Buckets:

hf-doc-build
/

doc

Files
xet
hf-doc-build/doc / diffusers /v0.43.0 /pt /optimization /memory.html
mishig
2 days ago
download
raw
102 kB
<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;Reduzir o uso de memória&quot;,&quot;local&quot;:&quot;reduce-memory-usage&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Várias GPUs&quot;,&quot;local&quot;:&quot;multiple-gpus&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Pontos de verificação fragmentados&quot;,&quot;local&quot;:&quot;sharded-checkpoints&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Colocação de dispositivo&quot;,&quot;local&quot;:&quot;device-placement&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3}],&quot;depth&quot;:2},{&quot;title&quot;:&quot;VAE slicing&quot;,&quot;local&quot;:&quot;vae-slicing&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Tiling do VAE&quot;,&quot;local&quot;:&quot;vae-tiling&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Offloading&quot;,&quot;local&quot;:&quot;offloading&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Offloading de CPU&quot;,&quot;local&quot;:&quot;cpu-offloading&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Deslocamento de modelos&quot;,&quot;local&quot;:&quot;model-offloading&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Descarregamento em grupo&quot;,&quot;local&quot;:&quot;group-offloading&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;CUDA stream&quot;,&quot;local&quot;:&quot;cuda-stream&quot;,&quot;sections&quot;:[],&quot;depth&quot;:4},{&quot;title&quot;:&quot;Offloading para o disco&quot;,&quot;local&quot;:&quot;offloading-to-disk&quot;,&quot;sections&quot;:[],&quot;depth&quot;:4}],&quot;depth&quot;:3}],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Casting por camada&quot;,&quot;local&quot;:&quot;layerwise-casting&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;torch.channels_last&quot;,&quot;local&quot;:&quot;torchchannelslast&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Atenção eficiente em memória&quot;,&quot;local&quot;:&quot;memory-efficient-attention&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2}],&quot;depth&quot;:1}"/>
		<link href="/docs/diffusers/main/en/_app/immutable/entry/start.0S0_iEoG.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/chunks/I9eooN77.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/chunks/Bupkjr9f.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/entry/app.CPdOxCbw.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/chunks/CyvZ3FS5.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/chunks/l0QnnOED.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/chunks/DnVIJfC5.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/chunks/DsnmJJEf.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/nodes/0.DteTvmQZ.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/chunks/CoSrdQax.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/nodes/339.DJYaCibf.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/chunks/DSPDl1gz.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/chunks/D5eijdnr.js" rel="modulepreload">
		<!--13d35im--><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;Reduce memory usage&quot;,&quot;local&quot;:&quot;reduce-memory-usage&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Multiple GPUs&quot;,&quot;local&quot;:&quot;multiple-gpus&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Sharded checkpoints&quot;,&quot;local&quot;:&quot;sharded-checkpoints&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Device placement&quot;,&quot;local&quot;:&quot;device-placement&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3}],&quot;depth&quot;:2},{&quot;title&quot;:&quot;VAE slicing&quot;,&quot;local&quot;:&quot;vae-slicing&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;VAE tiling&quot;,&quot;local&quot;:&quot;vae-tiling&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Offloading&quot;,&quot;local&quot;:&quot;offloading&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;CPU offloading&quot;,&quot;local&quot;:&quot;cpu-offloading&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Model offloading&quot;,&quot;local&quot;:&quot;model-offloading&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Group offloading&quot;,&quot;local&quot;:&quot;group-offloading&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;CUDA stream&quot;,&quot;local&quot;:&quot;cuda-stream&quot;,&quot;sections&quot;:[],&quot;depth&quot;:4},{&quot;title&quot;:&quot;Offloading to disk&quot;,&quot;local&quot;:&quot;offloading-to-disk&quot;,&quot;sections&quot;:[],&quot;depth&quot;:4}],&quot;depth&quot;:3}],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Layerwise casting&quot;,&quot;local&quot;:&quot;layerwise-casting&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;torch.channels_last&quot;,&quot;local&quot;:&quot;torchchannelslast&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Memory-efficient attention&quot;,&quot;local&quot;:&quot;memory-efficient-attention&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2}],&quot;depth&quot;:1}"/><!---->
		<link href="/docs/diffusers/main/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="reduce-memory-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="#reduce-memory-usage"><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><!---->Reduzir o uso de memória<!----></span></h1><!--]--><!----> <p>Modelos de difusão modernos, como <a href="../api/pipelines/flux">Flux</a> e <a href="../api/pipelines/wan">Wan</a>, possuem bilhões de parâmetros que ocupam uma grande quantidade de memória no seu hardware para inferência. Isso é desafiador, pois GPUs comuns frequentemente não possuem memória suficiente. Para superar as limitações de memória, você pode usar mais de uma GPU (se disponível), descarregar alguns dos componentes do pipeline para a CPU, entre outras opções.</p> <p>Este guia mostrará como reduzir o uso de memória.</p> <blockquote class="tip"><p>Lembre-se de que essas técnicas podem precisar ser ajustadas dependendo do modelo. Por exemplo, um modelo de difusão baseado em transformadores pode não se beneficiar igualmente dessas otimizações de memória em comparação com um modelo baseado em UNet.</p></blockquote> <!--[1--><h2 class="relative group"><a id="multiple-gpus" 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="#multiple-gpus"><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><!---->Várias GPUs<!----></span></h2><!--]--><!----> <p>Se você tiver acesso a mais de uma GPU, há algumas opções para carregar e distribuir um modelo grande de forma eficiente entre seu hardware. Esses recursos são suportados pela biblioteca <a href="https://huggingface.co/docs/accelerate/index" rel="nofollow">Accelerate</a>, portanto, certifique-se de que ela esteja instalada primeiro.</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-bash "><!---->pip install -U accelerate<!----></pre></div><!----> <!--[2--><h3 class="relative group"><a id="sharded-checkpoints" 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="#sharded-checkpoints"><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><!---->Pontos de verificação fragmentados<!----></span></h3><!--]--><!----> <p>Carregar checkpoints grandes em vários shards é útil porque os shards são carregados um de cada vez. Isso mantém o uso de memória baixo, exigindo apenas memória suficiente para o tamanho do modelo e o tamanho do maior shard. Recomendamos o uso de sharding quando o checkpoint em fp32 for maior que 5GB. O tamanho padrão do shard é 5GB.</p> <p>Divida um checkpoint em <a href="/docs/diffusers/v0.43.0/pt/api/pipelines/overview#diffusers.DiffusionPipeline.save_pretrained">save_pretrained()</a> com o parâmetro <code>max_shard_size</code>.</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">from</span> diffusers <span class="hljs-keyword">import</span> AutoModel

unet = AutoModel.from_pretrained(
    <span class="hljs-string">&quot;stabilityai/stable-diffusion-xl-base-1.0&quot;</span>, subfolder=<span class="hljs-string">&quot;unet&quot;</span>
)
unet.save_pretrained(<span class="hljs-string">&quot;sdxl-unet-sharded&quot;</span>, max_shard_size=<span class="hljs-string">&quot;5GB&quot;</span>)<!----></pre></div><!----> <p>Agora você pode usar o checkpoint fragmentado, em vez do checkpoint regular, para economizar memória.</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> AutoModel, StableDiffusionXLPipeline

