Buckets:

hf-doc-build
/

doc

Files
xet
hf-doc-build/doc / diffusers /v0.43.0 /pt /optimization /tensor_parallelism.html
mishig
2 days ago
download
raw
39 kB
<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;Paralelismo de tensor&quot;,&quot;local&quot;:&quot;tensor-parallelism&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Carregar um modelo fragmentado&quot;,&quot;local&quot;:&quot;load-a-sharded-model&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Desempenho de carregamento&quot;,&quot;local&quot;:&quot;loading-performance&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Limitações&quot;,&quot;local&quot;:&quot;limitations&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Adicionar paralelismo de tensor a um modelo&quot;,&quot;local&quot;:&quot;add-tensor-parallelism-to-a-model&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Colwise e rowwise&quot;,&quot;local&quot;:&quot;colwise-and-rowwise&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Projeções fundidas&quot;,&quot;local&quot;:&quot;fused-projections&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;O que deixar de fora&quot;,&quot;local&quot;:&quot;what-to-leave-out&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Restrições e verificação&quot;,&quot;local&quot;:&quot;constraints-and-verification&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3}],&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/347.DnqzUbRP.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/chunks/DSPDl1gz.js" rel="modulepreload">
		<!--17gkrup--><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;Tensor parallelism&quot;,&quot;local&quot;:&quot;tensor-parallelism&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Load a sharded model&quot;,&quot;local&quot;:&quot;load-a-sharded-model&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Loading performance&quot;,&quot;local&quot;:&quot;loading-performance&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Limitations&quot;,&quot;local&quot;:&quot;limitations&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Add tensor parallelism to a model&quot;,&quot;local&quot;:&quot;add-tensor-parallelism-to-a-model&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Colwise and rowwise&quot;,&quot;local&quot;:&quot;colwise-and-rowwise&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Fused projections&quot;,&quot;local&quot;:&quot;fused-projections&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;What to leave out&quot;,&quot;local&quot;:&quot;what-to-leave-out&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Constraints and verification&quot;,&quot;local&quot;:&quot;constraints-and-verification&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3}],&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="tensor-parallelism" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#tensor-parallelism"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span><!---->Paralelismo de tensor<!----></span></h1><!--]--><!----> <p><a href="https://huggingface.co/spaces/nanotron/ultrascale-playbook?section=tensor_parallelism" rel="nofollow">Tensor parallelism</a> distribui as matrizes de pesos de um modelo entre os dispositivos. Cada dispositivo mantém uma fatia por coluna (<code>"colwise"</code>) ou por linha (<code>"rowwise"</code>) de cada camada e calcula um resultado parcial. Um <code>AllReduce</code> ou <code>AllGather</code> na fronteira da camada combina os resultados parciais para formar a saída completa. Diferentemente do <a href="./context_parallelism">context parallelism</a>, ele reduz a memória de <em>weight</em> por dispositivo, o que ajuda quando um modelo não cabe em um único dispositivo. Consulte <a href="../training/distributed_inference">Distributed inference</a> para compará-lo com as outras estratégias.</p> <!--[1--><h2 class="relative group"><a id="load-a-sharded-model" 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="#load-a-sharded-model"><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><!---->Carregar um modelo fragmentado<!----></span></h2><!--]--><!----> <p>Passe um <a href="/docs/diffusers/v0.43.0/pt/api/parallel#diffusers.TensorParallelConfig">TensorParallelConfig</a> para o argumento <code>parallel_config</code> do <a href="/docs/diffusers/v0.43.0/pt/api/models/overview#diffusers.ModelMixin.from_pretrained">from_pretrained()</a> do modelo. <code>tp_degree</code> é o número de dispositivos para os quais o modelo será dividido e deve dividir o número de cabeças de atenção do modelo. A classe do modelo deve definir um <code>_tp_plan</code>, que mapeia nomes de módulos para um estilo de divisão. Para verificar se um modelo suporta paralelismo de tensor, procure por esse atributo em sua classe, como <a href="/docs/diffusers/v0.43.0/pt/api/models/flux2_transformer#diffusers.Flux2Transformer2DModel">Flux2Transformer2DModel</a>.</p> <p>Para adicionar um plano a outro modelo, consulte <a href="#add-tensor-parallelism-to-a-model">Adicionar paralelismo de tensor a um modelo</a>.</p> <p>Carregar dessa forma faz o checkpoint ser fragmentado conforme é lido. Cada rank lê apenas sua própria fatia de cada peso e a coloca diretamente em sua GPU, de modo que nenhum rank mantém um peso de tamanho completo. A memória por rank diminui à medida que <code>tp_degree</code> aumenta.</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-py "><!----><span class="hljs-keyword">import</span> os
<span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> torch <span class="hljs-keyword">import</span> distributed <span class="hljs-keyword">as</span> dist
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> DiffusionPipeline, Flux2Transformer2DModel, TensorParallelConfig

