Hugging Face's logo

Buckets:
hf-doc-build
/
doc

Files
xet
hf-doc-build/doc / diffusers /v0.43.0 /fr /quantization /torchao.html
mishig's picture
mishig
download
raw
34.2 kB
<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;torchao&quot;,&quot;local&quot;:&quot;torchao&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;torch.compile&quot;,&quot;local&quot;:&quot;torchcompile&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Types de quantification pris en charge&quot;,&quot;local&quot;:&quot;supported-quantization-types&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Sérialisation et désérialisation des modèles quantifiés&quot;,&quot;local&quot;:&quot;serializing-and-deserializing-quantized-models&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Ressources&quot;,&quot;local&quot;:&quot;resources&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2}],&quot;depth&quot;:1}"/>
<link href="/docs/diffusers/main/en/_app/immutable/entry/start.0S0_iEoG.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/chunks/I9eooN77.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/chunks/Bupkjr9f.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/entry/app.CPdOxCbw.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/chunks/CyvZ3FS5.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/chunks/l0QnnOED.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/chunks/DnVIJfC5.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/chunks/DsnmJJEf.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/nodes/0.DteTvmQZ.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/chunks/CoSrdQax.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/nodes/359.DjT-ulWG.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/chunks/DSPDl1gz.js" rel="modulepreload">
<!--1q5ej5h--><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;torchao&quot;,&quot;local&quot;:&quot;torchao&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;torch.compile&quot;,&quot;local&quot;:&quot;torchcompile&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Supported quantization types&quot;,&quot;local&quot;:&quot;supported-quantization-types&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Serializing and Deserializing quantized models&quot;,&quot;local&quot;:&quot;serializing-and-deserializing-quantized-models&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Resources&quot;,&quot;local&quot;:&quot;resources&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2}],&quot;depth&quot;:1}"/><!---->
<link href="/docs/diffusers/main/en/_app/immutable/assets/0.tn0RQdqM.css" rel="modulepreload"> <!--[--><!--[0--><!--[--><!--[0--><!--[--><p></p> <div class="items-center shrink-0 min-w-[100px] max-sm:min-w-[50px] justify-end ml-auto flex" style="float: inline-end; margin-inline-start: 10px; display: inline-flex; position: relative; z-index: 10;"><div dir="ltr" class="inline-flex rounded-md max-sm:rounded-sm"><button class="inline-flex items-center gap-1 h-7 max-sm:h-7 px-2 max-sm:px-1.5 text-sm font-medium text-gray-800 border border-r-0 rounded-l-md max-sm:rounded-l-sm border-gray-200 bg-white hover:shadow-inner dark:border-gray-850 dark:bg-gray-950 dark:text-gray-200 dark:hover:bg-gray-800" aria-live="polite"><span class="inline-flex items-center justify-center rounded-md p-0.5 max-sm:p-0 hover:text-gray-800 dark:hover:text-gray-200"><svg class="sm:size-3.5 size-3" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----></span> <span dir="auto">Copy page</span></button> <button class="inline-flex items-center justify-center w-6 max-sm:w-5 h-7 max-sm:h-7 disabled:pointer-events-none text-sm text-gray-500 hover:text-gray-700 dark:hover:text-white rounded-r-md max-sm:rounded-r-sm border border-l transition border-gray-200 bg-white hover:shadow-inner dark:border-gray-850 dark:bg-gray-950 dark:text-gray-200 dark:hover:bg-gray-800" aria-haspopup="menu" aria-expanded="false" aria-label="Open copy menu"><svg class="transition-transform text-gray-400 overflow-visible sm:size-3.5 size-3 rotate-0" width="1em" height="1em" viewBox="0 0 12 7" fill="none" xmlns="http://www.w3.org/2000/svg"><path d="M1 1L6 6L11 1" stroke="currentColor"></path></svg><!