Buckets:

hf-doc-build
/

doc

Files
xet
hf-doc-build/doc / diffusers /v0.43.0 /pt /optimization /cache.html
mishig
2 days ago
download
raw
39.2 kB
<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;Cache&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;Cache TaylorSeer&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><!---->Cache<!----></span></h1><!--]--><!----> <p>O cache acelera a inferência armazenando e reutilizando saídas intermediárias de diferentes camadas, como as camadas de atenção e feedforward, em vez de executar todo o cálculo em cada etapa de inferência. Ele melhora significativamente a velocidade de geração à custa de mais memória e não requer treinamento adicional.</p> <p>Este guia mostra como usar os métodos de cache suportados no Diffusers.</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><!---->Pyramid Attention Broadcast<!----></span></h2><!--]--><!----> <p><a href="https://huggingface.co/papers/2408.12588" rel="nofollow">Pyramid Attention Broadcast (PAB)</a> baseia-se na observação de que as saídas de atenção não diferem muito entre os timesteps sucessivos do processo de geração. As diferenças de atenção são menores nas camadas de atenção cruzada e, em geral, são armazenadas em cache por um intervalo de timesteps mais longo. Isso é seguido pelas camadas de atenção temporal e atenção espacial.</p> <blockquote class="tip"><p>Nem todos os modelos de vídeo possuem os três tipos de atenção (cross, temporal e espacial)!</p></blockquote> <p>O PAB pode ser combinado com outras técnicas, como paralelismo de sequência e paralelismo de orientação sem classificador (paralelismo de dados), para geração de vídeo em tempo quase real.</p> <p>Configure e passe um <a href="/docs/diffusers/v0.43.0/pt/api/cache#diffusers.PyramidAttentionBroadcastConfig">PyramidAttentionBroadcastConfig</a> para o transformer de um pipeline para ativá-lo. O <code>spatial_attention_block_skip_range</code> controla a frequência com que os cálculos de atenção são ignorados nos blocos de atenção espacial, e o <code>spatial_attention_timestep_skip_range</code> é o intervalo de timesteps a serem ignorados. Tenha cuidado ao escolher um intervalo apropriado, pois um intervalo menor pode levar a velocidades de inferência mais lentas e um intervalo maior pode resultar em menor qualidade de geração.</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> Copiado</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> faz o cache e reutiliza os recursos de atenção de forma semelhante ao <a href="#pyramid-attention-broadcast">PAB</a>, pois as diferenças de saída são pequenas para cada passo de tempo sucessivo.</p> <p>Este método também pode optar por pular a previsão de ramo incondicional, ao usar a orientação sem classificador para amostragem (comum na maioria dos modelos base), e estimá-la a partir da previsão de ramo condicional, caso haja redundância significativa nas saídas latentes previstas entre os passos de tempo sucessivos.</p> <p>Configure e passe um <a href="/docs/diffusers/v0.43.0/pt/api/cache#diffusers.FasterCacheConfig">FasterCacheConfig</a> para o transformer de um pipeline para ativá-lo.</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> Copiado</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> compara os indicadores Spectral-Evolution-Aware (SEA) entre as etapas sucessivas de remoção de ruído. Quando a mudança acumulada do indicador permanece abaixo de um limiar, ele ignora a pilha de blocos de transformer de alto custo e prevê sua saída a partir de resíduos em cache. Construa o indicador a partir dos latentes visuais que formam a saída gerada. Inclua os quadros de condicionamento limpos quando eles fazem parte dessa trajetória de saída, como na geração de imagem para vídeo e de vídeo para vídeo. Exclua dicas visuais separadas que condicionam a geração, mas que não fazem parte da saída. O condicionamento de texto é excluído porque não é um latente visual.</p> <p>Os pacotes de transferência do Cosmos 3 tratam as dicas de controle como sequências visuais separadas, de modo que seu adaptador as exclui do indicador.
Os ramos do Control-CFG comparam a mesma trajetória de saída, mas mantêm seus próprios resíduos em cache. As dicas de controle ainda
condicionam o transformer.</p> <p>A implementação fornece adaptadores integrados para os seguintes modelos:</p> <ul><li><strong>Cosmos 3</strong> é a integração principal otimizada e benchmarkada. Ele faz o cache da pilha completa do decodificador por meio de uma fronteira de pós-normalização.</li> <li><strong>Wan T2V</strong> usa o caminho de bloco repetido genérico em modo eager. Essa integração demonstra como outro
transformer de vídeo de fluxo único pode fornecer latentes de visão brutos ao SeaCache; não é uma afirmação de que os mesmos
parâmetros de cache sejam ótimos para o Wan ou de que outras variantes do Wan sejam suportadas.</li></ul> <p>Outros transformers de vídeo podem se integrar ao caminho genérico quando utilizam <code>CacheMixin</code>, expõem uma lista de blocos repetidos reconhecida e registram o layout de entrada/saída dos blocos em <code>TransformerBlockRegistry</code>. O pipeline deve entrar em um <code>cache_context</code> para cada chamada do transformer, anexar <code>step_index</code>, <code>sigma</code> e <code>num_inference_steps</code>, e usar nomes de contexto separados para trajetórias independentes, como o direcionamento condicional e incondicional. Passe um <code>raw_vision_callback</code> que retorne os latentes visuais que formam a saída gerada quando nenhum adaptador embutido estiver disponível. Valide a qualidade da saída e ajuste os parâmetros de cache para cada modelo e agendador; o suporte e os resultados de benchmark não são transferidos automaticamente do Cosmos 3.