Buckets:

hf-doc-build
/

doc

Files
xet
hf-doc-build/doc / diffusers /main /ko /quantization /nunchaku.html
mishig
about 18 hours ago
download
raw
32.4 kB
<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;Nunchaku Lite&quot;,&quot;local&quot;:&quot;nunchaku-lite&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;CUDA 커널 설치&quot;,&quot;local&quot;:&quot;install-the-cuda-kernels&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;양자화 파이프라인 로드&quot;,&quot;local&quot;:&quot;load-a-quantized-pipeline&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;지원되는 양자화 유형&quot;,&quot;local&quot;:&quot;supported-quantization-types&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;NunchakuLiteQuantizationConfig&quot;,&quot;local&quot;:&quot;nunchakulitequantizationconfig&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Fused 커널&quot;,&quot;local&quot;:&quot;fused-kernels&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;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.vZc-dOUR.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/chunks/Bafuik5c.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.B6c3Ml05.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.BQlx092V.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/347.BNLwZtJM.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/chunks/B4AMoTAu.js" rel="modulepreload">
		<!--p4df5g--><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;Nunchaku Lite&quot;,&quot;local&quot;:&quot;nunchaku-lite&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Install the CUDA kernels&quot;,&quot;local&quot;:&quot;install-the-cuda-kernels&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Load a quantized pipeline&quot;,&quot;local&quot;:&quot;load-a-quantized-pipeline&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Supported quantization types&quot;,&quot;local&quot;:&quot;supported-quantization-types&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;NunchakuLiteQuantizationConfig&quot;,&quot;local&quot;:&quot;nunchakulitequantizationconfig&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Fused kernels&quot;,&quot;local&quot;:&quot;fused-kernels&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;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="nunchaku-lite" 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="#nunchaku-lite"><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><!---->Nunchaku Lite<!----></span></h1><!--]--><!----> <p>Nunchaku Lite는 Diffusers에서 사전 양자화 체크포인트를 로드하기 위한 양자화 백엔드입니다. <a href="https://github.com/rootonchair/diffuse-compressor" rel="nofollow">diffuse-compressor</a>를 사용하여 호환되는 체크포인트를 생성합니다. 이 도구는 transformer를 양자화하고 내보낸 후, 이를 Diffusers 파이프라인으로 패키징합니다.</p> <p>Nunchaku Lite는 원래의 <a href="https://github.com/nunchaku-ai/nunchaku" rel="nofollow">Nunchaku</a> 추론 엔진, <a href="https://github.com/nunchaku-ai/deepcompressor" rel="nofollow">DeepCompressor</a> 양자화 라이브러리, <a href="https://arxiv.org/abs/2411.05007" rel="nofollow">SVDQuant 논문</a>을 기반으로 구축되었습니다.</p> <!--[1--><h2 class="relative group"><a id="install-the-cuda-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="#install-the-cuda-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><!---->CUDA 커널 설치<!----></span></h2><!--]--><!----> <p>kernels 패키지는 최적화된 CUDA 커널을 제공하며, 자동으로 로드됩니다. 먼저 설치하세요.</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 kernels<!----></pre></div><!----> <p>Nunchaku Lite는 <a href="https://huggingface.co/rootonchair/nunchaku-lite-kernels" rel="nofollow"><code>rootonchair/nunchaku-lite-kernels</code></a> 저장소에서 커널을 로드합니다. 해당 저장소의 게시자는 Hub에서 신뢰할 수 있는 커널 게시자가 아닙니다. 이를 로드하면 Hub에서 코드를 다운로드하여 실행하므로, Diffusers는 <code>DIFFUSERS_TRUST_REMOTE_KERNELS=true</code>를 설정하여 명시적으로 동의하도록 요구합니다. 자세한 내용은 <a href="../optimization/attention_backends#trusting-remote-kernels">원격 커널 신뢰</a>를 참조하세요.</p> <!--[1--><h2 class="relative group"><a id="load-a-quantized-pipeline" 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="#load-a-quantized-pipeline"><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><!---->양자화 파이프라인 로드<!----></span></h2><!--]--><!----> <p><code>config.json</code>에서 양자화 설정을 읽는 <a href="/docs/diffusers/main/ko/api/pipelines/overview#diffusers.DiffusionPipeline.from_pretrained">from_pretrained()</a>로 사전 양자화 파이프라인을 로드합니다.</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-python "><!----><span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> DiffusionPipeline

