Buckets:

hf-doc-build
/

doc

Files
xet
hf-doc-build/doc / diffusers /main /ja /optimization /memory.html
mishig
3 days ago
download
raw
107 kB
<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;メモリ使用量の削減&quot;,&quot;local&quot;:&quot;reduce-memory-usage&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;複数のGPU&quot;,&quot;local&quot;:&quot;multiple-gpus&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;シャード化されたチェックポイント&quot;,&quot;local&quot;:&quot;sharded-checkpoints&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;デバイスの配置&quot;,&quot;local&quot;:&quot;device-placement&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3}],&quot;depth&quot;:2},{&quot;title&quot;:&quot;VAEスライシング&quot;,&quot;local&quot;:&quot;vae-slicing&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;VAEのタイル分割&quot;,&quot;local&quot;:&quot;vae-tiling&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;オフローディング&quot;,&quot;local&quot;:&quot;offloading&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;CPUオフロード&quot;,&quot;local&quot;:&quot;cpu-offloading&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;モデルのオフロード&quot;,&quot;local&quot;:&quot;model-offloading&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;グループオフローディング&quot;,&quot;local&quot;:&quot;group-offloading&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;CUDA ストリーム&quot;,&quot;local&quot;:&quot;cuda-stream&quot;,&quot;sections&quot;:[],&quot;depth&quot;:4},{&quot;title&quot;:&quot;ディスクへのオフロード&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;レイヤーごとのキャスト&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;メモリ効率の高いアテンション&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><!---->メモリ使用量の削減<!----></span></h1><!--]--><!----> <p><a href="../api/pipelines/flux">Flux</a> や <a href="../api/pipelines/wan">Wan</a> などの最新の拡散モデルは数十億のパラメータを持ち、推論時にハードウェア上で大量のメモリを消費します。一般的な GPU には十分なメモリがないことが多いため、これは課題となります。メモリ制限を克服するには、複数の GPU を使用したり（利用可能な場合）、パイプラインの一部を CPU にオフロードしたり、その他の方法があります。</p> <p>このガイドでは、メモリ使用量を削減する方法について説明します。</p> <blockquote class="tip"><p>これらの手法は、モデルによって調整が必要な場合があります。例えば、Transformerベースの拡散モデルは、UNetベースのモデルと比べて、これらのメモリ最適化から同程度の恩恵を受けられない可能性があります。</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><!---->複数のGPU<!----></span></h2><!--]--><!----> <p>複数のGPUにアクセスできる場合、大規模なモデルをハードウェア全体に効率的にロードして分散するためのいくつかのオプションがあります。これらの機能は<a href="https://huggingface.co/docs/accelerate/index" rel="nofollow">Accelerate</a>ライブラリによってサポートされているため、まずインストールされていることを確認してください。</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> コピーしました</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><!---->シャード化されたチェックポイント<!----></span></h3><!--]--><!----> <p>複数のシャードに分割された大きなチェックポイントをロードすることは有用です。シャードは1つずつロードされるため、メモリ使用量を低く抑えられ、モデルサイズと最大のシャードサイズに必要なメモリだけで済みます。fp32のチェックポイントが5GBを超える場合は、シャード化することを推奨します。デフォルトのシャードサイズは5GBです。</p> <p><code>max_shard_size</code> パラメータを指定して、<a href="/docs/diffusers/main/ja/api/pipelines/overview#diffusers.DiffusionPipeline.save_pretrained">save_pretrained()</a> でチェックポイントをシャード化します。</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> コピーしました</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>これにより、通常のチェックポイントの代わりにシャード化されたチェックポイントを使用することで、メモリを節約できます。</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> コピーしました</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><!---->デバイスの配置<!----></span></h3><!--]--><!----> <blockquote class="warning"><p>デバイス配置は実験的な機能であり、APIが変更される可能性があります。現時点では、<code>balanced</code>戦略のみがサポートされています。今後、追加のマッピング戦略をサポートする予定です。</p></blockquote> <p><code>device_map</code> パラメータは、パイプライン内のモデルコンポーネント、または個別モデルのレイヤーがデバイス間でどのように分散されるかを制御します。