----></button></div> <!--[-1--><!--]--></div><!----> <!--[0--><h1 class="relative group"><a id="torchao" 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="#torchao"><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><!---->torchao<!----></span></h1><!--]--><!----> <p><a href="https://github.com/pytorch/ao" rel="nofollow">torchao</a> fournit des dtypes et des optimisations haute performance basés sur la quantification et la sparsité pour l’inférence et l’entraînement de modèles PyTorch. Il prend en charge tout modèle, quelle que soit sa modalité, à condition qu’il soit chargeable avec <a href="https://hf.co/docs/accelerate/index" rel="nofollow">Accelerate</a> et qu’il contienne des couches <code>torch.nn.Linear</code>.</p> <p>Vérifiez que Pytorch 2.5+ et torchao sont installés à l'aide de 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 "><!---->uv pip install -U torch torchao<!----></pre></div><!----> <p>Chaque type de quantisation est disponible en tant qu'instance distincte de la classe <a href="https://docs.pytorch.org/ao/main/api_reference/generated/torchao.core.config.AOBaseConfig.html" rel="nofollow">AOBaseConfig</a>. Cela offre des options de configuration plus flexibles en exposant un plus grand nombre d'arguments disponibles.</p> <p>Passez le <code>AOBaseConfig</code> d'un dtype de quantification, comme <a href="https://docs.pytorch.org/ao/main/api_reference/generated/torchao.quantization.Int4WeightOnlyConfig.html" rel="nofollow">Int4WeightOnlyConfig</a>, à <a href="/docs/diffusers/v0.43.0/fr/api/quantization#diffusers.TorchAoConfig">TorchAoConfig</a> dans <a href="/docs/diffusers/v0.43.0/fr/api/models/overview#diffusers.ModelMixin.from_pretrained">from_pretrained()</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-py "><!----><span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> DiffusionPipeline, PipelineQuantizationConfig, TorchAoConfig
<span class="hljs-keyword">from</span> torchao.quantization <span class="hljs-keyword">import</span> Int8WeightOnlyConfig
pipeline_quant_config = PipelineQuantizationConfig(
quant_mapping={<span class="hljs-string">&quot;transformer&quot;</span>: TorchAoConfig(Int8WeightOnlyConfig(group_size=<span class="hljs-number">128</span>, version=<span class="hljs-number">2</span>))}
)
pipeline = DiffusionPipeline.from_pretrained(
<span class="hljs-string">&quot;black-forest-labs/FLUX.1-dev&quot;</span>,
quantization_config=pipeline_quant_config,
dtype=torch.bfloat16,
device_map=<span class="hljs-string">&quot;cuda&quot;</span> <span class="hljs-comment"># or &quot;mps&quot;, &quot;xpu&quot;, &quot;cpu&quot;</span>
)<!----></pre></div><!----> <p><code>device_map="cuda"</code> quantise chaque couche sur le GPU au moment du chargement. C’est rapide, mais cela nécessite temporairement de la mémoire GPU supplémentaire pour les poids originaux et quantifiés. Si le modèle utilise déjà la majeure partie de votre mémoire GPU, le chargement peut échouer avec une erreur de mémoire insuffisante. Dans ce cas, supprimez <code>device_map</code> et déplacez le pipeline sur le GPU après le chargement :</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-py "><!---->pipeline = DiffusionPipeline.from_pretrained(
<span class="hljs-string">&quot;black-forest-labs/FLUX.1-dev&quot;</span>,
quantization_config=pipeline_quant_config,
torch_dtype=torch.bfloat16,
)
