Buckets:

hf-doc-build
/

doc

Files
xet
hf-doc-build/doc / diffusers /main /fr /optimization /memory.html
mishig
3 days ago
download
raw
104 kB
<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;Réduire l'utilisation de la mémoire&quot;,&quot;local&quot;:&quot;reduce-memory-usage&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Plusieurs GPU&quot;,&quot;local&quot;:&quot;multiple-gpus&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Points de contrôle fractionnés&quot;,&quot;local&quot;:&quot;sharded-checkpoints&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Placement sur l'appareil&quot;,&quot;local&quot;:&quot;device-placement&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3}],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Découpage du VAE&quot;,&quot;local&quot;:&quot;vae-slicing&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Tiling du 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;Déchargement sur CPU&quot;,&quot;local&quot;:&quot;cpu-offloading&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Déchargement du modèle&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;Flux CUDA&quot;,&quot;local&quot;:&quot;cuda-stream&quot;,&quot;sections&quot;:[],&quot;depth&quot;:4},{&quot;title&quot;:&quot;Déchargement sur disque&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 par couche&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;Attention efficace en mémoire&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><!---->Réduire l'utilisation de la mémoire<!----></span></h1><!--]--><!----> <p>Les modèles de diffusion modernes tels que <a href="../api/pipelines/flux">Flux</a> et <a href="../api/pipelines/wan">Wan</a> comptent des milliards de paramètres, ce qui requiert une grande quantité de mémoire sur votre matériel pour l’inférence. C’est un défi, car les GPU courants ne disposent souvent pas de mémoire suffisante. Pour contourner ces limitations de mémoire, vous pouvez utiliser plusieurs GPU (si disponibles), décharger certains composants du pipeline sur le CPU, et plus encore.</p> <p>Ce guide vous montre comment réduire votre consommation de mémoire.</p> <blockquote class="tip"><p>Gardez à l’esprit que ces techniques peuvent nécessiter des ajustements en fonction du modèle. Par exemple, un modèle de diffusion basé sur un transformer ne tirera pas forcément autant parti de ces optimisations de mémoire qu’un modèle basé sur un 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><!---->Plusieurs GPU<!----></span></h2><!--]--><!----> <p>Si vous disposez de plusieurs GPU, il existe plusieurs options pour charger et distribuer efficacement un grand modèle sur votre matériel. Ces fonctionnalités sont prises en charge par la bibliothèque <a href="https://huggingface.co/docs/accelerate/index" rel="nofollow">Accelerate</a>, assurez-vous donc qu'elle est installée au préalable.</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> Copié</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><!---->Points de contrôle fractionnés<!----></span></h3><!--]--><!----> <p>Charger de grands points de contrôle en plusieurs fragments est utile car les fragments sont chargés un par un. Cela maintient l'utilisation de la mémoire au minimum, ne nécessitant que suffisamment de mémoire pour la taille du modèle et la taille du plus grand fragment. Nous recommandons de fragmenter lorsque le point de contrôle fp32 dépasse 5 Go. La taille de fragment par défaut est de 5 Go.</p> <p>Fragmenter un point de contrôle dans <a href="/docs/diffusers/main/fr/api/pipelines/overview#diffusers.DiffusionPipeline.save_pretrained">save_pretrained()</a> avec le paramètre <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> Copié</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>Vous pouvez désormais utiliser le point de contrôle fragmenté, au lieu du point de contrôle standard, pour économiser de la mémoire.</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> Copié</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><!---->Placement sur l'appareil<!----></span></h3><!--]--><!----> <blockquote class="warning"><p>Le placement sur les appareils est une fonctionnalité expérimentale et l'API est susceptible d'évoluer. Seule la stratégie <code>balanced</code> est prise en charge pour le moment. Nous prévoyons de prendre en charge d'autres stratégies de correspondance à l'avenir.