<span class="hljs-keyword">def</span> <span class="hljs-title function_">main</span>():
    dist.init_process_group(backend=<span class="hljs-string">&quot;nccl&quot;</span>)
    rank, world_size = dist.get_rank(), dist.get_world_size()
    <span class="hljs-comment"># torchrun sets LOCAL_RANK to this process&#x27;s GPU index on its node</span>
    local_rank = <span class="hljs-built_in">int</span>(os.environ[<span class="hljs-string">&quot;LOCAL_RANK&quot;</span>])
    device = torch.device(<span class="hljs-string">f&quot;cuda:<span class="hljs-subst">{local_rank}</span>&quot;</span>)
    torch.cuda.set_device(device)

    <span class="hljs-comment"># Each rank reads only its own shard of every planned weight, straight onto its own GPU.</span>
    transformer = Flux2Transformer2DModel.from_pretrained(
        <span class="hljs-string">&quot;black-forest-labs/FLUX.2-dev&quot;</span>,
        subfolder=<span class="hljs-string">&quot;transformer&quot;</span>,
        dtype=torch.bfloat16,
        parallel_config=TensorParallelConfig(tp_degree=world_size),
    )

    pipeline = DiffusionPipeline.from_pretrained(
        <span class="hljs-string">&quot;black-forest-labs/FLUX.2-dev&quot;</span>, transformer=transformer, dtype=torch.bfloat16
    )
    <span class="hljs-comment"># The transformer is already on its device; move the remaining components individually. Do not call</span>
    <span class="hljs-comment"># `pipeline.to(device)` — that would move every rank&#x27;s shards onto the same device.</span>
    pipeline.text_encoder.to(device)
    pipeline.vae.to(device)

    generator = torch.Generator().manual_seed(<span class="hljs-number">42</span>)
    image = pipeline(prompt=<span class="hljs-string">&quot;a cat holding a sign that says hello&quot;</span>, generator=generator).images[<span class="hljs-number">0</span>]
    <span class="hljs-keyword">if</span> rank == <span class="hljs-number">0</span>:
        image.save(<span class="hljs-string">&quot;output.png&quot;</span>)
    dist.destroy_process_group()

