Buckets:

hf-doc-build
/

doc

Files
xet
hf-doc-build/doc / diffusers /main /pt /quantization /nunchaku.html
mishig
about 2 hours ago
download
raw
31.9 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;Instale os kernels CUDA&quot;,&quot;local&quot;:&quot;install-the-cuda-kernels&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Carregar um pipeline quantizado&quot;,&quot;local&quot;:&quot;load-a-quantized-pipeline&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Tipos de quantização suportados&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;Kernels fundidos&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;Recursos&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.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/17.BtT_NXXy.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/chunks/Dk44y1gU.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 Lite é um backend de quantização para carregar checkpoints pré-quantizados em Diffusers. Crie checkpoints compatíveis com <a href="https://github.com/rootonchair/diffuse-compressor" rel="nofollow">diffuse-compressor</a>. Ele quantiza e exporta um transformer, em seguida, o empacota como um pipeline do Diffusers.</p> <p>Nunchaku Lite é construído sobre o motor de inferência original <a href="https://github.com/nunchaku-ai/nunchaku" rel="nofollow">Nunchaku</a>, a biblioteca de quantização <a href="https://github.com/nunchaku-ai/deepcompressor" rel="nofollow">DeepCompressor</a> e o <a href="https://arxiv.org/abs/2411.05007" rel="nofollow">artigo SVDQuant</a>.</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><!---->Instale os kernels CUDA<!----></span></h2><!--]--><!----> <p>O pacote kernels fornece os kernels CUDA otimizados, que são carregados automaticamente. Instale-o primeiro.</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> Copiado</div><!----></button><!----></div> <pre class="language-bash "><!---->pip install -U kernels<!----></pre></div><!----> <p>Nunchaku Lite carrega seus kernels do repositório <a href="https://huggingface.co/kernels/rootonchair/nunchaku-lite-kernels" rel="nofollow"><code>rootonchair/nunchaku-lite-kernels</code></a>, cujo publicador não é um publicador de kernels confiável no Hub. Carregá-lo baixa e executa código do Hub, portanto, o Diffusers exige que você opte explicitamente por isso definindo <code>DIFFUSERS_TRUST_REMOTE_KERNELS=true</code>. Veja <a href="../optimization/attention_backends#trusting-remote-kernels">Confiança em kernels remotos</a> para mais detalhes.</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><!---->Carregar um pipeline quantizado<!----></span></h2><!--]--><!----> <p>Carregue o pipeline pré-quantizado com <a href="/docs/diffusers/main/pt/api/pipelines/overview#diffusers.DiffusionPipeline.from_pretrained">from_pretrained()</a>, que lê a configuração de quantização de <code>config.json</code>.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5   text-gray-600  " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Copiado</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>O state dict exportado deve corresponder exatamente à arquitetura do modelo Diffusers de destino. Por exemplo, um checkpoint quantizado com projeções QKV fundidas não será carregado em uma configuração de modelo que espera módulos de projeção Q, K e V separados.</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><!---->Tipos de quantização suportados<!----></span></h2><!--]--><!----> <p>O Nunchaku Lite suporta os seguintes formatos de camadas lineares quantizadas.</p> <blockquote class="tip"><p>Use <code>nvfp4</code> em GPUs Blackwell. Executar checkpoints <code>int4</code> no Blackwell pode ser mais lento do que <code>nvfp4</code>.</p></blockquote> <p>Os kernels CUDA atualmente suportam as seguintes arquiteturas de GPU da NVIDIA:</p> <ul><li><code>sm_75</code> (Turing, por exemplo, RTX 2080)</li> <li><code>sm_80</code> (Ampere, por exemplo, A100)</li> <li><code>sm_86</code> (Ampere, por exemplo, RTX 3090 e RTX A6000)</li> <li><code>sm_89</code> (Ada, por exemplo, RTX 4090)</li> <li><code>sm_120</code> (Blackwell, por exemplo, RTX 5090)</li></ul> <blockquote class="note"><p>As GPUs Hopper, como <code>sm_90</code> H100 e H200, não são suportadas no momento.</p></blockquote> <p>Os checkpoints <code>nvfp4</code> exigem uma GPU NVIDIA Blackwell ou mais recente. Em GPUs Blackwell, use PyTorch &gt;= 2.7 com CUDA &gt;= 12.8. Os checkpoints <code>int4</code> exigem uma GPU NVIDIA Turing ou mais recente.</p> <table><thead><tr><th>Método</th><th align="right">Precisão</th><th align="right">Tamanho do grupo</th><th>Observações</th></tr></thead><tbody><tr><td><code>svdq_w4a4</code></td><td align="right"><code>nvfp4</code></td><td align="right">16</td><td>Usa kernels de runtime NVFP4 com correção de baixo rank SVDQ.