Hugging Face's logo

Buckets:
hf-doc-build
/
doc

Files
xet
hf-doc-build/doc / diffusers /main /pt /quantization /modelopt.html
mishig's picture
mishig
download
raw
26.4 kB
<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;NVIDIA ModelOpt&quot;,&quot;local&quot;:&quot;nvidia-modelopt&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;NVIDIAModelOptConfig&quot;,&quot;local&quot;:&quot;nvidiamodeloptconfig&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Tipos de quantização suportados&quot;,&quot;local&quot;:&quot;supported-quantization-types&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Serialização e desserialização de modelos quantizados&quot;,&quot;local&quot;:&quot;serializing-and-deserializing-quantized-models&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2}],&quot;depth&quot;:1}"/>
<link href="/docs/diffusers/main/en/_app/immutable/entry/start.0S0_iEoG.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/chunks/I9eooN77.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/chunks/Bupkjr9f.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/entry/app.CPdOxCbw.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/chunks/CyvZ3FS5.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/chunks/l0QnnOED.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/chunks/DnVIJfC5.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/chunks/DsnmJJEf.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/nodes/0.DteTvmQZ.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/chunks/CoSrdQax.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/nodes/354.Da0b8rJZ.js" rel="modulepreload">
<link href="/docs/diffusers/main/en/_app/immutable/chunks/DSPDl1gz.js" rel="modulepreload">
<!--v66wdz--><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;NVIDIA ModelOpt&quot;,&quot;local&quot;:&quot;nvidia-modelopt&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;NVIDIAModelOptConfig&quot;,&quot;local&quot;:&quot;nvidiamodeloptconfig&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Supported quantization types&quot;,&quot;local&quot;:&quot;supported-quantization-types&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Serializing and Deserializing quantized models&quot;,&quot;local&quot;:&quot;serializing-and-deserializing-quantized-models&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2}],&quot;depth&quot;:1}"/><!---->
<link href="/docs/diffusers/main/en/_app/immutable/assets/0.tn0RQdqM.css" rel="modulepreload"> <!--[--><!--[0--><!--[--><!--[0--><!--[--><p></p> <div class="items-center shrink-0 min-w-[100px] max-sm:min-w-[50px] justify-end ml-auto flex" style="float: inline-end; margin-inline-start: 10px; display: inline-flex; position: relative; z-index: 10;"><div dir="ltr" class="inline-flex rounded-md max-sm:rounded-sm"><button class="inline-flex items-center gap-1 h-7 max-sm:h-7 px-2 max-sm:px-1.5 text-sm font-medium text-gray-800 border border-r-0 rounded-l-md max-sm:rounded-l-sm border-gray-200 bg-white hover:shadow-inner dark:border-gray-850 dark:bg-gray-950 dark:text-gray-200 dark:hover:bg-gray-800" aria-live="polite"><span class="inline-flex items-center justify-center rounded-md p-0.5 max-sm:p-0 hover:text-gray-800 dark:hover:text-gray-200"><svg class="sm:size-3.5 size-3" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----></span> <span dir="auto">Copy page</span></button> <button class="inline-flex items-center justify-center w-6 max-sm:w-5 h-7 max-sm:h-7 disabled:pointer-events-none text-sm text-gray-500 hover:text-gray-700 dark:hover:text-white rounded-r-md max-sm:rounded-r-sm border border-l transition border-gray-200 bg-white hover:shadow-inner dark:border-gray-850 dark:bg-gray-950 dark:text-gray-200 dark:hover:bg-gray-800" aria-haspopup="menu" aria-expanded="false" aria-label="Open copy menu"><svg class="transition-transform text-gray-400 overflow-visible sm:size-3.5 size-3 rotate-0" width="1em" height="1em" viewBox="0 0 12 7" fill="none" xmlns="http://www.w3.org/2000/svg"><path d="M1 1L6 6L11 1" stroke="currentColor"></path></svg><!----></button></div> <!--[-1--><!--]--></div><!----> <!--[0--><h1 class="relative group"><a id="nvidia-modelopt" 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="#nvidia-modelopt"><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><!---->NVIDIA ModelOpt<!----></span></h1><!--]--><!