Buckets:

hf-doc-build
/

doc

Files
xet
hf-doc-build/doc / diffusers /v0.43.0 /es /optimization /para_attn.html
mishig
2 days ago
download
raw
55.1 kB
<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;ParaAttention&quot;,&quot;local&quot;:&quot;paraattention&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Caché del primer bloque&quot;,&quot;local&quot;:&quot;first-block-cache&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;cuantización fp8&quot;,&quot;local&quot;:&quot;fp8-quantization&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Paralelismo de contexto&quot;,&quot;local&quot;:&quot;context-parallelism&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Benchmarks&quot;,&quot;local&quot;:&quot;benchmarks&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/344.CCPw-O4m.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">
		<!--fka9nn--><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;ParaAttention&quot;,&quot;local&quot;:&quot;paraattention&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;First Block Cache&quot;,&quot;local&quot;:&quot;first-block-cache&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;fp8 quantization&quot;,&quot;local&quot;:&quot;fp8-quantization&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Context Parallelism&quot;,&quot;local&quot;:&quot;context-parallelism&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Benchmarks&quot;,&quot;local&quot;:&quot;benchmarks&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="paraattention" 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="#paraattention"><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><!---->ParaAttention<!----></span></h1><!--]--><!----> <div class="flex justify-center"><img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/para-attn/flux-performance.png"/></div> <div class="flex justify-center"><img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/para-attn/hunyuan-video-performance.png"/></div> <p>Los modelos de generación de imágenes y videos de gran tamaño, como <a href="https://huggingface.co/black-forest-labs/FLUX.1-dev" rel="nofollow">FLUX.1-dev</a> y <a href="https://huggingface.co/tencent/HunyuanVideo" rel="nofollow">HunyuanVideo</a>, pueden suponer un reto de inferencia para aplicaciones en tiempo real y despliegue debido a su tamaño.</p> <p><a href="https://github.com/chengzeyi/ParaAttention" rel="nofollow">ParaAttention</a> es una biblioteca que implementa <strong>paralelismo de contexto</strong> y <strong>cache del primer bloque</strong>, y puede combinarse con otras técnicas (torch.compile, cuantización dinámica fp8) para acelerar la inferencia.</p> <p>Esta guía le mostrará cómo aplicar ParaAttention a FLUX.1-dev y HunyuanVideo en GPUs NVIDIA L20.
No se aplican optimizaciones para nuestro benchmark de línea base, excepto para HunyuanVideo, para evitar errores de memoria insuficiente.</p> <p>Nuestro benchmark de referencia muestra que FLUX.1-dev es capaz de generar una imagen de resolución 1024x1024 en 28 pasos en 26.36 segundos, y HunyuanVideo es capaz de generar 129 fotogramas a resolución 720p en 30 pasos en 3675.71 segundos.</p> <blockquote class="tip"><p>Para una inferencia aún más rápida con paralelismo de contexto, intente usar GPUs NVIDIA A100 o H100 (si están disponibles) con soporte para NVLink, especialmente cuando hay un gran número de GPUs.</p></blockquote> <!--[1--><h2 class="relative group"><a id="first-block-cache" 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="#first-block-cache"><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><!---->Caché del primer bloque<!----></span></h2><!--]--><!----> <p>Almacenar en caché la salida de los bloques de transformers del modelo y reutilizarla en los pasos de inferencia posteriores reduce el costo de cómputo y acelera la inferencia.</p> <p>Sin embargo, es difícil decidir cuándo reutilizar la caché para garantizar la calidad de las imágenes o videos generados. ParaAttention utiliza directamente la <strong>diferencia residual de la salida del primer bloque de transformer</strong> para aproximar la diferencia entre las salidas del modelo. Cuando la diferencia es lo suficientemente pequeña, se reutiliza la diferencia residual de los pasos de inferencia anteriores. En otras palabras, se omite el paso de desruido.</p> <p>Esto logra una aceleración de 2x en la inferencia de FLUX.1-dev y HunyuanVideo con una calidad muy buena.</p> <figure><img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/para-attn/ada-cache.png" alt="Cache in Diffusion Transformer"/> <figcaption>Cómo funciona AdaCache, First Block Cache es una variante de este</figcaption></figure> <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">FLUX-1.dev</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><p>Para aplicar la caché del primer bloque en FLUX.1-dev, llame a <code>apply_cache_on_pipe</code> como se muestra a continuación. 0.08 es el valor de diferencia residual predeterminado para los modelos FLUX.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5   text-gray-600  " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Copiado</div><!----></button><!----></div> <pre class="language-python "><!----><span class="hljs-keyword">import</span> time
<span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> FluxPipeline

