Buckets:

hf-doc-build
/

doc

Files
xet
hf-doc-build/doc / diffusers /v0.43.0 /fr /quantization /bitsandbytes.html
mishig
2 days ago
download
raw
68 kB
<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;bitsandbytes&quot;,&quot;local&quot;:&quot;bitsandbytes&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;8 bits (algorithme LLM.int8())&quot;,&quot;local&quot;:&quot;8-bit-llmint8-algorithm&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Seuil des valeurs aberrantes&quot;,&quot;local&quot;:&quot;outlier-threshold&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Passer la conversion du module&quot;,&quot;local&quot;:&quot;skip-module-conversion&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3}],&quot;depth&quot;:2},{&quot;title&quot;:&quot;4 bits (algorithme QLoRA)&quot;,&quot;local&quot;:&quot;4-bit-qlora-algorithm&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Type de données de calcul&quot;,&quot;local&quot;:&quot;compute-data-type&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Float 4 normal (NF4)&quot;,&quot;local&quot;:&quot;normal-float-4-nf4&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Quantification imbriquée&quot;,&quot;local&quot;:&quot;nested-quantization&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3}],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Déquantisation des modèles bitsandbytes&quot;,&quot;local&quot;:&quot;dequantizing-bitsandbytes-models&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;torch.compile&quot;,&quot;local&quot;:&quot;torchcompile&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/v0.43.0/en/_app/immutable/entry/start.1gaXe03k.js" rel="modulepreload">
		<link href="/docs/diffusers/v0.43.0/en/_app/immutable/chunks/BObs6oQO.js" rel="modulepreload">
		<link href="/docs/diffusers/v0.43.0/en/_app/immutable/chunks/DPnyVmXs.js" rel="modulepreload">
		<link href="/docs/diffusers/v0.43.0/en/_app/immutable/entry/app.BwdYRbCn.js" rel="modulepreload">
		<link href="/docs/diffusers/v0.43.0/en/_app/immutable/chunks/DSGUT_OW.js" rel="modulepreload">
		<link href="/docs/diffusers/v0.43.0/en/_app/immutable/chunks/C045MWWx.js" rel="modulepreload">
		<link href="/docs/diffusers/v0.43.0/en/_app/immutable/chunks/6tuJ3nyi.js" rel="modulepreload">
		<link href="/docs/diffusers/v0.43.0/en/_app/immutable/chunks/DsnmJJEf.js" rel="modulepreload">
		<link href="/docs/diffusers/v0.43.0/en/_app/immutable/nodes/0.CornNeQR.js" rel="modulepreload">
		<link href="/docs/diffusers/v0.43.0/en/_app/immutable/chunks/IaDvMxLQ.js" rel="modulepreload">
		<link href="/docs/diffusers/v0.43.0/en/_app/immutable/nodes/311.4lzQqrLa.js" rel="modulepreload">
		<link href="/docs/diffusers/v0.43.0/en/_app/immutable/chunks/bKdwVYpU.js" rel="modulepreload">
		<link href="/docs/diffusers/v0.43.0/en/_app/immutable/chunks/DTd9IGSp.js" rel="modulepreload">
		<!--169h6sl--><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;bitsandbytes&quot;,&quot;local&quot;:&quot;bitsandbytes&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;8-bit (LLM.int8() algorithm)&quot;,&quot;local&quot;:&quot;8-bit-llmint8-algorithm&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Outlier threshold&quot;,&quot;local&quot;:&quot;outlier-threshold&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Skip module conversion&quot;,&quot;local&quot;:&quot;skip-module-conversion&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3}],&quot;depth&quot;:2},{&quot;title&quot;:&quot;4-bit (QLoRA algorithm)&quot;,&quot;local&quot;:&quot;4-bit-qlora-algorithm&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Compute data type&quot;,&quot;local&quot;:&quot;compute-data-type&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Normal Float 4 (NF4)&quot;,&quot;local&quot;:&quot;normal-float-4-nf4&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Nested quantization&quot;,&quot;local&quot;:&quot;nested-quantization&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3}],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Dequantizing bitsandbytes models&quot;,&quot;local&quot;:&quot;dequantizing-bitsandbytes-models&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;torch.compile&quot;,&quot;local&quot;:&quot;torchcompile&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/v0.43.0/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="bitsandbytes" 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="#bitsandbytes"><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><!