diffusers/pr_12652/en/optimization/cache.html

<meta charset="utf-8" /><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;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;title&quot;:&quot;Usage&quot;,&quot;local&quot;:&quot;usage&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3}],&quot;depth&quot;:2}],&quot;depth&quot;:1}">
		<link href="/docs/diffusers/pr_12652/en/_app/immutable/assets/0.e3b0c442.css" rel="modulepreload">
		<link rel="modulepreload" href="/docs/diffusers/pr_12652/en/_app/immutable/entry/start.78b62fee.js">
		<link rel="modulepreload" href="/docs/diffusers/pr_12652/en/_app/immutable/chunks/scheduler.53228c21.js">
		<link rel="modulepreload" href="/docs/diffusers/pr_12652/en/_app/immutable/chunks/singletons.89d0b97a.js">
		<link rel="modulepreload" href="/docs/diffusers/pr_12652/en/_app/immutable/chunks/index.e93d0901.js">
		<link rel="modulepreload" href="/docs/diffusers/pr_12652/en/_app/immutable/chunks/paths.67f826e3.js">
		<link rel="modulepreload" href="/docs/diffusers/pr_12652/en/_app/immutable/entry/app.062e1615.js">
		<link rel="modulepreload" href="/docs/diffusers/pr_12652/en/_app/immutable/chunks/preload-helper.222e0275.js">
		<link rel="modulepreload" href="/docs/diffusers/pr_12652/en/_app/immutable/chunks/index.100fac89.js">
		<link rel="modulepreload" href="/docs/diffusers/pr_12652/en/_app/immutable/nodes/0.fe8af227.js">
		<link rel="modulepreload" href="/docs/diffusers/pr_12652/en/_app/immutable/chunks/each.e59479a4.js">
		<link rel="modulepreload" href="/docs/diffusers/pr_12652/en/_app/immutable/nodes/282.3dcb8a0d.js">
		<link rel="modulepreload" href="/docs/diffusers/pr_12652/en/_app/immutable/chunks/CopyLLMTxtMenu.50ab6782.js">
		<link rel="modulepreload" href="/docs/diffusers/pr_12652/en/_app/immutable/chunks/globals.7f7f1b26.js">
		<link rel="modulepreload" href="/docs/diffusers/pr_12652/en/_app/immutable/chunks/IconCopy.38cf8f56.js">
		<link rel="modulepreload" href="/docs/diffusers/pr_12652/en/_app/immutable/chunks/MermaidChart.svelte_svelte_type_style_lang.720a8c3c.js">
		<link rel="modulepreload" href="/docs/diffusers/pr_12652/en/_app/immutable/chunks/CodeBlock.d30a6509.js"><!-- HEAD_svelte-u9bgzb_START --><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;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;title&quot;:&quot;Usage&quot;,&quot;local&quot;:&quot;usage&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3}],&quot;depth&quot;:2}],&quot;depth&quot;:1}"><!-- HEAD_svelte-u9bgzb_END -->      <p></p>   <div class="items-center shrink-0 min-w-[100px] max-sm:min-w-[50px] justify-end ml-auto flex" style="float: right; margin-left: 10px; display: inline-flex; position: relative; z-index: 10;"><div 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>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> </div>  <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 class="" 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>Caching</span></h1> <p data-svelte-h="svelte-y872v2">Caching accelerates inference by storing and reusing intermediate outputs of different layers, such as attention and feedforward layers, instead of performing the entire computation at each inference step. It significantly improves generation speed at the expense of more memory and doesn’t require additional training.</p> <p data-svelte-h="svelte-67viyl">This guide shows you how to use the caching methods supported in Diffusers.</p>  <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 class="" 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>Pyramid Attention Broadcast</span></h2> <p data-svelte-h="svelte-175y28v"><a href="https://huggingface.co/papers/2408.12588" rel="nofollow">Pyramid Attention Broadcast (PAB)</a> is based on the observation that attention outputs aren’t that different between successive timesteps of the generation process. The attention differences are smallest in the cross attention layers and are generally cached over a longer timestep range. This is followed by temporal attention and spatial attention layers.</p> <blockquote class="tip" data-svelte-h="svelte-jobxuf"><p>Not all video models have three types of attention (cross, temporal, and spatial)!</p></blockquote> <p data-svelte-h="svelte-rlr2qi">PAB can be combined with other techniques like sequence parallelism and classifier-free guidance parallelism (data parallelism) for near real-time video generation.</p> <p data-svelte-h="svelte-1pz2r0k">Set up and pass a <a href="/docs/diffusers/pr_12652/en/api/cache#diffusers.PyramidAttentionBroadcastConfig">PyramidAttentionBroadcastConfig</a> to a pipeline’s transformer to enable it. The <code>spatial_attention_block_skip_range</code> controls how often to skip attention calculations in the spatial attention blocks and the <code>spatial_attention_timestep_skip_range</code> is the range of timesteps to skip. Take care to choose an appropriate range because a smaller interval can lead to slower inference speeds and a larger interval can result in lower generation quality.</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 class="" 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> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><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)<!-- HTML_TAG_END --></pre></div>  <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 class="" 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 data-svelte-h="svelte-1hv75yq"><a href="https://huggingface.co/papers/2410.19355" rel="nofollow">FasterCache</a> caches and reuses attention features similar to <a href="#pyramid-attention-broadcast">PAB</a> since output differences are small for each successive timestep.</p> <p data-svelte-h="svelte-1ekw8kg">This method may also choose to skip the unconditional branch prediction, when using classifier-free guidance for sampling (common in most base models), and estimate it from the conditional branch prediction if there is significant redundancy in the predicted latent outputs between successive timesteps.</p> <p data-svelte-h="svelte-1fu4r6d">Set up and pass a <a href="/docs/diffusers/pr_12652/en/api/cache#diffusers.FasterCacheConfig">FasterCacheConfig</a> to a pipeline’s transformer to enable it.</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 class="" 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> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><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)<!-- HTML_TAG_END --></pre></div>  <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 class="" 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 data-svelte-h="svelte-gd30ua"><a href="https://huggingface.co/docs/diffusers/main/en/api/cache#diffusers.FirstBlockCacheConfig" rel="nofollow">FirstBlock Cache</a> checks how much the early layers of the denoiser changes from one timestep to the next. If the change is small, the model skips the expensive later layers and reuses the previous output.</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 class="" 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> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><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>, torch_dtype=torch.bfloat16
