Hugging Face's logo

Buckets:
hf-doc-build
/
doc

Files
xet
hf-doc-build/doc / diffusers /main /pt /training /nemo_automodel.html
mishig's picture
mishig
download
raw
69.9 kB
<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;NeMo Automodel&quot;,&quot;local&quot;:&quot;nemo-automodel&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Modelos suportados&quot;,&quot;local&quot;:&quot;supported-models&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Instalação&quot;,&quot;local&quot;:&quot;installation&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Requisitos de hardware&quot;,&quot;local&quot;:&quot;hardware-requirements&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3}],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Preparação de dados&quot;,&quot;local&quot;:&quot;data-preparation&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Formato de saída&quot;,&quot;local&quot;:&quot;output-format&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3}],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Configuração de treinamento&quot;,&quot;local&quot;:&quot;training-configuration&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Referência de campos de configuração&quot;,&quot;local&quot;:&quot;config-field-reference&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3}],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Iniciar o treinamento&quot;,&quot;local&quot;:&quot;launch-training&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Geração&quot;,&quot;local&quot;:&quot;generation&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Integração com Diffusers&quot;,&quot;local&quot;:&quot;diffusers-integration&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Equipe NVIDIA&quot;,&quot;local&quot;:&quot;nvidia-team&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.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/375.boh_27Hd.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">
<!--1hvs45u--><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;NeMo Automodel&quot;,&quot;local&quot;:&quot;nemo-automodel&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Supported models&quot;,&quot;local&quot;:&quot;supported-models&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Installation&quot;,&quot;local&quot;:&quot;installation&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Hardware requirements&quot;,&quot;local&quot;:&quot;hardware-requirements&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3}],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Data preparation&quot;,&quot;local&quot;:&quot;data-preparation&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Output format&quot;,&quot;local&quot;:&quot;output-format&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3}],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Training configuration&quot;,&quot;local&quot;:&quot;training-configuration&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Config field reference&quot;,&quot;local&quot;:&quot;config-field-reference&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3}],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Launch training&quot;,&quot;local&quot;:&quot;launch-training&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Generation&quot;,&quot;local&quot;:&quot;generation&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Diffusers integration&quot;,&quot;local&quot;:&quot;diffusers-integration&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;NVIDIA Team&quot;,&quot;local&quot;:&quot;nvidia-team&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="nemo-automodel" 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="#nemo-automodel"><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><!---->NeMo Automodel<!----></span></h1><!--]--><!