Hugging Face's logo

Buckets:
hf-doc-build
/
doc

Files
xet
hf-doc-build/doc / diffusers /main /ko /optimization /fp16.html
mishig's picture
mishig
download
raw
56.3 kB
<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;정밀도와 컴파일&quot;,&quot;local&quot;:&quot;precision-and-compilation&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;모델 데이터 유형&quot;,&quot;local&quot;:&quot;model-data-type&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;스케일된 도트 프로덕트 어텐션&quot;,&quot;local&quot;:&quot;scaled-dot-product-attention&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;torch.compile&quot;,&quot;local&quot;:&quot;torchcompile&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;동적 형상 컴파일&quot;,&quot;local&quot;:&quot;dynamic-shape-compilation&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;지역별 컴파일&quot;,&quot;local&quot;:&quot;regional-compilation&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;그래프 브레이크&quot;,&quot;local&quot;:&quot;graph-breaks&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;GPU 동기화&quot;,&quot;local&quot;:&quot;gpu-sync&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;벤치마크&quot;,&quot;local&quot;:&quot;benchmarks&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3}],&quot;depth&quot;:2},{&quot;title&quot;:&quot;커널&quot;,&quot;local&quot;:&quot;kernels&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;동적 양자화&quot;,&quot;local&quot;:&quot;dynamic-quantization&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Fused projection matrices&quot;,&quot;local&quot;:&quot;fused-projection-matrices&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;다음 단계&quot;,&quot;local&quot;:&quot;next-steps&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2}],&quot;depth&quot;:1}"/>
<link href="/docs/diffusers/main/en/_app/immutable/entry/start.U6TUchwJ.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/chunks/yxSHiGDF.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/chunks/CPfOwUO_.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/entry/app.3qKhPoX-.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/chunks/BYe-f74j.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/chunks/D32hXVzh.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/nodes/0.DqpuBF3n.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/chunks/Dyp9xKq8.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/nodes/16.DxS2pBpm.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/chunks/Dk44y1gU.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/chunks/CYniLWKs.js" rel="modulepreload">
<!--9krdrc--><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;Precision and compilation&quot;,&quot;local&quot;:&quot;precision-and-compilation&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Model data type&quot;,&quot;local&quot;:&quot;model-data-type&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Scaled dot product attention&quot;,&quot;local&quot;:&quot;scaled-dot-product-attention&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;torch.compile&quot;,&quot;local&quot;:&quot;torchcompile&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Dynamic shape compilation&quot;,&quot;local&quot;:&quot;dynamic-shape-compilation&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Regional compilation&quot;,&quot;local&quot;:&quot;regional-compilation&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Graph breaks&quot;,&quot;local&quot;:&quot;graph-breaks&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;GPU sync&quot;,&quot;local&quot;:&quot;gpu-sync&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Benchmarks&quot;,&quot;local&quot;:&quot;benchmarks&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3}],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Kernels&quot;,&quot;local&quot;:&quot;kernels&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Dynamic