pipe = FluxPipeline.from_pretrained(
    <span class="hljs-string">&quot;black-forest-labs/FLUX.1-dev&quot;</span>,
    dtype=torch.bfloat16,
).to(<span class="hljs-string">&quot;cuda&quot;</span>)  <span class="hljs-comment"># or &quot;mps&quot;, &quot;xpu&quot;, &quot;cpu&quot;</span>

<span class="hljs-keyword">from</span> para_attn.first_block_cache.diffusers_adapters <span class="hljs-keyword">import</span> apply_cache_on_pipe

apply_cache_on_pipe(pipe, residual_diff_threshold=<span class="hljs-number">0.08</span>)

<span class="hljs-comment"># Enable memory savings</span>
<span class="hljs-comment"># pipe.enable_model_cpu_offload()</span>
<span class="hljs-comment"># pipe.enable_sequential_cpu_offload()</span>

begin = time.time()
image = pipe(
    <span class="hljs-string">&quot;A cat holding a sign that says hello world&quot;</span>,
    num_inference_steps=<span class="hljs-number">28</span>,
).images[<span class="hljs-number">0</span>]
end = time.time()
<span class="hljs-built_in">print</span>(<span class="hljs-string">f&quot;Time: <span class="hljs-subst">{end - begin:<span class="hljs-number">.2</span>f}</span>s&quot;</span>)

<span class="hljs-built_in">print</span>(<span class="hljs-string">&quot;Saving image to flux.png&quot;</span>)
image.save(<span class="hljs-string">&quot;flux.png&quot;</span>)<!----></pre></div><!----> <table><thead><tr><th>Optimizaciones</th><th>Original</th><th>FBCache rdt=0.06</th><th>FBCache rdt=0.08</th><th>FBCache rdt=0.10</th><th>FBCache rdt=0.12</th></tr></thead><tbody><tr><td>Vista previa</td><td><img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/para-attn/flux-original.png" alt="Original"/></td><td><img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/para-attn/flux-fbc-0.06.png" alt="FBCache rdt=0.06"/></td><td><img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/para-attn/flux-fbc-0.08.png" alt="FBCache rdt=0.08"/></td><td><img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/para-attn/flux-fbc-0.10.png" alt="FBCache rdt=0.10"/></td><td><img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/para-attn/flux-fbc-0.12.png" alt="FBCache rdt=0.12"/></td></tr><tr><td>Tiempo real (s)</td><td>26.36</td><td>21.83</td><td>17.01</td><td>16.00</td><td>13.78</td></tr></tbody></table> <p>First Block Cache redujo la velocidad de inferencia a 17,01 segundos en comparación con la línea base, es decir, 1,55 veces más rápido, mientras mantenía una pérdida de calidad casi nula.</p><!----></div><!----> <div hidden=""><p>Para aplicar First Block Cache en HunyuanVideo, use <code>apply_cache_on_pipe</code> como se muestra a continuación. 0.06 es el valor de diferencia residual predeterminado para los modelos de HunyuanVideo.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5   text-gray-600  " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Copiado</div><!----></button><!----></div> <pre class="language-python "><!----><span class="hljs-keyword">import</span> time
<span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> HunyuanVideoPipeline, HunyuanVideoTransformer3DModel
<span class="hljs-keyword">from</span> diffusers.utils <span class="hljs-keyword">import</span> export_to_video

model_id = <span class="hljs-string">&quot;tencent/HunyuanVideo&quot;</span>
transformer = HunyuanVideoTransformer3DModel.from_pretrained(
    model_id,
    subfolder=<span class="hljs-string">&quot;transformer&quot;</span>,
    dtype=torch.bfloat16,
    revision=<span class="hljs-string">&quot;refs/pr/18&quot;</span>,
)
pipe = HunyuanVideoPipeline.from_pretrained(
    model_id,
    transformer=transformer,
    dtype=torch.float16,
    revision=<span class="hljs-string">&quot;refs/pr/18&quot;</span>,
).to(<span class="hljs-string">&quot;cuda&quot;</span>)  <span class="hljs-comment"># or &quot;mps&quot;, &quot;xpu&quot;, &quot;cpu&quot;</span>