----> <p><a href="https://github.com/NVIDIA-NeMo/Automodel" rel="nofollow">NeMo Automodel</a> é uma biblioteca de treinamento nativa de PyTorch DTensor da NVIDIA para ajuste fino e pré-treino de modelos de difusão em escala. É nativa do Hugging Face — treine qualquer modelo no formato Diffusers do Hub sem conversão de checkpoints. A mesma receita YAML e o script de treinamento hackeável rodam em qualquer escala, de 1 GPU a centenas de nós, com treinamento distribuído <a href="https://pytorch.org/docs/stable/fsdp.html" rel="nofollow">FSDP2</a>, carregamento de dados em buckets multirresolução e treinamento em espaço latente pré-codificado para máxima utilização da GPU. Usa <a href="https://huggingface.co/papers/2210.02747" rel="nofollow">flow matching</a> para o treinamento e é totalmente open source (Apache 2.0), com suporte da NVIDIA e manutenção ativa.</p> <p>O NeMo Automodel se integra diretamente ao Diffusers. Ele carrega modelos pré-treinados do Hugging Face Hub usando as classes de modelo do Diffusers e gera saídas com o <a href="/docs/diffusers/main/pt/api/pipelines/overview#diffusers.DiffusionPipeline">DiffusionPipeline</a>.</p> <p>O fluxo de trabalho típico consiste em instalar o NeMo Automodel (pip ou Docker), preparar seus dados codificando-os em arquivos <code>.meta</code>, configurar uma receita YAML, iniciar o treinamento com <code>torchrun</code> e executar a inferência com o checkpoint resultante.</p> <!--[1--><h2 class="relative group"><a id="supported-models" 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-models"><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><!---->Modelos suportados<!----></span></h2><!--]--><!----> <table><thead><tr><th>Model</th><th>ID do Hugging Face</th><th>Tarefa</th><th>Parâmetros</th><th>Caso de uso</th></tr></thead><tbody><tr><td>Wan 2.1 T2V 1.3B</td><td><a href="https://huggingface.co/Wan-AI/Wan2.1-T2V-1.3B-Diffusers" rel="nofollow">Wan-AI/Wan2.1-T2V-1.3B-Diffusers</a></td><td>Texto para Vídeo</td><td>1.3B</td><td>geração de vídeo em hardware limitado (cabe em uma única A100 de 40GB)</td></tr><tr><td>FLUX.1-dev</td><td><a href="https://huggingface.co/black-forest-labs/FLUX.1-dev" rel="nofollow">black-forest-labs/FLUX.1-dev</a></td><td>Texto para Imagem</td><td>12B</td><td>geração de imagens de alta qualidade</td></tr><tr><td>HunyuanVideo 1.5</td><td><a href="https://huggingface.co/hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-720p_t2v" rel="nofollow">hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-720p_t2v</a></td><td>Texto para Vídeo</td><td>13B</td><td>geração de vídeo de alta qualidade</td></tr></tbody></table> <!--[1--><h2 class="relative group"><a id="installation" 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="#installation"><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><!---->Instalação<!----></span></h2><!--]--><!----> <!--[2--><h3 class="relative group"><a id="hardware-requirements" 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="#hardware-requirements"><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><!---->Requisitos de hardware<!----></span></h3><!--]--><!----> <table><thead><tr><th>Componente</th><th>Mínimo</th><th>Recomendado</th></tr></thead><tbody><tr><td>GPU</td><td>A100 40GB</td><td>A100 80GB / H100</td></tr><tr><td>GPUs</td><td>4</td><td>8+</td></tr><tr><td>RAM</td><td>128 GB</td><td>256 GB+</td></tr><tr><td>Armazenamento</td><td>SSD de 500 GB</td><td>2 TB NVMe</td></tr></tbody></table> <p>Instale o NeMo Automodel com pip. Para o conjunto completo de métodos de instalação (incluindo a partir do código-fonte), consulte o <a href="https://docs.nvidia.com/nemo/automodel/latest/guides/installation.html" rel="nofollow">guia de instalação do NeMo Automodel</a>.</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 "><!---->pip3 install nemo-automodel<!----></pre></div><!----> <p>Como alternativa, use o contêiner Docker pré-compilado, que inclui todas as dependências.</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 "><!---->docker pull nvcr.io/nvidia/nemo-automodel:26.02.00