model_id = <span class="hljs-string">&quot;rootonchair/ERNIE-Image-Turbo-nunchaku-lite-nvfp4&quot;</span>

pipe = DiffusionPipeline.from_pretrained(
    model_id, dtype=torch.bfloat16,
).to(<span class="hljs-string">&quot;cuda&quot;</span>)  <span class="hljs-comment"># or &quot;mps&quot;, &quot;xpu&quot;, &quot;cpu&quot;</span>

prompt = <span class="hljs-string">&quot;A modern red armchair in a quiet studio, soft window light, realistic product photography&quot;</span>
image = pipe(
    prompt=prompt,
    height=<span class="hljs-number">1024</span>,
    width=<span class="hljs-number">1024</span>,
    num_inference_steps=<span class="hljs-number">8</span>,
    guidance_scale=<span class="hljs-number">1.0</span>,
).images[<span class="hljs-number">0</span>]
image.save(<span class="hljs-string">&quot;ernie-image-turbo-nunchaku-lite.png&quot;</span>)<!----></pre></div><!----> <blockquote class="note"><p>내보낸 state dict는 대상 Diffusers 모델 아키텍처와 정확히 일치해야 합니다. 예를 들어, fused QKV projections으로 양자화된 체크포인트는 개별 Q, K, V projection 모듈을 기대하는 모델 구성에 로드되지 않습니다.</p></blockquote> <!--[1--><h2 class="relative group"><a id="supported-quantization-types" 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="#supported-quantization-types"><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><!---->지원되는 양자화 유형<!----></span></h2><!--]--><!----> <p>Nunchaku Lite는 다음 양자화 선형 레이어 형식을 지원합니다.</p> <blockquote class="tip"><p>Blackwell GPU에서 <code>nvfp4</code>를 사용하세요. Blackwell에서 <code>int4</code> 체크포인트를 실행하면 <code>nvfp4</code>보다 느릴 수 있습니다.</p></blockquote> <p>CUDA 커널은 현재 다음 NVIDIA GPU 아키텍처를 지원합니다:</p> <ul><li><code>sm_75</code> (예: RTX 2080을 위한 Turing)</li> <li><code>sm_80</code> (예: A100을 위한 Ampere)</li> <li><code>sm_86</code> (Ampere, 예: RTX 3090 및 RTX A6000)</li> <li><code>sm_89</code> (예: RTX 4090의 Ada)</li> <li><code>sm_120</code> (예: RTX 5090 Blackwell)</li></ul> <blockquote class="note"><p>H100, H200와 같은 <code>sm_90</code> Hopper GPU는 현재 지원되지 않습니다.</p></blockquote> <p><code>nvfp4</code> 체크포인트는 Blackwell 또는 그 이상의 NVIDIA GPU가 필요합니다. Blackwell GPU를 사용하는 경우 PyTorch &gt;= 2.7과 CUDA &gt;= 12.8을 사용하세요. <code>int4</code> 체크포인트는 Turing 또는 그 이상의 NVIDIA GPU가 필요합니다.</p> <table><thead><tr><th>메서드</th><th align="right">정밀도</th><th align="right">그룹 크기</th><th>The input is a single word "Notes".