----> <p><a href="https://github.com/NVIDIA/Model-Optimizer" rel="nofollow">NVIDIA-ModelOpt</a> é uma biblioteca unificada de técnicas de otimização de modelos de última geração, como quantização, poda, destilação, decodificação especulativa, etc. Ela comprime modelos de aprendizado profundo para frameworks de implantação downstream, como TensorRT-LLM ou TensorRT, a fim de otimizar a velocidade de inferência.</p> <p>Antes de começar, verifique se você tem o nvidia_modelopt instalado.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Copiado</div><!----></button><!----></div> <pre class="language-bash "><!---->pip install -U <span class="hljs-string">&quot;nvidia_modelopt[hf]&quot;</span><!----></pre></div><!----> <p>Quantize um modelo passando <code>NVIDIAModelOptConfig</code> para <a href="/docs/diffusers/main/pt/api/models/overview#diffusers.ModelMixin.from_pretrained">from_pretrained()</a> (você também pode carregar modelos pré-quantizados). Isso funciona para qualquer modelo em qualquer modalidade, desde que ele suporte carregamento com <a href="https://hf.co/docs/accelerate/index" rel="nofollow">Accelerate</a> e contenha camadas <code>torch.nn.Linear</code>.</p> <p>O exemplo abaixo apenas quantiza os pesos para FP8.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Copiado</div><!----></button><!----></div> <pre class="language-python "><!----><span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> AutoModel, SanaPipeline, NVIDIAModelOptConfig
model_id = <span class="hljs-string">&quot;Efficient-Large-Model/Sana_600M_1024px_diffusers&quot;</span>
dtype = torch.bfloat16
quantization_config = NVIDIAModelOptConfig(quant_type=<span class="hljs-string">&quot;FP8&quot;</span>, quant_method=<span class="hljs-string">&quot;modelopt&quot;</span>)
transformer = AutoModel.from_pretrained(
model_id,
subfolder=<span class="hljs-string">&quot;transformer&quot;</span>,
quantization_config=quantization_config,
dtype=dtype,
)
pipe = SanaPipeline.from_pretrained(
model_id,
transformer=transformer,
dtype=dtype,
)
pipe.to(<span class="hljs-string">&quot;cuda&quot;</span>) <span class="hljs-comment"># or &quot;mps&quot;, &quot;xpu&quot;, &quot;cpu&quot;</span>
<span class="hljs-built_in">print</span>(<span class="hljs-string">f&quot;Pipeline memory usage: <span class="hljs-subst">{torch.cuda.max_memory_reserved() / <span class="hljs-number">1024</span>**<span class="hljs-number">3</span>:<span class="hljs-number">.3</span>f}</span> GB&quot;</span>)
prompt = <span class="hljs-string">&quot;A cat holding a sign that says hello world&quot;</span>
image = pipe(
prompt, num_inference_steps=<span class="hljs-number">50</span>, guidance_scale=<span class="hljs-number">4.5</span>, max_sequence_length=<span class="hljs-number">512</span>
).images[<span class="hljs-number">0</span>]
image.save(<span class="hljs-string">&quot;output.png&quot;</span>)<!----></pre></div><!----> <blockquote><p><strong>Observação:</strong></p> <p>Os métodos de quantização no NVIDIA-ModelOpt são projetados para reduzir a pegada de memória dos pesos do modelo usando diversas técnicas de QAT (Quantization-Aware Training) e PTQ (Post-Training Quantization), mantendo o desempenho do modelo. No entanto, o ganho de desempenho real durante a inferência depende do framework de implantação (por exemplo, TRT-LLM, TensorRT) e da configuração de hardware específica.</p> <p>Mais detalhes podem ser encontrados <a href="https://github.com/NVIDIA/Model-Optimizer/tree/main/examples" rel="nofollow">aqui</a>.</p></blockquote> <!--[1--><h2 class="relative group"><a id="nvidiamodeloptconfig" 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="#nvidiamodeloptconfig"><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><!---->NVIDIAModelOptConfig<!----></span></h2><!--]--><!----> <p>A classe <code>NVIDIAModelOptConfig</code> aceita três parâmetros:</p> <ul><li><code>quant_type</code>: Um valor de string que menciona um dos tipos de quantização abaixo.</li> <li><code>modules_to_not_convert</code>: Uma lista de nomes completos ou parciais de módulos para os quais a quantização não deve ser realizada. Por exemplo, para não quantizar os blocos de projeção pos_embed do <a href="/docs/diffusers/main/pt/api/models/sd3_transformer2d#diffusers.SD3Transformer2DModel">SD3Transformer2DModel</a>, deve-se especificar: <code>modules_to_not_convert=["pos_embed.proj.weight"]</code>.