quantization&quot;,&quot;local&quot;:&quot;dynamic-quantization&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Fused projection matrices&quot;,&quot;local&quot;:&quot;fused-projection-matrices&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Next steps&quot;,&quot;local&quot;:&quot;next-steps&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="precision-and-compilation" 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="#precision-and-compilation"><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><!---->정밀도와 컴파일<!----></span></h1><!--]--><!----> <p>낮은 정밀도와 컴파일은 Diffusers 추론을 가속화하는 두 가지 방법입니다. 가중치를 <code>bfloat16</code> 또는 <code>float16</code>로 로드한 후, <code>torch.compile</code> 또는 지역적 컴파일을 사용하여 디노이저(UNet 또는 transformer)를 컴파일합니다.</p> <!--[1--><h2 class="relative group"><a id="model-data-type" 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-data-type"><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><!---->모델 데이터 유형<!----></span></h2><!--]--><!----> <p>모델 가중치의 정밀도와 데이터 타입은 추론 속도에 영향을 미칩니다. 정밀도가 높을수록 가중치를 로드하는 데 더 많은 메모리가 필요하고, 연산을 수행하는 데 더 많은 시간이 소요되기 때문입니다. <code>dtype</code>을 생략하면 Diffusers는 모델 가중치를 <code>float32</code>로 로드하므로, 데이터 타입을 변경하는 것은 추론 속도를 빠르게 하는 간단한 방법입니다.</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">bfloat16</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">float16</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">TensorFloat-32</div><!--]--></div> <div class="language-select"><div><p><code>bfloat16</code>은 <code>float16</code>과 유사하지만 수치 오류에 대해 더 강건합니다. <code>bfloat16</code>에 대한 하드웨어 지원은 환경에 따라 다르지만, 대부분의 최신 GPU는 <code>bfloat16</code>을 지원합니다.</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> 복사됨</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> StableDiffusionXLPipeline
pipeline = StableDiffusionXLPipeline.from_pretrained(
<span class="hljs-string">&quot;stabilityai/stable-diffusion-xl-base-1.0&quot;</span>, dtype=torch.bfloat16
).to(<span class="hljs-string">&quot;cuda&quot;</span>) <span class="hljs-comment"># or &quot;mps&quot;, &quot;xpu&quot;, &quot;cpu&quot;</span>
prompt = <span class="hljs-string">&quot;Astronaut in a jungle, cold color palette, muted colors, detailed, 8k&quot;</span>
pipeline(prompt, num_inference_steps=<span class="hljs-number">30</span>).images[<span class="hljs-number">0</span>]<!----></pre></div><!----><!----></div><!----> <div hidden=""><p><code>float16</code>은 <code>bfloat16</code>와 유사하지만 수치 오류가 발생할 가능성이 더 높을 수 있습니다.</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> 복사됨</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> 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>
prompt = <span class="hljs-string">&quot;Astronaut in a jungle, cold color palette, muted colors, detailed, 8k&quot;</span>
pipeline(prompt, num_inference_steps=<span class="hljs-number">30</span>).images[<span class="hljs-number">0</span>]<!----></pre></div><!----><!----></div><!----> <div hidden=""><p><a href="https://blogs.nvidia.com/blog/2020/05/14/tensorfloat-32-precision-format/" rel="nofollow">TensorFloat-32</a> (<code>tf32</code>) 모드는 NVIDIA Ampere 및 이후 GPU에서 지원되며, 컨볼루션 및 행렬 곱셈 연산을 <code>tf32</code>로 수행합니다. 저장 및 기타 연산은 <code>float32</code>로 유지됩니다. <code>float32</code>로 실행되는 연산의 속도를 높여 주므로, 파이프라인 전체 또는 일부가 <code>float32</code>로 유지될 때 가장 큰 도움이 됩니다.</p> <p>PyTorch는 기본적으로 합성곱(convolution)에 대해서만 <code>tf32</code> 모드를 활성화하며, 행렬 곱셈에 대해서는 명시적으로 활성화해야 합니다.</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> 복사됨</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> StableDiffusionXLPipeline
torch.backends.cuda.matmul.allow_tf32 = <span class="hljs-literal">True</span>
