Buckets:

hf-doc-build
/

doc

Files
xet
hf-doc-build/doc / diffusers /v0.43.0 /zh /optimization /context_parallelism.html
mishig
about 2 hours ago
download
raw
39.9 kB
<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;上下文并行&quot;,&quot;local&quot;:&quot;context-parallelism&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;环形注意力&quot;,&quot;local&quot;:&quot;ring-attention&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Ulysses 注意力&quot;,&quot;local&quot;:&quot;ulysses-attention&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;统一注意力&quot;,&quot;local&quot;:&quot;unified-attention&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Ulysses 任意注意力&quot;,&quot;local&quot;:&quot;ulysses-anything-attention&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Ring Anything Attention&quot;,&quot;local&quot;:&quot;ring-anything-attention&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;在加载时启用上下文并行&quot;,&quot;local&quot;:&quot;enable-context-parallelism-while-loading&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/333.OuTuM9lH.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/chunks/DSPDl1gz.js" rel="modulepreload">
		<!--ty2c99--><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;Context parallelism&quot;,&quot;local&quot;:&quot;context-parallelism&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Ring Attention&quot;,&quot;local&quot;:&quot;ring-attention&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Ulysses Attention&quot;,&quot;local&quot;:&quot;ulysses-attention&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Unified Attention&quot;,&quot;local&quot;:&quot;unified-attention&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Ulysses Anything Attention&quot;,&quot;local&quot;:&quot;ulysses-anything-attention&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Ring Anything Attention&quot;,&quot;local&quot;:&quot;ring-anything-attention&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Enable context parallelism while loading&quot;,&quot;local&quot;:&quot;enable-context-parallelism-while-loading&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="context-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="#context-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><!---->上下文并行<!----></span></h1><!--]--><!----> <p><a href="https://huggingface.co/spaces/nanotron/ultrascale-playbook?section=context_parallelism" rel="nofollow">上下文并行</a>将输入序列拆分到多个 GPU 上，每个 GPU 处理其对应的分片。这降低了每个 GPU 的内存占用，对于高分辨率图像和视频等长序列尤为有益。要将其与其他策略进行比较，请参阅<a href="../training/distributed_inference">分布式推理</a>。</p> <p>上下文并行在注意力后端内部运行，且只有部分后端支持该功能。如果您的后端不支持，<code>enable_parallelism</code> 将抛出一个错误，其中列出了兼容的后端。使用 <a href="/docs/diffusers/v0.43.0/zh/api/models/overview#diffusers.ModelMixin.set_attention_backend">set_attention_backend()</a> 切换后端，并参阅 <a href="./attention_backends#available-backends">可用后端</a> 以获取完整列表。</p> <p>使用 <a href="/docs/diffusers/v0.43.0/zh/api/parallel#diffusers.ContextParallelConfig">ContextParallelConfig</a> 配置以下每个变体。</p> <table><thead><tr><th>变体</th><th>配置</th><th>要求</th><th>使用时</th></tr></thead><tbody><tr><td><a href="#ring-attention">环形注意力</a></td><td><code>ring_degree > 1</code></td><td>序列长度必须能被<code>ring_degree</code>整除</td><td>序列较长且互连带宽有限</td></tr><tr><td><a href="#ulysses-attention">Ulysses 注意力</a></td><td><code>ulysses_degree > 1</code></td><td>序列长度和注意力头数必须能被<code>ulysses_degree</code>整除</td><td>您的 GPU 配备了高速互连</td></tr><tr><td><a href="#unified-attention">统一注意力</a></td><td><code>ring_degree > 1</code> 且 <code>ulysses_degree > 1</code></td><td>至少 4 个 GPU</td><td>你拥有的设备数量超过了 Ulysses Attention 单独能够使用的数量</td></tr><tr><td><a href="#ulysses-anything-attention">Ulysses 任意注意力</a></td><td><code>ulysses_degree > 1</code>, <code>ulysses_anything=True</code></td><td><code>ring_degree</code> 必须为 1</td><td>序列长度或头数无法整除</td></tr><tr><td><a href="#ring-anything-attention">Ring Anything Attention</a></td><td><code>ring_degree > 1</code>, <code>ring_anything=True</code></td><td><code>ulysses_degree</code> 必须为 1，仅用于推理，<code>attn_mask</code> 必须为 <code>None</code></td><td>序列长度无法被整除</td></tr></tbody></table> <!--[1--><h2 class="relative group"><a id="ring-attention" 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="#ring-attention"><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><!---->环形注意力<!----></span></h2><!--]--><!----> <p><a href="https://huggingface.co/papers/2310.01889" rel="nofollow">Ring Attention</a> 在环形排列的 GPU 之间传递键（K）和值（V）块。每个 GPU 计算其持有的 K/V 块的注意力，然后将该块传递给下一个 GPU。经过一轮完整的循环后，每个 GPU 都已处理过所有 token 的 K/V，且没有任何 GPU 持有完整序列。</p> <p>将 <code>ring_degree</code> 设置为 GPU 数量，并将该配置传递给 transformer 的 <code>enable_parallelism()</code> 方法。</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> 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, ContextParallelConfig

