Buckets:

hf-doc-build
/

doc

Files
xet
hf-doc-build/doc / diffusers /v0.43.0 /es /optimization /context_parallelism.html
mishig
2 days ago
download
raw
40.6 kB
<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;Paralelismo de contexto&quot;,&quot;local&quot;:&quot;context-parallelism&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Ring Attention&quot;,&quot;local&quot;:&quot;ring-attention&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Ulysses Attention&quot;,&quot;local&quot;:&quot;ulysses-attention&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Atención unificada&quot;,&quot;local&quot;:&quot;unified-attention&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Ulysses Anything Attention&quot;,&quot;local&quot;:&quot;ulysses-anything-attention&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Ring Anything Attention&quot;,&quot;local&quot;:&quot;ring-anything-attention&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Activar el paralelismo de contexto durante la carga&quot;,&quot;local&quot;:&quot;enable-context-parallelism-while-loading&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/333.OuTuM9lH.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/chunks/DSPDl1gz.js" rel="modulepreload">
		<!--ty2c99--><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;Context parallelism&quot;,&quot;local&quot;:&quot;context-parallelism&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Ring Attention&quot;,&quot;local&quot;:&quot;ring-attention&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Ulysses Attention&quot;,&quot;local&quot;:&quot;ulysses-attention&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Unified Attention&quot;,&quot;local&quot;:&quot;unified-attention&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Ulysses Anything Attention&quot;,&quot;local&quot;:&quot;ulysses-anything-attention&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Ring Anything Attention&quot;,&quot;local&quot;:&quot;ring-anything-attention&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Enable context parallelism while loading&quot;,&quot;local&quot;:&quot;enable-context-parallelism-while-loading&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="context-parallelism" 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="#context-parallelism"><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><!---->Paralelismo de contexto<!----></span></h1><!--]--><!----> <p><a href="https://huggingface.co/spaces/nanotron/ultrascale-playbook?section=context_parallelism" rel="nofollow">El paralelismo de contexto</a> divide la secuencia de entrada entre las GPU y cada GPU procesa su propia porción. Esto reduce la memoria por GPU, lo cual resulta más útil para secuencias largas, como imágenes de alta resolución y vídeo. Para compararlo con las demás estrategias, consulte <a href="../training/distributed_inference">Inferencia distribuida</a>.</p> <p>El paralelismo de contexto se ejecuta dentro del backend de atención, y solo algunos backends lo admiten. Si el suyo no lo hace, <code>enable_parallelism</code> lanza un error que enumera los backends compatibles. Cambie de backend con <a href="/docs/diffusers/v0.43.0/es/api/models/overview#diffusers.ModelMixin.set_attention_backend">set_attention_backend()</a> y consulte los <a href="./attention_backends#available-backends">backends disponibles</a> para ver la lista completa.</p> <p>Configure cada variante de abajo con <a href="/docs/diffusers/v0.43.0/es/api/parallel#diffusers.ContextParallelConfig">ContextParallelConfig</a>.</p> <table><thead><tr><th>Variante</th><th>Config</th><th>Requisitos</th><th>Uso</th></tr></thead><tbody><tr><td><a href="#ring-attention">Ring Attention</a></td><td><code>ring_degree > 1</code></td><td>la longitud de la secuencia debe ser divisible exactamente por <code>ring_degree</code></td><td>las secuencias son largas y el ancho de banda de interconexión es limitado</td></tr><tr><td><a href="#ulysses-attention">Atención Ulysses</a></td><td><code>ulysses_degree > 1</code></td><td>la longitud de la secuencia y el número de cabezas de atención deben ser divisibles exactamente por <code>ulysses_degree</code></td><td>sus GPUs tienen un interconexión rápida</td></tr><tr><td><a href="#unified-attention">Atención unificada</a></td><td><code>ring_degree > 1</code> y <code>ulysses_degree > 1</code></td><td>al menos 4 GPUs</td><td>tienes más dispositivos de los que Ulysses Attention puede usar por sí solo</td></tr><tr><td><a href="#ulysses-anything-attention">Ulysses Anything Attention</a></td><td><code>ulysses_degree > 1</code>, <code>ulysses_anything=True</code></td><td><code>ring_degree</code> debe ser 1</td><td>la longitud de la secuencia o el número de cabezas no se divide de forma uniforme</td></tr><tr><td><a href="#ring-anything-attention">Ring Anything Attention</a></td><td><code>ring_degree > 1</code>, <code>ring_anything=True</code></td><td><code>ulysses_degree</code> debe ser 1, solo inferencia, <code>attn_mask</code> debe ser <code>None</code></td><td>la longitud de la secuencia no se divide exactamente</td></tr></tbody></table> <!--[1--><h2 class="relative group"><a id="ring-attention" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#ring-attention"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span><!---->Ring Attention<!----></span></h2><!--]--><!----> <p><a href="https://huggingface.co/papers/2310.01889" rel="nofollow">Ring Attention</a> transmite los bloques de claves (K) y valores (V) entre las GPU dispuestas en un anillo. Cada GPU calcula la atención contra el bloque K/V que posee y luego pasa ese bloque a la GPU siguiente. Después de un ciclo completo, cada GPU ha visto los K/V de cada token y ninguna GPU almacena la secuencia completa.</p> <p>Establezca <code>ring_degree</code> en el número de GPUs y pase la configuración al método <code>enable_parallelism()</code> del transformer.</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-py "><!----><span class="hljs-keyword">import</span> os
<span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> torch <span class="hljs-keyword">import</span> distributed <span class="hljs-keyword">as</span> dist
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> DiffusionPipeline, ContextParallelConfig