</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">パイプラインレベル</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">モデルレベル</div><!--]--></div> <div class="language-select"><div><p><code>balanced</code> デバイス配置戦略では、パイプラインが利用可能なすべてのデバイスに均等に分割されます。</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> コピーしました</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><code>hf_device_map</code> を使用して、パイプラインのデバイスマップを確認できます。</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> コピーしました</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><code>device_map</code> は、12.5B パラメータを持つ Flux diffusion transformer のような大規模モデルの読み込みに便利です。<code>"auto"</code> に設定すると、モデルが最も高速なデバイスから順に、遅いデバイスへと自動的に分散配置されます。詳細については、<a href="#sharded-checkpoints">Sharded checkpoints</a> セクションを参照してください。</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> コピーしました</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><code>hf_device_map</code> を使用して、モデルのデバイスマップを確認できます。</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> コピーしました</div><!----></button><!----></div> <pre class="language-py "><!----><span class="hljs-built_in">print</span>(transformer.hf_device_map)<!----></pre></div><!----><!----></div><!----><!----></div><!----> <p>独自の<code>device_map</code>を設計する際は、モデルの特定のモジュール名またはレイヤーと、デバイス識別子（GPUの場合は整数、CPUの場合は<code>cpu</code>、ディスクの場合は<code>disk</code>）の辞書である必要があります。</p> <p>モデルに対して<code>hf_device_map</code>を呼び出して、モデルのレイヤーがどのように分散されているかを確認し、その後、独自の分散設計を行ってください。</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> コピーしました</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>例えば、以下の<code>device_map</code>では、<code>single_transformer_blocks.10</code>から<code>single_transformer_blocks.20</code>までを2番目のGPU(<code>1</code>)に配置します。</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> コピーしました</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>各デバイスの最大メモリ使用量をマッピングする辞書を渡すことで、上限を強制できます。<code>max_memory</code> に含まれていないデバイスは無視され、パイプラインのコンポーネントはそこに分配されません。</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> コピーしました</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 はデフォルトで全デバイスの最大メモリを使用しますが、GPU に収まらない場合は、以下の方法を使用して単一の GPU を使用し、CPU へオフロードする必要があります。</p> <ul><li><a href="/docs/diffusers/main/ja/api/pipelines/overview#diffusers.DiffusionPipeline.enable_model_cpu_offload">enable_model_cpu_offload()</a> は単一の GPU でのみ動作しますが、非常に大きなモデルは 1 つの GPU に収まらない場合があります。</li> <li><a href="/docs/diffusers/main/ja/api/pipelines/overview#diffusers.DiffusionPipeline.enable_sequential_cpu_offload">enable_sequential_cpu_offload()</a> は動作する場合がありますが、非常に低速であり、かつ単一の GPU に限定されます。</li></ul> <p><code>device_map</code>をリセットするには、<a href="/docs/diffusers/main/ja/api/pipelines/overview#diffusers.DiffusionPipeline.reset_device_map">reset_device_map()</a>メソッドを使用します。デバイスマップが設定されたパイプラインに対して<code>.to()</code>、<a href="/docs/diffusers/main/ja/api/pipelines/overview#diffusers.DiffusionPipeline.enable_sequential_cpu_offload">enable_sequential_cpu_offload()</a>、<a href="/docs/diffusers/main/ja/api/pipelines/overview#diffusers.DiffusionPipeline.enable_model_cpu_offload">enable_model_cpu_offload()</a>などのメソッドを使用する場合は、この処理が必要です。</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> コピーしました</div><!----></button><!----></div> <pre class="language-py "><!---->pipeline.reset_device_map()<!----></pre></div><!----> <!--[1--><h2 class="relative group"><a id="vae-slicing" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#vae-slicing"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span><!---->VAEスライシング<!----></span></h2><!