Hugging Face's logo

Buckets:
hf-doc-build
/
doc

Files
xet
hf-doc-build/doc / diffusers /main /ko /optimization /context_parallelism.html
mishig's picture
mishig
download
raw
41.1 kB
<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;컨텍스트 병렬성&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;통합 어텐션&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;…&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><!---->컨텍스트 병렬성<!----></span></h1><!--]--><!----> <p><a href="https://huggingface.co/spaces/nanotron/ultrascale-playbook?section=context_parallelism" rel="nofollow">Context parallelism</a>은 입력 시퀀스를 여러 GPU에 분할하여 각 GPU가 자신의 슬라이스를 처리합니다. 이를 통해 GPU당 메모리 사용량이 줄어들며, 고해상도 이미지나 비디오와 같은 긴 시퀀스에서 가장 큰 효과를 발휘합니다. 다른 전략과의 비교는 <a href="../training/distributed_inference">Distributed inference</a>을 참고하세요.</p> <p>Context parallelism은 attention backend 내부에서 실행되며, 일부 backend만 이를 지원합니다. 사용 중인 backend가 지원하지 않는 경우, <code>enable_parallelism</code>는 호환되는 backend 목록을 나열하는 오류를 발생시킵니다. <a href="/docs/diffusers/main/ko/api/models/overview#diffusers.ModelMixin.set_attention_backend">set_attention_backend()</a>를 사용하여 backend를 전환할 수 있으며, 전체 목록은 <a href="./attention_backends#available-backends">사용 가능한 backend</a>를 참고하세요.</p> <p>아래의 모든 변형을 <a href="/docs/diffusers/main/ko/api/parallel#diffusers.ContextParallelConfig">ContextParallelConfig</a>로 구성합니다.</p> <table><thead><tr><th>변형</th><th>설정</th><th>요구 사항</th><th>사용 시</th></tr></thead><tbody><tr><td><a href="#ring-attention">링 어텐션</a></td><td><code>ring_degree > 1</code></td><td>시퀀스 길이는 <code>ring_degree</code>로 나누어떨어져야 합니다</td><td>시퀀스가 길고 인터커넥트 대역폭이 제한되어 있습니다</td></tr><tr><td><a href="#ulysses-attention">Ulysses Attention</a></td><td><code>ulysses_degree > 1</code></td><td>시퀀스 길이와 어텐션 헤드 수는 <code>ulysses_degree</code>로 나누어떨어져야 합니다</td><td>GPU에 빠른 인터커넥트가 있습니다</td></tr><tr><td><a href="#unified-attention">통합 어텐션</a></td><td><code>ring_degree > 1</code> 및 <code>ulysses_degree > 1</code></td><td>최소 4개의 GPU</td><td>Ulysses Attention 단독으로 사용할 수 있는 장치보다 더 많은 장치를 보유하고 있습니다</td></tr><tr><td><a href="#ulysses-anything-attention">Ulysses Anything Attention</a></td><td><code>ulysses_degree > 1</code>, <code>ulysses_anything=True</code></td><td><code>ring_degree</code>은 1이어야 합니다</td><td>시퀀스 길이 또는 헤드 수가 균등하게 나누어지지 않는 경우</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>는 1이어야 하며, 추론 전용이고, <code>attn_mask</code>는 <code>None</code>이어야 합니다</td><td>시퀀스 길이가 균등하게 나누어지지 않는 경우</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><!---->Ring Attention<!----></span></h2><!--]--><!----> <p><a href="https://huggingface.co/papers/2310.01889" rel="nofollow">Ring Attention</a>은 링으로 연결된 GPU 사이에서 키(K)와 값(V) 블록을 전달합니다. 각 GPU는 자신이 보유하고 있는 K/V 블록에 대해 어텐션을 계산한 후, 해당 블록을 다음 GPU로 전달합니다. 한 바퀴를 완전히 돌면 모든 GPU가 모든 토큰의 K/V를 확인하게 되며, 어떤 GPU도 전체 시퀀스를 보유하지 않습니다.</p> <p><code>ring_degree</code>를 GPU 수로 설정하고, 해당 설정을 transformer의 <code>enable_parallelism()</code> 메서드에 전달합니다.</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> 복사됨</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>스크립트를 <code>ring_attention.py</code>로 저장한 후 <a href="https://docs.pytorch.org/docs/stable/elastic/run.html" rel="nofollow">torchrun</a> 또는 다른 PyTorch 호환 런처로 실행합니다. <code>--nproc-per-node</code>를 GPU 수로 설정합니다.