</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 é desativado por padrão. Ative-o no transformer; o loop de desruidificação do Cosmos 3 anexa o passo ativo do agendador, sigma e a contagem de passos a cada chamada de <code>cache_context</code>, portanto, nenhum cabeamento extra é necessário:</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> Copiado</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>Esta API de nível de modelo funciona com <a href="/docs/diffusers/v0.43.0/pt/api/pipelines/cosmos3#diffusers.Cosmos3OmniPipeline">Cosmos3OmniPipeline</a>, <a href="/docs/diffusers/v0.43.0/pt/api/pipelines/cosmos3#diffusers.Cosmos3OmniModularPipeline">Cosmos3OmniModularPipeline</a> e <a href="/docs/diffusers/v0.43.0/pt/api/pipelines/cosmos3#diffusers.Cosmos3DistilledModularPipeline">Cosmos3DistilledModularPipeline</a>. O SeaCache é uma otimização aproximada e pode alterar as saídas geradas. Chame <code>pipe.transformer.disable_cache()</code> quando for necessário que cada etapa de desruidamento execute o transformador completo.</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> verifica o quanto as camadas iniciais do denoiser mudam de um timestep para o próximo. Se a mudança for pequena, o modelo ignora as camadas posteriores, mais custosas, e reutiliza a saída anterior.</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> Copiado</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><!---->Cache TaylorSeer<!----></span></h2><!--]--><!----> <p><a href="https://huggingface.co/papers/2403.06923" rel="nofollow">TaylorSeer Cache</a> acelera a inferência de difusão usando expansões em série de Taylor para aproximar e armazenar em cache as ativações intermediárias entre as etapas de desruido. O método prevê saídas futuras com base em cálculos passados, reutilizando-as em intervalos especificados para reduzir cálculos redundantes.</p> <p>Esse mecanismo de cache entrega resultados robustos com um sobrecarga de memória adicional mínima. Para uma análise detalhada de desempenho, consulte <a href="https://github.com/huggingface/diffusers/pull/12648#issuecomment-3610615080" rel="nofollow">nossas descobertas aqui</a>.</p> <p>Para habilitar o TaylorSeer Cache, crie um <a href="/docs/diffusers/v0.43.0/pt/api/cache#diffusers.TaylorSeerCacheConfig">TaylorSeerCacheConfig</a> e passe-o para o transformer do seu pipeline:</p> <ul><li><code>cache_interval</code>: Número de etapas para reutilizar saídas em cache antes de executar uma passagem completa</li> <li><code>disable_cache_before_step</code>: Etapas iniciais que usam computações completas para coletar dados para aproximações</li> <li><code>max_order</code>: Precisão da aproximação (em teoria, valores maiores melhoram a qualidade, mas aumentam o uso de memória; recomendamos definir como <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> Copiado</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> acelera a inferência ao pular blocos de transformador com base na magnitude da atualização residual. Ele observa que a magnitude das atualizações (Output - Input) decai de forma previsível ao longo do processo de difusão. Ao acumular um “orçamento de erro” com base em razões de magnitude pré-computadas, ele decide dinamicamente quando pular o cálculo e reutilizar a residual anterior.</p> <p>MagCache depende de <strong>Magnitude Ratios</strong> (<code>mag_ratios</code>), que descrevem essa curva de decaimento. Essas proporções são específicas para o checkpoint do modelo e o scheduler. O <code>FLUX_MAG_RATIOS</code> incluído foi medido no FLUX.1; outros modelos, incluindo Flux2 e Flux2 Klein, precisam de sua própria execução de calibração.</p> <p>Para usar o MagCache, você geralmente segue um processo de duas etapas: <strong>Calibração</strong> e <strong>Inferência</strong>.</p> <ol><li><strong>Calibração</strong>: execute a inferência uma vez com <code>calibrate=True</code>. O hook medirá as magnitudes dos resíduos e imprimirá as proporções calculadas no console.</li> <li><strong>Inferência</strong>: Passe essas proporções para <code>MagCacheConfig</code> para habilitar a aceleração.</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> Copiado</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> representam a curva de decaimento de magnitude intrínseca do modelo. Proporções calibradas para um número alto de etapas (por exemplo, 50) podem ser reutilizadas para contagens de etapas menores (por exemplo, 20). A implementação usa interpolação para mapear a curva para o número atual de etapas de inferência.</p></blockquote> <blockquote class="tip"><p>Para pipelines que executam Classifier-Free Guidance de forma sequencial (como o Kandinsky 5.0), o log de calibração pode imprimir dois arrays: um para a passagem Conditional e outro para a passagem Unconditional. Na maioria dos casos, você deve usar o primeiro array (Conditional).</p></blockquote> <blockquote class="tip"><p>Para pipelines que executam Classifier-Free Guidance de forma <strong>em lote</strong> (como SDXL ou Flux), os <code>hidden_states</code> processados pelo modelo contêm tanto os ramos condicionais quanto os incondicionais concatenados. O processo de calibração leva isso automaticamente em consideração, produzindo um único array de proporções que representa o comportamento conjunto. Você pode usar esse array resultante diretamente, sem modificações.</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:: 39.2 kB
Xet hash:: d575af92e64ca921415bb1ac6c950b6074afa43bda5d6127560e409e69903603
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.