</td></tr><tr><td><code>svdq_w4a4</code></td><td align="right"><code>int4</code></td><td align="right">64</td><td>Usa kernels INT4 W4A4 com correção de baixo posto SVDQ.</td></tr><tr><td><code>awq_w4a16</code></td><td align="right"><code>int4</code></td><td align="right">64</td><td>Usa kernels AWQ-style apenas para pesos em INT4.</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>O arquivo <code>config.json</code> deve incluir um <a href="/docs/diffusers/main/pt/api/quantization#diffusers.NunchakuLiteQuantizationConfig">NunchakuLiteQuantizationConfig</a>. Ele define o <code>compute_dtype</code> de runtime e os módulos de destino para cada método de quantização Nunchaku Lite.</p> <ul><li><code>compute_dtype</code>: dtype de runtime para buffers de ponto flutuante em módulos quantizados, tipicamente <code>torch.bfloat16</code>.</li> <li><code>svdq_w4a4</code>: configuração de destino SVDQ W4A4 com <code>precision</code>, <code>group_size</code>, <code>rank</code> e <code>targets</code>.</li> <li><code>awq_w4a16</code>: configuração de destino AWQ W4A16 com <code>precision</code>, <code>group_size</code> e <code>targets</code>.</li></ul> <p>Cada entrada em <code>targets</code> deve apontar para uma camada linear. O Diffusers substitui cada alvo <code>svdq_w4a4</code> por uma camada SVDQ W4A4 e cada alvo <code>awq_w4a16</code> por uma camada AWQ W4A16. O exemplo abaixo mostra a
forma esperada com listas de alvos encurtadas.</p> <p>Liste cada módulo que deseja quantizar sob <code>svdq_w4a4</code> ou <code>awq_w4a16</code>. Um módulo pode usar apenas um método, portanto, não liste o mesmo destino em ambos.</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> Copiado</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><!---->Kernels fundidos<!----></span></h2><!--]--><!----> <p>O motor original <a href="https://github.com/nunchaku-ai/nunchaku" rel="nofollow">Nunchaku</a> obtém grande parte de sua velocidade de caminhos de execução fundidos específicos do modelo. Ele combina as projeções Q, K e V com RMSNorm e RoPE, e usa um kernel GELU fundido para o MLP. O Nunchaku Lite, por outro lado, usa o modelo padrão do Diffusers com camadas lineares quantizadas genéricas, portanto, não inclui essas fusões.</p> <p>As medições a seguir mostram o impacto na latência por etapa dessas fusões no motor Nunchaku original, medido com Flux Schnell em uma RTX 5090.</p> <table><thead><tr><th>Otimização</th><th>Latência com ele DESLIGADO</th><th>Latência com ele ATIVADO</th><th>Aceleração</th></tr></thead><tbody><tr><td>Fusão QKV+norm+RoPE (combinada)</td><td>1.057 s</td><td>0.834 s</td><td><strong>1.27×</strong></td></tr><tr><td>— agrupando Q/K/V em um único matmul, isoladamente</td><td>1.059 s</td><td>1.034 s</td><td>1.025×</td></tr><tr><td>— fusão do epílogo RMSNorm+RoPE, além do agrupamento</td><td>1.034 s</td><td>0.836 s</td><td>1.236×</td></tr><tr><td>Fusão GELU-MLP</td><td>0.766 s</td><td>0.722 s</td><td><strong>1.06×</strong></td></tr></tbody></table> <p>Sem essas fusões, a GPU inicia as operações de projeção, normalização, incorporação rotativa e MLP como kernels separados, o que adiciona sobrecarga de lançamento. <code>torch.compile</code> pode reduzir parte dessa sobrecarga, embora não reproduza necessariamente os kernels fundidos do engine original. Realize o benchmark do pipeline compilado para o seu modelo e carga de trabalho.</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>Os kernels do Nunchaku Lite e as camadas lineares quantizadas são compatíveis com <a href="../optimization/fp16#torchcompile"><code>torch.compile</code></a>.
Compile o transformer quantizado após carregar o pipeline para uma inferência mais rápida.</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> Copiado</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>O pipeline Nunchaku Lite NVFP4 compilado executa 1,8x mais rápido que o pipeline BF16 original (2,271s → 1,675s em uma RTX PRO 6000).</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><!---->Recursos<!----></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">Requisitos de instalação do Nunchaku</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_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, 17],
							data: [null,null],
							form: null,
							error: null
						});
					});
				}
			</script>
		
Xet Storage Details

Size:: 31.9 kB
Xet hash:: d234a2b0de570e1dae55758f9cdb83e40e6d1d90e8ffca0774254f3e182605ef
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.