pipeline = StableDiffusionXLPipeline.from_pretrained(
<span class="hljs-string">&quot;stabilityai/stable-diffusion-xl-base-1.0&quot;</span>, dtype=torch.float32
).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;Astronaut in a jungle, cold color palette, muted colors, detailed, 8k&quot;</span>
pipeline(prompt, num_inference_steps=<span class="hljs-number">30</span>).images[<span class="hljs-number">0</span>]<!----></pre></div><!----> <p>자세한 내용은 <a href="https://huggingface.co/docs/transformers/en/perf_train_gpu_one#mixed-precision" rel="nofollow">혼합 정밀도 학습</a> 문서를 참조하세요.</p><!----></div><!----><!----></div><!----> <!--[1--><h2 class="relative group"><a id="scaled-dot-product-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="#scaled-dot-product-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><!---->스케일된 도트 프로덕트 어텐션<!----></span></h2><!--]--><!----> <p><a href="https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html" rel="nofollow">Scaled dot product attention (SDPA)</a>는 PyTorch 2.0 이상에서 <code>AttnProcessor2_0</code> 또는 attention dispatcher의 <code>native</code> 백엔드를 통해 기본 attention으로 사용됩니다. PyTorch는 하드웨어에 맞는 SDPA 커널을 선택합니다. FlashAttention, SageAttention, xFormers, Hub 커널 및 기타 백엔드에 대해서는 <a href="./attention_backends">Attention backends</a>를 참고하세요.</p> <!--[1--><h2 class="relative group"><a id="torchcompile" 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="#torchcompile"><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.compile<!----></span></h2><!--]--><!----> <p><a href="https://pytorch.org/tutorials/intermediate/torch_compile_tutorial.html" rel="nofollow">torch.compile</a>은 PyTorch 코드와 연산을 최적화된 커널로 컴파일하여 추론을 가속화합니다. 일반적으로 런타임에 지배적인 역할을 하는 디노이저, 즉 <code>pipeline.unet</code> 또는 <code>pipeline.transformer</code>를 컴파일하며, 경우에 따라 VAE도 함께 컴파일합니다.</p> <p>최대 속도를 위해 다음 컴파일러 설정을 활성화하세요(추가 옵션은 <a href="https://github.com/pytorch/pytorch/blob/main/torch/_inductor/config.py" rel="nofollow">전체 목록</a>을 참조하세요).</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> 복사됨</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> StableDiffusionXLPipeline
torch._inductor.config.conv_1x1_as_mm = <span class="hljs-literal">True</span>
torch._inductor.config.coordinate_descent_tuning = <span class="hljs-literal">True</span>
torch._inductor.config.epilogue_fusion = <span class="hljs-literal">False</span>
torch._inductor.config.coordinate_descent_check_all_directions = <span class="hljs-literal">True</span><!----></pre></div><!----> <p>Denoiser와 VAE를 로드하고 컴파일합니다. Stable Diffusion XL과 같은 UNet 파이프라인에는 <code>pipeline.unet</code>를, Flux 및 DiT 스타일 파이프라인에는 <code>pipeline.transformer</code>를 사용합니다. 선택할 수 있는 여러 가지 모드가 있지만, <code>"max-autotune"</code>는 가장 빠른 커널을 검색하고 CUDA graphs를 사용합니다. CUDA graphs는 단일 CPU 연산을 통해 여러 GPU 연산을 실행함으로써 오버헤드를 줄입니다.</p> <p>메모리 레이아웃을 <a href="./memory#torchchannelslast">channels_last</a>로 변경하면 UNet, VAE와 같은 컨볼루션 중심 모델의 속도도 향상될 수 있습니다. 일부 모델은 오히려 느려질 수 있으므로 먼저 벤치마크를 실행해 보세요.</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> 복사됨</div><!----></button><!----></div> <pre class="language-py "><!---->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.unet.to(memory_format=torch.channels_last)
pipeline.unet = torch.<span class="hljs-built_in">compile</span>(
pipeline.unet, mode=<span class="hljs-string">&quot;max-autotune&quot;</span>, fullgraph=<span class="hljs-literal">True</span>
)
pipeline.vae.to(memory_format=torch.channels_last)
pipeline.vae.decode = torch.<span class="hljs-built_in">compile</span>(
pipeline.vae.decode,
mode=<span class="hljs-string">&quot;max-autotune&quot;</span>,
fullgraph=<span class="hljs-literal">True</span>
)
prompt = <span class="hljs-string">&quot;Astronaut in a jungle, cold color palette, muted colors, detailed, 8k&quot;</span>