unet = AutoModel.from_pretrained(
    <span class="hljs-string">&quot;username/sdxl-unet-sharded&quot;</span>, dtype=torch.float16
)
pipeline = StableDiffusionXLPipeline.from_pretrained(
    <span class="hljs-string">&quot;stabilityai/stable-diffusion-xl-base-1.0&quot;</span>,
    unet=unet,
    dtype=torch.float16
).to(<span class="hljs-string">&quot;cuda&quot;</span>)  <span class="hljs-comment"># or &quot;mps&quot;, &quot;xpu&quot;, &quot;cpu&quot;</span><!----></pre></div><!----> <!--[2--><h3 class="relative group"><a id="device-placement" 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="#device-placement"><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><!---->Colocação de dispositivo<!----></span></h3><!--]--><!----> <blockquote class="warning"><p>O posicionamento de dispositivos é um recurso experimental e a API pode mudar. No momento, apenas a estratégia <code>balanced</code> é suportada. Planejamos suportar estratégias de mapeamento adicionais no futuro.</p></blockquote> <p>O parâmetro <code>device_map</code> controla como os componentes do modelo em um pipeline ou as camadas em um modelo individual são distribuídos entre os dispositivos.</p> <div class="flex space-x-2 items-center my-1.5 mr-8 h-7 !pl-0 -mx-3 md:mx-0"><!--[--><div class="flex items-center border rounded-lg px-1.5 py-1 leading-none select-none text-smd border-gray-800 bg-black dark:bg-gray-700 text-white">nível de pipeline</div><div class="flex items-center border rounded-lg px-1.5 py-1 leading-none select-none text-smd text-gray-500 cursor-pointer opacity-90 hover:text-gray-700 dark:hover:text-gray-200 hover:shadow-sm">nível do modelo</div><!--]--></div> <div class="language-select"><div><p>A estratégia de posicionamento de dispositivos <code>balanced</code> divide o pipeline igualmente entre todos os dispositivos disponíveis.</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> AutoModel, StableDiffusionXLPipeline

pipeline = StableDiffusionXLPipeline.from_pretrained(
    <span class="hljs-string">&quot;stabilityai/stable-diffusion-xl-base-1.0&quot;</span>,
    dtype=torch.float16,
    device_map=<span class="hljs-string">&quot;balanced&quot;</span>
)<!----></pre></div><!----> <p>Você pode inspecionar o mapa de dispositivos de um pipeline com <code>hf_device_map</code>.</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-built_in">print</span>(pipeline.hf_device_map)
{<span class="hljs-string">&#x27;unet&#x27;</span>: <span class="hljs-number">1</span>, <span class="hljs-string">&#x27;vae&#x27;</span>: <span class="hljs-number">1</span>, <span class="hljs-string">&#x27;safety_checker&#x27;</span>: <span class="hljs-number">0</span>, <span class="hljs-string">&#x27;text_encoder&#x27;</span>: <span class="hljs-number">0</span>}<!----></pre></div><!----><!----></div><!----> <div hidden=""><p>O <code>device_map</code> é útil para carregar modelos grandes, como o Flux diffusion transformer, que possui 12,5 bilhões de parâmetros. Defina-o como <code>"auto"</code> para distribuir automaticamente um modelo primeiro no dispositivo mais rápido antes de passar para dispositivos mais lentos. Consulte a seção <a href="#sharded-checkpoints">Sharded checkpoints</a> para mais detalhes.</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> AutoModel

transformer = AutoModel.from_pretrained(
    <span class="hljs-string">&quot;black-forest-labs/FLUX.1-dev&quot;</span>, 
    subfolder=<span class="hljs-string">&quot;transformer&quot;</span>,
    device_map=<span class="hljs-string">&quot;auto&quot;</span>,
    dtype=torch.bfloat16
)<!----></pre></div><!----> <p>Você pode inspecionar o mapa de dispositivos de um modelo com <code>hf_device_map</code>.</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-built_in">print</span>(transformer.hf_device_map)<!----></pre></div><!----><!----></div><!----><!----></div><!----> <p>Ao projetar seu próprio <code>device_map</code>, ele deve ser um dicionário com o nome específico de um módulo ou camada do modelo e um identificador de dispositivo (um inteiro para GPUs, <code>cpu</code> para CPUs e <code>disk</code> para disco).</p> <p>Chame <code>hf_device_map</code> em um modelo para ver como as camadas do modelo são distribuídas e, em seguida, crie o seu próprio.</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-built_in">print</span>(transformer.hf_device_map)
{<span class="hljs-string">&#x27;pos_embed&#x27;</span>: <span class="hljs-number">0</span>, <span class="hljs-string">&#x27;time_text_embed&#x27;</span>: <span class="hljs-number">0</span>, <span class="hljs-string">&#x27;context_embedder&#x27;</span>: <span class="hljs-number">0</span>, <span class="hljs-string">&#x27;x_embedder&#x27;</span>: <span class="hljs-number">0</span>, <span class="hljs-string">&#x27;transformer_blocks&#x27;</span>: <span class="hljs-number">0</span>, <span class="hljs-string">&#x27;single_transformer_blocks.0&#x27;</span>: <span class="hljs-number">0</span>, <span class="hljs-string">&#x27;single_transformer_blocks.1&#x27;</span>: <span class="hljs-number">0</span>, <span class="hljs-string">&#x27;single_transformer_blocks.2&#x27;</span>: <span class="hljs-number">0</span>, <span class="hljs-string">&#x27;single_transformer_blocks.3&#x27;</span>: <span class="hljs-number">0</span>, <span class="hljs-string">&#x27;single_transformer_blocks.4&#x27;</span>: <span class="hljs-number">0</span>, <span class="hljs-string">&#x27;single_transformer_blocks.5&#x27;</span>: <span class="hljs-number">0</span>, <span class="hljs-string">&#x27;single_transformer_blocks.6&#x27;</span>: <span class="hljs-number">0</span>, <span class="hljs-string">&#x27;single_transformer_blocks.7&#x27;</span>: <span class="hljs-number">0</span>, <span class="hljs-string">&#x27;single_transformer_blocks.8&#x27;</span>: <span class="hljs-number">0</span>, <span class="hljs-string">&#x27;single_transformer_blocks.9&#x27;</span>: <span class="hljs-number">0</span>, <span class="hljs-string">&#x27;single_transformer_blocks.10&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>, <span class="hljs-string">&#x27;single_transformer_blocks.11&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>, <span class="hljs-string">&#x27;single_transformer_blocks.12&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>, <span class="hljs-string">&#x27;single_transformer_blocks.13&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>, <span class="hljs-string">&#x27;single_transformer_blocks.14&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>, <span class="hljs-string">&#x27;single_transformer_blocks.15&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>, <span class="hljs-string">&#x27;single_transformer_blocks.16&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>, <span class="hljs-string">&#x27;single_transformer_blocks.17&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>, <span class="hljs-string">&#x27;single_transformer_blocks.18&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>, <span class="hljs-string">&#x27;single_transformer_blocks.19&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>, <span class="hljs-string">&#x27;single_transformer_blocks.20&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>, <span class="hljs-string">&#x27;single_transformer_blocks.21&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>, <span class="hljs-string">&#x27;single_transformer_blocks.22&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>, <span class="hljs-string">&#x27;single_transformer_blocks.23&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>, <span class="hljs-string">&#x27;single_transformer_blocks.24&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>, <span class="hljs-string">&#x27;single_transformer_blocks.25&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>, <span class="hljs-string">&#x27;single_transformer_blocks.26&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>, <span class="hljs-string">&#x27;single_transformer_blocks.27&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>, <span class="hljs-string">&#x27;single_transformer_blocks.28&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>, <span class="hljs-string">&#x27;single_transformer_blocks.29&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>, <span class="hljs-string">&#x27;single_transformer_blocks.30&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>, <span class="hljs-string">&#x27;single_transformer_blocks.31&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>, <span class="hljs-string">&#x27;single_transformer_blocks.32&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>, <span class="hljs-string">&#x27;single_transformer_blocks.33&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>, <span class="hljs-string">&#x27;single_transformer_blocks.34&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>, <span class="hljs-string">&#x27;single_transformer_blocks.35&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>, <span class="hljs-string">&#x27;single_transformer_blocks.36&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>, <span class="hljs-string">&#x27;single_transformer_blocks.37&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>, <span class="hljs-string">&#x27;norm_out&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>, <span class="hljs-string">&#x27;proj_out&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>}<!----></pre></div><!----> <p>Por exemplo, o <code>device_map</code> abaixo coloca <code>single_transformer_blocks.10</code> até <code>single_transformer_blocks.20</code> em uma segunda GPU (<code>1</code>).</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> AutoModel