---->bitsandbytes<!----></span></h1><!--]--><!----> <p><a href="https://huggingface.co/docs/bitsandbytes/index" rel="nofollow">bitsandbytes</a> est l’option la plus simple pour quantiser un modèle en 8 et 4 bits. La quantification en 8 bits multiplie les valeurs aberrantes en fp16 par les valeurs non aberrantes en int8, convertit les valeurs non aberrantes de nouveau en fp16, puis les additionne pour restituer les poids en fp16. Cela réduit l’effet dégradant des valeurs aberrantes sur les performances du modèle.</p> <p>La quantification 4 bits comprime encore davantage un modèle et est couramment utilisée avec <a href="https://hf.co/papers/2305.14314" rel="nofollow">QLoRA</a> pour l’ajustement fin des LLM quantifiés.</p> <p>Ce guide montre comment la quantification permet d’exécuter <a href="https://huggingface.co/black-forest-labs/FLUX.1-dev" rel="nofollow">FLUX.1-dev</a> avec moins de 16 Go de VRAM, voire sur une instance Google Colab gratuite.</p> <p><img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/quant-bnb/comparison.png" alt="comparison image"/></p> <p>Pour utiliser bitsandbytes, assurez-vous d'avoir installé les bibliothèques suivantes :</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 "><!---->pip install diffusers transformers accelerate bitsandbytes -U<!----></pre></div><!----> <p>Vous pouvez désormais quantiser un modèle en passant un <a href="/docs/diffusers/v0.43.0/fr/api/quantization#diffusers.BitsAndBytesConfig">BitsAndBytesConfig</a> à <a href="/docs/diffusers/v0.43.0/fr/api/models/overview#diffusers.ModelMixin.from_pretrained">from_pretrained()</a>. Cela fonctionne pour n'importe quel modèle de n'importe quelle modalité, à condition qu'il prenne en charge le chargement 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> <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">8 bits</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">4 bits</div><!--]--></div> <div class="language-select"><div><p>La quantification d'un modèle en 8 bits réduit la consommation mémoire de moitié :</p> <p>bitsandbytes est pris en charge à la fois dans Transformers et Diffusers, vous permettant donc de quantifier à la fois <a href="/docs/diffusers/v0.43.0/fr/api/models/flux_transformer#diffusers.FluxTransformer2DModel">FluxTransformer2DModel</a> et <a href="https://huggingface.co/docs/transformers/v5.22.2/en/model_doc/t5#transformers.T5EncoderModel" rel="nofollow">T5EncoderModel</a>.</p> <p>Pour les GPU de la série Ada et supérieures, nous recommandons de changer <code>dtype</code> en <code>torch.bfloat16</code>.</p> <blockquote class="tip"><p>Les modèles <code>CLIPTextModel</code> et <a href="/docs/diffusers/v0.43.0/fr/api/models/autoencoderkl#diffusers.AutoencoderKL">AutoencoderKL</a> ne sont pas quantifiés car ils sont déjà de petite taille et parce que <a href="/docs/diffusers/v0.43.0/fr/api/models/autoencoderkl#diffusers.AutoencoderKL">AutoencoderKL</a> ne contient que quelques couches <code>torch.nn.Linear</code>.</p></blockquote> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5   text-gray-600  " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Copié</div><!----></button><!----></div> <pre class="language-py "><!----><span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> BitsAndBytesConfig <span class="hljs-keyword">as</span> DiffusersBitsAndBytesConfig
<span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> BitsAndBytesConfig <span class="hljs-keyword">as</span> TransformersBitsAndBytesConfig
<span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> AutoModel
<span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> T5EncoderModel