<span class="hljs-keyword">def</span> <span class="hljs-title function_">setup_distributed</span>():
    <span class="hljs-keyword">if</span> <span class="hljs-keyword">not</span> dist.is_initialized():
        dist.init_process_group(backend=<span class="hljs-string">&quot;nccl&quot;</span>)
    <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-keyword">return</span> device

<span class="hljs-keyword">def</span> <span class="hljs-title function_">main</span>():
    device = setup_distributed()
    world_size = dist.get_world_size()

    pipeline = DiffusionPipeline.from_pretrained(
        <span class="hljs-string">&quot;black-forest-labs/FLUX.1-dev&quot;</span>, dtype=torch.bfloat16
    ).to(device)
    pipeline.transformer.set_attention_backend(<span class="hljs-string">&quot;_native_cudnn&quot;</span>)

    cp_config = ContextParallelConfig(ring_degree=world_size)
    pipeline.transformer.enable_parallelism(config=cp_config)

    prompt = <span class="hljs-string">&quot;&quot;&quot;
    cinematic film still of a cat sipping a margarita in a pool in Palm Springs, California
    highly detailed, high budget hollywood movie, cinemascope, moody, epic, gorgeous, film grain
    &quot;&quot;&quot;</span>

    <span class="hljs-comment"># Must specify generator so all ranks start with same latents (or pass your own)</span>
    generator = torch.Generator().manual_seed(<span class="hljs-number">42</span>)
    image = pipeline(
        prompt,
        guidance_scale=<span class="hljs-number">3.5</span>,
        num_inference_steps=<span class="hljs-number">50</span>,
        generator=generator,
    ).images[<span class="hljs-number">0</span>]

    <span class="hljs-keyword">if</span> dist.get_rank() == <span class="hljs-number">0</span>:
        image.save(<span class="hljs-string">&quot;output.png&quot;</span>)

    <span class="hljs-keyword">if</span> dist.is_initialized():
        dist.destroy_process_group()