참고</th></tr></thead><tbody><tr><td><code>svdq_w4a4</code></td><td align="right"><code>nvfp4</code></td><td align="right">16</td><td>SVDQ 저순위 보정을 사용하는 NVFP4 런타임 커널을 사용합니다.</td></tr><tr><td><code>svdq_w4a4</code></td><td align="right"><code>int4</code></td><td align="right">64</td><td>SVDQ 저순위 보정을 사용하는 INT4 W4A4 커널을 사용합니다.</td></tr><tr><td><code>awq_w4a16</code></td><td align="right"><code>int4</code></td><td align="right">64</td><td>INT4 weight-only AWQ 스타일 커널을 사용합니다.</td></tr></tbody></table> <!--[1--><h2 class="relative group"><a id="nunchakulitequantizationconfig" 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="#nunchakulitequantizationconfig"><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><!---->NunchakuLiteQuantizationConfig<!----></span></h2><!--]--><!----> <p><code>config.json</code> 파일에는 <a href="/docs/diffusers/main/ko/api/quantization#diffusers.NunchakuLiteQuantizationConfig">NunchakuLiteQuantizationConfig</a>가 포함되어야 합니다. 이는 런타임 <code>compute_dtype</code>와 각 Nunchaku Lite 양자화 방식에 대한 대상 모듈을 정의합니다.</p> <ul><li><code>compute_dtype</code>: 양자화 모듈의 부동 소수점 버퍼에 사용되는 런타임 dtype으로, 일반적으로 <code>torch.bfloat16</code>입니다.</li> <li><code>svdq_w4a4</code>: <code>precision</code>, <code>group_size</code>, <code>rank</code>, <code>targets</code>를 포함한 SVDQ W4A4 대상 구성입니다.</li> <li><code>awq_w4a16</code>: <code>precision</code>, <code>group_size</code>, <code>targets</code>를 포함한 AWQ W4A16 대상 구성입니다.</li></ul> <p><code>targets</code>의 각 항목은 선형 레이어를 가리켜야 합니다. Diffusers는 각 <code>svdq_w4a4</code> 타깃을 SVDQ W4A4 레이어로, 각 <code>awq_w4a16</code> 타깃을 AWQ W4A16 레이어로 교체합니다. 아래 예제는 줄인 타깃 목록과 함께 예상되는 형태를 보여줍니다.</p> <p><code>svdq_w4a4</code> 또는 <code>awq_w4a16</code> 아래에 양자화하려는 각 모듈을 나열하세요. 모듈은 하나의 방법만 사용할 수 있으므로, 동일한 대상 모듈을 두 곳 모두에 나열하지 마세요.</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-json "><!----><span class="hljs-punctuation">{</span>
  <span class="hljs-attr">&quot;_class_name&quot;</span><span class="hljs-punctuation">:</span> <span class="hljs-string">&quot;ErnieImageTransformer2DModel&quot;</span><span class="hljs-punctuation">,</span>
  <span class="hljs-attr">&quot;quantization_config&quot;</span><span class="hljs-punctuation">:</span> <span class="hljs-punctuation">{</span>
    <span class="hljs-attr">&quot;quant_method&quot;</span><span class="hljs-punctuation">:</span> <span class="hljs-string">&quot;nunchaku_lite&quot;</span><span class="hljs-punctuation">,</span>
    <span class="hljs-attr">&quot;compute_dtype&quot;</span><span class="hljs-punctuation">:</span> <span class="hljs-string">&quot;bfloat16&quot;</span><span class="hljs-punctuation">,</span>
    <span class="hljs-attr">&quot;svdq_w4a4&quot;</span><span class="hljs-punctuation">:</span> <span class="hljs-punctuation">{</span>
      <span class="hljs-attr">&quot;precision&quot;</span><span class="hljs-punctuation">:</span> <span class="hljs-string">&quot;nvfp4&quot;</span><span class="hljs-punctuation">,</span>
      <span class="hljs-attr">&quot;group_size&quot;</span><span class="hljs-punctuation">:</span> <span class="hljs-number">16</span><span class="hljs-punctuation">,</span>
      <span class="hljs-attr">&quot;rank&quot;</span><span class="hljs-punctuation">:</span> <span class="hljs-number">32</span><span class="hljs-punctuation">,</span>
      <span class="hljs-attr">&quot;targets&quot;</span><span class="hljs-punctuation">:</span> <span class="hljs-punctuation">[</span><span class="hljs-string">&quot;layers.0.self_attention.to_q&quot;</span><span class="hljs-punctuation">]</span>
    <span class="hljs-punctuation">}</span><span class="hljs-punctuation">,</span>