device_map = {
    <span class="hljs-string">&#x27;pos_embed&#x27;</span>: <span class="hljs-number">0</span>, <span class="hljs-string">&#x27;time_text_embed&#x27;</span>: <span class="hljs-number">0</span>, <span class="hljs-string">&#x27;context_embedder&#x27;</span>: <span class="hljs-number">0</span>, <span class="hljs-string">&#x27;x_embedder&#x27;</span>: <span class="hljs-number">0</span>, <span class="hljs-string">&#x27;transformer_blocks&#x27;</span>: <span class="hljs-number">0</span>, <span class="hljs-string">&#x27;single_transformer_blocks.0&#x27;</span>: <span class="hljs-number">0</span>, <span class="hljs-string">&#x27;single_transformer_blocks.1&#x27;</span>: <span class="hljs-number">0</span>, <span class="hljs-string">&#x27;single_transformer_blocks.2&#x27;</span>: <span class="hljs-number">0</span>, <span class="hljs-string">&#x27;single_transformer_blocks.3&#x27;</span>: <span class="hljs-number">0</span>, <span class="hljs-string">&#x27;single_transformer_blocks.4&#x27;</span>: <span class="hljs-number">0</span>, <span class="hljs-string">&#x27;single_transformer_blocks.5&#x27;</span>: <span class="hljs-number">0</span>, <span class="hljs-string">&#x27;single_transformer_blocks.6&#x27;</span>: <span class="hljs-number">0</span>, <span class="hljs-string">&#x27;single_transformer_blocks.7&#x27;</span>: <span class="hljs-number">0</span>, <span class="hljs-string">&#x27;single_transformer_blocks.8&#x27;</span>: <span class="hljs-number">0</span>, <span class="hljs-string">&#x27;single_transformer_blocks.9&#x27;</span>: <span class="hljs-number">0</span>, <span class="hljs-string">&#x27;single_transformer_blocks.10&#x27;</span>: <span class="hljs-number">1</span>, <span class="hljs-string">&#x27;single_transformer_blocks.11&#x27;</span>: <span class="hljs-number">1</span>, <span class="hljs-string">&#x27;single_transformer_blocks.12&#x27;</span>: <span class="hljs-number">1</span>, <span class="hljs-string">&#x27;single_transformer_blocks.13&#x27;</span>: <span class="hljs-number">1</span>, <span class="hljs-string">&#x27;single_transformer_blocks.14&#x27;</span>: <span class="hljs-number">1</span>, <span class="hljs-string">&#x27;single_transformer_blocks.15&#x27;</span>: <span class="hljs-number">1</span>, <span class="hljs-string">&#x27;single_transformer_blocks.16&#x27;</span>: <span class="hljs-number">1</span>, <span class="hljs-string">&#x27;single_transformer_blocks.17&#x27;</span>: <span class="hljs-number">1</span>, <span class="hljs-string">&#x27;single_transformer_blocks.18&#x27;</span>: <span class="hljs-number">1</span>, <span class="hljs-string">&#x27;single_transformer_blocks.19&#x27;</span>: <span class="hljs-number">1</span>, <span class="hljs-string">&#x27;single_transformer_blocks.20&#x27;</span>: <span class="hljs-number">1</span>, <span class="hljs-string">&#x27;single_transformer_blocks.21&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>, <span class="hljs-string">&#x27;single_transformer_blocks.22&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>, <span class="hljs-string">&#x27;single_transformer_blocks.23&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>, <span class="hljs-string">&#x27;single_transformer_blocks.24&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>, <span class="hljs-string">&#x27;single_transformer_blocks.25&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>, <span class="hljs-string">&#x27;single_transformer_blocks.26&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>, <span class="hljs-string">&#x27;single_transformer_blocks.27&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>, <span class="hljs-string">&#x27;single_transformer_blocks.28&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>, <span class="hljs-string">&#x27;single_transformer_blocks.29&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>, <span class="hljs-string">&#x27;single_transformer_blocks.30&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>, <span class="hljs-string">&#x27;single_transformer_blocks.31&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>, <span class="hljs-string">&#x27;single_transformer_blocks.32&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>, <span class="hljs-string">&#x27;single_transformer_blocks.33&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>, <span class="hljs-string">&#x27;single_transformer_blocks.34&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>, <span class="hljs-string">&#x27;single_transformer_blocks.35&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>, <span class="hljs-string">&#x27;single_transformer_blocks.36&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>, <span class="hljs-string">&#x27;single_transformer_blocks.37&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>, <span class="hljs-string">&#x27;norm_out&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>, <span class="hljs-string">&#x27;proj_out&#x27;</span>: <span class="hljs-string">&#x27;cpu&#x27;</span>
}

transformer = AutoModel.from_pretrained(
    <span class="hljs-string">&quot;black-forest-labs/FLUX.1-dev&quot;</span>, 
    subfolder=<span class="hljs-string">&quot;transformer&quot;</span>,
    device_map=device_map,
    dtype=torch.bfloat16
)<!----></pre></div><!----> <p>Passe um dicionário mapeando o uso máximo de memória para cada dispositivo para impor um limite. Se um dispositivo não estiver em <code>max_memory</code>, ele será ignorado e os componentes do pipeline não serão distribuídos para ele.</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> AutoModel, StableDiffusionXLPipeline

