Buckets:

hf-doc-build
/

doc

Files
xet
hf-doc-build/doc / diffusers /v0.43.0 /zh /quantization /bitsandbytes.html
mishig
about 2 hours ago
download
raw
66.1 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 位（LLM.int8() 算法）&quot;,&quot;local&quot;:&quot;8-bit-llmint8-algorithm&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;异常值阈值&quot;,&quot;local&quot;:&quot;outlier-threshold&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;跳过模块转换&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 位（QLoRA 算法）&quot;,&quot;local&quot;:&quot;4-bit-qlora-algorithm&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;计算数据类型&quot;,&quot;local&quot;:&quot;compute-data-type&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;常规 Float 4 (NF4)&quot;,&quot;local&quot;:&quot;normal-float-4-nf4&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;嵌套量化&quot;,&quot;local&quot;:&quot;nested-quantization&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3}],&quot;depth&quot;:2},{&quot;title&quot;:&quot;反量化 bitsandbytes 模型&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;资源&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> 是将模型量化为 8 位和 4 位的最简单方案。8 位量化将 fp16 中的离群值与 int8 中的非离群值相乘，将非离群值转换回 fp16，然后将两者相加，从而以 fp16 格式返回权重。这降低了离群值对模型性能产生的负面影响。</p> <p>4 位量化可以进一步压缩模型，并且通常与 <a href="https://hf.co/papers/2305.14314" rel="nofollow">QLoRA</a> 结合使用，以微调量化后的大语言模型。</p> <p>本指南演示了如何通过量化，在不到 16GB 显存上运行 <a href="https://huggingface.co/black-forest-labs/FLUX.1-dev" rel="nofollow">FLUX.1-dev</a>，甚至可以在免费的 Google Colab 实例上运行。</p> <p><img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/quant-bnb/comparison.png" alt="comparison image"/></p> <p>要使用 bitsandbytes，请确保已安装以下库：</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> 已复制</div><!----></button><!----></div> <pre class="language-bash "><!---->pip install diffusers transformers accelerate bitsandbytes -U<!----></pre></div><!----> <p>现在，您可以通过向 <a href="/docs/diffusers/v0.43.0/zh/api/models/overview#diffusers.ModelMixin.from_pretrained">from_pretrained()</a> 传入 <a href="/docs/diffusers/v0.43.0/zh/api/quantization#diffusers.BitsAndBytesConfig">BitsAndBytesConfig</a> 来量化模型。只要模型支持使用 <a href="https://hf.co/docs/accelerate/index" rel="nofollow">Accelerate</a> 加载，并且包含 <code>torch.nn.Linear</code> 层，此方法就适用于任何模态下的任何模型。</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 位</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 位</div><!--]--></div> <div class="language-select"><div><p>将模型量化为 8 位可将内存占用减半：</p> <p>bitsandbytes 在 Transformers 和 Diffusers 中均受支持，因此您可以对 <a href="/docs/diffusers/v0.43.0/zh/api/models/flux_transformer#diffusers.FluxTransformer2DModel">FluxTransformer2DModel</a> 和 <a href="https://huggingface.co/docs/transformers/v5.22.2/en/model_doc/t5#transformers.T5EncoderModel" rel="nofollow">T5EncoderModel</a> 进行量化。</p> <p>对于 Ada 及更高系列的 GPU，我们建议将 <code>dtype</code> 更改为 <code>torch.bfloat16</code>。</p> <blockquote class="tip"><p><code>CLIPTextModel</code> 和 <a href="/docs/diffusers/v0.43.0/zh/api/models/autoencoderkl#diffusers.AutoencoderKL">AutoencoderKL</a> 未进行量化，因为它们本身体积已经很小，而且 <a href="/docs/diffusers/v0.43.0/zh/api/models/autoencoderkl#diffusers.AutoencoderKL">AutoencoderKL</a> 仅包含少数几个 <code>torch.nn.Linear</code> 层。</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> 已复制</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>默认情况下，<code>torch.nn.LayerNorm</code> 等其他所有模块都会被转换为 <code>torch.float16</code>。您可以通过 <code>dtype</code> 参数更改这些模块的数据类型。</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> 已复制</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>让我们使用量化后的模型生成一张图像。</p> <p>设置 <code>device_map="auto"</code> 会优先填满所有可用的 GPU 空间，然后是 CPU，最后，如果内存仍然不足，则使用硬盘（速度最慢的选项）。