<span class="hljs-keyword">if</span> __name__ == <span class="hljs-string">&quot;__main__&quot;</span>:
    main()<!----></pre></div><!----> <p>Salve o script como <code>tensor_parallel_flux2.py</code> e execute-o com <code>torchrun</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> Copiado</div><!----></button><!----></div> <pre class="language-shell "><!---->torchrun --nproc-per-node 4 tensor_parallel_flux2.py<!----></pre></div><!----> <p>O script define <code>tp_degree</code> como <code>world_size</code>, portanto <code>--nproc-per-node 4</code> distribui o transformer em 4 GPUs.</p> <p>Apenas o transformer é fragmentado. Cada rank mantém uma cópia completa do codificador de texto, portanto, cada GPU precisa de memória suficiente para o codificador de texto mais sua parte do transformer.</p> <!--[1--><h2 class="relative group"><a id="loading-performance" 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="#loading-performance"><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><!---->Desempenho de carregamento<!----></span></h2><!--]--><!----> <p>Carregar com <code>parallel_config</code> é mais rápido e usa menos memória de CPU por rank do que carregar o modelo completo e, em seguida, chamar <code>enable_parallelism()</code>. A tabela abaixo mostra os resultados para <a href="https://huggingface.co/black-forest-labs/FLUX.2-dev" rel="nofollow">black-forest-labs/FLUX.2-dev</a> (somente o transformer, 32B de parâmetros, bf16) em 4x A10G (23GB).</p> <table><thead><tr><th>tp_degree</th><th>método</th><th>tempo de carregamento</th><th>CPU de pico/por rank</th></tr></thead><tbody><tr><td>4</td><td><code>from_pretrained(parallel_config=...)</code></td><td>12.5s</td><td>6.8GB</td></tr><tr><td>4</td><td><code>from_pretrained</code> + <code>enable_parallelism</code></td><td>30.4s</td><td>64.1GB</td></tr></tbody></table> <!--[1--><h2 class="relative group"><a id="limitations" 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="#limitations"><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><!---->Limitações<!----></span></h2><!--]--><!----> <p>O paralelismo de tensor não funciona com algumas opções e recursos de carregamento, e não é possível salvar um modelo fragmentado.</p> <ul><li>Carregar com um <code>parallel_config</code> de paralelismo de tensor não funciona com <code>device_map</code>, <code>quantization_config</code>, <code>low_cpu_mem_usage=False</code>, <code>use_flashpack=True</code> ou pesos que não sejam safetensors. Cada um deles gera um erro em vez de carregar silenciosamente o checkpoint completo.</li> <li>A combinação de paralelismo de tensor com quantização, offloading ou adaptadores LoRA também gera um erro, independentemente de como o modelo é particionado.</li> <li>Para dividir um modelo que já está na memória, chame <code>enable_parallelism()</code> com a mesma configuração. Ele carrega o modelo completo primeiro e depois o divide, então cada rank precisa de memória para o checkpoint completo. Carregar com <code>parallel_config</code> evita esse custo.</li> <li>Salvar um modelo paralelo de tensor gera um erro em <a href="/docs/diffusers/v0.43.0/pt/api/models/overview#diffusers.ModelMixin.save_pretrained">save_pretrained()</a>. Salve o modelo antes de fragmentá-lo.</li></ul> <!--[1--><h2 class="relative group"><a id="add-tensor-parallelism-to-a-model" 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="#add-tensor-parallelism-to-a-model"><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><!---->Adicionar paralelismo de tensor a um modelo<!----></span></h2><!--]--><!----> <p>Esta seção é destinada a contribuidores que desejam adicionar suporte a paralelismo de tensor a um modelo do Diffusers.</p> <p>O paralelismo de tensores funciona apenas em modelos que definem um <code>_tp_plan</code>, um atributo de classe que mapeia nomes de módulos para um estilo de particionamento. Escrever um geralmente significa associar cada projeção que <em>expande</em> a dimensão oculta com a projeção que a <em>contrai</em> de volta.