pipeline(prompt, num_inference_steps=<span class="hljs-number">30</span>).images[<span class="hljs-number">0</span>]<!----></pre></div><!----> <p>최초 컴파일은 특히 <code>"max-autotune"</code>를 사용할 경우 느릴 수 있습니다. 컴파일러가 커널을 검색하기 때문이며, 이후에는 동일한 컴파일된 파이프라인 객체를 계속 사용합니다(캐싱 동작을 구성하는 방법은 <a href="https://docs.pytorch.org/tutorials/recipes/torch_compile_caching_tutorial.html" rel="nofollow">Compile Time Caching</a> 및 <a href="https://docs.pytorch.org/tutorials/recipes/torch_compile_caching_configuration_tutorial.html" rel="nofollow">Caching Configuration</a>을 참조하세요). 다른 이미지 크기로 컴파일된 파이프라인을 호출하면 다시 컴파일이 트리거됩니다.</p> <!--[2--><h3 class="relative group"><a id="dynamic-shape-compilation" 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="#dynamic-shape-compilation"><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><!---->동적 형상 컴파일<!----></span></h3><!--]--><!----> <p><code>torch.compile</code>은 입력 형상과 조건을 추적하며, 이들이 다르면 모델을 다시 컴파일합니다. 예를 들어, 1024x1024 해상도의 이미지로 컴파일된 모델이 다른 해상도의 이미지에 사용되면 재컴파일이 트리거됩니다.</p> <p>재컴파일링을 피하려면 <code>dynamic=True</code>를 추가하여 조건이 변경될 때 재컴파일링을 피할 수 있는 더 동적인 커널을 생성하도록 시도해 보세요.</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> 복사됨</div><!----></button><!----></div> <pre class="language-diff "><!----><span class="hljs-addition">+ torch.fx.experimental._config.use_duck_shape = False</span>
<span class="hljs-addition">+ pipeline.unet = torch.compile(</span>
pipeline.unet, fullgraph=True, dynamic=True
)<!----></pre></div><!----> <p><code>use_duck_shape=False</code>를 지정하면 컴파일러가 동일한 크기의 입력을 동일한 기호 변수로 표현할지 여부를 결정합니다. 자세한 내용은 이 <a href="https://github.com/huggingface/diffusers/pull/11327#discussion_r2047659790" rel="nofollow">댓글</a>을 확인하세요.</p> <p>모든 모델이 기본적으로 동적 컴파일의 이점을 누릴 수 있는 것은 아니며, 일부 모델은 변경이 필요할 수 있습니다. <a href="/docs/diffusers/main/ko/api/pipelines/aura_flow#diffusers.AuraFlowPipeline">AuraFlowPipeline</a> 구현을 개선하여 동적 컴파일의 이점을 누릴 수 있도록 한 이 <a href="https://github.com/huggingface/diffusers/pull/11297/" rel="nofollow">PR</a>을 참고하세요.</p> <p>Diffusers 모델에서 동적 컴파일이 예상대로 작동하지 않는 경우, 자유롭게 이슈를 열어 주세요.</p> <!--[2--><h3 class="relative group"><a id="regional-compilation" 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="#regional-compilation"><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><!---->지역별 컴파일<!----></span></h3><!--]--><!----> <p><a href="https://docs.pytorch.org/tutorials/recipes/regional_compilation.html" rel="nofollow">Regional compilation</a>은 모델의 <em>작고 반복적으로 나타나는 블록</em>, 일반적으로 transformer layer만 컴파일하여 콜드 스타트 지연을 줄이고, 이후 모든 발생 위치에서 컴파일된 산출물을 재사용할 수 있게 합니다.
많은 diffusion 아키텍처의 경우, 이는 전체 그래프 컴파일과 동등한 런타임 속도 향상을 제공하며 컴파일 시간을 최대 8~10배까지 단축할 수 있습니다.</p> <p>denoiser에 <code>torch.compile</code>을 래핑하는 헬퍼인 <a href="/docs/diffusers/main/ko/api/models/overview#diffusers.ModelMixin.compile_repeated_blocks">compile_repeated_blocks()</a> 메서드를 사용합니다. 파이프라인에 따라 <code>pipeline.unet</code> 또는 <code>pipeline.transformer</code>에 대해 호출합니다.</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> 복사됨</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> 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.unet.compile_repeated_blocks(fullgraph=<span class="hljs-literal">True</span>)<!----></pre></div><!----> <p>새 모델에 대해 지역 컴파일(regional compilation)을 활성화하려면 <code>_repeated_blocks</code>을 반복 블록의 클래스 이름(문자열)으로 설정합니다. SDXL의 UNet의 경우 <code>BasicTransformerBlock</code>을 사용합니다. Transformer 디노이저(denoiser)도 동일한 방식으로 자체 반복 블록 클래스 이름을 나열합니다.</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> 복사됨</div><!----></button><!----></div> <pre class="language-py "><!---->_repeated_blocks = [<span class="hljs-string">&quot;BasicTransformerBlock&quot;</span>]<!----></pre></div><!