pipeline.to(<span class="hljs-string">&quot;cuda&quot;</span>) <span class="hljs-comment"># or &quot;mps&quot;, &quot;xpu&quot;, &quot;cpu&quot;</span><!----></pre></div><!----> <p>Sans <code>device_map</code>, Diffusers quantise les couches sur le CPU. C'est plus lent, mais cela évite le pic temporaire de mémoire GPU lors de la quantisation. Pour réduire encore plus l'utilisation de la mémoire GPU, utilisez plutôt <a href="/docs/diffusers/v0.43.0/fr/api/pipelines/overview#diffusers.DiffusionPipeline.enable_model_cpu_offload">enable_model_cpu_offload()</a>. Vous pouvez également quantifier d'autres composants, tels que l'encodeur de texte.</p> <!--[1--><h2 class="relative group"><a id="torchcompile" 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="#torchcompile"><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><!---->torch.compile<!----></span></h2><!--]--><!----> <p>torchao prend en charge <a href="../optimization/fp16#torchcompile">torch.compile</a>, qui peut accélérer l'inférence avec une seule ligne de code.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Copié</div><!----></button><!----></div> <pre class="language-python "><!----><span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> DiffusionPipeline, PipelineQuantizationConfig, TorchAoConfig
<span class="hljs-keyword">from</span> torchao.quantization <span class="hljs-keyword">import</span> Int4WeightOnlyConfig
pipeline_quant_config = PipelineQuantizationConfig(
quant_mapping={<span class="hljs-string">&quot;transformer&quot;</span>: TorchAoConfig(Int4WeightOnlyConfig(group_size=<span class="hljs-number">128</span>))}
)
pipeline = DiffusionPipeline.from_pretrained(
<span class="hljs-string">&quot;black-forest-labs/FLUX.1-dev&quot;</span>,
quantization_config=pipeline_quant_config,
dtype=torch.bfloat16,
device_map=<span class="hljs-string">&quot;cuda&quot;</span> <span class="hljs-comment"># or &quot;mps&quot;, &quot;xpu&quot;, &quot;cpu&quot;</span>
)
pipeline.transformer.<span class="hljs-built_in">compile</span>(transformer, mode=<span class="hljs-string">&quot;max-autotune&quot;</span>, fullgraph=<span class="hljs-literal">True</span>)<!----></pre></div><!----> <p>Reportez-vous à ce <a href="https://github.com/huggingface/diffusers/pull/10009#issue-2688781450" rel="nofollow">tableau</a> pour les benchmarks de vitesse d’inférence et d’utilisation de la mémoire avec Flux et CogVideoX. D’autres benchmarks sur divers matériels sont également disponibles dans le <a href="https://github.com/pytorch/ao/tree/main/torchao/quantization#benchmarks" rel="nofollow">dépôt</a> torchao.</p> <blockquote class="tip"><p>Les schémas de quantification post-entraînement FP8 dans torchao sont efficaces pour les GPU ayant une capacité de calcul d'au moins 8.9 (RTX-4090, Hopper, etc.). FP8 offre souvent le meilleur compromis entre vitesse, mémoire et qualité lors de la génération d'images et de vidéos. Nous recommandons de combiner FP8 et torch.compile si votre GPU est compatible.</p></blockquote> <!--[1--><h2 class="relative group"><a id="supported-quantization-types" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#supported-quantization-types"><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><!---->Types de quantification pris en charge<!----></span></h2><!--]--><!----> <p>torchao prend en charge la quantification des poids uniquement ainsi que la quantification des poids et des activations dynamiques pour int8, float3-float8 et uint1-uint7.</p> <p>La quantification des poids uniquement stocke les poids du modèle dans un type de données à faible nombre de bits, mais effectue les calculs avec un type de données de plus haute précision, comme <code>bfloat16</code>. Cela réduit les besoins en mémoire liés aux poids du modèle, mais conserve les pics de mémoire pour le calcul des activations.</p> <p>La quantification dynamique des activations stocke les poids du modèle dans un type de données à faible nombre de bits, tout en quantifiant les activations à la volée pour économiser de la mémoire supplémentaire. Cela réduit les besoins en mémoire liés aux poids du modèle, ainsi que la surcharge mémoire liée aux calculs d'activation. Cependant, cela peut parfois entraîner un compromis en termes de qualité, il est donc recommandé de tester soigneusement différents modèles.