quant_config = TransformersBitsAndBytesConfig(load_in_8bit=<span class="hljs-literal">True</span>,)

text_encoder_2_8bit = T5EncoderModel.from_pretrained(
    <span class="hljs-string">&quot;black-forest-labs/FLUX.1-dev&quot;</span>,
    subfolder=<span class="hljs-string">&quot;text_encoder_2&quot;</span>,
    quantization_config=quant_config,
    dtype=torch.float16,
)

quant_config = DiffusersBitsAndBytesConfig(load_in_8bit=<span class="hljs-literal">True</span>,)

transformer_8bit = 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=quant_config,
    dtype=torch.float16,
)<!----></pre></div><!----> <p>Par défaut, tous les autres modules tels que <code>torch.nn.LayerNorm</code> sont convertis en <code>torch.float16</code>. Vous pouvez modifier le type de données de ces modules avec le paramètre <code>dtype</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-diff "><!---->transformer_8bit = AutoModel.from_pretrained(
    &quot;black-forest-labs/FLUX.1-dev&quot;,
    subfolder=&quot;transformer&quot;,
    quantization_config=quant_config,
<span class="hljs-addition">+   dtype=torch.float32,</span>
)<!----></pre></div><!----> <p>Générons une image à l'aide de nos modèles quantifiés.</p> <p>Définir <code>device_map="auto"</code> remplit automatiquement tout l’espace disponible sur le ou les GPU d’abord, puis sur le CPU, et enfin sur le disque dur (l’option la plus lente) s’il reste encore insuffisamment de mémoire.</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">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>,
    transformer=transformer_8bit,
    text_encoder_2=text_encoder_2_8bit,
    dtype=torch.float16,
    device_map=<span class="hljs-string">&quot;auto&quot;</span>,
)

pipe_kwargs = {
    <span class="hljs-string">&quot;prompt&quot;</span>: <span class="hljs-string">&quot;A cat holding a sign that says hello world&quot;</span>,
    <span class="hljs-string">&quot;height&quot;</span>: <span class="hljs-number">1024</span>,
    <span class="hljs-string">&quot;width&quot;</span>: <span class="hljs-number">1024</span>,
    <span class="hljs-string">&quot;guidance_scale&quot;</span>: <span class="hljs-number">3.5</span>,
    <span class="hljs-string">&quot;num_inference_steps&quot;</span>: <span class="hljs-number">50</span>,
    <span class="hljs-string">&quot;max_sequence_length&quot;</span>: <span class="hljs-number">512</span>,
}

image = pipe(**pipe_kwargs, generator=torch.manual_seed(<span class="hljs-number">0</span>),).images[<span class="hljs-number">0</span>]<!----></pre></div><!----> <div class="flex justify-center"><img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/quant-bnb/8bit.png"/></div> <p>Lorsque la mémoire est suffisante, vous pouvez également déplacer directement le pipeline sur le GPU avec <code>.to("cuda")</code> et appliquer <a href="/docs/diffusers/v0.43.0/fr/api/pipelines/overview#diffusers.DiffusionPipeline.enable_model_cpu_offload">enable_model_cpu_offload()</a> pour optimiser l'utilisation de la mémoire GPU.</p> <p>Une fois un modèle quantisé, vous pouvez le publier sur le Hub à l'aide de la méthode <a href="/docs/diffusers/v0.43.0/fr/api/schedulers/overview#diffusers.utils.PushToHubMixin.push_to_hub">push_to_hub()</a>. Le fichier de configuration de quantisation <code>config.json</code> est d'abord publié, suivi des poids du modèle quantisé. Vous pouvez également enregistrer localement les modèles 8 bits sérialisés avec <a href="/docs/diffusers/v0.43.0/fr/api/models/overview#diffusers.ModelMixin.save_pretrained">save_pretrained()</a>.</p><!----></div><!----> <div hidden=""><p>La quantification d'un modèle en 4 bits réduit votre consommation mémoire par 4 :</p> <p>bitsandbytes est pris en charge à la fois dans Transformers et Diffusers, vous permettant donc de quantifier à la fois le <a href="/docs/diffusers/v0.43.0/fr/api/models/flux_transformer#diffusers.FluxTransformer2DModel">FluxTransformer2DModel</a> et le <a href="https://huggingface.co/docs/transformers/v5.22.2/en/model_doc/t5#transformers.T5EncoderModel" rel="nofollow">T5EncoderModel</a>.</p> <p>Pour les GPU de la série Ada et supérieures, nous recommandons de changer <code>dtype</code> en <code>torch.bfloat16</code>.</p> <blockquote class="tip"><p>Les modèles <code>CLIPTextModel</code> et <a href="/docs/diffusers/v0.43.0/fr/api/models/autoencoderkl#diffusers.AutoencoderKL">AutoencoderKL</a> ne sont pas quantifiés car ils sont déjà de petite taille et parce que <a href="/docs/diffusers/v0.43.0/fr/api/models/autoencoderkl#diffusers.AutoencoderKL">AutoencoderKL</a> ne contient que quelques couches <code>torch.nn.Linear</code>.</p></blockquote> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5   text-gray-600  " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Copié</div><!----></button><!----></div> <pre class="language-py "><!----><span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> BitsAndBytesConfig <span class="hljs-keyword">as</span> DiffusersBitsAndBytesConfig
<span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> BitsAndBytesConfig <span class="hljs-keyword">as</span> TransformersBitsAndBytesConfig
<span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> AutoModel
<span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> T5EncoderModel