    <span class="hljs-attr">&quot;awq_w4a16&quot;</span><span class="hljs-punctuation">:</span> <span class="hljs-punctuation">{</span>
      <span class="hljs-attr">&quot;precision&quot;</span><span class="hljs-punctuation">:</span> <span class="hljs-string">&quot;int4&quot;</span><span class="hljs-punctuation">,</span>
      <span class="hljs-attr">&quot;group_size&quot;</span><span class="hljs-punctuation">:</span> <span class="hljs-number">64</span><span class="hljs-punctuation">,</span>
      <span class="hljs-attr">&quot;targets&quot;</span><span class="hljs-punctuation">:</span> <span class="hljs-punctuation">[</span><span class="hljs-string">&quot;final_linear&quot;</span><span class="hljs-punctuation">]</span>
    <span class="hljs-punctuation">}</span>
  <span class="hljs-punctuation">}</span>
<span class="hljs-punctuation">}</span><!----></pre></div><!----> <!--[1--><h2 class="relative group"><a id="fused-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="#fused-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><!---->Fused 커널<!----></span></h2><!--]--><!----> <p>원본 <a href="https://github.com/nunchaku-ai/nunchaku" rel="nofollow">Nunchaku</a> 엔진은 모델별 융합 실행 경로에서 상당한 속도를 얻습니다. Q, K, V 투영을 RMSNorm과 RoPE와 결합하고, MLP에는 융합된 GELU 커널을 사용합니다. 반면 Nunchaku Lite는 표준 Diffusers 모델과 범용 양자화 선형 레이어를 사용하므로 이러한 융합을 포함하지 않습니다.</p> <p>다음 측정값은 Flux Schnell을 RTX 5090에서 사용하여 원래 Nunchaku 엔진에서 이러한 융합의 단계별 지연 시간 영향을 측정한 결과입니다.</p> <table><thead><tr><th>최적화</th><th>OFF 상태의 지연 시간</th><th>ON 상태에서의 지연 시간</th><th>속도 향상</th></tr></thead><tbody><tr><td>QKV+norm+RoPE 융합(통합)</td><td>1.057 s</td><td>0.834 s</td><td><strong>1.27×</strong></td></tr><tr><td>— Q/K/V를 하나의 matmul로 그룹화, 단독으로</td><td>1.059 s</td><td>1.034 s</td><td>1.025×</td></tr><tr><td>— 그룹핑 위에 RMSNorm+RoPE 에필로그 퓨전</td><td>1.034 s</td><td>0.836 s</td><td>1.236×</td></tr><tr><td>GELU-MLP 융합</td><td>0.766 s</td><td>0.722 s</td><td><strong>1.06×</strong></td></tr></tbody></table> <p>이러한 퓨전이 없으면 GPU는 projection, normalization, rotary-embedding, MLP 연산을 개별 커널로 실행하여 런치 오버헤드가 발생합니다. <code>torch.compile</code>은 이러한 오버헤드를 일부 줄일 수 있지만, 원래 엔진의 퓨전 커널을 반드시 재현하지는 않습니다. 해당 모델과 워크로드에 대해 컴파일된 파이프라인의 벤치마크를 실행해 보세요.</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>Nunchaku Lite 커널과 양자화 선형 레이어는 <a href="../optimization/fp16#torchcompile"><code>torch.compile</code></a>과 호환됩니다.
더 빠른 추론을 위해 파이프라인을 로드한 후 양자화된 트랜스포머를 컴파일하세요.</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-python "><!---->pipe.transformer = torch.<span class="hljs-built_in">compile</span>(pipe.transformer, mode=<span class="hljs-string">&quot;default&quot;</span>, fullgraph=<span class="hljs-literal">True</span>)<!----></pre></div><!----> <p>컴파일된 Nunchaku Lite NVFP4 파이프라인은 원래의 BF16 파이프라인보다 1.8배 빠릅니다(RTX PRO 6000에서 2.271초 → 1.675초).</p> <!--[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><a href="https://github.com/rootonchair/diffuse-compressor" rel="nofollow">diffuse-compressor</a></li> <li><a href="https://nunchaku.tech/docs/nunchaku/installation/installation.html" rel="nofollow">Nunchaku 설치 요구 사항</a></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/quantization/nunchaku.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_1rjuniz = {
						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.vZc-dOUR.js"),
						import("/docs/diffusers/main/en/_app/immutable/entry/app.B6c3Ml05.js")
					]).then(([kit, app]) => {
						kit.start(app, element, {
							node_ids: [0, 347],
							data: [null,null],
							form: null,
							error: null
						});
					});
				}
			</script>
		
Xet Storage Details

Size:: 32.4 kB
Xet hash:: 7fbdf0e7e47cfe43aa01525a8ca8ab9975643fd9908f1ad32e8fc08e251f157d
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.