max_memory = {<span class="hljs-number">0</span>:<span class="hljs-string">&quot;1GB&quot;</span>, <span class="hljs-number">1</span>:<span class="hljs-string">&quot;1GB&quot;</span>}
pipeline = StableDiffusionXLPipeline.from_pretrained(
    <span class="hljs-string">&quot;stabilityai/stable-diffusion-xl-base-1.0&quot;</span>,
    dtype=torch.float16,
    device_map=<span class="hljs-string">&quot;balanced&quot;</span>,
    max_memory=max_memory
)<!----></pre></div><!----> <p>O Diffusers usa a memória máxima de todos os dispositivos por padrão, mas, se os modelos não couberem nas GPUs, será necessário usar uma única GPU e transferir para a CPU com os métodos abaixo.</p> <ul><li><a href="/docs/diffusers/v0.43.0/pt/api/pipelines/overview#diffusers.DiffusionPipeline.enable_model_cpu_offload">enable_model_cpu_offload()</a> funciona apenas em uma única GPU, mas um modelo muito grande pode não caber nela</li> <li><a href="/docs/diffusers/v0.43.0/pt/api/pipelines/overview#diffusers.DiffusionPipeline.enable_sequential_cpu_offload">enable_sequential_cpu_offload()</a> pode funcionar, mas é extremamente lento e também limitado a uma única GPU</li></ul> <p>Use o método <a href="/docs/diffusers/v0.43.0/pt/api/pipelines/overview#diffusers.DiffusionPipeline.reset_device_map">reset_device_map()</a> para redefinir o <code>device_map</code>. Isso é necessário se você quiser usar métodos como <code>.to()</code>, <a href="/docs/diffusers/v0.43.0/pt/api/pipelines/overview#diffusers.DiffusionPipeline.enable_sequential_cpu_offload">enable_sequential_cpu_offload()</a> e <a href="/docs/diffusers/v0.43.0/pt/api/pipelines/overview#diffusers.DiffusionPipeline.enable_model_cpu_offload">enable_model_cpu_offload()</a> em um pipeline que foi mapeado para dispositivos.</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 "><!---->pipeline.reset_device_map()<!----></pre></div><!----> <!--[1--><h2 class="relative group"><a id="vae-slicing" 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="#vae-slicing"><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><!---->VAE slicing<!----></span></h2><!--]--><!----> <p>O fatiamento do VAE economiza memória dividindo grandes lotes de entradas em um único lote de dados e processando-os separadamente. Este método funciona melhor ao gerar mais de uma imagem por vez.</p> <p>Por exemplo, se você estiver gerando 4 imagens de uma vez, a decodificação aumentaria a memória de ativação de pico em 4x. O fatiamento do VAE reduz isso decodificando apenas 1 imagem de cada vez, em vez de todas as 4 imagens de uma vez.</p> <p>Chame <code>enable_slicing()</code> para habilitar o VAE fatiado. Você pode esperar um pequeno aumento de desempenho ao decodificar lotes de múltiplas imagens e nenhum impacto no desempenho para lotes de imagem única.</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> AutoModel, StableDiffusionXLPipeline

pipeline = StableDiffusionXLPipeline.from_pretrained(
    <span class="hljs-string">&quot;stabilityai/stable-diffusion-xl-base-1.0&quot;</span>,
    dtype=torch.float16,
).to(<span class="hljs-string">&quot;cuda&quot;</span>)  <span class="hljs-comment"># or &quot;mps&quot;, &quot;xpu&quot;, &quot;cpu&quot;</span>
pipeline.vae.enable_slicing()
pipeline([<span class="hljs-string">&quot;An astronaut riding a horse on Mars&quot;</span>]*<span class="hljs-number">32</span>).images[<span class="hljs-number">0</span>]
<span class="hljs-built_in">print</span>(<span class="hljs-string">f&quot;Max memory reserved: <span class="hljs-subst">{torch.cuda.max_memory_allocated() / <span class="hljs-number">1024</span>**<span class="hljs-number">3</span>:<span class="hljs-number">.2</span>f}</span> GB&quot;</span>)<!----></pre></div><!----> <blockquote class="warning"><p>As classes <a href="/docs/diffusers/v0.43.0/pt/api/models/autoencoder_kl_wan#diffusers.AutoencoderKLWan">AutoencoderKLWan</a> e <a href="/docs/diffusers/v0.43.0/pt/api/models/asymmetricautoencoderkl#diffusers.AsymmetricAutoencoderKL">AsymmetricAutoencoderKL</a> não suportam slicing.</p></blockquote> <!--[1--><h2 class="relative group"><a id="vae-tiling" 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="#vae-tiling"><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><!---->Tiling do VAE<!----></span></h2><!--]--><!----> <p>O tiling do VAE economiza memória dividindo uma imagem em tiles menores sobrepostos, em vez de processar a imagem inteira de uma vez. Isso também reduz o uso de memória de pico, pois a GPU processa apenas um tile por vez.</p> <p>Chame <code>enable_tiling()</code> para habilitar o tiling do VAE. A imagem gerada pode apresentar alguma variação de tom de um tile para outro, pois eles são decodificados separadamente, mas não deve haver emendas óbvias entre os tiles. O tiling é desabilitado para resoluções menores que um limite pré-especificado (mas configurável). Por exemplo, esse limite é 512x512 para o VAE em <a href="/docs/diffusers/v0.43.0/pt/api/pipelines/stable_diffusion/text2img#diffusers.StableDiffusionPipeline">StableDiffusionPipeline</a>.</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> AutoPipelineForImage2Image
<span class="hljs-keyword">from</span> diffusers.utils <span class="hljs-keyword">import</span> load_image

pipeline = AutoPipelineForImage2Image.from_pretrained(
    <span class="hljs-string">&quot;stabilityai/stable-diffusion-xl-base-1.0&quot;</span>, dtype=torch.float16
).to(<span class="hljs-string">&quot;cuda&quot;</span>)  <span class="hljs-comment"># or &quot;mps&quot;, &quot;xpu&quot;, &quot;cpu&quot;</span>
pipeline.vae.enable_tiling()