<span class="hljs-keyword">def</span> <span class="hljs-title function_">setup_distributed</span>():
    <span class="hljs-keyword">if</span> <span class="hljs-keyword">not</span> dist.is_initialized():
        dist.init_process_group(backend=<span class="hljs-string">&quot;nccl&quot;</span>)
    <span class="hljs-comment"># torchrun sets LOCAL_RANK to this process&#x27;s GPU index on its node</span>
    local_rank = <span class="hljs-built_in">int</span>(os.environ[<span class="hljs-string">&quot;LOCAL_RANK&quot;</span>])
    device = torch.device(<span class="hljs-string">f&quot;cuda:<span class="hljs-subst">{local_rank}</span>&quot;</span>)
    torch.cuda.set_device(device)
    <span class="hljs-keyword">return</span> device

<span class="hljs-keyword">def</span> <span class="hljs-title function_">main</span>():
    device = setup_distributed()
    world_size = dist.get_world_size()

    pipeline = DiffusionPipeline.from_pretrained(
        <span class="hljs-string">&quot;black-forest-labs/FLUX.1-dev&quot;</span>, dtype=torch.bfloat16
    ).to(device)
    pipeline.transformer.set_attention_backend(<span class="hljs-string">&quot;_native_cudnn&quot;</span>)

    cp_config = ContextParallelConfig(ring_degree=world_size)
    pipeline.transformer.enable_parallelism(config=cp_config)

    prompt = <span class="hljs-string">&quot;&quot;&quot;
    cinematic film still of a cat sipping a margarita in a pool in Palm Springs, California
    highly detailed, high budget hollywood movie, cinemascope, moody, epic, gorgeous, film grain
    &quot;&quot;&quot;</span>

    <span class="hljs-comment"># Must specify generator so all ranks start with same latents (or pass your own)</span>
    generator = torch.Generator().manual_seed(<span class="hljs-number">42</span>)
    image = pipeline(
        prompt,
        guidance_scale=<span class="hljs-number">3.5</span>,
        num_inference_steps=<span class="hljs-number">50</span>,
        generator=generator,
    ).images[<span class="hljs-number">0</span>]

    <span class="hljs-keyword">if</span> dist.get_rank() == <span class="hljs-number">0</span>:
        image.save(<span class="hljs-string">&quot;output.png&quot;</span>)

    <span class="hljs-keyword">if</span> dist.is_initialized():
        dist.destroy_process_group()


