似优化，可能会改变生成的输出。当您需要每个去噪步骤都执行完整的 transformer 时，请调用 <code>pipe.transformer.disable_cache()</code>。</p> <!--[1--><h2 class="relative group"><a id="firstblockcache" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#firstblockcache"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span><!---->FirstBlockCache<!----></span></h2><!--]--><!----> <p><a href="https://huggingface.co/docs/diffusers/main/en/api/cache#diffusers.FirstBlockCacheConfig" rel="nofollow">FirstBlock Cache</a> 会检查去噪器（denoiser）的早期层在相邻时间步之间的变化程度。如果变化较小，模型将跳过代价高昂的后续层，并复用之前的输出。</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> 已复制</div><!----></button><!----></div> <pre class="language-py "><!----><span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> DiffusionPipeline
<span class="hljs-keyword">from</span> diffusers.hooks <span class="hljs-keyword">import</span> apply_first_block_cache, FirstBlockCacheConfig
pipeline = DiffusionPipeline.from_pretrained(
<span class="hljs-string">&quot;Qwen/Qwen-Image&quot;</span>, dtype=torch.bfloat16
)
apply_first_block_cache(pipeline.transformer, FirstBlockCacheConfig(threshold=<span class="hljs-number">0.2</span>))<!----></pre></div><!----> <!--[1--><h2 class="relative group"><a id="taylorseer-cache" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#taylorseer-cache"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span><!---->TaylorSeer 缓存<!----></span></h2><!--]--><!----> <p><a href="https://huggingface.co/papers/2403.06923" rel="nofollow">TaylorSeer Cache</a> 通过利用泰勒级数展开来近似并缓存去噪步骤之间的中间激活值，从而加速扩散推理。该方法基于过去的计算预测未来的输出，并在指定的间隔处复用这些结果，以减少冗余计算。</p> <p>该缓存机制在极小的额外内存开销下即可取得显著效果。有关详细的性能分析，请参阅<a href="https://github.com/huggingface/diffusers/pull/12648#issuecomment-3610615080" rel="nofollow">我们在此处的研究结果</a>。</p> <p>要启用 TaylorSeer Cache，请创建一个 <a href="/docs/diffusers/v0.43.0/zh/api/cache#diffusers.TaylorSeerCacheConfig">TaylorSeerCacheConfig</a> 并将其传递给您的 pipeline 中的 transformer：</p> <ul><li><code>cache_interval</code>：在执行完整前向传播之前，复用缓存输出的步数</li> <li><code>disable_cache_before_step</code>：初始步骤，使用完整计算来收集用于近似的数据</li> <li><code>max_order</code>：近似精度（理论上，较高的值可以提高质量，但会增加内存使用量，我们建议将其设置为 <code>1</code>）</li></ul> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> 已复制</div><!----></button><!----></div> <pre class="language-python "><!----><span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> FluxPipeline, TaylorSeerCacheConfig
pipe = FluxPipeline.from_pretrained(
<span class="hljs-string">&quot;black-forest-labs/FLUX.1-dev&quot;</span>,
dtype=torch.bfloat16,
).to(<span class="hljs-string">&quot;cuda&quot;</span>) <span class="hljs-comment"># or &quot;mps&quot;, &quot;xpu&quot;, &quot;cpu&quot;</span>
config = TaylorSeerCacheConfig(
cache_interval=<span class="hljs-number">5</span>,
max_order=<span class="hljs-number">1</span>,
disable_cache_before_step=<span class="hljs-number">10</span>,
taylor_factors_dtype=torch.bfloat16,
)
pipe.transformer.enable_cache(config)<!----></pre></div><!----> <!--[1--><h2 class="relative group"><a id="magcache" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#magcache"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span><!---->MagCache<!----></span></h2><!--]--><!----> <p><a href="https://github.com/Zehong-Ma/MagCache" rel="nofollow">MagCache</a> 通过根据残差更新的幅度来跳过 Transformer 块，从而加速推理。它观察到，在扩散过程中，更新幅度（Output - Input）会可预测地衰减。通过基于预先计算的幅度比率累积“误差预算”，它动态地决定何时跳过计算并复用之前的残差。</p> <p>MagCache 依赖于 <strong>幅度比率</strong>（<code>mag_ratios</code>），用于描述该衰减曲线。这些比率因模型检查点和调度器而异。内置的 <code>FLUX_MAG_RATIOS</code> 是在 FLUX.1 上测量得到的；其他模型（包括 Flux2 和 Flux2 Klein）需要各自进行校准运行。</p> <p>要使用 MagCache，通常需遵循两步流程：<strong>校准</strong>和<strong>推理</strong>。</p> <ol><li><strong>校准</strong>：使用 <code>calibrate=True</code> 运行一次推理。该 hook 会测量残差幅度，并将计算出的比率打印到控制台。</li> <li><strong>推理</strong>：将这些比率传递给 <code>MagCacheConfig</code> 以启用加速。</li></ol> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> 已复制</div><!----></button><!----></div> <pre class="language-python "><!----><span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> FluxPipeline, MagCacheConfig
