Buckets:

hf-doc-build
/

doc

Files
xet
hf-doc-build/doc / diffusers /v0.43.0 /ko /optimization /fp16.html
mishig
about 2 hours ago
download
raw
66.4 kB
<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;추론 가속&quot;,&quot;local&quot;:&quot;accelerate-inference&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;리소스&quot;,&quot;local&quot;:&quot;resources&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/337.9hsECfE2.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">
		<!--9krdrc--><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;Accelerate inference&quot;,&quot;local&quot;:&quot;accelerate-inference&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;Resources&quot;,&quot;local&quot;:&quot;resources&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="accelerate-inference" 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="#accelerate-inference"><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><!---->추론 가속<!----></span></h1><!--]--><!----> <p>Diffusion 모델은 추론이 느립니다. 생성은 반복적인 과정으로, 특정 수의 “스텝”에 걸쳐 노이즈가 점진적으로 이미지나 비디오로 정제되기 때문입니다. 이 과정을 가속화하려면 다양한 <a href="../api/schedulers/overview">스케줄러</a>를 실험해 보거나, 모델 가중치의 정밀도를 낮춰 계산 속도를 높이고, 더 메모리 효율적인 어텐션 메커니즘을 사용하는 등의 방법을 시도해 볼 수 있습니다.</p> <p>이러한 기법들을 결합하여 함께 사용하면, 단일 기법만 사용할 때보다 추론 속도를 더 빠르게 할 수 있습니다.</p> <p>이 가이드에서는 추론을 가속하는 방법을 다룹니다.</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 유형은 추론 속도에 영향을 미칩니다. 정밀도가 높을수록 가중치를 로드하는 데 더 많은 메모리가 필요하고, 연산을 수행하는 데 더 많은 시간이 소요되기 때문입니다. PyTorch는 기본적으로 모델 가중치를 float32 또는 전체 정밀도로 로드하므로, 데이터 유형을 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>bfloat16은 float16과 유사하지만 수치 오류에 대해 더 강건합니다. bfloat16에 대한 하드웨어 지원은 다양하지만, 대부분의 최신 GPU는 bfloat16을 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>float16은 bfloat16과 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 (tf32)</a> 모드는 NVIDIA Ampere GPU에서 지원되며, tf32로 컨볼루션 및 행렬 곱셈 연산을 수행합니다. 저장 및 기타 연산은 float32로 유지됩니다. bfloat16 또는 float16과 함께 사용하면 계산 속도가 크게 향상됩니다.</p> <p>PyTorch는 기본적으로 합성곱(convolution)에 대해서만 tf32 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.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><!----> <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><!--]--><!----> <blockquote class="tip"><p>메모리 효율적 어텐션은 추론 속도 <em>및</em> <a href="./memory#memory-efficient-attention">메모리 사용량</a>을 최적화합니다!</p></blockquote> <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 <a href="https://github.com/Dao-AILab/flash-attention" rel="nofollow">FlashAttention</a>, <a href="https://github.com/facebookresearch/xformers" rel="nofollow">xFormers</a> 및 네이티브 C++ 구현 등 여러 attention backends를 구현합니다. 하드웨어에 가장 최적화된 backends를 자동으로 선택합니다.</p> <p>PyTorch &gt;= 2.0을 사용하는 경우 SDPA가 기본적으로 활성화되며, 코드에 추가적인 변경 사항은 필요하지 않습니다. 원하시는 경우 다른 attention backends를 실험하여 직접 선택할 수 있습니다. 아래 예제는 <a href="https://pytorch.org/docs/stable/generated/torch.nn.attention.sdpa_kernel.html" rel="nofollow">torch.nn.attention.sdpa_kernel</a> 컨텍스트 매니저를 사용하여 효율적인 attention을 활성화합니다.</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">from</span> torch.nn.attention <span class="hljs-keyword">import</span> SDPBackend, sdpa_kernel
<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>