init_image = load_image(<span class="hljs-string">&quot;https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/img2img-sdxl-init.png&quot;</span>)
prompt = <span class="hljs-string">&quot;Astronaut in a jungle, cold color palette, muted colors, detailed, 8k&quot;</span>
pipeline(prompt, image=init_image, strength=<span class="hljs-number">0.5</span>).images[<span class="hljs-number">0</span>]
<span class="hljs-built_in">print</span>(<span class="hljs-string">f&quot;Max memory reserved: <span class="hljs-subst">{torch.cuda.max_memory_allocated() / <span class="hljs-number">1024</span>**<span class="hljs-number">3</span>:<span class="hljs-number">.2</span>f}</span> GB&quot;</span>)<!----></pre></div><!----> <blockquote class="warning"><p><a href="/docs/diffusers/v0.43.0/pt/api/models/autoencoder_kl_wan#diffusers.AutoencoderKLWan">AutoencoderKLWan</a> e <a href="/docs/diffusers/v0.43.0/pt/api/models/asymmetricautoencoderkl#diffusers.AsymmetricAutoencoderKL">AsymmetricAutoencoderKL</a> não suportam tiling.</p></blockquote> <p>Para ignorar completamente a memória do VAE na sua GPU, codifique e decodifique em um endpoint remoto com <a href="../hybrid_inference/overview">inferência remota</a>.</p> <!--[1--><h2 class="relative group"><a id="offloading" 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="#offloading"><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><!---->Offloading<!----></span></h2><!--]--><!----> <p>As estratégias de offloading movem camadas ou modelos que não estão ativos no momento para a CPU, a fim de evitar o aumento do uso de memória da GPU. Essas estratégias podem ser combinadas com quantização e torch.compile para equilibrar a velocidade de inferência e o uso de memória.</p> <p>Consulte o guia <a href="./speed-memory-optims">Compile and offloading quantized models</a> para mais detalhes.</p> <!--[2--><h3 class="relative group"><a id="cpu-offloading" 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="#cpu-offloading"><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><!---->Offloading de CPU<!----></span></h3><!--]--><!----> <p>O offloading de CPU move seletivamente os pesos da GPU para a CPU. Quando um componente é necessário, ele é transferido para a GPU e, quando não é necessário, é movido para a CPU. Este método funciona em submódulos, em vez de modelos inteiros. Ele economiza memória ao evitar armazenar o modelo inteiro na GPU.</p> <p>O offloading para CPU reduz drasticamente o uso de memória, mas também é <strong>extremamente lento</strong>, pois os submódulos são transferidos várias vezes entre os dispositivos. Muitas vezes, isso pode ser impraticável devido à lentidão.</p> <blockquote class="warning"><p>Não mova o pipeline para a CUDA antes de chamar <a href="/docs/diffusers/v0.43.0/pt/api/pipelines/overview#diffusers.DiffusionPipeline.enable_sequential_cpu_offload">enable_sequential_cpu_offload()</a>, caso contrário, a quantidade de memória economizada será mínima (consulte esta <a href="https://github.com/huggingface/diffusers/issues/1934" rel="nofollow">issue</a> para mais detalhes). Esta é uma operação com estado que instala hooks no modelo.</p></blockquote> <p>Chame <a href="/docs/diffusers/v0.43.0/pt/api/pipelines/overview#diffusers.DiffusionPipeline.enable_sequential_cpu_offload">enable_sequential_cpu_offload()</a> para ativá-lo em um pipeline.</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

pipeline = DiffusionPipeline.from_pretrained(
    <span class="hljs-string">&quot;black-forest-labs/FLUX.1-schnell&quot;</span>, dtype=torch.bfloat16
)
pipeline.enable_sequential_cpu_offload()

pipeline(
    prompt=<span class="hljs-string">&quot;An astronaut riding a horse on Mars&quot;</span>,
    guidance_scale=<span class="hljs-number">0.</span>,
    height=<span class="hljs-number">768</span>,
    width=<span class="hljs-number">1360</span>,
    num_inference_steps=<span class="hljs-number">4</span>,
    max_sequence_length=<span class="hljs-number">256</span>,
).images[<span class="hljs-number">0</span>]
<span class="hljs-built_in">print</span>(<span class="hljs-string">f&quot;Max memory reserved: <span class="hljs-subst">{torch.cuda.max_memory_allocated() / <span class="hljs-number">1024</span>**<span class="hljs-number">3</span>:<span class="hljs-number">.2</span>f}</span> GB&quot;</span>)<!----></pre></div><!----> <!--[2--><h3 class="relative group"><a id="model-offloading" 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="#model-offloading"><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><!---->Deslocamento de modelos<!----></span></h3><!--]--><!----> <p>O offloading de modelos move modelos inteiros para a GPU, em vez de mover seletivamente <em>algumas</em> camadas ou componentes do modelo. Um dos principais modelos do pipeline, geralmente o codificador de texto, UNet e VAE, é colocado na GPU, enquanto os outros componentes são mantidos na CPU. Componentes como o UNet, que são executados várias vezes, permanecem na GPU até que estejam completamente finalizados e não sejam mais necessários. Isso elimina o overhead de comunicação do <a href="#cpu-offloading">CPU offloading</a> e torna o offloading de modelos uma alternativa mais rápida. A compensação é que as economias de memória não serão tão grandes.</p> <blockquote class="warning"><p>Lembre-se de que, se os modelos forem reutilizados fora do pipeline após a instalação dos hooks (veja <a href="https://huggingface.co/docs/accelerate/en/package_reference/big_modeling#accelerate.hooks.remove_hook_from_module" rel="nofollow">Removendo Hooks</a> para mais detalhes), você precisa executar o pipeline inteiro e os modelos na ordem esperada para desativá-los corretamente. Esta é uma operação com estado que instala hooks no modelo.</p></blockquote> <p>Chame <a href="/docs/diffusers/v0.43.0/pt/api/pipelines/overview#diffusers.DiffusionPipeline.enable_model_cpu_offload">enable_model_cpu_offload()</a> para ativá-lo em um pipeline.</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

pipeline = DiffusionPipeline.from_pretrained(
    <span class="hljs-string">&quot;black-forest-labs/FLUX.1-schnell&quot;</span>, dtype=torch.bfloat16
)
pipeline.enable_model_cpu_offload()