</p></blockquote> <p>Le paramètre <code>device_map</code> contrôle la manière dont les composants du modèle dans un pipeline ou les couches d'un modèle individuel sont répartis sur les différents appareils.</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">niveau 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">niveau du modèle</div><!--]--></div> <div class="language-select"><div><p>La stratégie de placement sur les appareils <code>balanced</code> répartit uniformément le pipeline sur tous les appareils disponibles.</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> Copié</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>Vous pouvez inspecter la carte des appareils d’un pipeline avec <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> Copié</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>Le <code>device_map</code> est utile pour charger de grands modèles, comme le transformeur de diffusion Flux qui possède 12,5 milliards de paramètres. Définissez-le sur <code>"auto"</code> pour distribuer automatiquement un modèle sur l’appareil le plus rapide avant de passer aux appareils plus lents. Consultez la section <a href="#sharded-checkpoints">Points de contrôle fractionnés</a> pour plus de détails.</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> Copié</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>Vous pouvez inspecter la carte des appareils d’un modèle avec <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> Copié</div><!----></button><!----></div> <pre class="language-py "><!----><span class="hljs-built_in">print</span>(transformer.hf_device_map)<!----></pre></div><!----><!----></div><!----><!----></div><!----> <p>Lors de la conception de votre propre <code>device_map</code>, il doit s’agir d’un dictionnaire associant le nom spécifique d’un module ou d’une couche du modèle à un identifiant de périphérique (un entier pour les GPU, <code>cpu</code> pour les CPU et <code>disk</code> pour le disque).</p> <p>Appel de <code>hf_device_map</code> sur un modèle pour voir comment les couches du modèle sont distribuées, puis concevez la vôtre.</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> Copié</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>Par exemple, le <code>device_map</code> ci-dessous place <code>single_transformer_blocks.10</code> à <code>single_transformer_blocks.20</code> sur une seconde 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> Copié</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>Passez un dictionnaire associant l’utilisation mémoire maximale à chaque appareil pour imposer une limite. Si un appareil n’est pas présent dans <code>max_memory</code>, il est ignoré et les composants du pipeline ne seront pas répartis dessus.</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> Copié</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>Diffusers utilise par défaut la mémoire maximale de tous les appareils, mais si les modèles ne tiennent pas sur les GPU, vous devrez utiliser un seul GPU et décharger sur le CPU à l’aide des méthodes ci-dessous.</p> <ul><li><a href="/docs/diffusers/main/fr/api/pipelines/overview#diffusers.DiffusionPipeline.enable_model_cpu_offload">enable_model_cpu_offload()</a> ne fonctionne que sur un seul GPU, mais un modèle très volumineux peut ne pas y tenir</li> <li><a href="/docs/diffusers/main/fr/api/pipelines/overview#diffusers.DiffusionPipeline.enable_sequential_cpu_offload">enable_sequential_cpu_offload()</a> peut fonctionner, mais il est extrêmement lent et limité à une seule GPU</li></ul> <p>Utilisez la méthode <a href="/docs/diffusers/main/fr/api/pipelines/overview#diffusers.DiffusionPipeline.reset_device_map">reset_device_map()</a> pour réinitialiser le <code>device_map</code>. C'est nécessaire si vous souhaitez utiliser des méthodes comme <code>.to()</code>, <a href="/docs/diffusers/main/fr/api/pipelines/overview#diffusers.DiffusionPipeline.enable_sequential_cpu_offload">enable_sequential_cpu_offload()</a> et <a href="/docs/diffusers/main/fr/api/pipelines/overview#diffusers.DiffusionPipeline.enable_model_cpu_offload">enable_model_cpu_offload()</a> sur un pipeline qui a été mappé sur un appareil.