</li> <li><code>disable_conv_quantization</code>: Um valor booleano que, quando definido como <code>True</code>, desativa a quantização para todas as camadas convolucionais no modelo. Isso é útil, pois a quantização por canal e por bloco geralmente não funciona bem com camadas convolucionais (usado com INT4, NF4, NVFP4). Se você deseja desativar a quantização para camadas convolucionais específicas, use <code>modules_to_not_convert</code> em vez disso.</li> <li><code>algorithm</code>: O algoritmo a ser usado para determinar a escala, padrão é <code>"max"</code>. Você pode consultar a documentação do modelopt para mais algoritmos e detalhes.</li> <li><code>forward_loop</code>: A função do loop de forward a ser usada para calibrar a ativação durante a quantização. Se não for fornecida, ela se baseia em valores de escala estáticos calculados usando apenas os pesos.</li> <li><code>kwargs</code>: Um dicionário de argumentos por palavra-chave para passar ao método de quantização subjacente, que será chamado com base em <code>quant_type</code>.</li></ul> <!--[1--><h2 class="relative group"><a id="supported-quantization-types" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#supported-quantization-types"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span><!---->Tipos de quantização suportados<!----></span></h2><!--]--><!----> <p>O ModelOpt suporta quantização int8, fp8, int4, nf4 e nvfp4 apenas de pesos, por canal e por bloco. Os métodos de quantização são projetados para reduzir a pegada de memória dos pesos do modelo, mantendo o desempenho do modelo durante a inferência.</p> <p>A quantização apenas de pesos armazena os pesos do modelo em um tipo de dados de baixa bit, mas executa o cálculo com um tipo de dados de maior precisão, como <code>bfloat16</code>. Isso reduz os requisitos de memória dos pesos do modelo, mas mantém os picos de memória para o cálculo de ativações.</p> <p>Os métodos de quantização suportados são os seguintes:</p> <table><thead><tr><th><strong>Tipo de Quantização</strong></th><th><strong>Esquemas suportados</strong></th><th><strong>Kwargs obrigatórios</strong></th><th><strong>Observações adicionais</strong></th></tr></thead><tbody><tr><td><strong>INT8</strong></td><td><code>int8 weight only</code>, <code>int8 channel quantization</code>, <code>int8 block quantization</code></td><td><code>quant_type</code>, <code>quant_type + channel_quantize</code>, <code>quant_type + channel_quantize + block_quantize</code></td><td></td></tr><tr><td><strong>FP8</strong></td><td><code>fp8 weight only</code>, <code>fp8 channel quantization</code>, <code>fp8 block quantization</code></td><td><code>quant_type</code>, <code>quant_type + channel_quantize</code>, <code>quant_type + channel_quantize + block_quantize</code></td><td></td></tr><tr><td><strong>INT4</strong></td><td><code>int4 weight only</code>, <code>int4 block quantization</code></td><td><code>quant_type</code>, <code>quant_type + channel_quantize + block_quantize</code></td><td><code>channel_quantize = -1 is only supported for now</code></td></tr><tr><td><strong>NF4</strong></td><td><code>nf4 weight only</code>, <code>nf4 double block quantization</code></td><td><code>quant_type</code>, <code>quant_type + channel_quantize + block_quantize + scale_channel_quantize</code> + <code>scale_block_quantize</code></td><td><code>channel_quantize = -1 and scale_channel_quantize = -1 are only supported for now</code></td></tr><tr><td><strong>NVFP4</strong></td><td><code>nvfp4 weight only</code>, <code>nvfp4 block quantization</code></td><td><code>quant_type</code>, <code>quant_type + channel_quantize + block_quantize</code></td><td><code>channel_quantize = -1 is only supported for now</code></td></tr></tbody></table> <p>Consulte a <a href="https://nvidia.github.io/Model-Optimizer/" rel="nofollow">documentação oficial do modelopt</a> para uma melhor compreensão dos métodos de quantização disponíveis e da lista exaustiva de opções de configuração disponíveis.</p> <!--[1--><h2 class="relative group"><a id="serializing-and-deserializing-quantized-models" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#serializing-and-deserializing-quantized-models"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span><!---->Serialização e desserialização de modelos quantizados<!----></span></h2><!--]--><!----> <p>Para serializar um modelo quantizado em um dtype específico, primeiro carregue o modelo com o dtype de quantização desejado e, em seguida, salve-o usando o método <a href="/docs/diffusers/main/pt/api/models/overview#diffusers.ModelMixin.save_pretrained">save_pretrained()</a>.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Copiado</div><!----></button><!----></div> <pre class="language-python "><!----><span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> AutoModel, NVIDIAModelOptConfig