</p> <p>Reportez-vous à la <a href="https://docs.pytorch.org/ao/stable/index.html" rel="nofollow">documentation officielle de torchao</a> pour une meilleure compréhension des méthodes de quantification disponibles. Une liste exhaustive des options de configuration est disponible <a href="https://docs.pytorch.org/ao/main/workflows/inference.html#inference-workflows" rel="nofollow">ici</a>.</p> <p>Voici quelques exemples de configurations de quantification populaires :</p> <table><thead><tr><th><strong>Catégorie</strong></th><th><strong>Classes de configuration</strong></th></tr></thead><tbody><tr><td><strong>Quantification en entiers</strong></td><td><a href="https://docs.pytorch.org/ao/stable/api_reference/generated/torchao.quantization.Int4WeightOnlyConfig.html" rel="nofollow"><code>Int4WeightOnlyConfig</code></a>, <a href="https://docs.pytorch.org/ao/stable/api_reference/generated/torchao.quantization.Int8WeightOnlyConfig.html" rel="nofollow"><code>Int8WeightOnlyConfig</code></a>, <a href="https://docs.pytorch.org/ao/stable/api_reference/generated/torchao.quantization.Int8DynamicActivationInt8WeightConfig.html" rel="nofollow"><code>Int8DynamicActivationInt8WeightConfig</code></a></td></tr><tr><td><strong>Quantification en virgule flottante 8 bits</strong></td><td><a href="https://docs.pytorch.org/ao/stable/api_reference/generated/torchao.quantization.Float8WeightOnlyConfig.html" rel="nofollow"><code>Float8WeightOnlyConfig</code></a>, <a href="https://docs.pytorch.org/ao/stable/api_reference/generated/torchao.quantization.Float8DynamicActivationFloat8WeightConfig.html" rel="nofollow"><code>Float8DynamicActivationFloat8WeightConfig</code></a></td></tr><tr><td><strong>Quantification d'entiers non signés</strong></td><td><a href="https://docs.pytorch.org/ao/stable/api_reference/generated/torchao.quantization.IntxWeightOnlyConfig.html" rel="nofollow"><code>IntxWeightOnlyConfig</code></a></td></tr></tbody></table> <!--[1--><h2 class="relative group"><a id="serializing-and-deserializing-quantized-models" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#serializing-and-deserializing-quantized-models"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span><!---->Sérialisation et désérialisation des modèles quantifiés<!----></span></h2><!--]--><!----> <p>Pour sérialiser un modèle quantifié dans un dtype donné, chargez d'abord le modèle avec le dtype de quantification souhaité, puis enregistrez-le à l'aide de la méthode <a href="/docs/diffusers/v0.43.0/fr/api/models/overview#diffusers.ModelMixin.save_pretrained">save_pretrained()</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-python "><!----><span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> AutoModel, TorchAoConfig
<span class="hljs-keyword">from</span> torchao.quantization <span class="hljs-keyword">import</span> Int8WeightOnlyConfig
quantization_config = TorchAoConfig(Int8WeightOnlyConfig())
transformer = AutoModel.from_pretrained(
<span class="hljs-string">&quot;black-forest-labs/Flux.1-Dev&quot;</span>,
subfolder=<span class="hljs-string">&quot;transformer&quot;</span>,
quantization_config=quantization_config,
dtype=torch.bfloat16,
)
transformer.save_pretrained(<span class="hljs-string">&quot;/path/to/flux_int8wo&quot;</span>, safe_serialization=<span class="hljs-literal">False</span>)<!----></pre></div><!----> <p>Pour charger un modèle quantifié sérialisé, utilisez la méthode <a href="/docs/diffusers/v0.43.0/fr/api/models/overview#diffusers.ModelMixin.from_pretrained">from_pretrained()</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-python "><!----><span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> FluxPipeline, AutoModel
transformer = AutoModel.from_pretrained(<span class="hljs-string">&quot;/path/to/flux_int8wo&quot;</span>, dtype=torch.bfloat16, use_safetensors=<span class="hljs-literal">False</span>)
pipe = FluxPipeline.from_pretrained(<span class="hljs-string">&quot;black-forest-labs/Flux.1-Dev&quot;</span>, transformer=transformer, dtype=torch.bfloat16)