docker run --gpus all -it --<span class="hljs-built_in">rm</span> --shm-size=8g nvcr.io/nvidia/nemo-automodel:26.02.00<!----></pre></div><!----> <blockquote class="warning"><p>Os checkpoints são perdidos quando o container é encerrado, a menos que você monte o diretório de checkpoints no host. Por exemplo, adicione <code>-v /host/path/checkpoints:/workspace/checkpoints</code> ao comando <code>docker run</code>.</p></blockquote> <!--[1--><h2 class="relative group"><a id="data-preparation" 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="#data-preparation"><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><!---->Preparação de dados<!----></span></h2><!--]--><!----> <p>NeMo Automodel treina modelos de difusão no espaço latente. Imagens ou vídeos brutos devem ser pré-processados em arquivos <code>.meta</code> contendo latentes de VAE e incorporações de texto antes do treinamento. Isso evita a re-codificação em cada etapa de treinamento.</p> <p>Use a ferramenta de pré-processamento integrada para codificar seus dados. A ferramenta distribui automaticamente o trabalho entre todas as GPUs disponíveis.</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">pré-processamento de vídeo</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">pré-processamento de imagem</div><!--]--></div> <div class="language-select"><div><p>O comando de pré-processamento de vídeo é o mesmo para o Wan 2.1 e o HunyuanVideo, mas as flags diferem. O Wan 2.1 usa <code>--processor wan</code> com <code>--resolution_preset</code> e <code>--caption_format sidecar</code>, enquanto o HunyuanVideo usa <code>--processor hunyuan</code> com <code>--target_frames</code> para definir a contagem de quadros e <code>--caption_format meta_json</code>.</p> <p><strong>Wan 2.1:</strong></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 "><!---->python -m tools.diffusion.preprocessing_multiprocess video \
--video_dir /data/videos \
--output_dir /cache \
--processor wan \
--resolution_preset 512p \
--caption_format sidecar<!----></pre></div><!----> <p><strong>HunyuanVideo:</strong></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 "><!---->python -m tools.diffusion.preprocessing_multiprocess video \
--video_dir /data/videos \
--output_dir /cache \
--processor hunyuan \
--target_frames 121 \
--caption_format meta_json<!----></pre></div><!----><!----></div><!----> <div hidden=""><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 "><!---->python -m tools.diffusion.preprocessing_multiprocess image \
--image_dir /data/images \
--output_dir /cache \
--processor flux \
--resolution_preset 512p<!----></pre></div><!----></div><!----><!----></div><!----> <!--[2--><h3 class="relative group"><a id="output-format" 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="#output-format"><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><!---->Formato de saída<!----></span></h3><!--]--><!----> <p>O pré-processamento gera um diretório de cache organizado por bucket de resolução. O NeMo Automodel suporta treinamento multi-resolução por meio de amostragem em buckets. As amostras são agrupadas por resolução espacial para que cada lote contenha amostras do mesmo tamanho, evitando desperdício por padding.</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=" "><!----><span class="hljs-string">/cache/</span>
├── 512x512/ <span class="hljs-comment"># Resolution bucket</span>
│ ├── &lt;hash1&gt;<span class="hljs-string">.meta</span> <span class="hljs-comment"># VAE latents + text embeddings</span>
│ ├── &lt;hash2&gt;<span class="hljs-string">.meta</span>
│ └── <span class="hljs-string">...</span>
├── 832x480/ <span class="hljs-comment"># Another resolution bucket</span>
│ └── <span class="hljs-string">...</span>
├── metadata.json <span class="hljs-comment"># Global config (processor, model, total items)</span>
└── metadata_shard_0000.json <span class="hljs-comment"># Per-sample metadata (paths, resolutions, captions)</span><!----></pre></div><!----> <blockquote class="tip"><p>Consulte o guia <a href="https://docs.nvidia.com/nemo/automodel/latest/datasets/diffusion-dataset" rel="nofollow">Preparação de Dataset para Difusão</a> para formatos de legenda, requisitos de dados de entrada e todos os argumentos de pré-processamento disponíveis.</p></blockquote> <!--[1--><h2 class="relative group"><a id="training-configuration" 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="#training-configuration"><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><!---->Configuração de treinamento<!