--]--><!----> <p>VAE slicingは、大きなバッチの入力を単一のデータバッチに分割して個別に処理することでメモリを節約します。この手法は、一度に複数の画像を生成する場合に最も効果的です。</p> <p>例えば、一度に4枚の画像を生成する場合、デコードによってピーク時のアクティベーションメモリが4倍に増加します。VAEスライシングでは、4枚すべてを一度にデコードするのではなく、1枚ずつデコードすることで、この増加を抑制します。</p> <p>スライスされた VAE を有効にするには、<code>enable_slicing()</code> を呼び出します。複数画像のバッチをデコードする際にパフォーマンスがわずかに向上し、単一画像のバッチではパフォーマンスへの影響はありません。</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> コピーしました</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><a href="/docs/diffusers/main/ja/api/models/autoencoder_kl_wan#diffusers.AutoencoderKLWan">AutoencoderKLWan</a> クラスと <a href="/docs/diffusers/main/ja/api/models/asymmetricautoencoderkl#diffusers.AsymmetricAutoencoderKL">AsymmetricAutoencoderKL</a> クラスは、スライシングをサポートしていません。</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><!---->VAEのタイル分割<!----></span></h2><!--]--><!----> <p>VAE tilingは、画像全体を一度に処理するのではなく、より小さな重複するタイルに分割して処理することでメモリを節約します。これにより、GPUが一度に1つのタイルのみを処理するため、ピーク時のメモリ使用量も削減されます。</p> <p><code>enable_tiling()</code> を呼び出して VAE のタイル分割を有効にします。タイルごとに個別にデコードされるため、生成された画像にはタイル間でわずかなトーンのばらつきが生じることがありますが、タイルの境界に明らかな継ぎ目が出ることはありません。解像度が事前に指定された（ただし設定可能）閾値を下回る場合は、タイル分割は無効になります。例えば、<a href="/docs/diffusers/main/ja/api/pipelines/stable_diffusion/text2img#diffusers.StableDiffusionPipeline">StableDiffusionPipeline</a> の VAE では、この閾値は 512x512 です。</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> コピーしました</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/ja/api/models/autoencoder_kl_wan#diffusers.AutoencoderKLWan">AutoencoderKLWan</a> と <a href="/docs/diffusers/main/ja/api/models/asymmetricautoencoderkl#diffusers.AsymmetricAutoencoderKL">AsymmetricAutoencoderKL</a> は、タイル分割をサポートしていません。</p></blockquote> <p>GPU上のVAEメモリを完全にスキップするには、<a href="../hybrid_inference/overview">Remote inference</a>でリモートエンドポイント上でエンコードとデコードを実行します。</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><!---->オフローディング<!----></span></h2><!--]--><!----> <p>オフローディング戦略では、現在アクティブでないレイヤーやモデルをCPUに移動し、GPUメモリを増加させないようにします。これらの戦略は、量子化やtorch.compileと組み合わせることで、推論速度とメモリ使用量のバランスを取ることができます。</p> <p>詳細については、<a href="./speed-memory-optims">Compile and offloading quantized models</a>ガイドを参照してください。</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><!---->CPUオフロード<!----></span></h3><!--]--><!----> <p>CPU offloadingは、重みをGPUからCPUへ選択的に移動します。コンポーネントが必要になった場合はGPUに転送され、不要になった場合はCPUに移動します。この手法は、モデル全体ではなくサブモジュール単位で動作します。モデル全体をGPUに保持することを避けることで、メモリを節約します。</p> <p>CPU offloadingはメモリ使用量を大幅に削減しますが、サブモジュールがデバイス間で何度もやり取りされるため、<strong>非常に遅い</strong>です。その遅さゆえに、実用的でない場合もよくあります。</p> <blockquote class="warning"><p><a href="/docs/diffusers/main/ja/api/pipelines/overview#diffusers.DiffusionPipeline.enable_sequential_cpu_offload">enable_sequential_cpu_offload()</a>を呼び出す前に、パイプラインをCUDAに移動しないでください。そうしないと、メモリ削減の効果がほとんど得られません（詳細はこの<a href="https://github.com/huggingface/diffusers/issues/1934" rel="nofollow">issue</a>を参照してください）。これはステートフルな操作であり、モデルにフックをインストールします。</p></blockquote> <p>パイプラインでこれを有効にするには、<a href="/docs/diffusers/main/ja/api/pipelines/overview#diffusers.DiffusionPipeline.enable_sequential_cpu_offload">enable_sequential_cpu_offload()</a>を呼び出します。</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> コピーしました</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><!---->モデルのオフロード<!----></span></h3><!--]--><!----> <p>Model offloadingでは、<em>一部の</em>レイヤーやモデルコンポーネントを選択的に移動するのではなく、モデル全体をGPUに移動します。通常、テキストエンコーダー、UNet、VAEといった主要なパイプラインモデルの1つがGPUに配置され、その他のコンポーネントはCPU上に保持されます。UNetのように複数回実行されるコンポーネントは、完全に処理が完了して不要になるまでGPU上に保持されます。これにより、<a href="#cpu-offloading">CPU offloading</a>に伴う通信オーバーヘッドが排除され、Model offloadingはより高速な代替手段となります。