quant_config = TransformersBitsAndBytesConfig(load_in_4bit=<span class="hljs-literal">True</span>,)

text_encoder_2_4bit = T5EncoderModel.from_pretrained(
    <span class="hljs-string">&quot;black-forest-labs/FLUX.1-dev&quot;</span>,
    subfolder=<span class="hljs-string">&quot;text_encoder_2&quot;</span>,
    quantization_config=quant_config,
    dtype=torch.float16,
)

quant_config = DiffusersBitsAndBytesConfig(load_in_4bit=<span class="hljs-literal">True</span>,)

transformer_4bit = 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=quant_config,
    dtype=torch.float16,
)<!----></pre></div><!----> <p>Par défaut, tous les autres modules tels que <code>torch.nn.LayerNorm</code> sont convertis en <code>torch.float16</code>. Vous pouvez modifier le type de données de ces modules avec le paramètre <code>dtype</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-diff "><!---->transformer_4bit = AutoModel.from_pretrained(
    &quot;black-forest-labs/FLUX.1-dev&quot;,
    subfolder=&quot;transformer&quot;,
    quantization_config=quant_config,
<span class="hljs-addition">+   dtype=torch.float32,</span>
)<!----></pre></div><!----> <p>Générons une image à l'aide de nos modèles quantifiés.</p> <p>Définir <code>device_map="auto"</code> remplit automatiquement tout l’espace disponible sur les GPU d’abord, puis sur le CPU, et enfin sur le disque dur (l’option la plus lente) s’il reste encore insuffisamment de mémoire.</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">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>,
    transformer=transformer_4bit,
    text_encoder_2=text_encoder_2_4bit,
    dtype=torch.float16,
    device_map=<span class="hljs-string">&quot;auto&quot;</span>,
)

pipe_kwargs = {
    <span class="hljs-string">&quot;prompt&quot;</span>: <span class="hljs-string">&quot;A cat holding a sign that says hello world&quot;</span>,
    <span class="hljs-string">&quot;height&quot;</span>: <span class="hljs-number">1024</span>,
    <span class="hljs-string">&quot;width&quot;</span>: <span class="hljs-number">1024</span>,
    <span class="hljs-string">&quot;guidance_scale&quot;</span>: <span class="hljs-number">3.5</span>,
    <span class="hljs-string">&quot;num_inference_steps&quot;</span>: <span class="hljs-number">50</span>,
    <span class="hljs-string">&quot;max_sequence_length&quot;</span>: <span class="hljs-number">512</span>,
}