----></span></h2><!--]--><!----> <p>O fine-tuning é impulsionado por dois componentes:</p> <ol><li>Um script de receita (<a href="https://github.com/NVIDIA-NeMo/Automodel/blob/main/examples/diffusion/finetune/finetune.py" rel="nofollow">finetune.py</a>) é um ponto de entrada em Python que contém o loop de treinamento: carregamento do modelo, construção do dataloader, execução das passadas de forward/backward, cálculo da perda de flow matching, checkpointing e registro de logs.</li> <li>Um arquivo de configuração YAML especifica todas as configurações usadas pela receita: qual modelo fazer o fine-tuning, onde estão os dados, hiperparâmetros do otimizador, estratégia de paralelismo e mais. Você personaliza o treinamento editando esse arquivo, em vez de modificar o código, permitindo escalar de 1 para centenas de GPUs.</li></ol> <p>Qualquer campo YAML também pode ser sobrescrito pela CLI:</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 "><!---->torchrun --nproc-per-node=8 examples/diffusion/finetune/finetune.py \
-c examples/diffusion/finetune/wan2_1_t2v_flow.yaml \
--optim.learning_rate 1e-5 \
--step_scheduler.num_epochs 50<!----></pre></div><!----> <p>Abaixo está a configuração anotada para o ajuste fino do Wan 2.1 T2V 1.3B, com cada seção explicada.</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-yaml "><!----><span class="hljs-attr">seed:</span> <span class="hljs-number">42</span>
<span class="hljs-comment"># ── Experiment tracking (optional) ──────────────────────────────────────────</span>
<span class="hljs-comment"># Weights &amp; Biases integration for logging metrics, losses, and learning rates.</span>
<span class="hljs-comment"># Set mode: &quot;disabled&quot; to turn off.</span>
<span class="hljs-attr">wandb:</span>
<span class="hljs-attr">project:</span> <span class="hljs-string">wan-t2v-flow-matching</span>
<span class="hljs-attr">mode:</span> <span class="hljs-string">online</span>
<span class="hljs-attr">name:</span> <span class="hljs-string">wan2_1_t2v_fm</span>
<span class="hljs-comment"># ── Model ───────────────────────────────────────────────────────────────────</span>
<span class="hljs-comment"># pretrained_model_name_or_path: any Hugging Face model ID or local path.</span>
<span class="hljs-comment"># mode: &quot;finetune&quot; loads pretrained weights; &quot;pretrain&quot; trains from scratch.</span>
<span class="hljs-attr">model:</span>
<span class="hljs-attr">pretrained_model_name_or_path:</span> <span class="hljs-string">Wan-AI/Wan2.1-T2V-1.3B-Diffusers</span>
<span class="hljs-attr">mode:</span> <span class="hljs-string">finetune</span>
<span class="hljs-comment"># ── Training schedule ───────────────────────────────────────────────────────</span>
<span class="hljs-comment"># global_batch_size: effective batch across all GPUs.</span>
<span class="hljs-comment"># Gradient accumulation is computed automatically: global / (local × num_gpus).</span>
<span class="hljs-attr">step_scheduler:</span>
<span class="hljs-attr">global_batch_size:</span> <span class="hljs-number">8</span>
<span class="hljs-attr">local_batch_size:</span> <span class="hljs-number">1</span>
<span class="hljs-attr">ckpt_every_steps:</span> <span class="hljs-number">1000</span> <span class="hljs-comment"># Save a checkpoint every N steps</span>
<span class="hljs-attr">num_epochs:</span> <span class="hljs-number">100</span>
<span class="hljs-attr">log_every:</span> <span class="hljs-number">2</span> <span class="hljs-comment"># Log metrics every N steps</span>
<span class="hljs-comment"># ── Data ────────────────────────────────────────────────────────────────────</span>
<span class="hljs-comment"># _target_: the dataloader factory function.</span>
<span class="hljs-comment"># Use build_video_multiresolution_dataloader for video models (Wan, HunyuanVideo).</span>
<span class="hljs-comment"># Use build_text_to_image_multiresolution_dataloader for image models (FLUX).</span>
<span class="hljs-comment"># model_type: &quot;wan&quot; or &quot;hunyuan&quot; (selects the correct latent format).</span>
<span class="hljs-comment"># base_resolution: target resolution for multiresolution bucketing.</span>