その代わりとして、メモリ削減の幅はCPU offloadingほど大きくはなりません。</p> <blockquote class="warning"><p>フックがインストールされた後、パイプラインの外でモデルが再利用される場合（詳細は<a href="https://huggingface.co/docs/accelerate/en/package_reference/big_modeling#accelerate.hooks.remove_hook_from_module" rel="nofollow">フックの削除</a>を参照）、モデルを適切にオフロードするには、パイプライン全体とモデルを想定された順序で実行する必要があります。これはモデルにフックをインストールする状態付きの操作です。</p></blockquote> <p>パイプラインでこれを有効にするには、<a href="/docs/diffusers/main/ja/api/pipelines/overview#diffusers.DiffusionPipeline.enable_model_cpu_offload">enable_model_cpu_offload()</a>を呼び出します。</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> コピーしました</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/ja/api/pipelines/overview#diffusers.DiffusionPipeline.enable_model_cpu_offload">enable_model_cpu_offload()</a> は、<a href="/docs/diffusers/main/ja/api/pipelines/stable_diffusion/stable_diffusion_xl#diffusers.StableDiffusionXLPipeline.encode_prompt">encode_prompt()</a> メソッドを単独で使用してテキストエンコーダーの隠れ状態を生成する場合にも役立ちます。</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><!---->グループオフローディング<!----></span></h3><!--]--><!----> <p>Group offloadingは、内部レイヤーのグループ（<a href="https://pytorch.org/docs/stable/generated/torch.nn.ModuleList.html" rel="nofollow">torch.nn.ModuleList</a>または<a href="https://pytorch.org/docs/stable/generated/torch.nn.Sequential.html" rel="nofollow">torch.nn.Sequential</a>）をCPUに移動します。<a href="#model-offloading">model offloading</a>よりもメモリ使用量が少なく、通信オーバーヘッドを削減できるため、<a href="#cpu-offloading">CPU offloading</a>よりも高速です。</p> <blockquote class="warning"><p>グループオフローディングは、フォワード実装が重みに依存する入力のデバイスキャストを含む場合、グループオフローディングのデバイスキャスト機構と競合する可能性があるため、すべてのモデルで動作しない場合があります。</p></blockquote> <p><code>offload_type</code> パラメータを <code>block_level</code> または <code>leaf_level</code> に設定することで、グループオフロードを有効にできます。</p> <ul><li><code>block_level</code> は、<code>num_blocks_per_group</code> パラメータに基づいてレイヤーのグループをオフロードします。例えば、40 レイヤーのモデルで <code>num_blocks_per_group=2</code> の場合、一度に 2 レイヤーがオンロードおよびオフロードされます（オンロード/オフロードの総数は 20 回）。これにより、メモリ要件が大幅に削減されます。</li> <li><code>leaf_level</code> は最下位レベルで個々のレイヤーをオフロードし、<a href="#cpu-offloading">CPU オフロード</a> と同等です。ただし、推論速度を犠牲にすることなくストリームを使用することで、より高速にできます。</li></ul> <p>グループオフロードは、パイプライン全体または個々のモデルに対して適用できます。パイプライン全体にグループオフロードを適用するのが最も簡単ですが、個々のモデルに対して選択的に適用することで、ユーザーはモデルごとに異なるオフロード手法を使用する柔軟性を得られます。</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>パイプラインで<a href="/docs/diffusers/main/ja/api/pipelines/overview#diffusers.DiffusionPipeline.enable_group_offload">enable_group_offload()</a>を呼び出します。</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> コピーしました</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><a href="/docs/diffusers/main/ja/api/models/overview#diffusers.ModelMixin">ModelMixin</a> を継承する標準的な Diffusers のモデルコンポーネントに対して <a href="/docs/diffusers/main/ja/api/models/overview#diffusers.ModelMixin.enable_group_offload">enable_group_offload()</a> を呼び出します。<a href="/docs/diffusers/main/ja/api/models/overview#diffusers.ModelMixin">ModelMixin</a> を継承しない他のモデルコンポーネント、例えば汎用的な <a href="https://pytorch.org/docs/stable/generated/torch.nn.Module.html" rel="nofollow">torch.nn.Module</a> には、代わりに <a href="/docs/diffusers/main/ja/api/utilities#diffusers.hooks.apply_group_offloading">apply_group_offloading()</a> を使用します。</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> コピーしました</div><!----></button><!----></div> <pre class="language-py "><!----><span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> CogVideoXPipeline
<span class="hljs-keyword">from</span> diffusers.hooks <span class="hljs-keyword">import</span> apply_group_offloading