image = pipe(**pipe_kwargs, generator=torch.manual_seed(<span class="hljs-number">0</span>),).images[<span class="hljs-number">0</span>]<!----></pre></div><!----> <div class="flex justify-center"><img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/quant-bnb/4bit.png"/></div> <p>Lorsque la mémoire est suffisante, vous pouvez également déplacer directement le pipeline sur le GPU avec <code>.to("cuda")</code> et appliquer <a href="/docs/diffusers/v0.43.0/fr/api/pipelines/overview#diffusers.DiffusionPipeline.enable_model_cpu_offload">enable_model_cpu_offload()</a> pour optimiser l'utilisation de la mémoire GPU.</p> <p>Une fois un modèle quantifié, vous pouvez le publier sur le Hub à l'aide de la méthode <a href="/docs/diffusers/v0.43.0/fr/api/schedulers/overview#diffusers.utils.PushToHubMixin.push_to_hub">push_to_hub()</a>. Le fichier de configuration de quantisation <code>config.json</code> est d'abord publié, suivi des poids du modèle quantifié. Vous pouvez également enregistrer localement les modèles 4 bits sérialisés avec <a href="/docs/diffusers/v0.43.0/fr/api/models/overview#diffusers.ModelMixin.save_pretrained">save_pretrained()</a>.</p><!----></div><!----><!----></div><!----> <blockquote class="warning"><p>L'entraînement avec des poids en 8 bits et 4 bits n'est pris en charge que pour l'entraînement de paramètres <em>supplémentaires</em>.</p></blockquote> <p>Vérifiez votre empreinte mémoire avec la méthode <code>get_memory_footprint</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-py "><!----><span class="hljs-built_in">print</span>(model.get_memory_footprint())<!----></pre></div><!----> <p>Notez que cela ne vous indique que l’empreinte mémoire des paramètres du modèle et n’estime <em>pas</em> les besoins en mémoire pour l’inférence.</p> <p>Les modèles quantisés peuvent être chargés à partir de la méthode <a href="/docs/diffusers/v0.43.0/fr/api/models/overview#diffusers.ModelMixin.from_pretrained">from_pretrained()</a> sans avoir besoin de spécifier les paramètres <code>quantization_config</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-py "><!----><span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> AutoModel, BitsAndBytesConfig

quantization_config = BitsAndBytesConfig(load_in_4bit=<span class="hljs-literal">True</span>)

model_4bit = AutoModel.from_pretrained(
    <span class="hljs-string">&quot;hf-internal-testing/flux.1-dev-nf4-pkg&quot;</span>, subfolder=<span class="hljs-string">&quot;transformer&quot;</span>
)<!----></pre></div><!----> <!--[1--><h2 class="relative group"><a id="8-bit-llmint8-algorithm" 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="#8-bit-llmint8-algorithm"><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><!---->8 bits (algorithme LLM.int8())<!----></span></h2><!--]--><!----> <blockquote class="tip"><p>En savoir plus sur les détails de la quantification 8 bits dans ce <a href="https://huggingface.co/blog/hf-bitsandbytes-integration" rel="nofollow">article de blog</a> !</p></blockquote> <p>Cette section explore certaines fonctionnalités spécifiques des modèles 8 bits, telles que les seuils de valeurs aberrantes et le saut de la conversion de modules.</p> <!--[2--><h3 class="relative group"><a id="outlier-threshold" 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="#outlier-threshold"><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><!---->Seuil des valeurs aberrantes<!----></span></h3><!--]--><!----> <p>Un « outlier » est une valeur d’état caché supérieure à un certain seuil, et ces valeurs sont calculées en fp16. Bien que les valeurs soient généralement distribuées normalement ([-3.5, 3.5]), cette distribution peut être très différente pour les grands modèles ([-60, 6] ou [6, 60]). La quantification 8 bits fonctionne bien pour des valeurs ~5, mais au-delà, il y a une pénalité de performance significative. Une valeur de seuil par défaut appropriée est 6, mais un seuil plus bas peut être nécessaire pour les modèles plus instables (petits modèles ou finetuning).</p> <p>Pour trouver le meilleur seuil pour votre modèle, nous recommandons d'expérimenter avec le paramètre <code>llm_int8_threshold</code> dans <a href="/docs/diffusers/v0.43.0/fr/api/quantization#diffusers.BitsAndBytesConfig">BitsAndBytesConfig</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">from</span> diffusers <span class="hljs-keyword">import</span> AutoModel, BitsAndBytesConfig