<span class="hljs-attr">data:</span>
<span class="hljs-attr">dataloader:</span>
<span class="hljs-attr">_target_:</span> <span class="hljs-string">nemo_automodel.components.datasets.diffusion.build_video_multiresolution_dataloader</span>
<span class="hljs-attr">cache_dir:</span> <span class="hljs-string">PATH_TO_YOUR_DATA</span>
<span class="hljs-attr">model_type:</span> <span class="hljs-string">wan</span>
<span class="hljs-attr">base_resolution:</span> [<span class="hljs-number">512</span>, <span class="hljs-number">512</span>]
<span class="hljs-attr">dynamic_batch_size:</span> <span class="hljs-literal">false</span> <span class="hljs-comment"># When true, adjusts batch per bucket to maintain constant memory</span>
<span class="hljs-attr">shuffle:</span> <span class="hljs-literal">true</span>
<span class="hljs-attr">drop_last:</span> <span class="hljs-literal">false</span>
<span class="hljs-attr">num_workers:</span> <span class="hljs-number">0</span>
<span class="hljs-comment"># ── Optimizer ───────────────────────────────────────────────────────────────</span>
<span class="hljs-comment"># learning_rate: 5e-6 is a good starting point for fine-tuning.</span>
<span class="hljs-comment"># Adjust weight_decay and betas for your dataset.</span>
<span class="hljs-attr">optim:</span>
<span class="hljs-attr">learning_rate:</span> <span class="hljs-number">5e-6</span>
<span class="hljs-attr">optimizer:</span>
<span class="hljs-attr">weight_decay:</span> <span class="hljs-number">0.01</span>
<span class="hljs-attr">betas:</span> [<span class="hljs-number">0.9</span>, <span class="hljs-number">0.999</span>]
<span class="hljs-comment"># ── Learning rate scheduler ─────────────────────────────────────────────────</span>
<span class="hljs-comment"># Supports cosine, linear, and constant schedules.</span>
<span class="hljs-attr">lr_scheduler:</span>
<span class="hljs-attr">lr_decay_style:</span> <span class="hljs-string">cosine</span>
<span class="hljs-attr">lr_warmup_steps:</span> <span class="hljs-number">0</span>
<span class="hljs-attr">min_lr:</span> <span class="hljs-number">1e-6</span>
<span class="hljs-comment"># ── Flow matching ───────────────────────────────────────────────────────────</span>
<span class="hljs-comment"># adapter_type: model-specific adapter — must match the model:</span>
<span class="hljs-comment"># &quot;simple&quot; for Wan 2.1, &quot;flux&quot; for FLUX.1-dev, &quot;hunyuan&quot; for HunyuanVideo.</span>
<span class="hljs-comment"># timestep_sampling: &quot;uniform&quot; for Wan, &quot;logit_normal&quot; for FLUX and HunyuanVideo.</span>
<span class="hljs-comment"># flow_shift: shifts the flow schedule (model-dependent).</span>
<span class="hljs-comment"># i2v_prob: probability of image-to-video conditioning during training (video models).</span>
<span class="hljs-attr">flow_matching:</span>
<span class="hljs-attr">adapter_type:</span> <span class="hljs-string">&quot;simple&quot;</span>
<span class="hljs-attr">adapter_kwargs:</span> {}
<span class="hljs-attr">timestep_sampling:</span> <span class="hljs-string">&quot;uniform&quot;</span>
<span class="hljs-attr">logit_mean:</span> <span class="hljs-number">0.0</span>
<span class="hljs-attr">logit_std:</span> <span class="hljs-number">1.0</span>
<span class="hljs-attr">flow_shift:</span> <span class="hljs-number">3.0</span>
<span class="hljs-attr">num_train_timesteps:</span> <span class="hljs-number">1000</span>
<span class="hljs-attr">i2v_prob:</span> <span class="hljs-number">0.3</span>
<span class="hljs-attr">use_loss_weighting:</span> <span class="hljs-literal">true</span>
<span class="hljs-comment"># ── FSDP2 distributed training ──────────────────────────────────────────────</span>
<span class="hljs-comment"># dp_size: number of GPUs for data parallelism (typically = total GPUs on node).</span>
<span class="hljs-comment"># tp_size, cp_size, pp_size: tensor, context, and pipeline parallelism.</span>
<span class="hljs-comment"># For most fine-tuning, dp_size is all you need; leave others at 1.</span>
<span class="hljs-attr">fsdp:</span>