</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> 已复制</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>当内存充足时，您也可以使用 <code>.to("cuda")</code> 直接将 pipeline 移至 GPU，并应用 <a href="/docs/diffusers/v0.43.0/zh/api/pipelines/overview#diffusers.DiffusionPipeline.enable_model_cpu_offload">enable_model_cpu_offload()</a> 来优化 GPU 内存使用。</p> <p>Once a model is quantized, you can push the model to the Hub with the <a href="/docs/diffusers/v0.43.0/zh/api/schedulers/overview#diffusers.utils.PushToHubMixin.push_to_hub">push_to_hub()</a> method. The quantization <code>config.json</code> file is pushed first, followed by the quantized model weights. You can also save the serialized 8-bit models locally with <a href="/docs/diffusers/v0.43.0/zh/api/models/overview#diffusers.ModelMixin.save_pretrained">save_pretrained()</a>.</p><!----></div><!----> <div hidden=""><p>将模型量化为 4 位可将内存占用降低 4 倍：</p> <p>bitsandbytes 在 Transformers 和 Diffusers 中均受支持，因此您可以对 <a href="/docs/diffusers/v0.43.0/zh/api/models/flux_transformer#diffusers.FluxTransformer2DModel">FluxTransformer2DModel</a> 和 <a href="https://huggingface.co/docs/transformers/v5.22.2/en/model_doc/t5#transformers.T5EncoderModel" rel="nofollow">T5EncoderModel</a> 进行量化。</p> <p>对于 Ada 及更高系列的 GPU，我们建议将 <code>dtype</code> 更改为 <code>torch.bfloat16</code>。</p> <blockquote class="tip"><p><code>CLIPTextModel</code> 和 <a href="/docs/diffusers/v0.43.0/zh/api/models/autoencoderkl#diffusers.AutoencoderKL">AutoencoderKL</a> 未进行量化，因为它们本身体积已经很小，而且 <a href="/docs/diffusers/v0.43.0/zh/api/models/autoencoderkl#diffusers.AutoencoderKL">AutoencoderKL</a> 仅包含少数几个 <code>torch.nn.Linear</code> 层。</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> 已复制</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>默认情况下，<code>torch.nn.LayerNorm</code> 等其他所有模块都会被转换为 <code>torch.float16</code>。您可以通过 <code>dtype</code> 参数更改这些模块的数据类型。</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> 已复制</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>让我们使用量化后的模型生成一张图像。</p> <p>设置 <code>device_map="auto"</code> 会优先自动填满所有可用的 GPU 空间，然后是 CPU，最后，如果内存仍然不足，则使用硬盘（绝对最慢的选项）。</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> 已复制</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>当内存充足时，您也可以使用 <code>.to("cuda")</code> 直接将 pipeline 移至 GPU，并应用 <a href="/docs/diffusers/v0.43.0/zh/api/pipelines/overview#diffusers.DiffusionPipeline.enable_model_cpu_offload">enable_model_cpu_offload()</a> 来优化 GPU 内存使用。</p> <p>模型量化完成后，您可以使用 <a href="/docs/diffusers/v0.43.0/zh/api/schedulers/overview#diffusers.utils.PushToHubMixin.push_to_hub">push_to_hub()</a> 方法将模型推送到 Hub。量化 <code>config.json</code> 文件会先被推送，随后是量化后的模型权重。您也可以通过 <a href="/docs/diffusers/v0.43.0/zh/api/models/overview#diffusers.ModelMixin.save_pretrained">save_pretrained()</a> 将序列化后的 4-bit 模型保存到本地。</p><!----></div><!----><!----></div><!----> <blockquote class="warning"><p>使用 8 位和 4 位权重进行训练仅支持训练<em>额外</em>参数。</p></blockquote> <p>使用 <code>get_memory_footprint</code> 方法检查您的内存占用：</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> 已复制</div><!----></button><!----></div> <pre class="language-py "><!----><span class="hljs-built_in">print</span>(model.get_memory_footprint())<!----></pre></div><!----> <p>请注意，这仅能告知您模型参数的内存占用，而<em>不会</em>估算推理所需的内存。</p> <p>量化模型可以直接从 <a href="/docs/diffusers/v0.43.0/zh/api/models/overview#diffusers.ModelMixin.from_pretrained">from_pretrained()</a> 方法加载，无需指定 <code>quantization_config</code> 参数：</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> 已复制</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 位（LLM.int8() 算法）<!----></span></h2><!--]--><!----> <blockquote class="tip"><p>有关 8 位量化的详细信息，请参阅这篇<a href="https://huggingface.co/blog/hf-bitsandbytes-integration" rel="nofollow">博客文章</a>！</p></blockquote> <p>本节将探讨 8 位模型的一些特定功能，例如异常值阈值和跳过模块转换。</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><!---->异常值阈值<!----></span></h3><!--]--><!----> <p>“离群值”是指大于特定阈值的隐藏状态值，这些值以 fp16 计算。虽然这些值通常呈正态分布（[-3.5, 3.5]），但对于大型模型，其分布可能差异很大（[-60, 6] 或 [6, 60]）。