</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> Copié</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><!---->Découpage du VAE<!----></span></h2><!--]--><!----> <p>Le découpage du VAE économise de la mémoire en divisant de grands lots d'entrées en un seul lot de données et en les traitant séparément. Cette méthode fonctionne le mieux lors de la génération de plusieurs images à la fois.</p> <p>Par exemple, si vous générez 4 images en même temps, le décodage augmenterait la mémoire d’activation de pointe par 4. Le tranchage du VAE réduit cette augmentation en ne décodant qu’une image à la fois au lieu des 4 images simultanément.</p> <p>Appel de <code>enable_slicing()</code> pour activer le VAE par tranches. Vous pouvez vous attendre à une légère amélioration des performances lors du décodage de lots d'images multiples et à aucun impact sur les performances pour les lots d'images uniques.</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> Copié</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>Les classes <a href="/docs/diffusers/main/fr/api/models/autoencoder_kl_wan#diffusers.AutoencoderKLWan">AutoencoderKLWan</a> et <a href="/docs/diffusers/main/fr/api/models/asymmetricautoencoderkl#diffusers.AsymmetricAutoencoderKL">AsymmetricAutoencoderKL</a> ne prennent pas en charge le tranchage.</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 du VAE<!----></span></h2><!--]--><!----> <p>Le découpage en tuiles du VAE économise de la mémoire en divisant une image en tuiles plus petites et chevauchantes, plutôt qu’en traitant l’image entière d’un seul coup. Cela réduit également l’utilisation mémoire de pointe, car le GPU ne traite qu’une tuile à la fois.</p> <p>Appel de <code>enable_tiling()</code> pour activer le découpage en tuiles du VAE. L’image générée peut présenter de légères variations de tonalité d’une tuile à l’autre, car elles sont décodées séparément, mais il ne devrait y avoir aucune couture visible entre les tuiles. Le découpage est désactivé pour les résolutions inférieures à une limite prédéfinie (mais configurable). Par exemple, cette limite est de 512x512 pour le VAE de <a href="/docs/diffusers/main/fr/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> Copié</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/main/fr/api/models/autoencoder_kl_wan#diffusers.AutoencoderKLWan">AutoencoderKLWan</a> et <a href="/docs/diffusers/main/fr/api/models/asymmetricautoencoderkl#diffusers.AsymmetricAutoencoderKL">AsymmetricAutoencoderKL</a> ne prennent pas en charge le découpage en tuiles.</p></blockquote> <p>Pour éviter d'utiliser la mémoire du VAE sur votre GPU, effectuez l'encodage et le décodage sur un point de distanciation distant avec <a href="../hybrid_inference/overview">Remote inference</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>Les stratégies de déchargement déplacent les couches ou modèles non actifs vers le CPU afin d'éviter d'augmenter la mémoire GPU. Ces stratégies peuvent être combinées avec la quantification et torch.compile pour équilibrer la vitesse d'inférence et l'utilisation de la mémoire.</p> <p>Consultez le guide <a href="./speed-memory-optims">Compilation et offloading des modèles quantifiés</a> pour plus de détails.</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><!---->Déchargement sur CPU<!----></span></h3><!--]--><!----> <p>Le déchargement sur CPU (CPU offloading) déplace sélectivement les poids de la GPU vers la CPU. Lorsqu'un composant est requis, il est transféré vers la GPU ; lorsqu'il ne l'est pas, il est déplacé vers la CPU. Cette méthode fonctionne sur des sous-modules plutôt que sur des modèles entiers. Elle économise de la mémoire en évitant de stocker le modèle entier sur la GPU.</p> <p>Le déchargement sur CPU réduit considérablement l'utilisation de la mémoire, mais il est également <strong>extrêmement lent</strong> car les sous-modules sont transférés de nombreuses fois entre les appareils. Il peut souvent s'avérer impraticable en raison de sa lenteur.</p> <blockquote class="warning"><p>Ne déplacez pas le pipeline sur CUDA avant d’appeler <a href="/docs/diffusers/main/fr/api/pipelines/overview#diffusers.DiffusionPipeline.enable_sequential_cpu_offload">enable_sequential_cpu_offload()</a>, sinon la quantité de mémoire économisée sera minime (reportez-vous à cette <a href="https://github.com/huggingface/diffusers/issues/1934" rel="nofollow">issue</a> pour plus de détails). Il s’agit d’une opération avec état qui installe des hooks sur le modèle.