quantization_config = BitsAndBytesConfig(
    load_in_8bit=<span class="hljs-literal">True</span>, llm_int8_threshold=<span class="hljs-number">10</span>,
)

model_8bit = 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,
)<!----></pre></div><!----> <!--[2--><h3 class="relative group"><a id="skip-module-conversion" 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="#skip-module-conversion"><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><!---->Passer la conversion du module<!----></span></h3><!--]--><!----> <p>Pour certains modèles, il n’est pas nécessaire de quantiser chaque module en 8 bits, ce qui peut d’ailleurs provoquer une instabilité. Par exemple, pour les modèles de diffusion comme <a href="../api/pipelines/stable_diffusion/stable_diffusion_3">Stable Diffusion 3</a>, le module <code>proj_out</code> peut être ignoré à l’aide du paramètre <code>llm_int8_skip_modules</code> dans <a href="/docs/diffusers/v0.43.0/fr/api/quantization#diffusers.BitsAndBytesConfig">BitsAndBytesConfig</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">from</span> diffusers <span class="hljs-keyword">import</span> SD3Transformer2DModel, BitsAndBytesConfig

quantization_config = BitsAndBytesConfig(
    load_in_8bit=<span class="hljs-literal">True</span>, llm_int8_skip_modules=[<span class="hljs-string">&quot;proj_out&quot;</span>],
)

model_8bit = SD3Transformer2DModel.from_pretrained(
    <span class="hljs-string">&quot;stabilityai/stable-diffusion-3-medium-diffusers&quot;</span>,
    subfolder=<span class="hljs-string">&quot;transformer&quot;</span>,
    quantization_config=quantization_config,
)<!----></pre></div><!----> <!--[1--><h2 class="relative group"><a id="4-bit-qlora-algorithm" 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="#4-bit-qlora-algorithm"><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><!---->4 bits (algorithme QLoRA)<!----></span></h2><!--]--><!----> <blockquote class="tip"><p>En savoir plus sur ses détails dans ce <a href="https://huggingface.co/blog/4bit-transformers-bitsandbytes" rel="nofollow">article de blog</a>.</p></blockquote> <p>Cette section explore certaines fonctionnalités spécifiques des modèles 4 bits, telles que le changement du type de données de calcul, l'utilisation du type de données Normal Float 4 (NF4) et l'utilisation de la quantification imbriquée.</p> <!--[2--><h3 class="relative group"><a id="compute-data-type" 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="#compute-data-type"><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><!---->Type de données de calcul<!----></span></h3><!--]--><!----> <p>Pour accélérer le calcul, vous pouvez changer le type de données de float32 (la valeur par défaut) en bf16 à l'aide du paramètre <code>bnb_4bit_compute_dtype</code> dans <a href="/docs/diffusers/v0.43.0/fr/api/quantization#diffusers.BitsAndBytesConfig">BitsAndBytesConfig</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> BitsAndBytesConfig

quantization_config = BitsAndBytesConfig(load_in_4bit=<span class="hljs-literal">True</span>, bnb_4bit_compute_dtype=torch.bfloat16)<!----></pre></div><!----> <!--[2--><h3 class="relative group"><a id="normal-float-4-nf4" 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="#normal-float-4-nf4"><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><!---->Float 4 normal (NF4)<!----></span></h3><!--]--><!----> <p>NF4 est un type de données sur 4 bits issu de l'article <a href="https://hf.co/papers/2305.14314" rel="nofollow">QLoRA</a>, adapté aux poids initialisés à partir d'une distribution normale. Vous devez utiliser NF4 pour l'entraînement de modèles de base sur 4 bits. Cela peut être configuré avec le paramètre <code>bnb_4bit_quant_type</code> dans <a href="/docs/diffusers/v0.43.0/fr/api/quantization#diffusers.BitsAndBytesConfig">BitsAndBytesConfig</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">from</span> diffusers <span class="hljs-keyword">import</span> BitsAndBytesConfig <span class="hljs-keyword">as</span> DiffusersBitsAndBytesConfig
<span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> BitsAndBytesConfig <span class="hljs-keyword">as</span> TransformersBitsAndBytesConfig

