Buckets:

hf-doc-build
/

doc

Files
xet
hf-doc-build/doc / diffusers /main /fr /optimization /para_attn.html
mishig
3 days ago
download
raw
55.5 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;Cache du premier bloc&quot;,&quot;local&quot;:&quot;first-block-cache&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;quantification fp8&quot;,&quot;local&quot;:&quot;fp8-quantization&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Parallélisme de contexte&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>Les grands modèles de génération d’images et de vidéos, tels que <a href="https://huggingface.co/black-forest-labs/FLUX.1-dev" rel="nofollow">FLUX.1-dev</a> et <a href="https://huggingface.co/tencent/HunyuanVideo" rel="nofollow">HunyuanVideo</a>, peuvent poser un défi d’inférence pour les applications en temps réel et le déploiement en raison de leur taille.</p> <p><a href="https://github.com/chengzeyi/ParaAttention" rel="nofollow">ParaAttention</a> est une bibliothèque qui implémente le <strong>context parallelism</strong> et le <strong>first block cache</strong>, et peut être combinée avec d'autres techniques (torch.compile, quantification dynamique fp8) pour accélérer l'inférence.</p> <p>Ce guide vous montre comment appliquer ParaAttention à FLUX.1-dev et HunyuanVideo sur des GPU NVIDIA L20.
Aucune optimisation n’est appliquée pour notre benchmark de référence, à l’exception de HunyuanVideo afin d’éviter les erreurs de mémoire insuffisante.</p> <p>Notre benchmark de référence montre que FLUX.1-dev est capable de générer une image en résolution 1024x1024 en 28 étapes en 26,36 secondes, et que HunyuanVideo est capable de générer 129 images en résolution 720p en 30 étapes en 3675,71 secondes.</p> <blockquote class="tip"><p>Pour une inférence encore plus rapide avec le parallélisme de contexte, essayez d’utiliser des GPU NVIDIA A100 ou H100 (si disponibles) avec prise en charge de NVLink, en particulier lorsqu’il y a un grand nombre de GPU.</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><!---->Cache du premier bloc<!----></span></h2><!--]--><!----> <p>Mettre en cache la sortie des blocs de transformers du modèle et la réutiliser aux étapes d'inférence suivantes réduit le coût de calcul et accélère l'inférence.</p> <p>Cependant, il est difficile de déterminer quand réutiliser le cache afin de garantir la qualité des images ou vidéos générées. ParaAttention utilise directement la <strong>différence résiduelle de la sortie du premier bloc transformer</strong> pour approximer l'écart entre les sorties du modèle. Lorsque cet écart est suffisamment faible, la différence résiduelle des étapes d'inférence précédentes est réutilisée. Autrement dit, l'étape de débruitage est ignorée.</p> <p>Cela permet d'obtenir un gain de vitesse de 2x lors de l'inférence de FLUX.1-dev et HunyuanVideo, avec une très bonne qualité.</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>Fonctionnement d'AdaCache, First Block Cache en est une variante</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>Pour appliquer le cache du premier bloc sur FLUX.1-dev, appelez <code>apply_cache_on_pipe</code> comme indiqué ci-dessous. 0,08 est la valeur par défaut de la différence résiduelle pour les modèles 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> Copié</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>Optimisations</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>Aperçu</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>Temps réel (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 a réduit la vitesse d’inférence à 17,01 secondes par rapport à la référence, soit 1,55 fois plus rapide, tout en maintenant une perte de qualité quasi nulle.</p><!----></div><!----> <div hidden=""><p>Pour appliquer First Block Cache sur HunyuanVideo, utilisez <code>apply_cache_on_pipe</code> comme indiqué ci-dessous. 0,06 est la valeur par défaut de la différence résiduelle pour les modèles 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> Copié</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 a réduit la vitesse d’inférence à 2271,06 secondes par rapport à la référence, soit 1,62 fois plus rapide, tout en maintenant une perte de qualité quasi nulle.</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><!---->quantification fp8<!----></span></h2><!--]--><!----> <p>fp8 avec quantification dynamique accélère davantage l’inférence et réduit l’utilisation de la mémoire. Les activations et les poids doivent tous deux être quantifiés afin d’utiliser les <a href="https://www.nvidia.com/en-us/data-center/tensor-cores/" rel="nofollow">NVIDIA Tensor Cores</a> 8 bits.</p> <p>Utilisez <code>float8_weight_only</code> et <code>float8_dynamic_activation_float8_weight</code> pour quantifier le codeur de texte et le modèle transformer.</p> <p>La méthode de quantification par défaut est la quantification par tenseur, mais si votre GPU prend en charge la quantification par ligne, vous pouvez également l'essayer pour une meilleure précision.