)
apply_first_block_cache(pipeline.transformer, FirstBlockCacheConfig(threshold=<span class="hljs-number">0.2</span>))<!-- HTML_TAG_END --></pre></div>  <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 class="" 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 Cache</span></h2> <p data-svelte-h="svelte-zcsxd2"><a href="https://huggingface.co/papers/2403.06923" rel="nofollow">TaylorSeer Cache</a> accelerates diffusion inference by using Taylor series expansions to approximate and cache intermediate activations across denoising steps. The method predicts future outputs based on past computations, reusing them at specified intervals to reduce redundant calculations.</p> <p data-svelte-h="svelte-wgp4uh">This caching mechanism delivers strong results with minimal additional memory overhead. For detailed performance analysis, see <a href="https://github.com/huggingface/diffusers/pull/12648#issuecomment-3610615080" rel="nofollow">our findings here</a>.</p> <p data-svelte-h="svelte-q1omc6">To enable TaylorSeer Cache, create a <a href="/docs/diffusers/pr_12652/en/api/cache#diffusers.TaylorSeerCacheConfig">TaylorSeerCacheConfig</a> and pass it to your pipeline’s transformer:</p> <ul data-svelte-h="svelte-ilgde3"><li><code>cache_interval</code>: Number of steps to reuse cached outputs before performing a full forward pass</li> <li><code>disable_cache_before_step</code>: Initial steps that use full computations to gather data for approximations</li> <li><code>max_order</code>: Approximation accuracy (in theory, higher values improve quality but increase memory usage but we recommend it should be set to <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 class="" 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> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><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>,
    torch_dtype=torch.bfloat16,
).to(<span class="hljs-string">&quot;cuda&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)<!-- HTML_TAG_END --></pre></div>  <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 class="" 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 data-svelte-h="svelte-18b6w9g"><a href="https://github.com/Zehong-Ma/MagCache" rel="nofollow">MagCache</a> accelerates inference by skipping transformer blocks based on the magnitude of the residual update. It observes that the magnitude of updates (Output - Input) decays predictably over the diffusion process. By accumulating an “error budget” based on pre-computed magnitude ratios, it dynamically decides when to skip computation and reuse the previous residual.</p> <p data-svelte-h="svelte-8wc7v4">MagCache relies on <strong>Magnitude Ratios</strong> (<code>mag_ratios</code>), which describe this decay curve. These ratios are specific to the model checkpoint and scheduler.</p>  <h3 class="relative group"><a id="usage" 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="#usage"><span><svg class="" 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>Usage</span></h3> <p data-svelte-h="svelte-1igke45">To use MagCache, you typically follow a two-step process: <strong>Calibration</strong> and <strong>Inference</strong>.</p> <ol data-svelte-h="svelte-1ewqnvj"><li><strong>Calibration</strong>: Run inference once with <code>calibrate=True</code>. The hook will measure the residual magnitudes and print the calculated ratios to the console.</li> <li><strong>Inference</strong>: Pass these ratios to <code>MagCacheConfig</code> to enable acceleration.</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 class="" 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> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><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>,
    torch_dtype=torch.bfloat16
).to(<span class="hljs-string">&quot;cuda&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>]<!-- HTML_TAG_END --></pre></div> <blockquote class="note" data-svelte-h="svelte-cxe63a"><p><code>mag_ratios</code> represent the model’s intrinsic magnitude decay curve. Ratios calibrated for a high number of steps (e.g., 50) can be reused for lower step counts (e.g., 20). The implementation uses interpolation to map the curve to the current number of inference steps.</p></blockquote> <blockquote class="tip" data-svelte-h="svelte-1i0grgu"><p>For pipelines that run Classifier-Free Guidance sequentially (like Kandinsky 5.0), the calibration log might print two arrays: one for the Conditional pass and one for the Unconditional pass. In most cases, you should use the first array (Conditional).</p></blockquote> <blockquote class="tip" data-svelte-h="svelte-1gyu2q"><p>For pipelines that run Classifier-Free Guidance in a <strong>batched</strong> manner (like SDXL or Flux), the <code>hidden_states</code> processed by the model contain both conditional and unconditional branches concatenated together. The calibration process automatically accounts for this, producing a single array of ratios that represents the joint behavior. You can use this resulting array directly without modification.</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 data-svelte-h="svelte-zjs2n5"><span class="underline">Update</span> on GitHub</span></a>  <p></p> 
			
			<script>
				{
					__sveltekit_1te7xiv = {
						assets: "/docs/diffusers/pr_12652/en",
						base: "/docs/diffusers/pr_12652/en",
						env: {}
					};

					const element = document.currentScript.parentElement;

					const data = [null,null];

					Promise.all([
						import("/docs/diffusers/pr_12652/en/_app/immutable/entry/start.78b62fee.js"),
						import("/docs/diffusers/pr_12652/en/_app/immutable/entry/app.062e1615.js")
					]).then(([kit, app]) => {
						kit.start(app, element, {
							node_ids: [0, 282],
							data,
							form: null,
							error: null
						});
					});
				}
			</script>