<span class="hljs-keyword">from</span> diffusers.utils <span class="hljs-keyword">import</span> export_to_video

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

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

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

prompt = (
    <span class="hljs-string">&quot;A panda, dressed in a small, red jacket and a tiny hat, sits on a wooden stool in a serene bamboo forest. &quot;</span>
    <span class="hljs-string">&quot;The panda&#x27;s fluffy paws strum a miniature acoustic guitar, producing soft, melodic tunes. Nearby, a few other &quot;</span>
    <span class="hljs-string">&quot;pandas gather, watching curiously and some clapping in rhythm. Sunlight filters through the tall bamboo, &quot;</span>
    <span class="hljs-string">&quot;casting a gentle glow on the scene. The panda&#x27;s face is expressive, showing concentration and joy as it plays. &quot;</span>
    <span class="hljs-string">&quot;The background includes a small, flowing stream and vibrant green foliage, enhancing the peaceful and magical &quot;</span>
    <span class="hljs-string">&quot;atmosphere of this unique musical performance.&quot;</span>
)
video = pipeline(prompt=prompt, guidance_scale=<span class="hljs-number">6</span>, num_inference_steps=<span class="hljs-number">50</span>).frames[<span class="hljs-number">0</span>]
<span class="hljs-built_in">print</span>(<span class="hljs-string">f&quot;Max memory reserved: <span class="hljs-subst">{torch.cuda.max_memory_allocated() / <span class="hljs-number">1024</span>**<span class="hljs-number">3</span>:<span class="hljs-number">.2</span>f}</span> GB&quot;</span>)
export_to_video(video, <span class="hljs-string">&quot;output.mp4&quot;</span>, fps=<span class="hljs-number">8</span>)<!----></pre></div><!----><!----></div><!----><!----></div><!----> <!--[3--><h4 class="relative group"><a id="cuda-stream" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#cuda-stream"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span><!---->CUDA ストリーム<!----></span></h4><!--]--><!----> <p>CUDA デバイスで非同期データ転送ストリームがサポートされている場合、<code>use_stream</code> パラメータを有効にすることで、<a href="#cpu-offloading">CPU オフロード</a>と比較して全体の実行時間を短縮できます。レイヤープリフェッチを使用して、データ転送と計算をオーバーラップさせます。現在のレイヤーの実行中に、次に実行するレイヤーを GPU に読み込みます。CPU メモリが大幅に増加する可能性があるため、モデルサイズの 2 倍のメモリを確保してください。</p> <p>メモリ使用量がわずかに増加する代わりに、より大きな高速化を得るには、<code>record_stream=True</code> を設定してください。詳細については、<a href="https://pytorch.org/docs/stable/generated/torch.Tensor.record_stream.html" rel="nofollow">torch.Tensor.record_stream</a> のドキュメントを参照してください。</p> <blockquote class="tip"><p>タイル分割を有効にした VAE で <code>use_stream=True</code> を使用する場合、デバイス不一致エラーを回避するために、推論の前にダミーのフォワードパス（ダミー入力でも可）を実行してください。すべての実装で動作するとは限らないため、問題が発生した場合は、お気軽に issue を作成してください。</p></blockquote> <p><code>block_level</code> グループオフロードを <code>use_stream</code> を有効にして使用している場合、<code>num_blocks_per_group</code> パラメータは <code>1</code> に設定する必要があります。そうしないと、警告が発生します。</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> コピーしました</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><code>low_cpu_mem_usage</code> パラメータを <code>True</code> に設定すると、グループオフローディング時にストリームを使用している場合の CPU メモリ使用量を削減できます。<code>leaf_level</code> オフローディングや、CPU メモリがボトルネックになっている場合に特に有効です。ピン留めされたテンソルを事前に作成するのではなく、必要に応じてピン留めされたテンソルを動的に作成することでメモリを節約します。ただし、これにより全体の実行時間が長くなる可能性があります。</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><!---->ディスクへのオフロード<!----></span></h4><!--]--><!----> <p>Group offloading はモデルサイズによって、システムメモリを大量に消費する可能性があります。