----> <p><a href="https://huggingface.co/docs/accelerate/index" rel="nofollow">Accelerate</a>에는 모델에서 컴파일할 후보 블록을 자동으로 선택하는 <a href="https://github.com/huggingface/accelerate/blob/273799c85d849a1954a4f2e65767216eb37fa089/src/accelerate/utils/other.py#L78" rel="nofollow">compile_regions</a> 메서드도 있습니다. 나머지 그래프는 별도로 컴파일됩니다. 컴파일할 블록을 지정하거나 컴파일 플래그를 조정할 수 있는 옵션이 많지 않으므로, 빠른 실험에 유용합니다.</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> 복사됨</div><!----></button><!----></div> <pre class="language-bash "><!---->pip install -U accelerate<!----></pre></div><!----> <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> 복사됨</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> StableDiffusionXLPipeline
<span class="hljs-keyword">from</span> accelerate.utils <span class="hljs-keyword">import</span> compile_regions
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.unet = compile_regions(pipeline.unet, mode=<span class="hljs-string">&quot;reduce-overhead&quot;</span>, fullgraph=<span class="hljs-literal">True</span>)<!----></pre></div><!----> <p><a href="/docs/diffusers/main/ko/api/models/overview#diffusers.ModelMixin.compile_repeated_blocks">compile_repeated_blocks()</a>는 의도적으로 명시적으로 설계되었습니다. <code>_repeated_blocks</code>에 반복할 블록을 나열하면 헬퍼는 해당 블록만 컴파일합니다. 이는 한 줄의 코드로 캐시 재사용에 대한 예측 가능한 동작과 쉬운 추론을 제공합니다.</p> <!--[2--><h3 class="relative group"><a id="graph-breaks" 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="#graph-breaks"><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><!---->그래프 브레이크<!----></span></h3><!--]--><!----> <p><code>fullgraph=True</code>를 설정하면 torch.compile이 그래프 분할을 조용히 수행하여 속도 향상이 감소하는 대신, 그래프 브레이크 발생 시 오류를 발생시킵니다.</p> <!--[2--><h3 class="relative group"><a id="gpu-sync" 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="#gpu-sync"><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><!---->GPU 동기화<!----></span></h3><!--]--><!----> <p>각 denoiser 예측 후, 파이프라인은 <code>sigmas</code>를 인덱싱하는 스케줄러의 <code>step</code> 메서드를 호출합니다. <code>sigmas</code>를 GPU에 유지하면 모든 단계에서 CPU↔GPU 동기화가 강제될 수 있습니다. 이 비용은 denoiser가 컴파일되기 전까지는 쉽게 간과할 수 있습니다. <code>sigmas</code>를 CPU에 유지하는 것이 좋습니다(Euler와 같은 Diffusers 스케줄러는 이미 이렇게 처리합니다).</p> <!--[2--><h3 class="relative group"><a id="benchmarks" 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="#benchmarks"><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><!---->벤치마크<!----></span></h3><!--]--><!----> <p>컴파일된 파이프라인의 추론 지연 시간 및 메모리 사용량 데이터를 확인하려면 <a href="https://huggingface.co/datasets/diffusers/benchmarks" rel="nofollow">diffusers/benchmarks</a> 데이터셋을 참조하세요.</p> <p><a href="https://github.com/sayakpaul/diffusers-torchao#benchmarking-results" rel="nofollow">diffusers-torchao</a> 저장소에는 Flux와 CogVideoX의 컴파일된 버전 벤치마킹 결과도 포함되어 있습니다.</p> <!--[1--><h2 class="relative group"><a id="kernels" 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="#kernels"><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><!---->커널<!----></span></h2><!--]--><!----> <p><a href="https://huggingface.co/docs/kernels/index" rel="nofollow">Kernels</a>는 <a href="https://huggingface.co/kernels-community" rel="nofollow">Hub</a>에서 최적화된 compute kernel을 빌드, 배포 및 로드하기 위한 라이브러리입니다. RMSNorm, GEGLU, RoPE, AdaLN과 같은 연산을 위한 <a href="./attention_backends#set-a-backend-on-the-model">attention</a> kernel 및 사용자 정의 CUDA kernel을 지원합니다.</p> <p>The <a href="https://github.com/huggingface/kernels/blob/main/kernel-builder/skills/cuda-kernels/references/diffusers-integration.md" rel="nofollow">Diffusers Pipeline Integration</a> guide shows how to integrate a kernel with the <a href="https://github.com/huggingface/kernels/blob/main/kernel-builder/skills/cuda-kernels/SKILL.md" rel="nofollow">add cuda-kernels</a> skill. This skill enables an agent, like Claude or Codex, to write custom kernels targeted towards a specific model and your hardware. The <a href="https://huggingface.co/blog/custom-cuda-kernels-agent-skills" rel="nofollow">Custom kernels for all from Codex and Claude</a> post has more detail.