8 位量化在数值约为 5 时效果良好，但超出该范围后，性能会显著下降。一个好的默认阈值是 6，但对于更不稳定的模型（小型模型或微调），可能需要更低的阈值。</p> <p>要找到适合您模型的最佳阈值，我们建议尝试调整 <a href="/docs/diffusers/v0.43.0/zh/api/quantization#diffusers.BitsAndBytesConfig">BitsAndBytesConfig</a> 中的 <code>llm_int8_threshold</code> 参数：</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> 已复制</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><!---->跳过模块转换<!----></span></h3><!--]--><!----> <p>对于某些模型，您无需将所有模块都量化为 8 位，因为这实际上可能会导致不稳定性。例如，对于 <a href="../api/pipelines/stable_diffusion/stable_diffusion_3">Stable Diffusion 3</a> 等扩散模型，可以使用 <a href="/docs/diffusers/v0.43.0/zh/api/quantization#diffusers.BitsAndBytesConfig">BitsAndBytesConfig</a> 中的 <code>llm_int8_skip_modules</code> 参数来跳过 <code>proj_out</code> 模块：</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> 已复制</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 位（QLoRA 算法）<!----></span></h2><!--]--><!----> <blockquote class="tip"><p>有关其详细信息，请参阅此<a href="https://huggingface.co/blog/4bit-transformers-bitsandbytes" rel="nofollow">博客文章</a>。</p></blockquote> <p>本节将探讨 4 位模型的一些特定功能，例如更改计算数据类型、使用 Normal Float 4 (NF4) 数据类型以及使用嵌套量化。</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><!---->计算数据类型<!----></span></h3><!--]--><!----> <p>为了加速计算，您可以使用 <a href="/docs/diffusers/v0.43.0/zh/api/quantization#diffusers.BitsAndBytesConfig">BitsAndBytesConfig</a> 中的 <code>bnb_4bit_compute_dtype</code> 参数，将数据类型从 float32（默认值）更改为 bf16：</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> 已复制</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><!---->常规 Float 4 (NF4)<!----></span></h3><!--]--><!----> <p>NF4 是 <a href="https://hf.co/papers/2305.14314" rel="nofollow">QLoRA</a> 论文中提出的一种 4 位数据类型，专为从正态分布初始化的权重而设计。在训练 4 位基础模型时，应使用 NF4。可以通过 <a href="/docs/diffusers/v0.43.0/zh/api/quantization#diffusers.BitsAndBytesConfig">BitsAndBytesConfig</a> 中的 <code>bnb_4bit_quant_type</code> 参数进行配置：</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> 已复制</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>在推理阶段，<code>bnb_4bit_quant_type</code> 对性能的影响并不大。但为了与模型权重保持一致，您应使用 <code>bnb_4bit_compute_dtype</code> 和 <code>dtype</code> 的值。</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><!---->嵌套量化<!----></span></h3><!--]--><!----> <p>嵌套量化是一种可以在不增加额外性能开销的情况下进一步节省内存的技术。该功能会对已量化的权重进行第二次量化，从而额外节省 0.4 位/参数。</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> 已复制</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><!---->反量化 bitsandbytes 模型<!----></span></h2><!--]--><!----> <p>量化完成后，您可以将模型反量化回其原始精度，但这可能会导致轻微的质量损失。请确保您的 GPU 显存足以容纳反量化后的模型。</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> 已复制</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>使用 <code>torch.compile</code> 加速推理。请确保已安装最新版本的 <code>bitsandbytes</code>，同时我们建议安装 <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 位</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 位</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> 已复制</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> 已复制</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>在配备编译功能的 RTX 4090 上，4 位 Flux 生成耗时 25.809 秒，而未启用编译时耗时 32.570 秒。</p> <p>有关更多详细信息，请查看 <a href="https://gist.github.com/sayakpaul/0db9d8eeeb3d2a0e5ed7cf0d9ca19b7d" rel="nofollow">基准测试脚本</a>。</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><!---->资源<!----></span></h2><!--]--><!----> <ul><li><a href="https://gist.github.com/sayakpaul/c76bd845b48759e11687ac550b99d8b4" rel="nofollow">展示在免费层 Colab 中进行 Flux.1 Dev 推理的端到端笔记本</a></li> <li><a href="https://github.com/huggingface/diffusers/blob/8c661ea586bf11cb2440da740dd3c4cf84679b85/examples/dreambooth/README_hidream.md#using-quantization" rel="nofollow">训练</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:: 66.1 kB
Xet hash:: c262dabdf898dd7b55ff40600bb267e0bb25c878632cf68b4923cd215a575851
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.