</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> 복사됨</div><!----></button><!----></div> <pre class="language-bash "><!---->torchrun --nproc-per-node 2 ring_attention.py<!----></pre></div><!----> <p>다른 변형을 사용해 보려면 <code>ring_attention.py</code>의 <code>enable_parallelism</code> 호출을 해당 섹션의 호출로 교체하세요.</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 Attention<!----></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><code>ulysses_degree</code>를 GPU 수로 설정하고, 해당 설정을 <code>enable_parallelism()</code>에 전달합니다.</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> 복사됨</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><!---->통합 어텐션<!----></span></h2><!--]--><!----> <p><a href="https://huggingface.co/papers/2405.07719" rel="nofollow">Unified Sequence Parallelism</a>은 Ring Attention과 Ulysses Attention을 결합합니다. all-to-all을 실행하여 head와 token을 재분배한 후, 그 결과에 Ring Attention을 실행하고, 다시 all-to-all을 역방향으로 실행하여 원래의 레이아웃을 복원합니다. Ulysses는 attention head를 따라 병렬화하고 Ring은 시퀀스를 따라 병렬화하므로, 두 기법을 함께 사용하면 Ulysses 단독 사용 시보다 더 많은 GPU를 활용할 수 있습니다.</p> <p>GPU가 4개 이상인 경우 Unified Attention을 사용하세요. <code>ulysses_degree</code>와 <code>ring_degree</code>를 모두 1보다 큰 값으로 설정합니다. GPU는 <code>ulysses_degree * ring_degree</code>개의 장치로 구성된 2D 그리드를 형성하며, 각 GPU는 하나의 Ulysses 그룹과 하나의 Ring 그룹에 속합니다.</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> 복사됨</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>아래 벤치마크는 <a href="https://github.com/huggingface/diffusers/pull/12693#issuecomment-3694727532" rel="nofollow">이 스크립트</a>를 사용하여 4개의 H100 GPU에서 Ulysses, Ring, Unified Attention을 비교합니다. 각 이미지는 50개의 추론 스텝을 사용합니다.</p> <table><thead><tr><th>CP 백엔드</th><th>시간 / 이미지 (ms)</th><th>초당 스텝</th><th>피크 메모리 (GB)</th></tr></thead><tbody><tr><td>ulysses</td><td>6670.789</td><td>7.50</td><td>33.85</td></tr><tr><td>링</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>Ulysses는 가장 높은 처리량을 제공하지만, 모델의 어텐션 헤드 수보다 많은 GPU를 사용할 수 없습니다. Unified Attention은 이러한 제한을 제거합니다.</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 Anything Attention<!----></span></h2><!--]--><!----> <p>Ulysses Attention은 시퀀스 길이와 어텐션 헤드 수가 <code>ulysses_degree</code>로 나누어떨어져야 합니다. <a href="https://github.com/huggingface/diffusers/pull/12996" rel="nofollow">Ulysses Anything Attention</a>은 이 두 가지 요구사항을 모두 제거합니다. Unified Attention과 함께 사용할 수 없으므로 <code>ring_degree</code>는 1이어야 합니다.</p> <p><code>ulysses_degree</code>를 1보다 큰 값으로 설정하고 <code>ulysses_anything=True</code>를 지정하세요.</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> 복사됨</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><a href="https://docs.pytorch.org/docs/stable/distributed.html#torch.distributed.init_process_group" rel="nofollow">init_process_group</a>에 CPU 텐서를 위한 <code>gloo</code> 백엔드를 추가합니다. 이를 통해 호스트-디바이스 및 디바이스-호스트 전송으로 인한 강제 CUDA 동기화를 피할 수 있습니다.</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> 복사됨</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>아래 벤치마크는 <a href="https://github.com/huggingface/diffusers/pull/12996#issuecomment-3797695999" rel="nofollow">이 스크립트</a>를 사용하여 4개의 L20 GPU에서 FLUX.1-dev를 실행합니다.