</p> <p>Installez <a href="https://github.com/pytorch/ao/tree/main" rel="nofollow">torchao</a> avec la commande ci-dessous.</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> Copié</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> avec <code>mode="max-autotune-no-cudagraphs"</code> ou <code>mode="max-autotune"</code> sélectionne le meilleur noyau pour les performances. La compilation peut prendre beaucoup de temps s’il s’agit de la première fois que le modèle est appelé, mais cela en vaut la peine une fois le modèle compilé.</p> <p>Cet exemple ne quantise que le modèle transformer, mais vous pouvez également quantiser l’encodeur de texte pour réduire encore davantage l’utilisation de la mémoire.</p> <blockquote class="tip"><p>La quantification dynamique peut modifier considérablement la distribution de la sortie du modèle, il est donc nécessaire d’augmenter la valeur de <code>residual_diff_threshold</code> pour que cela prenne effet.</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> Copié</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 quantification dynamique fp8 et torch.compile ont réduit le temps d’inférence à 7,56 secondes par rapport à la référence, soit 3,48 fois plus rapide.</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> Copié</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>Un GPU NVIDIA L20 ne dispose que de 48 Go de mémoire et peut rencontrer des erreurs de mémoire insuffisante (OOM) après la compilation si <code>enable_model_cpu_offload</code> n’est pas appelé, car HunyuanVideo génère des tenseurs d’activation très volumineux lors de l’exécution en haute résolution et avec un grand nombre d’images. Pour les GPU disposant de moins de 80 Go de mémoire, vous pouvez essayer de réduire la résolution et le nombre d’images afin d’éviter les erreurs OOM.</p> <p>Les grands modèles de génération vidéo sont généralement limités par les calculs d’attention plutôt que par les couches entièrement connectées. Ces modèles ne bénéficient pas significativement de la quantification et 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><!---->Parallélisme de contexte<!----></span></h2><!--]--><!----> <p>Le parallélisme de contexte (Context Parallelism) parallélise l'inférence et s'échelonne sur plusieurs GPU. La conception compositionnelle de ParaAttention vous permet de combiner le parallélisme de contexte avec le cache du premier bloc et la quantification dynamique.</p> <blockquote class="tip"><p>Reportez-vous au dépôt <a href="https://github.com/chengzeyi/ParaAttention/tree/main" rel="nofollow">ParaAttention</a> pour des instructions détaillées et des exemples sur la manière de mettre à l’échelle l’inférence avec plusieurs GPU.</p></blockquote> <p>Si le processus d’inférence doit être persistant et accessible en tant que service, il est recommandé d’utiliser <a href="https://pytorch.org/docs/stable/multiprocessing.html" rel="nofollow">torch.multiprocessing</a> pour écrire votre propre processeur d’inférence. Cela permet d’éliminer le surcoût lié au lancement du processus ainsi qu’au chargement et à la recompilation du modèle.</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>L'extrait de code ci-dessous combine First Block Cache, quantification dynamique fp8, torch.compile et Context Parallelism pour la vitesse d'inférence la plus rapide.</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> Copié</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>Enregistrez dans <code>run_flux.py</code> et lancez-le avec <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> Copié</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 vitesse d’inférence est réduite à 8,20 secondes par rapport à la référence, soit 3,21 fois plus rapide, avec 2 GPU NVIDIA L20. Sur 4 L20, la vitesse d’inférence est de 3,90 secondes, soit 6,75 fois plus rapide.</p><!----></div><!----> <div hidden=""><p>L'extrait de code ci-dessous combine First Block Cache et Context Parallelism pour obtenir la vitesse d'inférence la plus rapide.</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> Copié</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>Enregistrez dans <code>run_hunyuan_video.py</code> et lancez-le avec <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> Copié</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 vitesse d'inférence est réduite à 649,23 secondes par rapport à la référence, soit 5,66 fois plus rapide, avec 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>Type de GPU</th><th>Nombre de GPU</th><th>Optimisations</th><th>Temps réel (s)</th><th>Accélération</th></tr></thead><tbody><tr><td>NVIDIA L20</td><td>1</td><td>Ligne de 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>Type de GPU</th><th>Nombre de GPU</th><th>Optimisations</th><th>Temps réel (s)</th><th>Accélération</th></tr></thead><tbody><tr><td>NVIDIA L20</td><td>1</td><td>Ligne de 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:: 55.5 kB
Xet hash:: 781dd0bdbcb7de0a64e64bbfaf10cbe94094fb2387c0d318c00ff867272cb682
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.