<span class="hljs-keyword">if</span> __name__ == <span class="hljs-string">&quot;__main__&quot;</span>:
    main()<!----></pre></div><!----> <p>将脚本保存为 <code>ring_attention.py</code>，并使用 <a href="https://docs.pytorch.org/docs/stable/elastic/run.html" rel="nofollow">torchrun</a> 或其他兼容 PyTorch 的启动器运行。将 <code>--nproc-per-node</code> 设置为 GPU 数量。</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 "><!---->torchrun --nproc-per-node 2 ring_attention.py<!----></pre></div><!----> <p>若要尝试其他变体，请将 <code>ring_attention.py</code> 中的 <code>enable_parallelism</code> 调用替换为对应章节中的调用。</p> <!--[1--><h2 class="relative group"><a id="ulysses-attention" 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="#ulysses-attention"><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><!---->Ulysses 注意力<!----></span></h2><!--]--><!----> <p><a href="https://huggingface.co/papers/2309.14509" rel="nofollow">Ulysses Attention</a> uses <em>all-to-all</em> communication, where every GPU exchanges data with every other GPU. The first all-to-all regroups the data so each GPU holds every token for a subset of attention heads. Each GPU computes attention for its heads, and a second all-to-all regroups the results by token for the next layer.</p> <p>将 <code>ulysses_degree</code> 设置为 GPU 数量，并将该配置传递给 <code>enable_parallelism()</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> 已复制</div><!----></button><!----></div> <pre class="language-py "><!---->pipeline.transformer.enable_parallelism(config=ContextParallelConfig(ulysses_degree=<span class="hljs-number">2</span>))<!----></pre></div><!----> <!--[1--><h2 class="relative group"><a id="unified-attention" 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="#unified-attention"><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><!---->统一注意力<!----></span></h2><!--]--><!----> <p><a href="https://huggingface.co/papers/2405.07719" rel="nofollow">Unified Sequence Parallelism</a> 结合了 Ring Attention 和 Ulysses Attention。它首先执行 all-to-all 以重新分配 heads 和 tokens，然后对结果运行 Ring Attention，最后反向执行 all-to-all 以恢复原始布局。Ulysses 在 attention heads 维度上进行并行，而 Ring 在序列维度上进行并行，因此两者结合后可以比单独使用 Ulysses 利用更多的 GPU。</p> <p>当拥有至少 4 个 GPU 时，请使用 Unified Attention。将 <code>ulysses_degree</code> 和 <code>ring_degree</code> 均设置为大于 1 的值。这些 GPU 会组成一个由 <code>ulysses_degree * ring_degree</code> 个设备构成的二维网格，每个 GPU 属于一个 Ulysses 组和一个 Ring 组。</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 "><!---->pipeline.transformer.enable_parallelism(config=ContextParallelConfig(ulysses_degree=<span class="hljs-number">2</span>, ring_degree=<span class="hljs-number">2</span>))<!----></pre></div><!----> <p>下方的基准测试使用 <a href="https://github.com/huggingface/diffusers/pull/12693#issuecomment-3694727532" rel="nofollow">此脚本</a> 在 4 块 H100 GPU 上比较了 Ulysses、Ring 和 Unified Attention。每张图像使用 50 个推理步骤。</p> <table><thead><tr><th>CP 后端</th><th>时间 / 图像 (毫秒)</th><th>步数 / 秒</th><th>峰值内存（GB）</th></tr></thead><tbody><tr><td>ulysses</td><td>6670.789</td><td>7.50</td><td>33.85</td></tr><tr><td>ring</td><td>13076.492</td><td>3.82</td><td>56.02</td></tr><tr><td>unified_balanced</td><td>11068.705</td><td>4.52</td><td>33.85</td></tr></tbody></table> <p>Ulysses 的吞吐量最高，但它使用的 GPU 数量不能超过模型的注意力头数。Unified Attention 消除了这一限制。</p> <!--[1--><h2 class="relative group"><a id="ulysses-anything-attention" 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="#ulysses-anything-attention"><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><!---->Ulysses 任意注意力<!----></span></h2><!--]--><!----> <p>Ulysses Attention 要求序列长度和注意力头数都能被 <code>ulysses_degree</code> 整除。<a href="https://github.com/huggingface/diffusers/pull/12996" rel="nofollow">Ulysses Anything Attention</a> 移除了这两项要求。它不能与 Unified Attention 组合使用，因此 <code>ring_degree</code> 必须为 1。</p> <p>将 <code>ulysses_degree</code> 设置为大于 1，并将 <code>ulysses_anything=True</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> 已复制</div><!----></button><!----></div> <pre class="language-py "><!---->pipeline.transformer.enable_parallelism(config=ContextParallelConfig(ulysses_degree=<span class="hljs-number">2</span>, ulysses_anything=<span class="hljs-literal">True</span>))<!----></pre></div><!----> <blockquote class="tip"><p>为 <a href="https://docs.pytorch.org/docs/stable/distributed.html#torch.distributed.init_process_group" rel="nofollow">init_process_group</a> 添加一个用于 CPU 张量的 <code>gloo</code> 后端。它可以避免主机到设备和设备到主机传输所导致的强制 CUDA 同步。</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">import</span> torch.distributed <span class="hljs-keyword">as</span> dist