</p> <p>예를 들어, <code>add cuda-kernels</code> 스킬로 생성된 커스텀 RMSNorm 커널에 <a href="#torchcompile">torch.compile</a>을 적용하면 H100에서 LTX-Video 생성 속도가 1.43배 빨라집니다.</p> <iframe src="https://huggingface.co/datasets/docs-benchmarks/kernel-ltx-video/embed/viewer/default/train" frameborder="0" width="100%" height="560px"></iframe> <!--[1--><h2 class="relative group"><a id="dynamic-quantization" 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="#dynamic-quantization"><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><!---->동적 양자화<!----></span></h2><!--]--><!----> <p>동적 int8 활성화 양자화는 <a href="../quantization/torchao">torchao</a>에서 사용할 수 있습니다. <a href="/docs/diffusers/main/ko/api/quantization#diffusers.TorchAoConfig">TorchAoConfig</a>를 <a href="https://docs.pytorch.org/ao/stable/api_reference/generated/torchao.quantization.Int8DynamicActivationInt8WeightConfig.html" rel="nofollow"><code>Int8DynamicActivationInt8WeightConfig</code></a>와 같은 설정과 함께 사용하거나, torchao의 <code>quantize_</code> API를 사용하세요. 이전의 <code>apply_dynamic_quant</code> 헬퍼는 사용하지 마세요.</p> <!--[1--><h2 class="relative group"><a id="fused-projection-matrices" 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="#fused-projection-matrices"><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><!---->Fused projection matrices<!----></span></h2><!--]--><!----> <blockquote class="warning"><p><code>~DiffusionPipeline.fuse_qkv_projections</code>는 실험적입니다.</p></blockquote> <p>Q, K, V 투영 행렬로 표현되는 세 개의 부분 공간으로 입력이 투영되는 attention 블록에서, 이러한 투영은 일반적으로 개별적으로 계산됩니다. 하지만 이를 가로로 결합하여 단일 행렬로 만들고, 한 단계에서 투영을 수행할 수 있습니다. 이렇게 하면 입력 투영의 행렬 곱셈 크기가 증가하고, 양자화의 효과도 향상됩니다.</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> 복사됨</div><!----></button><!----></div> <pre class="language-py "><!---->pipeline.fuse_qkv_projections()<!----></pre></div><!----> <!--[1--><h2 class="relative group"><a id="next-steps" 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="#next-steps"><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><!---->다음 단계<!----></span></h2><!--]--><!----> <ul><li><p><a href="https://pytorch.org/blog/presenting-flux-fast-making-flux-go-brrr-on-h100s/" rel="nofollow">Presenting Flux Fast: Making Flux go brrr on H100s</a> 블로그 포스트를 읽어 <a href="https://github.com/huggingface/flux-fast" rel="nofollow">flux-fast</a>의 레시피를 사용하여 <a href="https://docs.pytorch.org/docs/stable/torch.compiler.html" rel="nofollow">TorchInductor</a>와 <a href="https://docs.pytorch.org/docs/stable/torch.compiler_aot_inductor.html" rel="nofollow">AOTInductor</a>와 이 모든 최적화를 결합하여 약 2.5배의 속도 향상을 달성하는 방법에 대해 자세히 알아보세요.</p> <p>이 레시피는 AMD 하드웨어와 <a href="https://huggingface.co/black-forest-labs/FLUX.1-Kontext-dev" rel="nofollow">Flux.1 Kontext Dev</a>를 지원합니다.</p></li> <li><p>Refer to the <a href="https://pytorch.org/blog/torch-compile-and-diffusers-a-hands-on-guide-to-peak-performance/" rel="nofollow">torch.compile and Diffusers: A Hands-On Guide to Peak Performance</a> blog post for maximizing performance with <code>torch.compile</code> for diffusion models.</p></li></ul> <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/fp16.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_gxjg5b = {
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.U6TUchwJ.js"),
import("/docs/diffusers/main/en/_app/immutable/entry/app.3qKhPoX-.js")
]).then(([kit, app]) => {
kit.start(app, element, {
node_ids: [0, 16],
data: [null,null],
form: null,
error: null
});
});
}
</script>

Xet Storage Details

Size:
56.3 kB
·
Xet hash:
5552b73d10d47a991db8c989067d281bb949107c00ecccc4176bd7634fe9eb86

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.