Hugging Face's logo

Buckets:
hf-doc-build
/
doc

Files
xet
hf-doc-build/doc / diffusers /main /es /training /nemo_automodel.html
mishig's picture
mishig
download
raw
70.3 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 compatibles&quot;,&quot;local&quot;:&quot;supported-models&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Instalación&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;Preparación de datos&quot;,&quot;local&quot;:&quot;data-preparation&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Formato de salida&quot;,&quot;local&quot;:&quot;output-format&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3}],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Configuración de entrenamiento&quot;,&quot;local&quot;:&quot;training-configuration&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Referencia de campos de configuración&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 el entrenamiento&quot;,&quot;local&quot;:&quot;launch-training&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Generación&quot;,&quot;local&quot;:&quot;generation&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Integración de Diffusers&quot;,&quot;local&quot;:&quot;diffusers-integration&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Equipo de 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> es una biblioteca de entrenamiento nativa de PyTorch DTensor de NVIDIA para el ajuste fino y el preentrenamiento de modelos de difusión a gran escala. Es nativa de Hugging Face: entrena cualquier modelo en formato Diffusers del Hub sin necesidad de convertir puntos de control. La misma receta YAML y el mismo script de entrenamiento modificable se ejecutan a cualquier escala, desde 1 GPU hasta cientos de nodos, con entrenamiento distribuido <a href="https://pytorch.org/docs/stable/fsdp.html" rel="nofollow">FSDP2</a>, carga de datos por cubos multiresolución y entrenamiento en el espacio latente precodificado para una máxima utilización de la GPU. Utiliza <a href="https://huggingface.co/papers/2210.02747" rel="nofollow">flow matching</a> para el entrenamiento y es completamente de código abierto (Apache 2.0), con soporte de NVIDIA y mantenimiento activo.</p> <p>NeMo Automodel se integra directamente con Diffusers. Carga modelos preentrenados desde Hugging Face Hub utilizando las clases de modelos de Diffusers y genera salidas con <a href="/docs/diffusers/main/es/api/pipelines/overview#diffusers.DiffusionPipeline">DiffusionPipeline</a>.</p> <p>El flujo de trabajo típico consiste en instalar NeMo Automodel (pip o Docker), preparar los datos codificándolos en archivos <code>.meta</code>, configurar una receta YAML, iniciar el entrenamiento con <code>torchrun</code> y ejecutar la inferencia con el punto de control 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 compatibles<!----></span></h2><!--]--><!----> <table><thead><tr><th>Modelo</th><th>ID de Hugging Face</th><th>Tarea</th><th>Parámetros</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 a video</td><td>1.3B</td><td>generación de video en hardware limitado (cabe en una sola A100 de 40 GB)</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 a imagen</td><td>12B</td><td>generación de imágenes de alta calidad</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 a video</td><td>13B</td><td>generación de video de alta calidad</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><!---->Instalación<!----></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>Almacenamiento</td><td>SSD de 500 GB</td><td>2 TB NVMe</td></tr></tbody></table> <p>Instale NeMo Automodel con pip. Para el conjunto completo de métodos de instalación (incluido desde el código fuente), consulte la <a href="https://docs.nvidia.com/nemo/automodel/latest/guides/installation.html" rel="nofollow">guía de instalación de 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 el contenedor de Docker precompilado que incluye todas las dependencias.</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>Los puntos de control se pierden al salir del contenedor, a menos que se monte el directorio de puntos de control en el host mediante un bind mount. Por ejemplo, añade <code>-v /host/path/checkpoints:/workspace/checkpoints</code> al 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><!---->Preparación de datos<!----></span></h2><!--]--><!----> <p>NeMo Automodel entrena modelos de difusión en el espacio latente. Las imágenes o videos en bruto deben preprocesarse en archivos <code>.meta</code> que contienen latentes de VAE y embeddings de texto antes del entrenamiento. Esto evita re-codificar en cada paso de entrenamiento.</p> <p>Utilice la herramienta de preprocesamiento integrada para codificar sus datos. La herramienta distribuye automáticamente el trabajo entre todas las GPU disponibles.