<span class="hljs-keyword">from</span> para_attn.first_block_cache.diffusers_adapters <span class="hljs-keyword">import</span> apply_cache_on_pipe

apply_cache_on_pipe(pipe, residual_diff_threshold=<span class="hljs-number">0.6</span>)

pipe.vae.enable_tiling()

begin = time.time()
output = pipe(
    prompt=<span class="hljs-string">&quot;A cat walks on the grass, realistic&quot;</span>,
    height=<span class="hljs-number">720</span>,
    width=<span class="hljs-number">1280</span>,
    num_frames=<span class="hljs-number">129</span>,
    num_inference_steps=<span class="hljs-number">30</span>,
).frames[<span class="hljs-number">0</span>]
end = time.time()
<span class="hljs-built_in">print</span>(<span class="hljs-string">f&quot;Time: <span class="hljs-subst">{end - begin:<span class="hljs-number">.2</span>f}</span>s&quot;</span>)

<span class="hljs-built_in">print</span>(<span class="hljs-string">&quot;Saving video to hunyuan_video.mp4&quot;</span>)
export_to_video(output, <span class="hljs-string">&quot;hunyuan_video.mp4&quot;</span>, fps=<span class="hljs-number">15</span>)<!----></pre></div><!----> <video controls=""><source src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/para-attn/hunyuan-video-original.mp4" type="video/mp4"/> Your browser does not support the video tag.</video> <small>HunyuanVideo without FBCache</small> <video controls=""><source src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/para-attn/hunyuan-video-fbc.mp4" type="video/mp4"/> Your browser does not support the video tag.</video> <small>HunyuanVideo with FBCache</small> <p>First Block Cache redujo la velocidad de inferencia a 2271.06 segundos en comparación con la línea base, es decir, 1.62 veces más rápido, mientras mantenía una pérdida de calidad casi nula.</p><!----></div><!----><!----></div><!----> <!--[1--><h2 class="relative group"><a id="fp8-quantization" 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="#fp8-quantization"><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><!---->cuantización fp8<!----></span></h2><!--]--><!----> <p>fp8 con cuantización dinámica acelera aún más la inferencia y reduce el uso de memoria. Tanto las activaciones como los pesos deben estar cuantizados para poder utilizar los <a href="https://www.nvidia.com/en-us/data-center/tensor-cores/" rel="nofollow">NVIDIA Tensor Cores</a> de 8 bits.</p> <p>Use <code>float8_weight_only</code> y <code>float8_dynamic_activation_float8_weight</code> para cuantizar el codificador de texto y el modelo transformer.</p> <p>El método de cuantización predeterminado es la cuantización por tensor, pero si tu GPU admite la cuantización por fila, también puedes probarla para obtener una mayor precisión.</p> <p>Instala <a href="https://github.com/pytorch/ao/tree/main" rel="nofollow">torchao</a> con el siguiente comando.</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 -U torch torchao<!----></pre></div><!----> <p><a href="https://pytorch.org/tutorials/intermediate/torch_compile_tutorial.html" rel="nofollow">torch.compile</a> con <code>mode="max-autotune-no-cudagraphs"</code> o <code>mode="max-autotune"</code> selecciona el mejor kernel para el rendimiento. La compilación puede tardar mucho si es la primera vez que se llama al modelo, pero vale la pena una vez que el modelo ha sido compilado.</p> <p>Este ejemplo solo cuantiza el modelo transformer, pero también puedes cuantizar el codificador de texto para reducir aún más el uso de memoria.</p> <blockquote class="tip"><p>La cuantización dinámica puede cambiar significativamente la distribución de la salida del modelo, por lo que es necesario cambiar <code>residual_diff_threshold</code> a un valor mayor para que surta efecto.</p></blockquote> <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">FLUX-1.dev</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-python "><!----><span class="hljs-keyword">import</span> time
<span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> FluxPipeline

pipe = FluxPipeline.from_pretrained(
    <span class="hljs-string">&quot;black-forest-labs/FLUX.1-dev&quot;</span>,
    dtype=torch.bfloat16,
).to(<span class="hljs-string">&quot;cuda&quot;</span>)  <span class="hljs-comment"># or &quot;mps&quot;, &quot;xpu&quot;, &quot;cpu&quot;</span>