pipeline(
    prompt=<span class="hljs-string">&quot;An astronaut riding a horse on Mars&quot;</span>,
    guidance_scale=<span class="hljs-number">0.</span>,
    height=<span class="hljs-number">768</span>,
    width=<span class="hljs-number">1360</span>,
    num_inference_steps=<span class="hljs-number">4</span>,
    max_sequence_length=<span class="hljs-number">256</span>,
).images[<span class="hljs-number">0</span>]
<span class="hljs-built_in">print</span>(<span class="hljs-string">f&quot;Max memory reserved: <span class="hljs-subst">{torch.cuda.max_memory_allocated() / <span class="hljs-number">1024</span>**<span class="hljs-number">3</span>:<span class="hljs-number">.2</span>f}</span> GB&quot;</span>)<!----></pre></div><!----> <p><a href="/docs/diffusers/v0.43.0/pt/api/pipelines/overview#diffusers.DiffusionPipeline.enable_model_cpu_offload">enable_model_cpu_offload()</a> também ajuda quando você usa o método <a href="/docs/diffusers/v0.43.0/pt/api/pipelines/stable_diffusion/stable_diffusion_xl#diffusers.StableDiffusionXLPipeline.encode_prompt">encode_prompt()</a> isoladamente para gerar o estado oculto dos codificadores de texto.</p> <!--[2--><h3 class="relative group"><a id="group-offloading" 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="#group-offloading"><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><!---->Descarregamento em grupo<!----></span></h3><!--]--><!----> <p>O Group offloading move grupos de camadas internas (<a href="https://pytorch.org/docs/stable/generated/torch.nn.ModuleList.html" rel="nofollow">torch.nn.ModuleList</a> ou <a href="https://pytorch.org/docs/stable/generated/torch.nn.Sequential.html" rel="nofollow">torch.nn.Sequential</a>) para a CPU. Ele usa menos memória que o <a href="#model-offloading">model offloading</a> e é mais rápido que o <a href="#cpu-offloading">CPU offloading</a> porque reduz o overhead de comunicação.</p> <blockquote class="warning"><p>O offloading por grupo pode não funcionar com todos os modelos se a implementação do forward contiver o casting de dispositivos dos inputs dependente dos pesos, pois isso pode conflitar com o mecanismo de casting de dispositivos do offloading por grupo.</p></blockquote> <p>Ative o offloading de grupo configurando o parâmetro <code>offload_type</code> para <code>block_level</code> ou <code>leaf_level</code>.</p> <ul><li><code>block_level</code> descarrega grupos de camadas com base no parâmetro <code>num_blocks_per_group</code>. Por exemplo, se <code>num_blocks_per_group=2</code> em um modelo com 40 camadas, 2 camadas são carregadas e descarregadas de cada vez (20 carregamentos/descarregamentos no total). Isso reduz drasticamente os requisitos de memória.</li> <li><code>leaf_level</code> descarrega camadas individuais no nível mais baixo e é equivalente ao <a href="#cpu-offloading">descarregamento para CPU</a>. No entanto, pode ser acelerado se você usar streams sem abrir mão da velocidade de inferência.</li></ul> <p>O offloading em grupo é suportado para pipelines inteiros ou para modelos individuais. Aplicar o offloading em grupo a todo o pipeline é a opção mais simples, enquanto aplicá-lo seletivamente a modelos individuais oferece aos usuários mais flexibilidade para usar diferentes técnicas de offloading para diferentes modelos.</p> <div class="flex space-x-2 items-center my-1.5 mr-8 h-7 !pl-0 -mx-3 md:mx-0"><!--[--><div class="flex items-center border rounded-lg px-1.5 py-1 leading-none select-none text-smd border-gray-800 bg-black dark:bg-gray-700 text-white">pipeline</div><div class="flex items-center border rounded-lg px-1.5 py-1 leading-none select-none text-smd text-gray-500 cursor-pointer opacity-90 hover:text-gray-700 dark:hover:text-gray-200 hover:shadow-sm">model</div><!--]--></div> <div class="language-select"><div><p>Chame <a href="/docs/diffusers/v0.43.0/pt/api/pipelines/overview#diffusers.DiffusionPipeline.enable_group_offload">enable_group_offload()</a> em um pipeline.</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> CogVideoXPipeline
<span class="hljs-keyword">from</span> diffusers.hooks <span class="hljs-keyword">import</span> apply_group_offloading
<span class="hljs-keyword">from</span> diffusers.utils <span class="hljs-keyword">import</span> export_to_video

onload_device = torch.device(<span class="hljs-string">&quot;cuda&quot;</span>)
offload_device = torch.device(<span class="hljs-string">&quot;cpu&quot;</span>)

pipeline = CogVideoXPipeline.from_pretrained(<span class="hljs-string">&quot;THUDM/CogVideoX-5b&quot;</span>, dtype=torch.bfloat16)
pipeline.enable_group_offload(
    onload_device=onload_device,
    offload_device=offload_device,
    offload_type=<span class="hljs-string">&quot;leaf_level&quot;</span>,
    use_stream=<span class="hljs-literal">True</span>
)

prompt = (
    <span class="hljs-string">&quot;A panda, dressed in a small, red jacket and a tiny hat, sits on a wooden stool in a serene bamboo forest. &quot;</span>
    <span class="hljs-string">&quot;The panda&#x27;s fluffy paws strum a miniature acoustic guitar, producing soft, melodic tunes. Nearby, a few other &quot;</span>
    <span class="hljs-string">&quot;pandas gather, watching curiously and some clapping in rhythm. Sunlight filters through the tall bamboo, &quot;</span>
    <span class="hljs-string">&quot;casting a gentle glow on the scene. The panda&#x27;s face is expressive, showing concentration and joy as it plays. &quot;</span>
    <span class="hljs-string">&quot;The background includes a small, flowing stream and vibrant green foliage, enhancing the peaceful and magical &quot;</span>
    <span class="hljs-string">&quot;atmosphere of this unique musical performance.&quot;</span>
)
video = pipeline(prompt=prompt, guidance_scale=<span class="hljs-number">6</span>, num_inference_steps=<span class="hljs-number">50</span>).frames[<span class="hljs-number">0</span>]
<span class="hljs-built_in">print</span>(<span class="hljs-string">f&quot;Max memory reserved: <span class="hljs-subst">{torch.cuda.max_memory_allocated() / <span class="hljs-number">1024</span>**<span class="hljs-number">3</span>:<span class="hljs-number">.2</span>f}</span> GB&quot;</span>)
export_to_video(video, <span class="hljs-string">&quot;output.mp4&quot;</span>, fps=<span class="hljs-number">8</span>)<!----></pre></div><!----><!----></div><!----> <div hidden=""><p>Chame <a href="/docs/diffusers/v0.43.0/pt/api/models/overview#diffusers.ModelMixin.enable_group_offload">enable_group_offload()</a> nos componentes padrão de modelos do Diffusers que herdam de <a href="/docs/diffusers/v0.43.0/pt/api/models/overview#diffusers.ModelMixin">ModelMixin</a>. Para outros componentes de modelo que não herdam de <a href="/docs/diffusers/v0.43.0/pt/api/models/overview#diffusers.ModelMixin">ModelMixin</a>, como um <a href="https://pytorch.org/docs/stable/generated/torch.nn.Module.html" rel="nofollow">torch.nn.Module</a> genérico, use <a href="/docs/diffusers/v0.43.0/pt/api/utilities#diffusers.hooks.apply_group_offloading">apply_group_offloading()</a> em vez disso.</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> CogVideoXPipeline
<span class="hljs-keyword">from</span> diffusers.hooks <span class="hljs-keyword">import</span> apply_group_offloading
<span class="hljs-keyword">from</span> diffusers.utils <span class="hljs-keyword">import</span> export_to_video

onload_device = torch.device(<span class="hljs-string">&quot;cuda&quot;</span>)
offload_device = torch.device(<span class="hljs-string">&quot;cpu&quot;</span>)
pipeline = CogVideoXPipeline.from_pretrained(<span class="hljs-string">&quot;THUDM/CogVideoX-5b&quot;</span>, dtype=torch.bfloat16)

<span class="hljs-comment"># Use the enable_group_offload method for Diffusers model implementations</span>
pipeline.transformer.enable_group_offload(onload_device=onload_device, offload_device=offload_device, offload_type=<span class="hljs-string">&quot;leaf_level&quot;</span>)
pipeline.vae.enable_group_offload(onload_device=onload_device, offload_type=<span class="hljs-string">&quot;leaf_level&quot;</span>)