</p> <p>Cada chave pode conter no máximo um <code>*</code>, e a parte antes do <code>*</code> deve resolver para um <a href="https://pytorch.org/docs/stable/generated/torch.nn.ModuleList.html" rel="nofollow"><code>nn.ModuleList</code></a>, de modo que uma entrada cubra todos os blocos. Uma chave sem <code>*</code> se aplica ao próprio modelo. Os caminhos são relativos ao modelo.</p> <!--[2--><h3 class="relative group"><a id="colwise-and-rowwise" 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="#colwise-and-rowwise"><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><!---->Colwise e rowwise<!----></span></h3><!--]--><!----> <table><thead><tr><th>Estilo</th><th>Shards</th><th>Cada rank</th><th>Usar para</th></tr></thead><tbody><tr><td><code>"colwise"</code></td><td>características de saída (dim 0 de <code>weight</code>)</td><td>calcula uma fatia da saída</td><td><code>to_q</code>, <code>to_k</code>, <code>to_v</code>, in-projection do FFN</td></tr><tr><td><code>"rowwise"</code></td><td>recursos de entrada (dim 1 de <code>weight</code>)</td><td>calcula uma soma parcial</td><td><code>to_out.0</code>, projeção de saída do FFN</td></tr><tr><td><code>"rowwise_split_input"</code></td><td>recursos de entrada (dim 1 de <code>weight</code>)</td><td>calcula uma soma parcial a partir de sua fatia da entrada replicada</td><td>uma projeção independente cuja entrada é replicada, como uma projeção de modulação AdaLN que lê o embedding do timestep</td></tr></tbody></table> <p>Sempre combine <code>"colwise"</code> e <code>"rowwise"</code> nessa ordem. Uma projeção <code>"colwise"</code> deixa sua saída fragmentada, a projeção <code>"rowwise"</code> subsequente consome esse fragmento diretamente e um único <code>AllReduce</code> na fronteira do bloco reconstrói o resultado. Fragmentar o par de qualquer outra forma força um gather no meio e comunica muito mais.</p> <p>Na atenção, as projeções de Q, K e V por coluna atribuem a cada rank um subconjunto de cabeças, portanto <code>tp_degree</code> deve dividir o número de cabeças. Os duplicados do stream do codificador (<code>add_q_proj</code>, <code>to_add_out</code>, <code>ff_context</code>) seguem o mesmo padrão de seus equivalentes no stream de imagem.</p> <!--[2--><h3 class="relative group"><a id="fused-projections" 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="#fused-projections"><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><!---->Projeções fundidas<!----></span></h3><!--]--><!----> <p>Quando um <code>Linear</code> empacota vários tensores lógicos ao longo da dimensão que está sendo fragmentada, fatias simples de <code>"colwise"</code>/<code>"rowwise"</code> atravessam a concatenação e desalinharam as peças. Use <code>PackedColwiseParallel</code>/<code>PackedRowwiseParallel</code> em vez disso, que fragmenta cada bloco empacotado de forma independente.</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-py "><!----><span class="hljs-comment"># in src/diffusers/models/transformers/your_model.py</span>
<span class="hljs-keyword">from</span> ...hooks.tensor_parallel <span class="hljs-keyword">import</span> PackedColwiseParallel, PackedRowwiseParallel<!----></pre></div><!----> <p>Pass <code>blocks</code>, uma lista de inteiros proporcionais cuja soma divide a dimensão empacotada. Por exemplo, use <code>[1, 1]</code> para uma projeção de gate e up do SwiGLU com metades iguais, ou <code>[1, 1, 1, 3, 3]</code> para uma projeção fundida de Q, K, V, gate e up com <code>mlp_ratio=3</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> Copiado</div><!----></button><!----></div> <pre class="language-py "><!----><span class="hljs-string">&quot;transformer_blocks.*.ff.linear_in&quot;</span>: PackedColwiseParallel([<span class="hljs-number">1</span>, <span class="hljs-number">1</span>]),<!----></pre></div><!----> <p>Quando os tamanhos dos blocos são conhecidos apenas a partir da configuração, omita <code>blocks</code> e armazene os tamanhos absolutos no <code>Linear</code> durante o <code>__init__</code>, como <code>_tp_packed_col_blocks</code> ou <code>_tp_packed_row_blocks</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> Copiado</div><!----></button><!----></div> <pre class="language-py "><!----><span class="hljs-comment"># in the attention module&#x27;s __init__</span>