<span class="hljs-attr">tp_size:</span> <span class="hljs-number">1</span>
<span class="hljs-attr">cp_size:</span> <span class="hljs-number">1</span>
<span class="hljs-attr">pp_size:</span> <span class="hljs-number">1</span>
<span class="hljs-attr">dp_replicate_size:</span> <span class="hljs-number">1</span>
<span class="hljs-attr">dp_size:</span> <span class="hljs-number">8</span>
<span class="hljs-comment"># ── Checkpointing ──────────────────────────────────────────────────────────</span>
<span class="hljs-comment"># checkpoint_dir: where to save checkpoints (use a persistent path with Docker).</span>
<span class="hljs-comment"># restore_from: path to resume training from a previous checkpoint.</span>
<span class="hljs-attr">checkpoint:</span>
<span class="hljs-attr">enabled:</span> <span class="hljs-literal">true</span>
<span class="hljs-attr">checkpoint_dir:</span> <span class="hljs-string">PATH_TO_YOUR_CKPT_DIR</span>
<span class="hljs-attr">model_save_format:</span> <span class="hljs-string">torch_save</span>
<span class="hljs-attr">save_consolidated:</span> <span class="hljs-literal">false</span>
<span class="hljs-attr">restore_from:</span> <span class="hljs-literal">null</span><!----></pre></div><!----> <!--[2--><h3 class="relative group"><a id="config-field-reference" 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="#config-field-reference"><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><!---->Referência de campos de configuração<!----></span></h3><!--]--><!----> <p>A tabela abaixo lista as configurações mínimas necessárias. Consulte os <a href="https://github.com/NVIDIA-NeMo/Automodel/tree/main/examples/diffusion/finetune" rel="nofollow">exemplos do NeMo Automodel</a> para obter configurações de exemplo completas para todos os modelos.</p> <table><thead><tr><th>Seção</th><th>Obrigatório?</th><th>O que alterar</th></tr></thead><tbody><tr><td><code>model</code></td><td>Sim</td><td>Defina <code>pretrained_model_name_or_path</code> como o ID do modelo no Hugging Face. Defina <code>mode: finetune</code> ou <code>mode: pretrain</code>.</td></tr><tr><td><code>step_scheduler</code></td><td>Sim</td><td><code>global_batch_size</code> é o tamanho efetivo do lote em todas as GPUs. <code>ckpt_every_steps</code> controla a frequência dos checkpoints. A acumulação de gradientes é calculada automaticamente.</td></tr><tr><td><code>data</code></td><td>Sim</td><td>Defina <code>cache_dir</code> como o caminho que contém seus arquivos <code>.meta</code> pré-processados. Altere <code>_target_</code> e <code>model_type</code> para diferentes modelos.</td></tr><tr><td><code>optim</code></td><td>Sim</td><td><code>learning_rate: 5e-6</code> é um valor padrão adequado para ajuste fino. Ajuste conforme seu conjunto de dados e modelo.</td></tr><tr><td><code>lr_scheduler</code></td><td>Sim</td><td>Escolha <code>cosine</code>, <code>linear</code> ou <code>constant</code> para <code>lr_decay_style</code>. Defina <code>lr_warmup_steps</code> para um aquecimento gradual.</td></tr><tr><td><code>flow_matching</code></td><td>Sim</td><td><code>adapter_type</code> deve corresponder ao modelo (<code>simple</code> para Wan, <code>flux</code> para FLUX, <code>hunyuan</code> para HunyuanVideo). Consulte as configurações específicas do modelo para <code>adapter_kwargs</code>.</td></tr><tr><td><code>fsdp</code></td><td>Sim</td><td>Defina <code>dp_size</code> para o número de GPUs. Para multi-nó, defina para o total de GPUs em todos os nós.</td></tr><tr><td><code>checkpoint</code></td><td>Recomendado</td><td>Defina <code>checkpoint_dir</code> para um caminho persistente, especialmente em Docker. Use <code>restore_from</code> para retomar de um checkpoint anterior.</td></tr><tr><td><code>wandb</code></td><td>Opcional</td><td>Configure para habilitar o rastreamento de experimentos do Weights &amp; Biases. Defina <code>mode: disabled</code> para desativar.</td></tr></tbody></table> <!--[1--><h2 class="relative group"><a id="launch-training" 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="#launch-training"><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><!---->Iniciar o treinamento<!----></span></h2><!--]--><!----> <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">nó único</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">multi-nó</div><!--]--></div> <div class="language-select"><div><div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Copiado</div><!----></button><!----></div> <pre class="language-bash "><!---->torchrun --nproc-per-node=8 \