</p></blockquote> <p>Appel de <a href="/docs/diffusers/main/fr/api/pipelines/overview#diffusers.DiffusionPipeline.enable_sequential_cpu_offload">enable_sequential_cpu_offload()</a> pour l’activer sur un 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> Copié</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><!---->Déchargement du modèle<!----></span></h3><!--]--><!----> <p>Le déchargement de modèle (model offloading) déplace des modèles entiers vers le GPU, au lieu de déplacer sélectivement <em>certaines</em> couches ou composants du modèle. L’un des principaux modèles du pipeline, généralement l’encodeur de texte, le UNet et le VAE, est placé sur le GPU, tandis que les autres composants sont conservés sur le CPU. Les composants comme le UNet, qui s’exécutent plusieurs fois, restent sur le GPU jusqu’à ce qu’ils soient totalement terminés et non plus nécessaires. Cela élimine la surcharge de communication liée au <a href="#cpu-offloading">déchargement sur CPU (CPU offloading)</a> et rend le déchargement de modèle une alternative plus rapide. Le compromis est que les économies de mémoire ne seront pas aussi importantes.</p> <blockquote class="warning"><p>Gardez à l’esprit que si les modèles sont réutilisés en dehors du pipeline après l’installation des hooks (voir <a href="https://huggingface.co/docs/accelerate/en/package_reference/big_modeling#accelerate.hooks.remove_hook_from_module" rel="nofollow">Removing Hooks</a> pour plus de détails), vous devez exécuter l’ensemble du pipeline et les modèles dans l’ordre attendu afin de les décharger correctement. Il s’agit d’une opération avec état qui installe des hooks sur le modèle.</p></blockquote> <p>Appel de <a href="/docs/diffusers/main/fr/api/pipelines/overview#diffusers.DiffusionPipeline.enable_model_cpu_offload">enable_model_cpu_offload()</a> pour l’activer sur un 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> Copié</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/main/fr/api/pipelines/overview#diffusers.DiffusionPipeline.enable_model_cpu_offload">enable_model_cpu_offload()</a> est également utile lorsque vous utilisez la méthode <a href="/docs/diffusers/main/fr/api/pipelines/stable_diffusion/stable_diffusion_xl#diffusers.StableDiffusionXLPipeline.encode_prompt">encode_prompt()</a> seule pour générer l’état caché des encodeurs de texte.</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><!---->Group offloading<!----></span></h3><!--]--><!----> <p>Le group offloading déplace des groupes de couches internes (<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>) vers le CPU. Il utilise moins de mémoire que le <a href="#model-offloading">model offloading</a> et est plus rapide que le <a href="#cpu-offloading">CPU offloading</a> car il réduit la surcharge de communication.</p> <blockquote class="warning"><p>Le déchargement par groupe peut ne pas fonctionner avec tous les modèles si l’implémentation du calcul avant (forward) contient un cast de dispositif des entrées dépendant des poids, car cela peut entrer en conflit avec le mécanisme de cast de dispositif du déchargement par groupe.</p></blockquote> <p>Activez le déchargement par groupe en configurant le paramètre <code>offload_type</code> sur <code>block_level</code> ou <code>leaf_level</code>.</p> <ul><li><code>block_level</code> décharge des groupes de couches en fonction du paramètre <code>num_blocks_per_group</code>. Par exemple, si <code>num_blocks_per_group=2</code> sur un modèle de 40 couches, 2 couches sont chargées et déchargées à la fois (20 chargements/déchargements au total). Cela réduit considérablement les besoins en mémoire.</li> <li><code>leaf_level</code> décharge les couches individuelles au niveau le plus bas et est équivalent à <a href="#cpu-offloading">la décharge CPU</a>. Mais il peut être accéléré si vous utilisez des flux sans sacrifier la vitesse d’inférence.</li></ul> <p>Le group offloading est pris en charge pour des pipelines entiers ou pour des modèles individuels. Appliquer le group offloading à l’ensemble du pipeline est l’option la plus simple, tandis que son application sélective à des modèles individuels offre aux utilisateurs une plus grande flexibilité pour utiliser différentes techniques d’offloading pour différents modèles.