<span class="hljs-variable language_">self</span>.to_out._tp_packed_row_blocks = [<span class="hljs-variable language_">self</span>.inner_dim, <span class="hljs-variable language_">self</span>.mlp_hidden_dim]<!----></pre></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> Copiado</div><!----></button><!----></div> <pre class="language-py "><!----><span class="hljs-comment"># in the model&#x27;s _tp_plan</span>
<span class="hljs-string">&quot;single_transformer_blocks.*.attn.to_out&quot;</span>: PackedRowwiseParallel(),<!----></pre></div><!----> <p>O plano em <a href="https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/transformers/transformer_flux2.py" rel="nofollow"><code>transformer_flux2.py</code></a> usa ambas as formas. O par SwiGLU de duplo fluxo passa <code>blocks</code> diretamente. A projeção QKV e MLP de fluxo único fundida armazena seus tamanhos no <code>Linear</code> em <code>Flux2ParallelSelfAttention.__init__</code>, porque eles dependem de <code>mlp_ratio</code> e <code>mlp_mult_factor</code>.</p> <!--[2--><h3 class="relative group"><a id="what-to-leave-out" 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="#what-to-leave-out"><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><!---->O que deixar de fora<!----></span></h3><!--]--><!----> <p>Módulos excluídos do plano são replicados em cada rank. Exclua as camadas de normalização, as incorporações de patch e de texto, e o <code>norm_out</code> e <code>proj_out</code> finais. Eles são pequenos, portanto, fragmentá-los poupa pouco memória e adiciona comunicação. A modulação AdaLN (<code>img_mod</code> e <code>txt_mod</code>) geralmente é pequena o suficiente para também ser excluída. Quando é grande, como o <code>adaln_proj.linear</code> do MiniMax-H3, fragmente-a com <code>"rowwise_split_input"</code>.</p> <!--[2--><h3 class="relative group"><a id="constraints-and-verification" 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="#constraints-and-verification"><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><!---->Restrições e verificação<!----></span></h3><!--]--><!----> <ul><li><code>tp_degree</code> deve dividir <code>config.num_attention_heads</code>.</li> <li>Cada bloco empacotado deve ser <em>individualmente</em> divisível por <code>tp_degree</code>, e não apenas sua soma.</li></ul> <p><a href="/docs/diffusers/v0.43.0/pt/api/models/overview#diffusers.ModelMixin.from_pretrained">from_pretrained()</a> e <code>enable_parallelism()</code> verificam ambos antes de carregar ou dividir qualquer peso.</p> <p>Valide um novo plano numericamente, em vez de a olho nu: gere com uma semente fixa em um único dispositivo e, em seguida, novamente sob paralelismo de tensor, e compare as saídas. Um <code>"colwise"</code>/<code>"rowwise"</code> mal posicionado geralmente ainda executa e produz uma imagem plausível, porém incorreta.</p> <blockquote class="tip"><p>Comece a partir de um plano existente para uma arquitetura semelhante.</p> <ul><li><a href="https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/transformers/transformer_qwenimage.py" rel="nofollow"><code>transformer_qwenimage.py</code></a> não possui projeções fundidas, portanto, cada entrada é apenas <code>"colwise"</code> ou <code>"rowwise"</code>.</li> <li><a href="https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/transformers/transformer_flux.py" rel="nofollow"><code>transformer_flux.py</code></a> adiciona uma projeção empacotada linha a linha.</li> <li><a href="https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/transformers/transformer_flux2.py" rel="nofollow"><code>transformer_flux2.py</code></a> usa ambos os estilos empacotados.</li></ul></blockquote> <a class="!text-gray-400 !no-underline text-sm flex items-center not-prose mt-4" href="https://github.com/huggingface/diffusers/blob/main/docs/source/en/optimization/tensor_parallelism.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, 347],
							data: [null,null],
							form: null,
							error: null
						});
					});
				}
			</script>
		
Xet Storage Details

Size:: 39 kB
Xet hash:: c6a22cd4c2a9e24855f12279a67c29c891a58b0a2bd9f014a5c22bb6fb762c6c
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.