<span class="hljs-keyword">from</span> para_attn.first_block_cache.diffusers_adapters <span class="hljs-keyword">import</span> apply_cache_on_pipe

apply_cache_on_pipe(
    pipe,
    residual_diff_threshold=<span class="hljs-number">0.12</span>,  <span class="hljs-comment"># Use a larger value to make the cache take effect</span>
)

<span class="hljs-keyword">from</span> torchao.quantization <span class="hljs-keyword">import</span> quantize_, float8_dynamic_activation_float8_weight, float8_weight_only

quantize_(pipe.text_encoder, float8_weight_only())
quantize_(pipe.transformer, float8_dynamic_activation_float8_weight())
pipe.transformer = torch.<span class="hljs-built_in">compile</span>(
   pipe.transformer, mode=<span class="hljs-string">&quot;max-autotune-no-cudagraphs&quot;</span>,
)

<span class="hljs-comment"># Enable memory savings</span>
<span class="hljs-comment"># pipe.enable_model_cpu_offload()</span>
<span class="hljs-comment"># pipe.enable_sequential_cpu_offload()</span>

<span class="hljs-keyword">for</span> i <span class="hljs-keyword">in</span> <span class="hljs-built_in">range</span>(<span class="hljs-number">2</span>):
    begin = time.time()
    image = pipe(
        <span class="hljs-string">&quot;A cat holding a sign that says hello world&quot;</span>,
        num_inference_steps=<span class="hljs-number">28</span>,
    ).images[<span class="hljs-number">0</span>]
    end = time.time()
    <span class="hljs-keyword">if</span> i == <span class="hljs-number">0</span>:
        <span class="hljs-built_in">print</span>(<span class="hljs-string">f&quot;Warm up time: <span class="hljs-subst">{end - begin:<span class="hljs-number">.2</span>f}</span>s&quot;</span>)
    <span class="hljs-keyword">else</span>:
        <span class="hljs-built_in">print</span>(<span class="hljs-string">f&quot;Time: <span class="hljs-subst">{end - begin:<span class="hljs-number">.2</span>f}</span>s&quot;</span>)

<span class="hljs-built_in">print</span>(<span class="hljs-string">&quot;Saving image to flux.png&quot;</span>)
image.save(<span class="hljs-string">&quot;flux.png&quot;</span>)<!----></pre></div><!----> <p>La cuantización dinámica fp8 y torch.compile redujeron la velocidad de inferencia a 7,56 segundos en comparación con la línea base, es decir, 3,48 veces más rápido.</p><!----></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-python "><!----><span class="hljs-keyword">import</span> time
<span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> HunyuanVideoPipeline, HunyuanVideoTransformer3DModel
<span class="hljs-keyword">from</span> diffusers.utils <span class="hljs-keyword">import</span> export_to_video

model_id = <span class="hljs-string">&quot;tencent/HunyuanVideo&quot;</span>
transformer = HunyuanVideoTransformer3DModel.from_pretrained(
    model_id,
    subfolder=<span class="hljs-string">&quot;transformer&quot;</span>,
    dtype=torch.bfloat16,
    revision=<span class="hljs-string">&quot;refs/pr/18&quot;</span>,
)
pipe = HunyuanVideoPipeline.from_pretrained(
    model_id,
    transformer=transformer,
    dtype=torch.float16,
    revision=<span class="hljs-string">&quot;refs/pr/18&quot;</span>,
).to(<span class="hljs-string">&quot;cuda&quot;</span>)  <span class="hljs-comment"># or &quot;mps&quot;, &quot;xpu&quot;, &quot;cpu&quot;</span>

<span class="hljs-keyword">from</span> para_attn.first_block_cache.diffusers_adapters <span class="hljs-keyword">import</span> apply_cache_on_pipe

apply_cache_on_pipe(pipe)

<span class="hljs-keyword">from</span> torchao.quantization <span class="hljs-keyword">import</span> quantize_, float8_dynamic_activation_float8_weight, float8_weight_only

quantize_(pipe.text_encoder, float8_weight_only())
quantize_(pipe.transformer, float8_dynamic_activation_float8_weight())
pipe.transformer = torch.<span class="hljs-built_in">compile</span>(
   pipe.transformer, mode=<span class="hljs-string">&quot;max-autotune-no-cudagraphs&quot;</span>,
)