</p> <table><thead><tr><th>CP 백엔드</th><th>시간 / 반복 (ms)</th><th>초당 스텝</th><th>피크 메모리 (GB)</th><th>Shape (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>링</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>실패</td><td>실패</td><td>실패</td><td>1008x1008</td></tr><tr><td>링</td><td>실패</td><td>실패</td><td>실패</td><td>1008x1008</td></tr><tr><td>unified_balanced</td><td>실패</td><td>실패</td><td>실패</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>1008x1008에서는 시퀀스 길이가 균등하게 나누어지지 않으므로 Ulysses Anything Attention만 실행됩니다. 1024x1024에서는 표준 Ulysses Attention과 동일한 성능을 보입니다.</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 Attention은 시퀀스 길이가 <code>ring_degree</code>로 나누어떨어져야 합니다. <a href="https://github.com/huggingface/diffusers/pull/13545#issuecomment-4302195582" rel="nofollow">Ring Anything Attention</a>은 모든 시퀀스 길이를 지원합니다. 각 GPU의 로컬 K/V를 가장 긴 로컬 길이로 패딩하고, 패딩된 K/V를 all-gather한 뒤, attention을 계산하기 전에 각 블록을 원래 길이로 슬라이싱합니다.</p> <p>Unified Attention과 함께 사용할 수 없으므로 <code>ulysses_degree</code>는 1이어야 합니다. <code>ring_degree</code>를 1보다 큰 값으로 설정하고 <code>ring_anything=True</code>를 지정하세요.</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> 복사됨</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 Attention은 또한 <a href="#ulysses-anything-attention">Ulysses Anything Attention</a>에 표시된 <code>gloo</code> 백엔드에서도 이점을 얻습니다.</p> <blockquote class="note"><p>Ring Anything Attention은 추론만 지원하며, <code>attn_mask</code>는 <code>None</code>이어야 합니다.</p></blockquote> <p>아래 벤치마크는 4개의 RTX 4090 (48GB) GPU에서 FLUX.1-dev를 실행합니다.</p> <table><thead><tr><th>CP 백엔드</th><th>시간 / 반복 (ms)</th><th>초당 스텝</th><th>피크 메모리 (GB)</th><th>Shape (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>링</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>실패</td><td>실패</td><td>실패</td><td>1008x1008</td></tr><tr><td>링</td><td>실패</td><td>실패</td><td>실패</td><td>1008x1008</td></tr><tr><td>unified_balanced</td><td>실패</td><td>실패</td><td>실패</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 Attention은 1008x1008에서 실행되며 표준 Ring Attention과 거의 동일한 성능을 보입니다. 동기와 트레이드오프에 대해서는 <a href="https://github.com/huggingface/diffusers/pull/13545#issuecomment-4304104462" rel="nofollow">이 댓글</a>을 참고하세요.</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><!---->…<!----></span></h2><!--]--><!----> <p><code>enable_parallelism</code>를 로드 후에 호출하는 대신, <a href="/docs/diffusers/main/ko/api/parallel#diffusers.ContextParallelConfig">ContextParallelConfig</a>를 <a href="/docs/diffusers/main/ko/api/models/overview#diffusers.ModelMixin.from_pretrained">from_pretrained()</a>의 <code>parallel_config</code> 인자로 전달하세요. <code>from_pretrained</code>는 <code>torch.distributed</code>가 초기화되지 않은 경우 오류를 발생시키므로, 먼저 프로세스 그룹을 초기화해야 합니다. <a href="#ring-attention">Ring Attention</a> 예제와 마찬가지로 스크립트를 <code>torchrun</code>으로 실행하세요.</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> 복사됨</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:
41.1 kB
·
Xet hash:
72bf278d52d3d27ee04e4d987fd69639ad81f7af58fbfa6d73a9fcd41cce92b6

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