<span class="hljs-keyword">if</span> __name__ == <span class="hljs-string">&quot;__main__&quot;</span>:
    main()<!----></pre></div><!----> <p>Guarde el script como <code>ring_attention.py</code> y ejecútelo con <a href="https://docs.pytorch.org/docs/stable/elastic/run.html" rel="nofollow">torchrun</a> o cualquier otro lanzador compatible con PyTorch. Establezca <code>--nproc-per-node</code> en el número de GPUs.</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 2 ring_attention.py<!----></pre></div><!----> <p>Para probar las otras variantes, reemplace la llamada a <code>enable_parallelism</code> en <code>ring_attention.py</code> con la de su sección.</p> <!--[1--><h2 class="relative group"><a id="ulysses-attention" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#ulysses-attention"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span><!---->Ulysses Attention<!----></span></h2><!--]--><!----> <p><a href="https://huggingface.co/papers/2309.14509" rel="nofollow">Ulysses Attention</a> utiliza comunicación <em>all-to-all</em>, donde cada GPU intercambia datos con todas las demás GPUs. El primer all-to-all reagrupa los datos para que cada GPU contenga todos los tokens de un subconjunto de cabezas de atención. Cada GPU calcula la atención para sus cabezas, y un segundo all-to-all reagrupa los resultados por token para la siguiente capa.</p> <p>Establezca <code>ulysses_degree</code> en el número de GPUs y pase la configuración a <code>enable_parallelism()</code>.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5   text-gray-600  " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Copiado</div><!----></button><!----></div> <pre class="language-py "><!---->pipeline.transformer.enable_parallelism(config=ContextParallelConfig(ulysses_degree=<span class="hljs-number">2</span>))<!----></pre></div><!----> <!--[1--><h2 class="relative group"><a id="unified-attention" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#unified-attention"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span><!---->Atención unificada<!----></span></h2><!--]--><!----> <p><a href="https://huggingface.co/papers/2405.07719" rel="nofollow">Unified Sequence Parallelism</a> combina Ring Attention y Ulysses Attention. Ejecuta un all-to-all para redistribuir las cabezas y los tokens, ejecuta Ring Attention sobre el resultado y luego invierte el all-to-all para restaurar el diseño original. Ulysses paraleliza entre las cabezas de atención y Ring paraleliza a lo largo de la secuencia, por lo que, juntos, pueden utilizar más GPUs que Ulysses por sí solo.</p> <p>Use Unified Attention cuando tenga al menos 4 GPUs. Establezca tanto <code>ulysses_degree</code> como <code>ring_degree</code> en un valor mayor que 1. Las GPUs forman una cuadrícula 2D de <code>ulysses_degree * ring_degree</code> dispositivos, y cada GPU pertenece a un grupo Ulysses y a un grupo Ring.</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-py "><!---->pipeline.transformer.enable_parallelism(config=ContextParallelConfig(ulysses_degree=<span class="hljs-number">2</span>, ring_degree=<span class="hljs-number">2</span>))<!----></pre></div><!----> <p>El siguiente benchmark compara Ulysses, Ring y Unified Attention en 4 GPUs H100 con <a href="https://github.com/huggingface/diffusers/pull/12693#issuecomment-3694727532" rel="nofollow">este script</a>. Cada imagen utiliza 50 pasos de inferencia.</p> <table><thead><tr><th>Backend de CP</th><th>Tiempo / Imagen (ms)</th><th>Pasos / Seg</th><th>Memoria máxima (GB)</th></tr></thead><tbody><tr><td>ulysses</td><td>6670.789</td><td>7.50</td><td>33.85</td></tr><tr><td>ring</td><td>13076.492</td><td>3.82</td><td>56.02</td></tr><tr><td>unified_balanced</td><td>11068.705</td><td>4.52</td><td>33.85</td></tr></tbody></table> <p>Ulysses tiene el mayor rendimiento, pero no puede usar más GPUs que las cabezas de atención que tiene el modelo. Unified Attention elimina ese límite.</p> <!--[1--><h2 class="relative group"><a id="ulysses-anything-attention" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#ulysses-anything-attention"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span><!---->Ulysses Anything Attention<!----></span></h2><!--]--><!----> <p>Ulysses Attention requiere que la longitud de la secuencia y el número de cabezas de atención sean divisibles exactamente por <code>ulysses_degree</code>. <a href="https://github.com/huggingface/diffusers/pull/12996" rel="nofollow">Ulysses Anything Attention</a> elimina ambos requisitos. No se puede combinar con Unified Attention, por lo que <code>ring_degree</code> debe ser 1.</p> <p>Establezca <code>ulysses_degree</code> en un valor mayor que 1 y <code>ulysses_anything=True</code>.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5   text-gray-600  " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Copiado</div><!----></button><!----></div> <pre class="language-py "><!---->pipeline.transformer.enable_parallelism(config=ContextParallelConfig(ulysses_degree=<span class="hljs-number">2</span>, ulysses_anything=<span class="hljs-literal">True</span>))<!----></pre></div><!----> <blockquote class="tip"><p>Agregue un backend <code>gloo</code> para tensores de CPU a <a href="https://docs.pytorch.org/docs/stable/distributed.html#torch.distributed.init_process_group" rel="nofollow">init_process_group</a>. Esto evita las sincronizaciones forzadas de CUDA que provocan las transferencias de host a dispositivo y de dispositivo a host.</p></blockquote> <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-py "><!----><span class="hljs-keyword">import</span> torch.distributed <span class="hljs-keyword">as</span> dist