<span class="hljs-comment"># Enable memory savings</span>
pipe.vae.enable_tiling()
<span class="hljs-comment"># pipe.enable_model_cpu_offload()</span>
<span class="hljs-comment"># pipe.enable_sequential_cpu_offload()</span>

<span class="hljs-keyword">for</span> i <span class="hljs-keyword">in</span> <span class="hljs-built_in">range</span>(<span class="hljs-number">2</span>):
    begin = time.time()
    output = pipe(
        prompt=<span class="hljs-string">&quot;A cat walks on the grass, realistic&quot;</span>,
        height=<span class="hljs-number">720</span>,
        width=<span class="hljs-number">1280</span>,
        num_frames=<span class="hljs-number">129</span>,
        num_inference_steps=<span class="hljs-number">1</span> <span class="hljs-keyword">if</span> i == <span class="hljs-number">0</span> <span class="hljs-keyword">else</span> <span class="hljs-number">30</span>,
    ).frames[<span class="hljs-number">0</span>]
    end = time.time()
    <span class="hljs-keyword">if</span> i == <span class="hljs-number">0</span>:
        <span class="hljs-built_in">print</span>(<span class="hljs-string">f&quot;Warm up time: <span class="hljs-subst">{end - begin:<span class="hljs-number">.2</span>f}</span>s&quot;</span>)
    <span class="hljs-keyword">else</span>:
        <span class="hljs-built_in">print</span>(<span class="hljs-string">f&quot;Time: <span class="hljs-subst">{end - begin:<span class="hljs-number">.2</span>f}</span>s&quot;</span>)

<span class="hljs-built_in">print</span>(<span class="hljs-string">&quot;Saving video to hunyuan_video.mp4&quot;</span>)
export_to_video(output, <span class="hljs-string">&quot;hunyuan_video.mp4&quot;</span>, fps=<span class="hljs-number">15</span>)<!----></pre></div><!----> <p>Una GPU NVIDIA L20 solo tiene 48 GB de memoria y podría presentar errores de memoria insuficiente (OOM) después de la compilación si no se llama a <code>enable_model_cpu_offload</code>, ya que HunyuanVideo tiene tensores de activación muy grandes al ejecutarse con alta resolución y un gran número de fotogramas. Para GPUs con menos de 80 GB de memoria, puede intentar reducir la resolución y el número de fotogramas para evitar errores de OOM.</p> <p>Los modelos de generación de video a gran escala suelen estar limitados por los cálculos de atención en lugar de por las capas totalmente conectadas. Estos modelos no se benefician significativamente de la cuantización ni de torch.compile.</p><!----></div><!----><!----></div><!----> <!--[1--><h2 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></h2><!--]--><!----> <p>El paralelismo de contexto (Context Parallelism) paraleliza la inferencia y escala con múltiples GPUs. El diseño composicional de ParaAttention le permite combinar el paralelismo de contexto con la caché del primer bloque (First Block Cache) y la cuantización dinámica.</p> <blockquote class="tip"><p>Consulte el repositorio <a href="https://github.com/chengzeyi/ParaAttention/tree/main" rel="nofollow">ParaAttention</a> para obtener instrucciones detalladas y ejemplos sobre cómo escalar la inferencia con varias GPU.</p></blockquote> <p>Si el proceso de inferencia debe ser persistente y servible, se sugiere usar <a href="https://pytorch.org/docs/stable/multiprocessing.html" rel="nofollow">torch.multiprocessing</a> para escribir su propio procesador de inferencia. Esto puede eliminar la sobrecarga de iniciar el proceso y cargar y recompilar el modelo.</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">FLUX-1.dev</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><p>La muestra de código siguiente combina First Block Cache, cuantización dinámica fp8, torch.compile y Context Parallelism para la velocidad de inferencia más rápida.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5   text-gray-600  " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Copiado</div><!----></button><!----></div> <pre class="language-python "><!----><span class="hljs-keyword">import</span> time
<span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">import</span> torch.distributed <span class="hljs-keyword">as</span> dist
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> FluxPipeline

dist.init_process_group()

torch.cuda.set_device(dist.get_rank())

pipe = FluxPipeline.from_pretrained(
    <span class="hljs-string">&quot;black-forest-labs/FLUX.1-dev&quot;</span>,
    dtype=torch.bfloat16,
).to(<span class="hljs-string">&quot;cuda&quot;</span>)  <span class="hljs-comment"># or &quot;mps&quot;, &quot;xpu&quot;, &quot;cpu&quot;</span>