</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>Appeler <a href="/docs/diffusers/main/fr/api/pipelines/overview#diffusers.DiffusionPipeline.enable_group_offload">enable_group_offload()</a> sur un 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> Copié</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>Appel <a href="/docs/diffusers/main/fr/api/models/overview#diffusers.ModelMixin.enable_group_offload">enable_group_offload()</a> sur les composants de modèles Diffusers standard qui héritent de <a href="/docs/diffusers/main/fr/api/models/overview#diffusers.ModelMixin">ModelMixin</a>. Pour les autres composants de modèles qui n’héritent pas de <a href="/docs/diffusers/main/fr/api/models/overview#diffusers.ModelMixin">ModelMixin</a>, tels qu’un <a href="https://pytorch.org/docs/stable/generated/torch.nn.Module.html" rel="nofollow">torch.nn.Module</a> générique, utilisez plutôt <a href="/docs/diffusers/main/fr/api/utilities#diffusers.hooks.apply_group_offloading">apply_group_offloading()</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> Copié</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><!---->Flux CUDA<!----></span></h4><!--]--><!----> <p>Le paramètre <code>use_stream</code> peut être activé sur les appareils CUDA qui prennent en charge les flux de transfert de données asynchrones afin de réduire le temps d’exécution global par rapport au <a href="#cpu-offloading">déchargement sur CPU</a>. Il superpose le transfert de données et le calcul en utilisant la prélecture des couches. La couche suivante à exécuter est chargée sur le GPU pendant que la couche actuelle est encore en cours d’exécution. Cela peut augmenter considérablement la mémoire CPU, assurez-vous donc de disposer de deux fois la taille du modèle en mémoire.</p> <p>Définissez <code>record_stream=True</code> pour un gain de vitesse plus important, au prix d’une légère augmentation de l’utilisation de la mémoire. Consultez la documentation de <a href="https://pytorch.org/docs/stable/generated/torch.Tensor.record_stream.html" rel="nofollow">torch.Tensor.record_stream</a> pour en savoir plus.</p> <blockquote class="tip"><p>Lorsque <code>use_stream=True</code> est utilisé sur des VAE avec le tiling activé, assurez-vous d'effectuer un passage avant factice (possible également avec des entrées factices) avant l'inférence afin d'éviter les erreurs de correspondance de périphérique. Cela peut ne pas fonctionner sur toutes les implémentations, n'hésitez donc pas à ouvrir un ticket si vous rencontrez des problèmes.</p></blockquote> <p>Si vous utilisez le déchargement de groupes au niveau des blocs (<code>block_level</code>) avec <code>use_stream</code> activé, le paramètre <code>num_blocks_per_group</code> doit être défini sur <code>1</code>, sinon un avertissement sera émis.</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> Copié</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>Le paramètre <code>low_cpu_mem_usage</code> peut être défini sur <code>True</code> pour réduire l’utilisation de la mémoire CPU lors de l’utilisation de flux pendant le déchargement par groupes. C’est particulièrement adapté au déchargement <code>leaf_level</code> et lorsque la mémoire CPU est un goulot d’étranglement. La mémoire est économisée en créant des tenseurs épinglés à la volée plutôt qu’en les épinglant au préalable. Toutefois, cela peut augmenter le temps d’exécution global.</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><!---->Déchargement sur disque<!----></span></h4><!--]--><!----> <p>Le group offloading peut consommer une quantité importante de mémoire système en fonction de la taille du modèle. Sur les systèmes à mémoire limitée, essayez le group offloading sur le disque en tant que mémoire secondaire.</p> <p>Définissez l'argument <code>offload_to_disk_path</code> dans <a href="/docs/diffusers/main/fr/api/models/overview#diffusers.ModelMixin.enable_group_offload">enable_group_offload()</a> ou <a href="/docs/diffusers/main/fr/api/utilities#diffusers.hooks.apply_group_offloading">apply_group_offloading()</a> pour décharger le modèle sur le disque.