メモリが限られているシステムでは、セカンダリメモリとしてディスクへの group offloading を試してください。</p> <p>モデルをディスクにオフロードするには、<a href="/docs/diffusers/main/ja/api/models/overview#diffusers.ModelMixin.enable_group_offload">enable_group_offload()</a> または <a href="/docs/diffusers/main/ja/api/utilities#diffusers.hooks.apply_group_offloading">apply_group_offloading()</a> のいずれかで <code>offload_to_disk_path</code> 引数を設定します。</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> コピーしました</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>速度とメモリ使用量のトレードオフを比較するには、これらの<a href="https://github.com/huggingface/diffusers/pull/11682#issue-3129365363" rel="nofollow">2つの</a><a href="https://github.com/huggingface/diffusers/pull/11682#issuecomment-2955715126" rel="nofollow">テーブル</a>を参照してください。</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><!---->レイヤーごとのキャスト<!----></span></h2><!--]--><!----> <blockquote class="tip"><p>レイヤーごとのキャストを<a href="#group-offloading">グループオフローディング</a>と組み合わせることで、さらにメモリを節約できます。</p></blockquote> <p>Layerwise castingは、重みをより小さいデータ形式（例えば、<code>torch.float8_e4m3fn</code>や<code>torch.float8_e5m2</code>）で保存することでメモリ使用量を削減し、計算時には<code>torch.float16</code>や<code>torch.bfloat16</code>などの高精度な形式にアップキャストします。正規化やモジュレーションに関連する重みを含む特定のレイヤーは、fp8で保存すると生成品質が低下する可能性があるため、スキップされます。</p> <blockquote class="warning"><p>レイヤーワイズキャストは、フォワード実装が重みの内部型キャストを含む場合、すべてのモデルで動作しない可能性があります。現在のレイヤーワイズキャストの実装は、フォワードパスが重みの精度に依存しないこと、および入力データ型が常に<code>compute_dtype</code>で指定されることを前提としています（非互換な実装の例は<a href="https://github.com/huggingface/transformers/blob/7f5077e53682ca855afc826162b204ebf809f1f9/src/transformers/models/t5/modeling_t5.py#L294-L299" rel="nofollow">こちら</a>を参照）。</p> <p><a href="https://huggingface.co/docs/peft/index" rel="nofollow">PEFT</a>レイヤーを含むカスタムモデリング実装では、レイヤーごとのキャストが失敗する可能性があります。いくつかのチェックは用意されていますが、これらは十分にテストされておらず、すべてのケースで動作が保証されるものではありません。</p></blockquote> <p><a href="/docs/diffusers/main/ja/api/models/overview#diffusers.ModelMixin.enable_layerwise_casting">enable_layerwise_casting()</a>を呼び出して、ストレージと計算のデータ型を設定します。</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> コピーしました</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>より細かな制御と柔軟性が必要な場合は、<a href="/docs/diffusers/main/ja/api/utilities#diffusers.hooks.apply_layerwise_casting">apply_layerwise_casting()</a> メソッドも使用できます。特定の内部モジュールに対して呼び出すことで、モデルのレイヤーに部分的に適用できます。正規化レイヤーやモジュレーションレイヤーなど、回避するモジュールを指定するには、<code>skip_modules_pattern</code> または <code>skip_modules_classes</code> パラメータを使用します。</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> コピーしました</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> は、テンソルの保存形式を <code>(batch size, channels, height, width)</code> から <code>(batch size, height, width, channels)</code> に反転します。これにより、テンソルはメモリに保存されたテンソルにハードウェアがシーケンシャルにアクセスする方式と整合し、メモリ内でピクセル値にアクセスするために前後に飛び回ることを回避できます。</p> <p>現在、すべてのオペレーターが channels-last 形式をサポートしているわけではなく、パフォーマンスが最悪になる場合もありますが、試してみる価値はあります。</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> コピーしました</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><!---->メモリ効率の高いアテンション<!----></span></h2><!--]--><!----> <p>Diffusers は <a href="/docs/diffusers/main/ja/api/models/overview#diffusers.ModelMixin.set_attention_backend">set_attention_backend()</a> を通じて、複数のメモリ効率に優れたアテンションバックエンド（FlashAttention、xFormers、SageAttention など）をサポートしています。切り替え方については、<a href="./attention_backends">Attention backends</a> ガイドを参照してください。</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:: 107 kB
Xet hash:: 7f4e70e31715acd32cbc02f6c03a0907889665f43ec858375604bc8aa194f4e9
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.