examples/diffusion/finetune/finetune.py \
-c examples/diffusion/finetune/wan2_1_t2v_flow.yaml<!----></pre></div><!----></div><!----> <div hidden=""><p>Execute o seguinte em cada nó, definindo <code>NODE_RANK</code> de acordo:</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 "><!----><span class="hljs-built_in">export</span> MASTER_ADDR=node0.hostname
<span class="hljs-built_in">export</span> MASTER_PORT=29500
<span class="hljs-built_in">export</span> NODE_RANK=0 <span class="hljs-comment"># 0 on master, 1 on second node, etc.</span>
torchrun \
--nnodes=2 \
--nproc-per-node=8 \
--node_rank=<span class="hljs-variable">${NODE_RANK}</span> \
--rdzv_backend=c10d \
--rdzv_endpoint=<span class="hljs-variable">${MASTER_ADDR}</span>:<span class="hljs-variable">${MASTER_PORT}</span> \
examples/diffusion/finetune/finetune.py \
-c examples/diffusion/finetune/wan2_1_t2v_flow_multinode.yaml<!----></pre></div><!----> <blockquote class="note"><p>Para treinamento em múltiplos nós, defina <code>fsdp.dp_size</code> no YAML com o número <strong>total</strong> de GPUs em todos os nós (por exemplo, 16 para 2 nós com 8 GPUs cada).</p></blockquote><!----></div><!----><!----></div><!----> <!--[1--><h2 class="relative group"><a id="generation" 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="#generation"><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><!---->Geração<!----></span></h2><!--]--><!----> <p>Após o treinamento, gere vídeos ou imagens a partir de prompts de texto usando o checkpoint ajustado.</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">Wan 2.1</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">FLUX</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">HunyuanVideo</div><!--]--></div> <div class="language-select"><div><div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Copiado</div><!----></button><!----></div> <pre class="language-bash "><!---->python examples/diffusion/generate/generate.py \
-c examples/diffusion/generate/configs/generate_wan.yaml<!----></pre></div><!----> <p>Com um checkpoint ajustado finamente:</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 "><!---->python examples/diffusion/generate/generate.py \
-c examples/diffusion/generate/configs/generate_wan.yaml \
--model.checkpoint ./checkpoints/step_1000 \
--inference.prompts <span class="hljs-string">&#x27;[&quot;A dog running on a beach&quot;]&#x27;</span><!----></pre></div><!----><!----></div><!----> <div hidden=""><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 "><!---->python examples/diffusion/generate/generate.py \
-c examples/diffusion/generate/configs/generate_flux.yaml<!----></pre></div><!----> <p>Com um checkpoint ajustado finamente:</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 "><!---->python examples/diffusion/generate/generate.py \
-c examples/diffusion/generate/configs/generate_flux.yaml \
--model.checkpoint ./checkpoints/step_1000 \
--inference.prompts <span class="hljs-string">&#x27;[&quot;A dog running on a beach&quot;]&#x27;</span><!----></pre></div><!----><!----></div><!----> <div hidden=""><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 "><!---->python examples/diffusion/generate/generate.py \
-c examples/diffusion/generate/configs/generate_hunyuan.yaml<!----></pre></div><!----> <p>Com um checkpoint ajustado finamente:</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 "><!---->python examples/diffusion/generate/generate.py \
-c examples/diffusion/generate/configs/generate_hunyuan.yaml \
--model.checkpoint ./checkpoints/step_1000 \