dist.init_process_group(backend=<span class="hljs-string">&quot;cpu:gloo,cuda:nccl&quot;</span>)<!----></pre></div><!----> <p>El benchmark siguiente ejecuta FLUX.1-dev en 4 GPUs L20 con <a href="https://github.com/huggingface/diffusers/pull/12996#issuecomment-3797695999" rel="nofollow">este script</a>.</p> <table><thead><tr><th>Backend de CP</th><th>Tiempo / Iteración (ms)</th><th>Pasos / Seg</th><th>Memoria máxima (GB)</th><th>Forma (HxW)</th></tr></thead><tbody><tr><td>ulysses</td><td>281.07</td><td>3.56</td><td>37.11</td><td>1024x1024</td></tr><tr><td>ring</td><td>351.34</td><td>2.85</td><td>37.01</td><td>1024x1024</td></tr><tr><td>unified_balanced</td><td>324.37</td><td>3.08</td><td>37.16</td><td>1024x1024</td></tr><tr><td>ulysses_anything</td><td>280.94</td><td>3.56</td><td>37.11</td><td>1024x1024</td></tr><tr><td>ulysses</td><td>fallido</td><td>fallido</td><td>fallido</td><td>1008x1008</td></tr><tr><td>ring</td><td>fallido</td><td>fallido</td><td>fallido</td><td>1008x1008</td></tr><tr><td>unified_balanced</td><td>fallido</td><td>fallido</td><td>fallido</td><td>1008x1008</td></tr><tr><td>ulysses_anything</td><td>278.40</td><td>3.59</td><td>36.99</td><td>1008x1008</td></tr></tbody></table> <p>En 1008x1008, donde la longitud de la secuencia no se divide exactamente, solo se ejecuta Ulysses Anything Attention. En 1024x1024, su rendimiento es el mismo que el de Ulysses Attention estándar.</p> <!--[1--><h2 class="relative group"><a id="ring-anything-attention" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#ring-anything-attention"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span><!---->Ring Anything Attention<!----></span></h2><!--]--><!----> <p>Ring Attention requiere que la longitud de la secuencia sea divisible exactamente por <code>ring_degree</code>. <a href="https://github.com/huggingface/diffusers/pull/13545#issuecomment-4302195582" rel="nofollow">Ring Anything Attention</a> admite cualquier longitud de secuencia. Rellena los K/V locales de cada GPU hasta la longitud local más larga, realiza un all-gather de los K/V rellenos y recorta cada bloque a su longitud real antes de calcular la atención.</p> <p>No se puede combinar con Unified Attention, por lo que <code>ulysses_degree</code> debe ser 1. Establezca <code>ring_degree</code> en un valor mayor que 1 y <code>ring_anything=True</code>.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5   text-gray-600  " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Copiado</div><!----></button><!----></div> <pre class="language-py "><!---->pipeline.transformer.enable_parallelism(config=ContextParallelConfig(ring_degree=<span class="hljs-number">2</span>, ring_anything=<span class="hljs-literal">True</span>))<!----></pre></div><!----> <p>Ring Anything Attention también se beneficia del backend <code>gloo</code> mostrado en <a href="#ulysses-anything-attention">Ulysses Anything Attention</a>.</p> <blockquote class="note"><p>Ring Anything Attention solo admite inferencia y <code>attn_mask</code> debe ser <code>None</code>.</p></blockquote> <p>El benchmark siguiente ejecuta FLUX.1-dev en 4 GPUs RTX 4090 (48GB).</p> <table><thead><tr><th>Backend de CP</th><th>Tiempo / Iteración (ms)</th><th>Pasos / Seg</th><th>Memoria máxima (GB)</th><th>Forma (HxW)</th></tr></thead><tbody><tr><td>ulysses</td><td>259.07</td><td>3.86</td><td>33.83</td><td>1024x1024</td></tr><tr><td>ring</td><td>338.98</td><td>2.95</td><td>33.83</td><td>1024x1024</td></tr><tr><td>unified_balanced</td><td>321.54</td><td>3.11</td><td>33.83</td><td>1024x1024</td></tr><tr><td>ulysses_anything</td><td>259.07</td><td>3.86</td><td>33.83</td><td>1024x1024</td></tr><tr><td>ring_anything</td><td>340.14</td><td>2.94</td><td>33.83</td><td>1024x1024</td></tr><tr><td>ulysses</td><td>fallido</td><td>fallido</td><td>fallido</td><td>1008x1008</td></tr><tr><td>ring</td><td>fallido</td><td>fallido</td><td>fallido</td><td>1008x1008</td></tr><tr><td>unified_balanced</td><td>fallido</td><td>fallido</td><td>fallido</td><td>1008x1008</td></tr><tr><td>ulysses_anything</td><td>253.16</td><td>3.95</td><td>33.75</td><td>1008x1008</td></tr><tr><td>ring_anything</td><td>335.57</td><td>2.98</td><td>33.75</td><td>1008x1008</td></tr></tbody></table> <p>Ring Anything Attention se ejecuta a 1008x1008 y rinde aproximadamente igual que el Ring Attention estándar. Para conocer la motivación y las compensaciones, consulte <a href="https://github.com/huggingface/diffusers/pull/13545#issuecomment-4304104462" rel="nofollow">este comentario</a>.</p> <!--[1--><h2 class="relative group"><a id="enable-context-parallelism-while-loading" 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="#enable-context-parallelism-while-loading"><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><!---->Activar el paralelismo de contexto durante la carga<!----></span></h2><!--]--><!----> <p>En lugar de llamar a <code>enable_parallelism</code> después de cargar, pase el <a href="/docs/diffusers/v0.43.0/es/api/parallel#diffusers.ContextParallelConfig">ContextParallelConfig</a> al argumento <code>parallel_config</code> de <a href="/docs/diffusers/v0.43.0/es/api/models/overview#diffusers.ModelMixin.from_pretrained">from_pretrained()</a>. Inicialice primero el grupo de procesos, ya que <code>from_pretrained</code> genera un error si <code>torch.distributed</code> no está inicializado. Ejecute el script con <code>torchrun</code>, como en el ejemplo de <a href="#ring-attention">Ring Attention</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-py "><!----><span class="hljs-keyword">import</span> os
<span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> torch <span class="hljs-keyword">import</span> distributed <span class="hljs-keyword">as</span> dist
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> AutoModel, ContextParallelConfig, DiffusionPipeline