</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> Copié</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>Consultez ces <a href="https://github.com/huggingface/diffusers/pull/11682#issue-3129365363" rel="nofollow">deux</a> <a href="https://github.com/huggingface/diffusers/pull/11682#issuecomment-2955715126" rel="nofollow">tableaux</a> pour comparer les compromis entre vitesse et mémoire.</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 par couche<!----></span></h2><!--]--><!----> <blockquote class="tip"><p>Combinez le <a href="#group-offloading">group offloading</a> avec le casting par couche pour des économies de mémoire encore plus importantes.</p></blockquote> <p>Le casting par couche stocke les poids dans un format de données plus compact (par exemple, <code>torch.float8_e4m3fn</code> et <code>torch.float8_e5m2</code>) afin de réduire la mémoire utilisée, puis les convertit en une précision supérieure comme <code>torch.float16</code> ou <code>torch.bfloat16</code> pour le calcul. Certaines couches (les poids liés à la normalisation et à la modulation) sont ignorées, car les stocker en fp8 peut dégrader la qualité de la génération.</p> <blockquote class="warning"><p>Le castage par couche peut ne pas fonctionner avec tous les modèles si l’implémentation du passage avant (forward) contient un castage interne des poids. L’implémentation actuelle du castage par couche suppose que le passage avant est indépendant de la précision des poids et que les types de données d’entrée sont toujours spécifiés dans <code>compute_dtype</code> (voir <a href="https://github.com/huggingface/transformers/blob/7f5077e53682ca855afc826162b204ebf809f1f9/src/transformers/models/t5/modeling_t5.py#L294-L299" rel="nofollow">ici</a> pour une implémentation incompatible).</p> <p>Le castage par couche peut également échouer sur des implémentations de modélisation personnalisées avec des couches <a href="https://huggingface.co/docs/peft/index" rel="nofollow">PEFT</a>. Des vérifications sont disponibles, mais elles ne sont pas testées de manière exhaustive ni garanties de fonctionner dans tous les cas.</p></blockquote> <p>Appel <a href="/docs/diffusers/main/fr/api/models/overview#diffusers.ModelMixin.enable_layerwise_casting">enable_layerwise_casting()</a> pour définir les types de données de stockage et de calcul.</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> Copié</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>La méthode <a href="/docs/diffusers/main/fr/api/utilities#diffusers.hooks.apply_layerwise_casting">apply_layerwise_casting()</a> peut également être utilisée si vous avez besoin de plus de contrôle et de flexibilité. Elle peut être appliquée partiellement aux couches du modèle en l'appelant sur des modules internes spécifiques. Utilisez les paramètres <code>skip_modules_pattern</code> ou <code>skip_modules_classes</code> pour spécifier les modules à éviter, tels que les couches de normalisation et de modulation.</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> Copié</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> inverse la façon dont les tenseurs sont stockés, passant de <code>(batch size, channels, height, width)</code> à <code>(batch size, height, width, channels)</code>. Cela aligne les tenseurs sur la manière dont le matériel accède séquentiellement aux tenseurs stockés en mémoire et évite de sauter d’un endroit à l’autre en mémoire pour accéder aux valeurs des pixels.</p> <p>Tous les opérateurs ne prennent pas encore en charge le format channels-last et peuvent entraîner des performances dégradées, mais il vaut toujours la peine d’essayer.</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> Copié</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><!---->Attention efficace en mémoire<!----></span></h2><!--]--><!----> <p>Diffusers prend en charge plusieurs backends d'attention efficaces en mémoire (FlashAttention, xFormers, SageAttention, et plus encore) via <a href="/docs/diffusers/main/fr/api/models/overview#diffusers.ModelMixin.set_attention_backend">set_attention_backend()</a>. Consultez le guide <a href="./attention_backends">Backends d'attention</a> pour apprendre à basculer entre eux.</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:: 104 kB
Xet hash:: 2d94e15c5d68e2de6f8345b2b0c810c97429c1e9816c3cf820b676b573c6a7e0
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.