--inference.prompts <span class="hljs-string">&#x27;[&quot;A dog running on a beach&quot;]&#x27;</span><!----></pre></div><!----><!----></div><!----><!----></div><!----> <!--[1--><h2 class="relative group"><a id="diffusers-integration" 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="#diffusers-integration"><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><!---->Integração com Diffusers<!----></span></h2><!--]--><!----> <p>O NeMo Automodel é construído sobre o Diffusers e o utiliza como base para o carregamento de modelos e inferência. Ele carrega modelos diretamente do Hugging Face Hub usando classes de modelos do Diffusers, como <a href="/docs/diffusers/main/pt/api/models/wan_transformer_3d#diffusers.WanTransformer3DModel">WanTransformer3DModel</a>, <a href="/docs/diffusers/main/pt/api/models/flux_transformer#diffusers.FluxTransformer2DModel">FluxTransformer2DModel</a> e <a href="/docs/diffusers/main/pt/api/models/hunyuan_video_transformer_3d#diffusers.HunyuanVideoTransformer3DModel">HunyuanVideoTransformer3DModel</a>, e gera saídas por meio de pipelines do Diffusers, como <a href="/docs/diffusers/main/pt/api/pipelines/wan#diffusers.WanPipeline">WanPipeline</a> e <a href="/docs/diffusers/main/pt/api/pipelines/flux#diffusers.FluxPipeline">FluxPipeline</a>.</p> <p>Esta integração oferece vários benefícios para os usuários do Diffusers:</p> <ul><li><strong>Na conversão de checkpoint</strong>: os pesos pré-treinados do Hub funcionam imediatamente. Aponte <code>pretrained_model_name_or_path</code> para qualquer ID de modelo no formato Diffusers e comece a treinar na hora.</li> <li><strong>Suporte a modelos no dia zero</strong>: quando um novo modelo de difusão é adicionado ao Diffusers e enviado ao Hub, ele pode ser ajustado com o NeMo Automodel sem esperar por um script de treinamento dedicado.</li> <li><strong>Saídas compatíveis com Pipeline</strong>: os checkpoints de fine-tuning são salvos em um formato que pode ser carregado diretamente de volta nos pipelines do Diffusers para inferência, compartilhamento no Hub ou otimização adicional com ferramentas como quantização e compilação.</li> <li><strong>Escalonável treinamento para modelos Diffusers</strong>: NeMo Automodel adiciona capacidades de treinamento distribuído (FSDP2, multi-nó, bucketing multirresolução) que vão além do que os scripts de treinamento integrados do Diffusers oferecem, mantendo as mesmas interfaces de modelo e pipeline.</li> <li><strong>Ecossistema compartilhado</strong>: qualquer modelo, adaptador LoRA ou componente de pipeline do ecossistema Diffusers permanece compatível em todo o fluxo de trabalho de treinamento e inferência.</li></ul> <!--[1--><h2 class="relative group"><a id="nvidia-team" 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="#nvidia-team"><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><!---->Equipe NVIDIA<!----></span></h2><!--]--><!----> <ul><li>Pranav Prashant Thombre, <a href="mailto:pthombre@nvidia.com">pthombre@nvidia.com</a></li> <li>Linnan Wang, <a href="mailto:linnanw@nvidia.com">linnanw@nvidia.com</a></li> <li>Alexandros Koumparoulis, <a href="mailto:akoumparouli@nvidia.com">akoumparouli@nvidia.com</a></li></ul> <!--[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/NVIDIA-NeMo/Automodel" rel="nofollow">NeMo Automodel GitHub</a></li> <li><a href="https://docs.nvidia.com/nemo/automodel/latest/recipes-e2e-examples/diffusion-fine-tuning" rel="nofollow">Guia de Ajuste Fino de Difusão</a></li> <li><a href="https://docs.nvidia.com/nemo/automodel/latest/datasets/diffusion-dataset" rel="nofollow">Preparação do Conjunto de Dados de Difusão</a></li> <li><a href="https://docs.nvidia.com/nemo/automodel/latest/model-coverage/diffusion.html" rel="nofollow">Cobertura de Modelos de Difusão</a></li> <li><a href="https://huggingface.co/docs/transformers/en/community_integrations/nemo_automodel_finetuning" rel="nofollow">NeMo Automodel para Transformers (fine-tuning de LLM/VLM)</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/training/nemo_automodel.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, 375],
data: [null,null],
form: null,
error: null
});
});
}
</script>

Xet Storage Details

Size:
69.9 kB
·
Xet hash:
1ceab7c825019d4ef3b6c14ac8caa4d1a4db0da1ee37a14fbc0fc5cd194446ff

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