</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">preprocesamiento de video</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">preprocesamiento de imágenes</div><!--]--></div> <div class="language-select"><div><p>El comando de preprocesamiento de video es el mismo para Wan 2.1 y HunyuanVideo, pero las banderas difieren. Wan 2.1 utiliza <code>--processor wan</code> con <code>--resolution_preset</code> y <code>--caption_format sidecar</code>, mientras que HunyuanVideo utiliza <code>--processor hunyuan</code> con <code>--target_frames</code> para establecer el número de fotogramas y <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 salida<!----></span></h3><!--]--><!----> <p>El preprocesamiento genera un directorio de caché organizado por cubo de resolución. NeMo Automodel admite el entrenamiento multi-resolución mediante muestreo por cubos. Las muestras se agrupan por resolución espacial para que cada lote contenga muestras del mismo tamaño, evitando el desperdicio por relleno.</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 la guía de <a href="https://docs.nvidia.com/nemo/automodel/latest/datasets/diffusion-dataset" rel="nofollow">Preparación de conjuntos de datos de difusión</a> para conocer los formatos de descripción, los requisitos de los datos de entrada y todos los argumentos de preprocesamiento disponibles.</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><!---->Configuración de entrenamiento<!----></span></h2><!--]--><!----> <p>El ajuste fino se basa en dos componentes:</p> <ol><li>Un script de receta (<a href="https://github.com/NVIDIA-NeMo/Automodel/blob/main/examples/diffusion/finetune/finetune.py" rel="nofollow">finetune.py</a>) es un punto de entrada en Python que contiene el bucle de entrenamiento: carga del modelo, construcción del dataloader, ejecución de las pasadas hacia adelante y hacia atrás, cálculo de la pérdida de flow matching, creación de checkpoints y registro de logs.</li> <li>Un archivo de configuración YAML especifica todas las configuraciones que utiliza la receta: qué modelo se debe ajustar, dónde se encuentra los datos, los hiperparámetros del optimizador, la estrategia de paralelismo y más. Personaliza el entrenamiento editando este archivo en lugar de modificar el código, lo que te permite escalar desde 1 hasta cientos de GPUs.</li></ol> <p>Cualquier campo YAML también se puede sobrescribir desde la 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>A continuación se muestra la configuración anotada para el ajuste fino de Wan 2.1 T2V 1.3B, con cada sección 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><!---->Referencia de campos de configuración<!----></span></h3><!--]--><!----> <p>La tabla siguiente enumera las configuraciones mínimas requeridas. Consulte los <a href="https://github.com/NVIDIA-NeMo/Automodel/tree/main/examples/diffusion/finetune" rel="nofollow">ejemplos de NeMo Automodel</a> para ver configuraciones de ejemplo completas para todos los modelos.</p> <table><thead><tr><th>Sección</th><th>¿Obligatorio?</th><th>Qué cambiar</th></tr></thead><tbody><tr><td><code>model</code></td><td>Sí</td><td>Establezca <code>pretrained_model_name_or_path</code> en el ID del modelo de Hugging Face. Establezca <code>mode: finetune</code> o <code>mode: pretrain</code>.</td></tr><tr><td><code>step_scheduler</code></td><td>Sí</td><td><code>global_batch_size</code> es el tamaño de lote efectivo en todas las GPU. <code>ckpt_every_steps</code> controla la frecuencia de los puntos de control. La acumulación de gradientes se calcula automáticamente.</td></tr><tr><td><code>data</code></td><td>Sí</td><td>Establezca <code>cache_dir</code> en la ruta que contiene sus archivos <code>.meta</code> preprocesados. Cambie <code>_target_</code> y <code>model_type</code> para diferentes modelos.</td></tr><tr><td><code>optim</code></td><td>Sí</td><td><code>learning_rate: 5e-6</code> es un valor predeterminado adecuado para el ajuste fino. Ajusta este valor según tu conjunto de datos y modelo.</td></tr><tr><td><code>lr_scheduler</code></td><td>Sí</td><td>Elija <code>cosine</code>, <code>linear</code> o <code>constant</code> para <code>lr_decay_style</code>. Establezca <code>lr_warmup_steps</code> para un calentamiento gradual.</td></tr><tr><td><code>flow_matching</code></td><td>Sí</td><td><code>adapter_type</code> debe coincidir con el modelo (<code>simple</code> para Wan, <code>flux</code> para FLUX, <code>hunyuan</code> para HunyuanVideo). Consulte las configuraciones específicas del modelo para <code>adapter_kwargs</code>.</td></tr><tr><td><code>fsdp</code></td><td>Sí</td><td>Establezca <code>dp_size</code> en el número de GPUs. Para varios nodos, establézcalo en el número total de GPUs de todos los nodos.