<span class="hljs-keyword">with</span> sdpa_kernel(SDPBackend.EFFICIENT_ATTENTION):
  image = pipeline(prompt, num_inference_steps=<span class="hljs-number">30</span>).images[<span class="hljs-number">0</span>]<!----></pre></div><!----> <!--[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. Diffusers는 일반적으로 UNet, transformer, VAE와 같이 계산 집약적인 모델을 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>UNet과 VAE를 로드하고 컴파일합니다. 선택할 수 있는 여러 가지 모드가 있지만, <code>"max-autotune"</code>은 CUDA graph로 컴파일하여 가장 빠른 속도를 최적화합니다. CUDA graphs는 단일 CPU 연산을 통해 여러 GPU 연산을 실행함으로써 오버헤드를 효과적으로 줄여줍니다.</p> <blockquote class="tip"><p>PyTorch 2.3.1에서는 torch.compile의 캐싱 동작을 제어할 수 있습니다. 이는 여러 컴파일 플래그에 대해 그리드 서치를 수행하여 최적의 구성을 찾는 <code>"max-autotune"</code>와 같은 컴파일 모드에서 특히 유용합니다. <a href="https://pytorch.org/tutorials/recipes/torch_compile_caching_tutorial.html" rel="nofollow">torch.compile의 Compile Time Caching</a> 튜토리얼에서 자세히 알아보세요.</p></blockquote> <p>메모리 레이아웃을 <a href="./memory#torchchannelslast">channels_last</a>로 변경하면 메모리와 추론 속도도 최적화됩니다.</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.vae.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.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>컴파일은 첫 번째 실행 시 느리지만, 한 번 컴파일된 후에는 훨씬 빠릅니다. 동일한 유형의 추론 작업에만 컴파일된 파이프라인을 사용하도록 하십시오. 다른 이미지 크기로 컴파일된 파이프라인을 호출하면 컴파일이 다시 트리거되며, 이는 느리고 비효율적입니다.</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><!--]--><!----> <blockquote class="tip"><p>더 나은 지원을 위해 항상 PyTorch의 nightly 버전을 사용하도록 하십시오.</p></blockquote> <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/v0.43.0/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 실행되는 블록</em>, 일반적으로 transformer layer만 컴파일하여 콜드 스타트 지연 시간을 줄이고, 이후 모든 발생 위치에서 컴파일된 산출물을 재사용할 수 있게 합니다.

많은 diffusion 아키텍처의 경우, 이는 전체 그래프 컴파일과 동일한 런타임 속도 향상을 제공하면서 컴파일 시간을 8~10배 줄여줍니다.</p> <p>아래와 같이 transformer 모델과 같은 컴포넌트에 <code>torch.compile</code>을 래핑하는 헬퍼인 <a href="/docs/diffusers/v0.43.0/ko/api/models/overview#diffusers.ModelMixin.compile_repeated_blocks">compile_repeated_blocks()</a> 메서드를 사용합니다.</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-comment"># pip install -U diffusers</span>
<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>

<span class="hljs-comment"># compile only the repeated transformer layers inside the UNet</span>
pipeline.unet.compile_repeated_blocks(fullgraph=<span class="hljs-literal">True</span>)<!----></pre></div><!----> <p>새 모델에 대한 지역 컴파일을 활성화하려면, 컴파일할 블록의 클래스 이름(문자열)을 포함하는 <code>_repeated_blocks</code> 속성을 모델 클래스에 추가하세요.</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">class</span> <span class="hljs-title class_">MyUNet</span>(<span class="hljs-title class_ inherited__">ModelMixin</span>):
    _repeated_blocks = (<span class="hljs-string">&quot;Transformer2DModel&quot;</span>,)  <span class="hljs-comment"># ← compiled by default</span><!----></pre></div><!----> <blockquote class="tip"><p>지역별 컴파일 예제가 더 필요하면 <a href="https://github.com/huggingface/diffusers/pull/11705" rel="nofollow">PR</a>을 참고하세요.</p></blockquote> <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-py "><!----><span class="hljs-comment"># pip install -U accelerate</span>
<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/v0.43.0/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 torch.compile에 지정하는 것이 중요합니다. 이를 통해 기반 모델에 그래프 브레이크가 발생하지 않도록 보장할 수 있으며, 성능 저하 없이 torch.compile의 이점을 활용할 수 있습니다. UNet과 VAE의 경우, 반환 변수에 접근하는 방식이 달라집니다.</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-deletion">- latents = unet(</span>
<span class="hljs-deletion">-   latents, timestep=timestep, encoder_hidden_states=prompt_embeds</span>
<span class="hljs-deletion">-).sample</span>