<span class="hljs-comment"># Use the apply_group_offloading method for other model components</span>
apply_group_offloading(pipeline.text_encoder, onload_device=onload_device, offload_type=<span class="hljs-string">&quot;block_level&quot;</span>, num_blocks_per_group=<span class="hljs-number">2</span>)

prompt = (
    <span class="hljs-string">&quot;A panda, dressed in a small, red jacket and a tiny hat, sits on a wooden stool in a serene bamboo forest. &quot;</span>
    <span class="hljs-string">&quot;The panda&#x27;s fluffy paws strum a miniature acoustic guitar, producing soft, melodic tunes. Nearby, a few other &quot;</span>
    <span class="hljs-string">&quot;pandas gather, watching curiously and some clapping in rhythm. Sunlight filters through the tall bamboo, &quot;</span>
    <span class="hljs-string">&quot;casting a gentle glow on the scene. The panda&#x27;s face is expressive, showing concentration and joy as it plays. &quot;</span>
    <span class="hljs-string">&quot;The background includes a small, flowing stream and vibrant green foliage, enhancing the peaceful and magical &quot;</span>
    <span class="hljs-string">&quot;atmosphere of this unique musical performance.&quot;</span>
)
video = pipeline(prompt=prompt, guidance_scale=<span class="hljs-number">6</span>, num_inference_steps=<span class="hljs-number">50</span>).frames[<span class="hljs-number">0</span>]
<span class="hljs-built_in">print</span>(<span class="hljs-string">f&quot;Max memory reserved: <span class="hljs-subst">{torch.cuda.max_memory_allocated() / <span class="hljs-number">1024</span>**<span class="hljs-number">3</span>:<span class="hljs-number">.2</span>f}</span> GB&quot;</span>)
export_to_video(video, <span class="hljs-string">&quot;output.mp4&quot;</span>, fps=<span class="hljs-number">8</span>)<!----></pre></div><!----><!----></div><!----><!----></div><!----> <!--[3--><h4 class="relative group"><a id="cuda-stream" 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="#cuda-stream"><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><!---->CUDA stream<!----></span></h4><!--]--><!----> <p>O parâmetro <code>use_stream</code> pode ser ativado para dispositivos CUDA que suportam streams de transferência de dados assíncronos para reduzir o tempo total de execução em comparação com <a href="#cpu-offloading">CPU offloading</a>. Ele sobrepõe a transferência de dados e o cálculo usando prefetching de camadas. A próxima camada a ser executada é carregada na GPU enquanto a camada atual ainda está sendo executada. Isso pode aumentar significativamente a memória da CPU, portanto, certifique-se de ter 2x a quantidade de memória do tamanho do modelo.</p> <p>Defina <code>record_stream=True</code> para obter mais aceleração, com um custo de uso de memória ligeiramente maior. Consulte a documentação de <a href="https://pytorch.org/docs/stable/generated/torch.Tensor.record_stream.html" rel="nofollow">torch.Tensor.record_stream</a> para saber mais.</p> <blockquote class="tip"><p>Quando <code>use_stream=True</code> em VAEs com tiling habilitado, certifique-se de executar uma passagem de forward fictícia (possível também com entradas fictícias) antes da inferência para evitar erros de incompatibilidade de dispositivo. Isso pode não funcionar em todas as implementações, então sinta-se à vontade para abrir uma issue se encontrar algum problema.</p></blockquote> <p>Se você estiver usando offloading de grupo em nível de <code>block_level</code> com <code>use_stream</code> habilitado, o parâmetro <code>num_blocks_per_group</code> deve ser definido como <code>1</code>; caso contrário, um aviso será emitido.</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 "><!---->pipeline.transformer.enable_group_offload(onload_device=onload_device, offload_device=offload_device, offload_type=<span class="hljs-string">&quot;leaf_level&quot;</span>, use_stream=<span class="hljs-literal">True</span>, record_stream=<span class="hljs-literal">True</span>)<!----></pre></div><!----> <p>O parâmetro <code>low_cpu_mem_usage</code> pode ser definido como <code>True</code> para reduzir o uso de memória de CPU ao usar streams durante o offloading de grupos. É ideal para offloading de <code>leaf_level</code> e quando a memória de CPU é o gargalo. A memória é economizada criando tensores fixados (pinned) sob demanda, em vez de fixá-los previamente. No entanto, isso pode aumentar o tempo total de execução.</p> <!--[3--><h4 class="relative group"><a id="offloading-to-disk" 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="#offloading-to-disk"><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><!---->Offloading para o disco<!----></span></h4><!--]--><!----> <p>O offloading de grupos pode consumir uma quantidade significativa de memória do sistema, dependendo do tamanho do modelo. Em sistemas com memória limitada, tente realizar o offloading de grupos para o disco como memória secundária.</p> <p>Defina o argumento <code>offload_to_disk_path</code> em <a href="/docs/diffusers/v0.43.0/pt/api/models/overview#diffusers.ModelMixin.enable_group_offload">enable_group_offload()</a> ou <a href="/docs/diffusers/v0.43.0/pt/api/utilities#diffusers.hooks.apply_group_offloading">apply_group_offloading()</a> para descarregar o modelo no disco.</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 "><!---->pipeline.transformer.enable_group_offload(onload_device=onload_device, offload_device=offload_device, offload_type=<span class="hljs-string">&quot;leaf_level&quot;</span>, offload_to_disk_path=<span class="hljs-string">&quot;path/to/disk&quot;</span>)

apply_group_offloading(pipeline.text_encoder, onload_device=onload_device, offload_type=<span class="hljs-string">&quot;block_level&quot;</span>, num_blocks_per_group=<span class="hljs-number">2</span>, offload_to_disk_path=<span class="hljs-string">&quot;path/to/disk&quot;</span>)<!----></pre></div><!----> <p>Consulte estas <a href="https://github.com/huggingface/diffusers/pull/11682#issue-3129365363" rel="nofollow">duas</a> <a href="https://github.com/huggingface/diffusers/pull/11682#issuecomment-2955715126" rel="nofollow">tabelas</a> para comparar os compromissos entre velocidade e memória.</p> <!--[1--><h2 class="relative group"><a id="layerwise-casting" 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="#layerwise-casting"><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><!---->Casting por camada<!----></span></h2><!--]--><!----> <blockquote class="tip"><p>Combine o casting por camada com <a href="#group-offloading">group offloading</a> para economizar ainda mais memória.</p></blockquote> <p>O casting por camada armazena os pesos em um formato de dados menor (por exemplo, <code>torch.float8_e4m3fn</code> e <code>torch.float8_e5m2</code>) para usar menos memória e faz o upcast desses pesos para uma precisão maior, como <code>torch.float16</code> ou <code>torch.bfloat16</code>, para o cálculo. Certas camadas (pesos relacionados a normalização e modulação) são ignoradas, pois armazená-los em fp8 pode degradar a qualidade da geração.</p> <blockquote class="warning"><p>O casting por camada pode não funcionar com todos os modelos se a implementação do forward contiver conversão de tipos interna dos pesos. A implementação atual do casting por camada pressupõe que a passagem forward seja independente da precisão dos pesos e que os tipos de dados de entrada sejam sempre especificados em <code>compute_dtype</code> (veja <a href="https://github.com/huggingface/transformers/blob/7f5077e53682ca855afc826162b204ebf809f1f9/src/transformers/models/t5/modeling_t5.py#L294-L299" rel="nofollow">aqui</a> para uma implementação incompatível).</p> <p>O casting por camada também pode falhar em implementações de modelagem personalizadas com camadas de <a href="https://huggingface.co/docs/peft/index" rel="nofollow">PEFT</a>. Existem algumas verificações disponíveis, mas elas não são amplamente testadas nem garantidas para funcionar em todos os casos.</p></blockquote> <p>Chame <a href="/docs/diffusers/v0.43.0/pt/api/models/overview#diffusers.ModelMixin.enable_layerwise_casting">enable_layerwise_casting()</a> para definir os tipos de dados de armazenamento e computação.</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> CogVideoXPipeline, CogVideoXTransformer3DModel
<span class="hljs-keyword">from</span> diffusers.utils <span class="hljs-keyword">import</span> export_to_video