<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> AutoModel
<span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> T5EncoderModel

quant_config = TransformersBitsAndBytesConfig(
    load_in_4bit=<span class="hljs-literal">True</span>,
    bnb_4bit_quant_type=<span class="hljs-string">&quot;nf4&quot;</span>,
)

text_encoder_2_4bit = T5EncoderModel.from_pretrained(
    <span class="hljs-string">&quot;black-forest-labs/FLUX.1-dev&quot;</span>,
    subfolder=<span class="hljs-string">&quot;text_encoder_2&quot;</span>,
    quantization_config=quant_config,
    dtype=torch.float16,
)

quant_config = DiffusersBitsAndBytesConfig(
    load_in_4bit=<span class="hljs-literal">True</span>,
    bnb_4bit_quant_type=<span class="hljs-string">&quot;nf4&quot;</span>,
)

transformer_4bit = 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=quant_config,
    dtype=torch.float16,
)<!----></pre></div><!----> <p>Pour l'inférence, le <code>bnb_4bit_quant_type</code> n'a pas d'impact majeur sur les performances. Toutefois, pour rester cohérent avec les poids du modèle, vous devez utiliser les valeurs de <code>bnb_4bit_compute_dtype</code> et <code>dtype</code>.</p> <!--[2--><h3 class="relative group"><a id="nested-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="#nested-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 imbriquée<!----></span></h3><!--]--><!----> <p>La quantification imbriquée est une technique qui permet d’économiser de la mémoire supplémentaire sans coût de performance additionnel. Cette fonctionnalité effectue une seconde quantification des poids déjà quantifiés afin d’économiser 0,4 bit/paramètre supplémentaire.</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">from</span> diffusers <span class="hljs-keyword">import</span> BitsAndBytesConfig <span class="hljs-keyword">as</span> DiffusersBitsAndBytesConfig
<span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> BitsAndBytesConfig <span class="hljs-keyword">as</span> TransformersBitsAndBytesConfig

<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> AutoModel
<span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> T5EncoderModel

quant_config = TransformersBitsAndBytesConfig(
    load_in_4bit=<span class="hljs-literal">True</span>,
    bnb_4bit_use_double_quant=<span class="hljs-literal">True</span>,
)

text_encoder_2_4bit = T5EncoderModel.from_pretrained(
    <span class="hljs-string">&quot;black-forest-labs/FLUX.1-dev&quot;</span>,
    subfolder=<span class="hljs-string">&quot;text_encoder_2&quot;</span>,
    quantization_config=quant_config,
    dtype=torch.float16,
)

quant_config = DiffusersBitsAndBytesConfig(
    load_in_4bit=<span class="hljs-literal">True</span>,
    bnb_4bit_use_double_quant=<span class="hljs-literal">True</span>,
)

transformer_4bit = 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=quant_config,
    dtype=torch.float16,
)<!----></pre></div><!----> <!--[1--><h2 class="relative group"><a id="dequantizing-bitsandbytes-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="#dequantizing-bitsandbytes-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><!---->Déquantisation des modèles bitsandbytes<!----></span></h2><!--]--><!----> <p>Une fois quantisé, vous pouvez déquantiser un modèle pour revenir à sa précision d’origine, mais cela peut entraîner une légère perte de qualité. Assurez-vous de disposer de suffisamment de RAM GPU pour charger le modèle déquantisé.</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">from</span> diffusers <span class="hljs-keyword">import</span> BitsAndBytesConfig <span class="hljs-keyword">as</span> DiffusersBitsAndBytesConfig
<span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> BitsAndBytesConfig <span class="hljs-keyword">as</span> TransformersBitsAndBytesConfig