dist.init_process_group(backend=<span class="hljs-string">&quot;cpu:gloo,cuda:nccl&quot;</span>)<!----></pre></div><!----> <p>以下基准测试在 4 块 L20 GPU 上使用<a href="https://github.com/huggingface/diffusers/pull/12996#issuecomment-3797695999" rel="nofollow">此脚本</a>运行 FLUX.1-dev。</p> <table><thead><tr><th>CP 后端</th><th>时间 / 迭代次数（毫秒）</th><th>步数 / 秒</th><th>峰值内存（GB）</th><th>形状 (HxW)</th></tr></thead><tbody><tr><td>ulysses</td><td>281.07</td><td>3.56</td><td>37.11</td><td>1024x1024</td></tr><tr><td>ring</td><td>351.34</td><td>2.85</td><td>37.01</td><td>1024x1024</td></tr><tr><td>unified_balanced</td><td>324.37</td><td>3.08</td><td>37.16</td><td>1024x1024</td></tr><tr><td>ulysses_anything</td><td>280.94</td><td>3.56</td><td>37.11</td><td>1024x1024</td></tr><tr><td>ulysses</td><td>失败</td><td>失败</td><td>失败</td><td>1008x1008</td></tr><tr><td>ring</td><td>失败</td><td>失败</td><td>失败</td><td>1008x1008</td></tr><tr><td>unified_balanced</td><td>失败</td><td>失败</td><td>失败</td><td>1008x1008</td></tr><tr><td>ulysses_anything</td><td>278.40</td><td>3.59</td><td>36.99</td><td>1008x1008</td></tr></tbody></table> <p>在 1008x1008 分辨率下，由于序列长度无法整除，仅 Ulysses Anything Attention 可以运行。在 1024x1024 分辨率下，其性能与标准 Ulysses Attention 相同。</p> <!--[1--><h2 class="relative group"><a id="ring-anything-attention" 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="#ring-anything-attention"><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><!---->Ring Anything Attention<!----></span></h2><!--]--><!----> <p>Ring Attention 要求序列长度能被 <code>ring_degree</code> 整除。<a href="https://github.com/huggingface/diffusers/pull/13545#issuecomment-4302195582" rel="nofollow">Ring Anything Attention</a> 支持任意序列长度。它会将每个 GPU 上的本地 K/V 填充至最长的本地长度，对所有填充后的 K/V 执行 all-gather，然后在计算注意力之前将每个块切回其真实长度。</p> <p>它无法与 Unified Attention 结合使用，因此 <code>ulysses_degree</code> 必须为 1。请将 <code>ring_degree</code> 设置为大于 1，并设置 <code>ring_anything=True</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> 已复制</div><!----></button><!----></div> <pre class="language-py "><!---->pipeline.transformer.enable_parallelism(config=ContextParallelConfig(ring_degree=<span class="hljs-number">2</span>, ring_anything=<span class="hljs-literal">True</span>))<!----></pre></div><!----> <p>Ring Anything Attention 同样受益于 <a href="#ulysses-anything-attention">Ulysses Anything Attention</a> 中展示的 <code>gloo</code> 后端。</p> <blockquote class="note"><p>Ring Anything Attention 仅支持推理，且 <code>attn_mask</code> 必须为 <code>None</code>。</p></blockquote> <p>以下基准测试在 4 块 RTX 4090 (48GB) GPU 上运行 FLUX.1-dev。