<span class="hljs-keyword">from</span> para_attn.context_parallel <span class="hljs-keyword">import</span> init_context_parallel_mesh
<span class="hljs-keyword">from</span> para_attn.context_parallel.diffusers_adapters <span class="hljs-keyword">import</span> parallelize_pipe
<span class="hljs-keyword">from</span> para_attn.parallel_vae.diffusers_adapters <span class="hljs-keyword">import</span> parallelize_vae

mesh = init_context_parallel_mesh(
    pipe.device.<span class="hljs-built_in">type</span>,
    max_ring_dim_size=<span class="hljs-number">2</span>,
)
parallelize_pipe(
    pipe,
    mesh=mesh,
)
parallelize_vae(pipe.vae, mesh=mesh._flatten())

<span class="hljs-keyword">from</span> para_attn.first_block_cache.diffusers_adapters <span class="hljs-keyword">import</span> apply_cache_on_pipe

apply_cache_on_pipe(
    pipe,
    residual_diff_threshold=<span class="hljs-number">0.12</span>,  <span class="hljs-comment"># Use a larger value to make the cache take effect</span>
)

<span class="hljs-keyword">from</span> torchao.quantization <span class="hljs-keyword">import</span> quantize_, float8_dynamic_activation_float8_weight, float8_weight_only

quantize_(pipe.text_encoder, float8_weight_only())
quantize_(pipe.transformer, float8_dynamic_activation_float8_weight())
torch._inductor.config.reorder_for_compute_comm_overlap = <span class="hljs-literal">True</span>
pipe.transformer = torch.<span class="hljs-built_in">compile</span>(
   pipe.transformer, mode=<span class="hljs-string">&quot;max-autotune-no-cudagraphs&quot;</span>,
)

<span class="hljs-comment"># Enable memory savings</span>
<span class="hljs-comment"># pipe.enable_model_cpu_offload(gpu_id=dist.get_rank())</span>
<span class="hljs-comment"># pipe.enable_sequential_cpu_offload(gpu_id=dist.get_rank())</span>

<span class="hljs-keyword">for</span> i <span class="hljs-keyword">in</span> <span class="hljs-built_in">range</span>(<span class="hljs-number">2</span>):
    begin = time.time()
    image = pipe(
        <span class="hljs-string">&quot;A cat holding a sign that says hello world&quot;</span>,
        num_inference_steps=<span class="hljs-number">28</span>,
        output_type=<span class="hljs-string">&quot;pil&quot;</span> <span class="hljs-keyword">if</span> dist.get_rank() == <span class="hljs-number">0</span> <span class="hljs-keyword">else</span> <span class="hljs-string">&quot;pt&quot;</span>,
    ).images[<span class="hljs-number">0</span>]
    end = time.time()
    <span class="hljs-keyword">if</span> dist.get_rank() == <span class="hljs-number">0</span>:
        <span class="hljs-keyword">if</span> i == <span class="hljs-number">0</span>:
            <span class="hljs-built_in">print</span>(<span class="hljs-string">f&quot;Warm up time: <span class="hljs-subst">{end - begin:<span class="hljs-number">.2</span>f}</span>s&quot;</span>)
        <span class="hljs-keyword">else</span>:
            <span class="hljs-built_in">print</span>(<span class="hljs-string">f&quot;Time: <span class="hljs-subst">{end - begin:<span class="hljs-number">.2</span>f}</span>s&quot;</span>)

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

dist.destroy_process_group()<!----></pre></div><!----> <p>Guárdalo en <code>run_flux.py</code> y ejecútalo con <a href="https://pytorch.org/docs/stable/elastic/run.html" rel="nofollow">torchrun</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 "><!----><span class="hljs-comment"># Use --nproc_per_node to specify the number of GPUs</span>
torchrun --nproc_per_node=2 run_flux.py<!----></pre></div><!----> <p>La velocidad de inferencia se reduce a 8,20 segundos en comparación con la línea base, es decir, es 3,21 veces más rápida, con 2 GPU NVIDIA L20. En 4 L20, la velocidad de inferencia es de 3,90 segundos, o 6,75 veces más rápida.</p><!----></div><!----> <div hidden=""><p>La muestra de código siguiente combina First Block Cache y Context Parallelism para la velocidad de inferencia más rápida.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5   text-gray-600  " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Copiado</div><!----></button><!----></div> <pre class="language-python "><!----><span class="hljs-keyword">import</span> time
<span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">import</span> torch.distributed <span class="hljs-keyword">as</span> dist
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> HunyuanVideoPipeline, HunyuanVideoTransformer3DModel
<span class="hljs-keyword">from</span> diffusers.utils <span class="hljs-keyword">import</span> export_to_video