</td></tr><tr><td><code>checkpoint</code></td><td>Recomendado</td><td>Establezca <code>checkpoint_dir</code> en una ruta persistente, especialmente en Docker. Use <code>restore_from</code> para reanudar desde un punto de control anterior.</td></tr><tr><td><code>wandb</code></td><td>Opcional</td><td>Configure para habilitar el seguimiento de experimentos de Weights &amp; Biases. Establezca <code>mode: disabled</code> para desactivarlo.</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 el entrenamiento<!----></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">nodo 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-nodo</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>Ejecute lo siguiente en cada nodo, estableciendo <code>NODE_RANK</code> en consecuencia:</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 el entrenamiento multinodo, establezca <code>fsdp.dp_size</code> en el YAML con el número <strong>total</strong> de GPUs en todos los nodos (por ejemplo, 16 para 2 nodos con 8 GPUs cada uno).</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><!---->Generación<!----></span></h2><!--]--><!----> <p>Después del entrenamiento, genera videos o imágenes a partir de prompts de texto utilizando el punto de control 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>Con un punto de control ajustado:</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>Con un punto de control ajustado:</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>Con un punto de control ajustado:</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><!---->Integración de Diffusers<!----></span></h2><!--]--><!----> <p>NeMo Automodel se construye sobre Diffusers y lo utiliza como base para la carga de modelos y la inferencia. Carga los modelos directamente desde Hugging Face Hub mediante clases de modelos de Diffusers como <a href="/docs/diffusers/main/es/api/models/wan_transformer_3d#diffusers.WanTransformer3DModel">WanTransformer3DModel</a>, <a href="/docs/diffusers/main/es/api/models/flux_transformer#diffusers.FluxTransformer2DModel">FluxTransformer2DModel</a> y <a href="/docs/diffusers/main/es/api/models/hunyuan_video_transformer_3d#diffusers.HunyuanVideoTransformer3DModel">HunyuanVideoTransformer3DModel</a>, y genera salidas a través de pipelines de Diffusers como <a href="/docs/diffusers/main/es/api/pipelines/wan#diffusers.WanPipeline">WanPipeline</a> y <a href="/docs/diffusers/main/es/api/pipelines/flux#diffusers.FluxPipeline">FluxPipeline</a>.</p> <p>Esta integración ofrece varios beneficios para los usuarios de Diffusers:</p> <ul><li><strong>Sin conversión de puntos de control</strong>: los pesos preentrenados del Hub funcionan directamente. Apunta <code>pretrained_model_name_or_path</code> a cualquier ID de modelo en formato Diffusers y comienza a entrenar de inmediato.</li> <li><strong>Soporte de modelos desde el día 0</strong>: cuando se añade un nuevo modelo de difusión a Diffusers y se sube al Hub, se puede ajustar con NeMo Automodel sin esperar a un guion de entrenamiento dedicado.</li> <li><strong>Salidas compatibles con Pipeline</strong>: los puntos de control ajustados se guardan en un formato que puede cargarse directamente en las pipelines de Diffusers para inferencia, para compartir en el Hub o para una optimización adicional con herramientas como la cuantización y la compilación.</li> <li><strong>Entrenamiento escalable para modelos de Diffusers</strong>: NeMo Automodel añade capacidades de entrenamiento distribuido (FSDP2, multinodo, bucketing multiresolución) que van más allá de lo que ofrecen los scripts de entrenamiento integrados de Diffusers, manteniendo las mismas interfaces de modelo y pipeline.</li> <li><strong>Ecosistema compartido</strong>: cualquier modelo, adaptador LoRA o componente de pipeline del ecosistema de Diffusers permanece compatible durante todo el flujo de trabajo de entrenamiento e inferencia.</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><!---->Equipo de 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">Guía de ajuste fino de difusión</a></li> <li><a href="https://docs.nvidia.com/nemo/automodel/latest/datasets/diffusion-dataset" rel="nofollow">Preparación del conjunto de datos de difusión</a></li> <li><a href="https://docs.nvidia.com/nemo/automodel/latest/model-coverage/diffusion.html" rel="nofollow">Cobertura de modelos de difusión</a></li> <li><a href="https://huggingface.co/docs/transformers/en/community_integrations/nemo_automodel_finetuning" rel="nofollow">NeMo Automodel para Transformers (ajuste fino 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:
70.3 kB
·
Xet hash:
752808fbb108b9407df5bf06a35b920154b2e161916ec7c8617a9d8049531400

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