CKPT_ID = <span class="hljs-string">&quot;black-forest-labs/FLUX.1-dev&quot;</span>

dist.init_process_group(backend=<span class="hljs-string">&quot;nccl&quot;</span>)
local_rank = <span class="hljs-built_in">int</span>(os.environ[<span class="hljs-string">&quot;LOCAL_RANK&quot;</span>])
device = torch.device(<span class="hljs-string">f&quot;cuda:<span class="hljs-subst">{local_rank}</span>&quot;</span>)
torch.cuda.set_device(device)

cp_config = ContextParallelConfig(ring_degree=dist.get_world_size())
transformer = AutoModel.from_pretrained(
    CKPT_ID,
    subfolder=<span class="hljs-string">&quot;transformer&quot;</span>,
    dtype=torch.bfloat16,
    parallel_config=cp_config
)

pipeline = DiffusionPipeline.from_pretrained(
    CKPT_ID, transformer=transformer, dtype=torch.bfloat16,
).to(device)<!----></pre></div><!----> <a class="!text-gray-400 !no-underline text-sm flex items-center not-prose mt-4" href="https://github.com/huggingface/diffusers/blob/main/docs/source/en/optimization/context_parallelism.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, 333],
							data: [null,null],
							form: null,
							error: null
						});
					});
				}
			</script>
		
Xet Storage Details

Size:: 40.6 kB
Xet hash:: d863c77dc82cf8b5f91fc3b8c0dc4135ee31aa458362852a3264d6bb958c0eb3
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.