<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> AutoModel
<span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> T5EncoderModel

quant_config = TransformersBitsAndBytesConfig(
    load_in_4bit=<span class="hljs-literal">True</span>,
    bnb_4bit_use_double_quant=<span class="hljs-literal">True</span>,
)

text_encoder_2_4bit = T5EncoderModel.from_pretrained(
    <span class="hljs-string">&quot;black-forest-labs/FLUX.1-dev&quot;</span>,
    subfolder=<span class="hljs-string">&quot;text_encoder_2&quot;</span>,
    quantization_config=quant_config,
    dtype=torch.float16,
)

quant_config = DiffusersBitsAndBytesConfig(
    load_in_4bit=<span class="hljs-literal">True</span>,
    bnb_4bit_use_double_quant=<span class="hljs-literal">True</span>,
)

transformer_4bit = 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=quant_config,
    dtype=torch.float16,
)

text_encoder_2_4bit.dequantize()
transformer_4bit.dequantize()<!----></pre></div><!----> <!--[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>Accélérez l'inférence avec <code>torch.compile</code>. Assurez-vous d'avoir la dernière version de <code>bitsandbytes</code> installée et nous recommandons également d'installer <a href="https://pytorch.org/get-started/locally/" rel="nofollow">PyTorch nightly</a>.</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">8 bits</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">4 bits</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-py "><!---->torch._dynamo.config.capture_dynamic_output_shape_ops = <span class="hljs-literal">True</span>

quant_config = DiffusersBitsAndBytesConfig(load_in_8bit=<span class="hljs-literal">True</span>)
transformer_4bit = 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=quant_config,
    dtype=torch.float16,
)
transformer_4bit.<span class="hljs-built_in">compile</span>(fullgraph=<span class="hljs-literal">True</span>)<!----></pre></div><!----></div><!----> <div hidden=""><div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5   text-gray-600  " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Copié</div><!----></button><!----></div> <pre class="language-py "><!---->quant_config = DiffusersBitsAndBytesConfig(load_in_4bit=<span class="hljs-literal">True</span>)
transformer_4bit = 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=quant_config,
    dtype=torch.float16,
)
transformer_4bit.<span class="hljs-built_in">compile</span>(fullgraph=<span class="hljs-literal">True</span>)<!----></pre></div><!----></div><!----><!----></div><!----> <p>Sur une RTX 4090 avec compilation, la génération Flux en 4 bits s’est terminée en 25,809 secondes contre 32,570 secondes sans compilation.</p> <p>Consultez le <a href="https://gist.github.com/sayakpaul/0db9d8eeeb3d2a0e5ed7cf0d9ca19b7d" rel="nofollow">script de benchmarking</a> pour plus de détails.</p> <!--[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://gist.github.com/sayakpaul/c76bd845b48759e11687ac550b99d8b4" rel="nofollow">Carnet de notes de bout en bout montrant l'inférence de Flux.1 Dev dans un Colab gratuit</a></li> <li><a href="https://github.com/huggingface/diffusers/blob/8c661ea586bf11cb2440da740dd3c4cf84679b85/examples/dreambooth/README_hidream.md#using-quantization" rel="nofollow">Entraînement</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/bitsandbytes.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_ywq9g6 = {
						base: "/docs/diffusers/v0.43.0/en",
						assets: "/docs/diffusers/v0.43.0/en"
					};
					const element = document.currentScript.parentElement;
					Promise.all([
						import("/docs/diffusers/v0.43.0/en/_app/immutable/entry/start.1gaXe03k.js"),
						import("/docs/diffusers/v0.43.0/en/_app/immutable/entry/app.BwdYRbCn.js")
					]).then(([kit, app]) => {
						kit.start(app, element, {
							node_ids: [0, 311],
							data: [null,null],
							form: null,
							error: null
						});
					});
				}
			</script>
		
Xet Storage Details

Size:: 68 kB
Xet hash:: 43432593d2dd7cce33f9adb4330d04a11e3e00b69d7a413791633e0887896de7
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.