pipe.to(<span class="hljs-string">&quot;cuda&quot;</span>) <span class="hljs-comment"># or &quot;mps&quot;, &quot;xpu&quot;, &quot;cpu&quot;</span>
prompt = <span class="hljs-string">&quot;A cat holding a sign that says hello world&quot;</span>
image = pipe(prompt, num_inference_steps=<span class="hljs-number">30</span>, guidance_scale=<span class="hljs-number">7.0</span>).images[<span class="hljs-number">0</span>]
image.save(<span class="hljs-string">&quot;output.png&quot;</span>)<!----></pre></div><!----> <p>Si vous utilisez <code>torch&lt;=2.6.0</code>, certaines méthodes de quantification, telles que <code>uint4</code> weight-only, ne peuvent pas être chargées directement et peuvent entraîner une <code>UnpicklingError</code> lors de la tentative de chargement des modèles, mais fonctionnent comme prévu lors de leur sauvegarde. Pour contourner ce problème, il est possible de charger manuellement le state dict dans le modèle. Notez toutefois que cela nécessite d’utiliser <code>weights_only=False</code> dans <code>torch.load</code>, il ne faut donc l’exécuter que si les poids proviennent d’une source de confiance.</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> torch
<span class="hljs-keyword">from</span> accelerate <span class="hljs-keyword">import</span> init_empty_weights
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> FluxPipeline, AutoModel, TorchAoConfig
<span class="hljs-keyword">from</span> torchao.quantization <span class="hljs-keyword">import</span> IntxWeightOnlyConfig
<span class="hljs-comment"># Serialize the model</span>
transformer = AutoModel.from_pretrained(
<span class="hljs-string">&quot;black-forest-labs/Flux.1-Dev&quot;</span>,
subfolder=<span class="hljs-string">&quot;transformer&quot;</span>,
quantization_config=TorchAoConfig(IntxWeightOnlyConfig(dtype=torch.uint4)),
dtype=torch.bfloat16,
)
transformer.save_pretrained(<span class="hljs-string">&quot;/path/to/flux_uint4wo&quot;</span>, safe_serialization=<span class="hljs-literal">False</span>, max_shard_size=<span class="hljs-string">&quot;50GB&quot;</span>)
<span class="hljs-comment"># ...</span>
<span class="hljs-comment"># Load the model</span>
state_dict = torch.load(<span class="hljs-string">&quot;/path/to/flux_uint4wo/diffusion_pytorch_model.bin&quot;</span>, weights_only=<span class="hljs-literal">False</span>, map_location=<span class="hljs-string">&quot;cpu&quot;</span>)
<span class="hljs-keyword">with</span> init_empty_weights():
transformer = AutoModel.from_config(<span class="hljs-string">&quot;/path/to/flux_uint4wo/config.json&quot;</span>)
transformer.load_state_dict(state_dict, strict=<span class="hljs-literal">True</span>, assign=<span class="hljs-literal">True</span>)<!----></pre></div><!----> <blockquote class="tip"><p>L'API <a href="/docs/diffusers/v0.43.0/fr/api/models/auto_model#diffusers.AutoModel">AutoModel</a> est prise en charge pour PyTorch &gt;= 2.6, comme le montrent les exemples ci-dessous.</p></blockquote> <!--[1--><h2 class="relative group"><a id="resources" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#resources"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span><!---->Ressources<!----></span></h2><!--]--><!----> <ul><li><a href="https://docs.pytorch.org/ao/stable/index.html" rel="nofollow">API de quantification TorchAO</a></li> <li><a href="https://github.com/sayakpaul/diffusers-torchao" rel="nofollow">Exemples Diffusers-TorchAO</a></li></ul> <a class="!text-gray-400 !no-underline text-sm flex items-center not-prose mt-4" href="https://github.com/huggingface/diffusers/blob/main/docs/source/en/quantization/torchao.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, 359],
data: [null,null],
form: null,
error: null
});
});
}
</script>

Xet Storage Details

Size:
34.2 kB
·
Xet hash:
4b45a291bb5b81f86dff463a636c082efc26dce180159d47b516564ed69ef535

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