dist.init_process_group()

torch.cuda.set_device(dist.get_rank())

model_id = <span class="hljs-string">&quot;tencent/HunyuanVideo&quot;</span>
transformer = HunyuanVideoTransformer3DModel.from_pretrained(
    model_id,
    subfolder=<span class="hljs-string">&quot;transformer&quot;</span>,
    dtype=torch.bfloat16,
    revision=<span class="hljs-string">&quot;refs/pr/18&quot;</span>,
)
pipe = HunyuanVideoPipeline.from_pretrained(
    model_id,
    transformer=transformer,
    dtype=torch.float16,
    revision=<span class="hljs-string">&quot;refs/pr/18&quot;</span>,
).to(<span class="hljs-string">&quot;cuda&quot;</span>)  <span class="hljs-comment"># or &quot;mps&quot;, &quot;xpu&quot;, &quot;cpu&quot;</span>

<span class="hljs-keyword">from</span> para_attn.context_parallel <span class="hljs-keyword">import</span> init_context_parallel_mesh
<span class="hljs-keyword">from</span> para_attn.context_parallel.diffusers_adapters <span class="hljs-keyword">import</span> parallelize_pipe
<span class="hljs-keyword">from</span> para_attn.parallel_vae.diffusers_adapters <span class="hljs-keyword">import</span> parallelize_vae

mesh = init_context_parallel_mesh(
    pipe.device.<span class="hljs-built_in">type</span>,
)
parallelize_pipe(
    pipe,
    mesh=mesh,
)
parallelize_vae(pipe.vae, mesh=mesh._flatten())

<span class="hljs-keyword">from</span> para_attn.first_block_cache.diffusers_adapters <span class="hljs-keyword">import</span> apply_cache_on_pipe

apply_cache_on_pipe(pipe)

<span class="hljs-comment"># from torchao.quantization import quantize_, float8_dynamic_activation_float8_weight, float8_weight_only</span>
<span class="hljs-comment">#</span>
<span class="hljs-comment"># torch._inductor.config.reorder_for_compute_comm_overlap = True</span>
<span class="hljs-comment">#</span>
<span class="hljs-comment"># quantize_(pipe.text_encoder, float8_weight_only())</span>
<span class="hljs-comment"># quantize_(pipe.transformer, float8_dynamic_activation_float8_weight())</span>
<span class="hljs-comment"># pipe.transformer = torch.compile(</span>
<span class="hljs-comment">#    pipe.transformer, mode=&quot;max-autotune-no-cudagraphs&quot;,</span>
<span class="hljs-comment"># )</span>

<span class="hljs-comment"># Enable memory savings</span>
pipe.vae.enable_tiling()
<span class="hljs-comment"># pipe.enable_model_cpu_offload(gpu_id=dist.get_rank())</span>
<span class="hljs-comment"># pipe.enable_sequential_cpu_offload(gpu_id=dist.get_rank())</span>

<span class="hljs-keyword">for</span> i <span class="hljs-keyword">in</span> <span class="hljs-built_in">range</span>(<span class="hljs-number">2</span>):
    begin = time.time()
    output = pipe(
        prompt=<span class="hljs-string">&quot;A cat walks on the grass, realistic&quot;</span>,
        height=<span class="hljs-number">720</span>,
        width=<span class="hljs-number">1280</span>,
        num_frames=<span class="hljs-number">129</span>,
        num_inference_steps=<span class="hljs-number">1</span> <span class="hljs-keyword">if</span> i == <span class="hljs-number">0</span> <span class="hljs-keyword">else</span> <span class="hljs-number">30</span>,
        output_type=<span class="hljs-string">&quot;pil&quot;</span> <span class="hljs-keyword">if</span> dist.get_rank() == <span class="hljs-number">0</span> <span class="hljs-keyword">else</span> <span class="hljs-string">&quot;pt&quot;</span>,
    ).frames[<span class="hljs-number">0</span>]
    end = time.time()
    <span class="hljs-keyword">if</span> dist.get_rank() == <span class="hljs-number">0</span>:
        <span class="hljs-keyword">if</span> i == <span class="hljs-number">0</span>:
            <span class="hljs-built_in">print</span>(<span class="hljs-string">f&quot;Warm up time: <span class="hljs-subst">{end - begin:<span class="hljs-number">.2</span>f}</span>s&quot;</span>)
        <span class="hljs-keyword">else</span>:
            <span class="hljs-built_in">print</span>(<span class="hljs-string">f&quot;Time: <span class="hljs-subst">{end - begin:<span class="hljs-number">.2</span>f}</span>s&quot;</span>)