</p> <table><thead><tr><th>CP 后端</th><th>时间 / 迭代次数（毫秒）</th><th>步数 / 秒</th><th>峰值内存（GB）</th><th>形状 (HxW)</th></tr></thead><tbody><tr><td>ulysses</td><td>259.07</td><td>3.86</td><td>33.83</td><td>1024x1024</td></tr><tr><td>ring</td><td>338.98</td><td>2.95</td><td>33.83</td><td>1024x1024</td></tr><tr><td>unified_balanced</td><td>321.54</td><td>3.11</td><td>33.83</td><td>1024x1024</td></tr><tr><td>ulysses_anything</td><td>259.07</td><td>3.86</td><td>33.83</td><td>1024x1024</td></tr><tr><td>ring_anything</td><td>340.14</td><td>2.94</td><td>33.83</td><td>1024x1024</td></tr><tr><td>ulysses</td><td>失败</td><td>失败</td><td>失败</td><td>1008x1008</td></tr><tr><td>ring</td><td>失败</td><td>失败</td><td>失败</td><td>1008x1008</td></tr><tr><td>unified_balanced</td><td>失败</td><td>失败</td><td>失败</td><td>1008x1008</td></tr><tr><td>ulysses_anything</td><td>253.16</td><td>3.95</td><td>33.75</td><td>1008x1008</td></tr><tr><td>ring_anything</td><td>335.57</td><td>2.98</td><td>33.75</td><td>1008x1008</td></tr></tbody></table> <p>Ring Anything Attention 在 1008x1008 分辨率下运行，性能与标准 Ring Attention 大致相当。有关动机和权衡，请参阅 <a href="https://github.com/huggingface/diffusers/pull/13545#issuecomment-4304104462" rel="nofollow">此评论</a>。</p> <!--[1--><h2 class="relative group"><a id="enable-context-parallelism-while-loading" 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="#enable-context-parallelism-while-loading"><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><!---->在加载时启用上下文并行<!----></span></h2><!--]--><!----> <p>与其在加载后调用 <code>enable_parallelism</code>，不如将 <a href="/docs/diffusers/v0.43.0/zh/api/parallel#diffusers.ContextParallelConfig">ContextParallelConfig</a> 传递给 <a href="/docs/diffusers/v0.43.0/zh/api/models/overview#diffusers.ModelMixin.from_pretrained">from_pretrained()</a> 的 <code>parallel_config</code> 参数。请先初始化进程组，因为如果 <code>torch.distributed</code> 未初始化，<code>from_pretrained</code> 会引发错误。请像 <a href="#ring-attention">Ring Attention</a> 示例中那样，使用 <code>torchrun</code> 启动脚本。</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> 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> AutoModel, ContextParallelConfig, DiffusionPipeline

CKPT_ID = <span class="hljs-string">&quot;black-forest-labs/FLUX.1-dev&quot;</span>

dist.init_process_group(backend=<span class="hljs-string">&quot;nccl&quot;</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)

cp_config = ContextParallelConfig(ring_degree=dist.get_world_size())
transformer = AutoModel.from_pretrained(
    CKPT_ID,
    subfolder=<span class="hljs-string">&quot;transformer&quot;</span>,
    dtype=torch.bfloat16,
    parallel_config=cp_config
)

pipeline = DiffusionPipeline.from_pretrained(
    CKPT_ID, transformer=transformer, dtype=torch.bfloat16,
).to(device)<!----></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/optimization/context_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, 333],
							data: [null,null],
							form: null,
							error: null
						});
					});
				}
			</script>
		
Xet Storage Details

Size:: 39.9 kB
Xet hash:: 17c0b7e07217d6212aa5dc5002d882225ab6ba60b4317f6791305d98a70261ac
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.