<span class="hljs-addition">+ latents = unet(</span>
<span class="hljs-addition">+   latents, timestep=timestep, encoder_hidden_states=prompt_embeds, return_dict=False</span>
<span class="hljs-addition">+)[0]</span><!----></pre></div><!----> <!--[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><code>step()</code> 함수는 denoiser가 예측을 할 때마다 scheduler에서 <a href="https://github.com/huggingface/diffusers/blob/1d686bac8146037e97f3fd8c56e4063230f71751/src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl.py#L1228" rel="nofollow">호출</a>되며, <code>sigmas</code> 변수는 <a href="https://github.com/huggingface/diffusers/blob/1d686bac8146037e97f3fd8c56e4063230f71751/src/diffusers/schedulers/scheduling_euler_discrete.py#L476" rel="nofollow">인덱싱</a>됩니다. GPU에 배치되면 CPU와 GPU 간의 통신 동기화 때문에 지연이 발생합니다. denoiser가 이미 컴파일된 경우 이 지연이 더 뚜렷하게 나타납니다.</p> <p>일반적으로 통신 동기화 및 지연을 피하기 위해 <code>sigmas</code>는 <a href="https://github.com/huggingface/diffusers/blob/35a969d297cba69110d175ee79c59312b9f49e1e/src/diffusers/schedulers/scheduling_euler_discrete.py#L240" rel="nofollow">CPU에 유지</a>되어야 합니다.</p> <blockquote class="tip"><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></blockquote> <!--[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#setattentionbackend">attention</a> kernel 및 사용자 정의 CUDA kernel을 지원합니다.</p> <p><a href="https://github.com/huggingface/kernels/blob/main/kernel-builder/skills/cuda-kernels/references/diffusers-integration.md" rel="nofollow">Diffusers Pipeline Integration</a> 가이드는 커널을 <a href="https://github.com/huggingface/kernels/blob/main/kernel-builder/skills/cuda-kernels/SKILL.md" rel="nofollow">add cuda-kernels</a> 스킬과 통합하는 방법을 보여줍니다. 이 스킬을 사용하면 Claude나 Codex와 같은 에이전트가 특정 모델과 사용자의 하드웨어를 대상으로 하는 커스텀 커널을 작성할 수 있습니다.</p> <blockquote class="tip"><p>Install the <a href="https://github.com/huggingface/kernels/blob/main/kernel-builder/skills/cuda-kernels/SKILL.md" rel="nofollow">add cuda-kernels</a> skill to teach an agent how to write a kernel. The <a href="https://huggingface.co/blog/custom-cuda-kernels-agent-skills" rel="nofollow">Custom kernels for all from Codex and Claude</a> blog post covers this in more detail.</p></blockquote> <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><a href="https://docs.pytorch.org/ao/main/api_reference/generated/torchao.quantization.Int8DynamicActivationInt8WeightConfig.html" rel="nofollow">Dynamic quantization</a>은 정밀도를 낮추어 더 빠른 수학 연산을 가능하게 함으로써 추론 속도를 향상시킵니다. 이 특정 양자화 방식은 고정된 스케일링 계수를 사용하는 대신 런타임의 데이터에 기반하여 활성화(activation)의 스케일링 방식을 결정합니다. 그 결과, 스케일링 계수가 데이터에 더 정확하게 맞춰집니다.</p> <p>아래 예시에서는 <a href="../quantization/torchao">torchao</a> 라이브러리를 사용하여 UNet과 VAE에 <a href="https://docs.pytorch.org/ao/main/api_reference/generated/torchao.quantization.Int8DynamicActivationInt8WeightConfig.html" rel="nofollow">동적 int8 양자화</a>를 적용합니다.</p> <blockquote class="tip"><p>Diffusers의 torchao 통합 사용법에 대해 자세히 알아보려면 <a href="../quantization/torchao">torchao</a> 문서를 참조하세요.</p></blockquote> <p>최대 속도를 위해 컴파일러 태그를 구성합니다.</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> torchao <span class="hljs-keyword">import</span> apply_dynamic_quant
<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>
torch._inductor.config.force_fuse_int_mm_with_mul = <span class="hljs-literal">True</span>
torch._inductor.config.use_mixed_mm = <span class="hljs-literal">True</span><!----></pre></div><!----> <p>동적 양자화에서 이점을 얻지 못하는 UNet과 VAE의 일부 선형 레이어를 <a href="https://github.com/huggingface/diffusion-fast/blob/0f169640b1db106fe6a479f78c1ed3bfaeba3386/utils/pipeline_utils.py#L16" rel="nofollow">dynamic_quant_filter_fn</a>으로 필터링합니다.</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.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>

apply_dynamic_quant(pipeline.unet, dynamic_quant_filter_fn)
apply_dynamic_quant(pipeline.vae, dynamic_quant_filter_fn)

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><!----> <!--[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><a href="https://github.com/huggingface/diffusers/blob/58431f102cf39c3c8a569f32d71b2ea8caa461e1/src/diffusers/pipelines/pipeline_utils.py#L2034" rel="nofollow">fuse_qkv_projections</a> 메서드는 실험적이며 지원 범위가 주로 Stable Diffusion 파이프라인으로 제한됩니다. 다른 파이프라인에서 이를 활성화하는 방법에 대해서는 이 <a href="https://github.com/huggingface/diffusers/pull/6179" rel="nofollow">PR</a>을 참고하세요.</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="resources" 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="#resources"><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><!---->리소스<!----></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><code>torch.compile</code> 사용 시 성능을 극대화하려면 <a href="https://pytorch.org/blog/torch-compile-and-diffusers-a-hands-on-guide-to-peak-performance/" rel="nofollow">torch.compile 및 Diffusers: 피크 성능을 위한 실전 가이드</a> 블로그 글을 읽어보세요.</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_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, 337],
							data: [null,null],
							form: null,
							error: null
						});
					});
				}
			</script>
		
Xet Storage Details

Size:: 66.4 kB
Xet hash:: 5f00beeca49b93cf3d0b4e892209ba60cde765aed2ebe3d9973a80b8241a66fb
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.