<span class="hljs-keyword">if</span> dist.get_rank() == <span class="hljs-number">0</span>:
    <span class="hljs-built_in">print</span>(<span class="hljs-string">&quot;Saving video to hunyuan_video.mp4&quot;</span>)
    export_to_video(output, <span class="hljs-string">&quot;hunyuan_video.mp4&quot;</span>, fps=<span class="hljs-number">15</span>)

dist.destroy_process_group()<!----></pre></div><!----> <p>Guárdalo en <code>run_hunyuan_video.py</code> y ejecútalo con <a href="https://pytorch.org/docs/stable/elastic/run.html" rel="nofollow">torchrun</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 "><!----><span class="hljs-comment"># Use --nproc_per_node to specify the number of GPUs</span>
torchrun --nproc_per_node=8 run_hunyuan_video.py<!----></pre></div><!----> <p>La velocidad de inferencia se reduce a 649,23 segundos en comparación con la línea base, o sea, 5,66 veces más rápido, con 8 GPU NVIDIA L20.</p><!----></div><!----><!----></div><!----> <!--[1--><h2 class="relative group"><a id="benchmarks" 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="#benchmarks"><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><!---->Benchmarks<!----></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">FLUX-1.dev</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><table><thead><tr><th>Tipo de GPU</th><th>Número de GPUs</th><th>Optimizaciones</th><th>Tiempo real (s)</th><th>Aceleración</th></tr></thead><tbody><tr><td>NVIDIA L20</td><td>1</td><td>Línea base</td><td>26.36</td><td>1.00x</td></tr><tr><td>NVIDIA L20</td><td>1</td><td>FBCache (rdt=0.08)</td><td>17.01</td><td>1.55x</td></tr><tr><td>NVIDIA L20</td><td>1</td><td>FP8 DQ</td><td>13.40</td><td>1.96x</td></tr><tr><td>NVIDIA L20</td><td>1</td><td>FBCache (rdt=0.12) + FP8 DQ</td><td>7.56</td><td>3.48x</td></tr><tr><td>NVIDIA L20</td><td>2</td><td>FBCache (rdt=0.12) + FP8 DQ + CP</td><td>4.92</td><td>5.35x</td></tr><tr><td>NVIDIA L20</td><td>4</td><td>FBCache (rdt=0.12) + FP8 DQ + CP</td><td>3.90</td><td>6.75x</td></tr></tbody></table><!----></div><!----> <div hidden=""><table><thead><tr><th>Tipo de GPU</th><th>Número de GPUs</th><th>Optimizaciones</th><th>Tiempo real (s)</th><th>Aceleración</th></tr></thead><tbody><tr><td>NVIDIA L20</td><td>1</td><td>Línea base</td><td>3675.71</td><td>1.00x</td></tr><tr><td>NVIDIA L20</td><td>1</td><td>FBCache</td><td>2271.06</td><td>1.62x</td></tr><tr><td>NVIDIA L20</td><td>2</td><td>FBCache + CP</td><td>1132.90</td><td>3.24x</td></tr><tr><td>NVIDIA L20</td><td>4</td><td>FBCache + CP</td><td>718.15</td><td>5.12x</td></tr><tr><td>NVIDIA L20</td><td>8</td><td>FBCache + CP</td><td>649.23</td><td>5.66x</td></tr></tbody></table><!----></div><!----><!----></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/para_attn.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, 344],
							data: [null,null],
							form: null,
							error: null
						});
					});
				}
			</script>
		
Xet Storage Details

Size:: 55.1 kB
Xet hash:: 78df04d3f5d6d4a238627a813e59e9296a54bc4a8f0da635975582ddb7d144dc
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.