pipe = FluxPipeline.from_pretrained(
<span class="hljs-string">&quot;black-forest-labs/FLUX.1-schnell&quot;</span>,
dtype=torch.bfloat16
).to(<span class="hljs-string">&quot;cuda&quot;</span>) <span class="hljs-comment"># or &quot;mps&quot;, &quot;xpu&quot;, &quot;cpu&quot;</span>
<span class="hljs-comment"># 1. Calibration Step</span>
<span class="hljs-comment"># Run full inference to measure model behavior.</span>
calib_config = MagCacheConfig(calibrate=<span class="hljs-literal">True</span>, num_inference_steps=<span class="hljs-number">4</span>)
pipe.transformer.enable_cache(calib_config)
<span class="hljs-comment"># Run a prompt to trigger calibration</span>
pipe(<span class="hljs-string">&quot;A cat playing chess&quot;</span>, num_inference_steps=<span class="hljs-number">4</span>)
<span class="hljs-comment"># Logs will print something like: &quot;MagCache Calibration Results: [1.0, 1.37, 0.97, 0.87]&quot;</span>
<span class="hljs-comment"># 2. Inference Step</span>
<span class="hljs-comment"># Apply the specific ratios obtained from calibration for optimized speed.</span>
<span class="hljs-comment"># Note: For Flux models, you can also import defaults: </span>
<span class="hljs-comment"># from diffusers.hooks.mag_cache import FLUX_MAG_RATIOS</span>
mag_config = MagCacheConfig(
mag_ratios=[<span class="hljs-number">1.0</span>, <span class="hljs-number">1.37</span>, <span class="hljs-number">0.97</span>, <span class="hljs-number">0.87</span>],
num_inference_steps=<span class="hljs-number">4</span>
)
pipe.transformer.enable_cache(mag_config)
image = pipe(<span class="hljs-string">&quot;A cat playing chess&quot;</span>, num_inference_steps=<span class="hljs-number">4</span>).images[<span class="hljs-number">0</span>]<!----></pre></div><!----> <blockquote class="note"><p><code>mag_ratios</code> 表示模型固有的幅度衰减曲线。针对较高步数（例如 50 步）校准的比率可复用于较低的步数（例如 20 步）。该实现使用插值将曲线映射到当前的推理步数。</p></blockquote> <blockquote class="tip"><p>对于按顺序执行 Classifier-Free Guidance 的流水线（例如 Kandinsky 5.0），校准日志可能会打印两个数组：一个用于 Conditional 通道，另一个用于 Unconditional 通道。在大多数情况下，应使用第一个数组（Conditional）。</p></blockquote> <blockquote class="tip"><p>对于以<strong>批量</strong>方式运行 Classifier-Free Guidance 的流水线（例如 SDXL 或 Flux），模型处理的<code>hidden_states</code>包含拼接在一起的有条件分支和无条件分支。校准过程会自动处理这一点，生成一个表示联合行为的比率数组。您可以直接使用该生成的数组，无需任何修改。</p></blockquote> <a class="!text-gray-400 !no-underline text-sm flex items-center not-prose mt-4" href="https://github.com/huggingface/diffusers/blob/main/docs/source/en/optimization/cache.md" target="_blank"><svg class="mr-1" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M31,16l-7,7l-1.41-1.41L28.17,16l-5.58-5.59L24,9l7,7z"></path><path d="M1,16l7-7l1.41,1.41L3.83,16l5.58,5.59L8,23l-7-7z"></path><path d="M12.419,25.484L17.639,6.552l1.932,0.518L14.351,26.002z"></path></svg><!----> <span><span class="underline">Update</span> on GitHub</span></a><!----> <p></p><!--]--><!----><!--]--><!--]--><!--]--> <!--[-1--><!--]--><!--]-->
<script>
{
__sveltekit_ywq9g6 = {
base: "/docs/diffusers/v0.43.0/en",
assets: "/docs/diffusers/v0.43.0/en"
};
const element = document.currentScript.parentElement;
Promise.all([
import("/docs/diffusers/v0.43.0/en/_app/immutable/entry/start.1gaXe03k.js"),
import("/docs/diffusers/v0.43.0/en/_app/immutable/entry/app.BwdYRbCn.js")
]).then(([kit, app]) => {
kit.start(app, element, {
node_ids: [0, 308],
data: [null,null],
form: null,
error: null
});
});
}
</script>

Xet Storage Details

Size:
37.7 kB
·
Xet hash:
fc9dd800e433cb1c4fce6c9fa7e342977cd67d5b8469168db9388909623e48d0

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.