transformer = CogVideoXTransformer3DModel.from_pretrained(
    <span class="hljs-string">&quot;THUDM/CogVideoX-5b&quot;</span>,
    subfolder=<span class="hljs-string">&quot;transformer&quot;</span>,
    dtype=torch.bfloat16
)
transformer.enable_layerwise_casting(storage_dtype=torch.float8_e4m3fn, compute_dtype=torch.bfloat16)

pipeline = CogVideoXPipeline.from_pretrained(<span class="hljs-string">&quot;THUDM/CogVideoX-5b&quot;</span>,
    transformer=transformer,
    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>
prompt = (
    <span class="hljs-string">&quot;A panda, dressed in a small, red jacket and a tiny hat, sits on a wooden stool in a serene bamboo forest. &quot;</span>
    <span class="hljs-string">&quot;The panda&#x27;s fluffy paws strum a miniature acoustic guitar, producing soft, melodic tunes. Nearby, a few other &quot;</span>
    <span class="hljs-string">&quot;pandas gather, watching curiously and some clapping in rhythm. Sunlight filters through the tall bamboo, &quot;</span>
    <span class="hljs-string">&quot;casting a gentle glow on the scene. The panda&#x27;s face is expressive, showing concentration and joy as it plays. &quot;</span>
    <span class="hljs-string">&quot;The background includes a small, flowing stream and vibrant green foliage, enhancing the peaceful and magical &quot;</span>
    <span class="hljs-string">&quot;atmosphere of this unique musical performance.&quot;</span>
)
video = pipeline(prompt=prompt, guidance_scale=<span class="hljs-number">6</span>, num_inference_steps=<span class="hljs-number">50</span>).frames[<span class="hljs-number">0</span>]
<span class="hljs-built_in">print</span>(<span class="hljs-string">f&quot;Max memory reserved: <span class="hljs-subst">{torch.cuda.max_memory_allocated() / <span class="hljs-number">1024</span>**<span class="hljs-number">3</span>:<span class="hljs-number">.2</span>f}</span> GB&quot;</span>)
export_to_video(video, <span class="hljs-string">&quot;output.mp4&quot;</span>, fps=<span class="hljs-number">8</span>)<!----></pre></div><!----> <p>O método <a href="/docs/diffusers/v0.43.0/pt/api/utilities#diffusers.hooks.apply_layerwise_casting">apply_layerwise_casting()</a> também pode ser usado se você precisar de mais controle e flexibilidade. Ele pode ser aplicado parcialmente às camadas do modelo chamando-o em módulos internos específicos. Use os parâmetros <code>skip_modules_pattern</code> ou <code>skip_modules_classes</code> para especificar os módulos a serem evitados, como as camadas de normalização e modulação.</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> CogVideoXTransformer3DModel
<span class="hljs-keyword">from</span> diffusers.hooks <span class="hljs-keyword">import</span> apply_layerwise_casting

transformer = CogVideoXTransformer3DModel.from_pretrained(
    <span class="hljs-string">&quot;THUDM/CogVideoX-5b&quot;</span>,
    subfolder=<span class="hljs-string">&quot;transformer&quot;</span>,
    dtype=torch.bfloat16
)

<span class="hljs-comment"># skip the normalization layer</span>
apply_layerwise_casting(
    transformer,
    storage_dtype=torch.float8_e4m3fn,
    compute_dtype=torch.bfloat16,
    skip_modules_classes=[<span class="hljs-string">&quot;norm&quot;</span>],
    non_blocking=<span class="hljs-literal">True</span>,
)<!----></pre></div><!----> <!--[1--><h2 class="relative group"><a id="torchchannelslast" 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="#torchchannelslast"><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><!---->torch.channels_last<!----></span></h2><!--]--><!----> <p><a href="https://pytorch.org/tutorials/intermediate/memory_format_tutorial.html" rel="nofollow">torch.channels_last</a> inverte a forma como os tensors são armazenados, de <code>(batch size, channels, height, width)</code> para <code>(batch size, height, width, channels)</code>. Isso alinha os tensors com a forma como o hardware acessa sequencialmente os tensors armazenados na memória e evita saltos na memória para acessar os valores dos pixels.</p> <p>Nem todos os operadores suportam atualmente o formato channels-last e podem resultar em pior desempenho, mas ainda vale a pena tentar.</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-built_in">print</span>(pipeline.unet.conv_out.state_dict()[<span class="hljs-string">&quot;weight&quot;</span>].stride())  <span class="hljs-comment"># (2880, 9, 3, 1)</span>
pipeline.unet.to(memory_format=torch.channels_last)  <span class="hljs-comment"># in-place operation</span>
<span class="hljs-built_in">print</span>(
    pipeline.unet.conv_out.state_dict()[<span class="hljs-string">&quot;weight&quot;</span>].stride()
)  <span class="hljs-comment"># (2880, 1, 960, 320) having a stride of 1 for the 2nd dimension proves that it works</span><!----></pre></div><!----> <!--[1--><h2 class="relative group"><a id="memory-efficient-attention" 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="#memory-efficient-attention"><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><!---->Atenção eficiente em memória<!----></span></h2><!--]--><!----> <p>O Diffusers suporta vários backends de atenção eficientes em memória (FlashAttention, xFormers, SageAttention, entre outros) por meio de <a href="/docs/diffusers/v0.43.0/pt/api/models/overview#diffusers.ModelMixin.set_attention_backend">set_attention_backend()</a>. Consulte o guia <a href="./attention_backends">Attention backends</a> para aprender a alternar entre eles.</p> <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/memory.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_1qsdbwf = {
						base: "/docs/diffusers/main/en",
						assets: "/docs/diffusers/main/en"
					};
					const element = document.currentScript.parentElement;
					Promise.all([
						import("/docs/diffusers/main/en/_app/immutable/entry/start.0S0_iEoG.js"),
						import("/docs/diffusers/main/en/_app/immutable/entry/app.CPdOxCbw.js")
					]).then(([kit, app]) => {
						kit.start(app, element, {
							node_ids: [0, 339],
							data: [null,null],
							form: null,
							error: null
						});
					});
				}
			</script>
		
Xet Storage Details

Size:: 102 kB
Xet hash:: 6cc550991ff471476dfc2909fb3c507ba121d467f37e27e6a6528693a2062155
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.