<span class="hljs-keyword">from</span> modelopt.torch.opt <span class="hljs-keyword">import</span> enable_huggingface_checkpointing
enable_huggingface_checkpointing()
model_id = <span class="hljs-string">&quot;Efficient-Large-Model/Sana_600M_1024px_diffusers&quot;</span>
quant_config_fp8 = {<span class="hljs-string">&quot;quant_type&quot;</span>: <span class="hljs-string">&quot;FP8&quot;</span>, <span class="hljs-string">&quot;quant_method&quot;</span>: <span class="hljs-string">&quot;modelopt&quot;</span>}
quant_config_fp8 = NVIDIAModelOptConfig(**quant_config_fp8)
model = AutoModel.from_pretrained(
model_id,
subfolder=<span class="hljs-string">&quot;transformer&quot;</span>,
quantization_config=quant_config_fp8,
dtype=torch.bfloat16,
)
model.save_pretrained(<span class="hljs-string">&#x27;path/to/sana_fp8&#x27;</span>, safe_serialization=<span class="hljs-literal">False</span>)<!----></pre></div><!----> <p>Para carregar um modelo quantizado serializado, use o método <a href="/docs/diffusers/main/pt/api/models/overview#diffusers.ModelMixin.from_pretrained">from_pretrained()</a>.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Copiado</div><!----></button><!----></div> <pre class="language-python "><!----><span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> AutoModel, NVIDIAModelOptConfig, SanaPipeline
<span class="hljs-keyword">from</span> modelopt.torch.opt <span class="hljs-keyword">import</span> enable_huggingface_checkpointing
enable_huggingface_checkpointing()
quantization_config = NVIDIAModelOptConfig(quant_type=<span class="hljs-string">&quot;FP8&quot;</span>, quant_method=<span class="hljs-string">&quot;modelopt&quot;</span>)
transformer = AutoModel.from_pretrained(
<span class="hljs-string">&quot;path/to/sana_fp8&quot;</span>,
subfolder=<span class="hljs-string">&quot;transformer&quot;</span>,
quantization_config=quantization_config,
dtype=torch.bfloat16,
)
pipe = SanaPipeline.from_pretrained(
<span class="hljs-string">&quot;Efficient-Large-Model/Sana_600M_1024px_diffusers&quot;</span>,
transformer=transformer,
dtype=torch.bfloat16,
)
pipe.to(<span class="hljs-string">&quot;cuda&quot;</span>) <span class="hljs-comment"># or &quot;mps&quot;, &quot;xpu&quot;, &quot;cpu&quot;</span>
prompt = <span class="hljs-string">&quot;A cat holding a sign that says hello world&quot;</span>
image = pipe(
prompt, num_inference_steps=<span class="hljs-number">50</span>, guidance_scale=<span class="hljs-number">4.5</span>, max_sequence_length=<span class="hljs-number">512</span>
).images[<span class="hljs-number">0</span>]
image.save(<span class="hljs-string">&quot;output.png&quot;</span>)<!----></pre></div><!----> <a class="!text-gray-400 !no-underline text-sm flex items-center not-prose mt-4" href="https://github.com/huggingface/diffusers/blob/main/docs/source/en/quantization/modelopt.md" target="_blank"><svg class="mr-1" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M31,16l-7,7l-1.41-1.41L28.17,16l-5.58-5.59L24,9l7,7z"></path><path d="M1,16l7-7l1.41,1.41L3.83,16l5.58,5.59L8,23l-7-7z"></path><path d="M12.419,25.484L17.639,6.552l1.932,0.518L14.351,26.002z"></path></svg><!----> <span><span class="underline">Update</span> on GitHub</span></a><!----> <p></p><!--]--><!----><!--]--><!--]--><!--]--> <!--[-1--><!--]--><!--]-->
<script>
{
__sveltekit_1qsdbwf = {
base: "/docs/diffusers/main/en",
assets: "/docs/diffusers/main/en"
};
const element = document.currentScript.parentElement;
Promise.all([
import("/docs/diffusers/main/en/_app/immutable/entry/start.0S0_iEoG.js"),
import("/docs/diffusers/main/en/_app/immutable/entry/app.CPdOxCbw.js")
]).then(([kit, app]) => {
kit.start(app, element, {
node_ids: [0, 354],
data: [null,null],
form: null,
error: null
});
});
}
</script>

Xet Storage Details

Size:
26.4 kB
·
Xet hash:
ef77c4bbd691ccc46f369e11217000912ce09740db8b798778b57afbfe3bf327

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