Buckets:
| <meta charset="utf-8" /><meta name="hf:doc:metadata" content="{"title":"Параллелизм контекста","local":"context-parallelism","sections":[{"title":"Ring Attention","local":"ring-attention","sections":[],"depth":2},{"title":"Ulysses Attention","local":"ulysses-attention","sections":[],"depth":2},{"title":"Единое внимание","local":"unified-attention","sections":[],"depth":2},{"title":"Ulysses Anything Attention","local":"ulysses-anything-attention","sections":[],"depth":2},{"title":"Ring Anything Attention","local":"ring-anything-attention","sections":[],"depth":2},{"title":"Включить параллелизм контекста при загрузке","local":"enable-context-parallelism-while-loading","sections":[],"depth":2}],"depth":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="{"title":"Context parallelism","local":"context-parallelism","sections":[{"title":"Ring Attention","local":"ring-attention","sections":[],"depth":2},{"title":"Ulysses Attention","local":"ulysses-attention","sections":[],"depth":2},{"title":"Unified Attention","local":"unified-attention","sections":[],"depth":2},{"title":"Ulysses Anything Attention","local":"ulysses-anything-attention","sections":[],"depth":2},{"title":"Ring Anything Attention","local":"ring-anything-attention","sections":[],"depth":2},{"title":"Enable context parallelism while loading","local":"enable-context-parallelism-while-loading","sections":[],"depth":2}],"depth":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><!---->Параллелизм контекста<!----></span></h1><!--]--><!----> <p><a href="https://huggingface.co/spaces/nanotron/ultrascale-playbook?section=context_parallelism" rel="nofollow">Контекстная параллельность</a> распределяет входную последовательность между GPU, и каждый GPU обрабатывает свой срез. Это снижает потребление памяти на каждом GPU, что особенно полезно для длинных последовательностей, таких как изображения высокого разрешения и видео. Для сравнения с другими стратегиями см. <a href="../training/distributed_inference">Распределённый инференс</a>.</p> <p>Контекстный параллелизм выполняется внутри бэкенда внимания, и поддерживается только некоторыми из них. Если ваш бэкенд не поддерживает контекстный параллелизм, <code>enable_parallelism</code> вызывает ошибку, в которой перечислены совместимые бэкенды. Переключить бэкенд можно с помощью <a href="/docs/diffusers/v0.43.0/ru/api/models/overview#diffusers.ModelMixin.set_attention_backend">set_attention_backend()</a>, а полный список приведён в разделе <a href="./attention_backends#available-backends">доступные бэкенды</a>.</p> <p>Настройте каждый из вариантов ниже с помощью <a href="/docs/diffusers/v0.43.0/ru/api/parallel#diffusers.ContextParallelConfig">ContextParallelConfig</a>.</p> <table><thead><tr><th>Вариант</th><th>Конфигурация</th><th>Требования</th><th>Когда использовать</th></tr></thead><tbody><tr><td><a href="#ring-attention">Ring Attention</a></td><td><code>ring_degree > 1</code></td><td>длина последовательности должна делиться нацело на <code>ring_degree</code></td><td>последовательности длинные, а пропускная способность межсоединений ограничена</td></tr><tr><td><a href="#ulysses-attention">Ulysses Attention</a></td><td><code>ulysses_degree > 1</code></td><td>длина последовательности и количество головок внимания должны делиться нацело на <code>ulysses_degree</code></td><td>ваши GPU имеют высокоскоростное соединение</td></tr><tr><td><a href="#unified-attention">Единое внимание</a></td><td><code>ring_degree > 1</code> и <code>ulysses_degree > 1</code></td><td>не менее 4 GPU</td><td>у вас больше устройств, чем может использовать Ulysses Attention в одиночку</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> должен быть равен 1</td><td>длина последовательности или количество головок не делятся нацело</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> должен быть равен 1, только для инференса, <code>attn_mask</code> должен быть <code>None</code></td><td>длина последовательности не делится нацело</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> передаёт блоки ключей (K) и значений (V) между GPU, организованными в кольцо. Каждый GPU вычисляет внимание по своему блоку K/V, а затем передаёт этот блок следующему GPU. После одного полного цикла каждый GPU обработал K/V всех токенов, при этом ни один GPU не хранит полную последовательность.</p> <p>Установите <code>ring_degree</code> равным количеству GPU и передайте конфигурацию в метод <code>enable_parallelism()</code> трансформера.</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> Скопировано</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">"nccl"</span>) | |
| <span class="hljs-comment"># torchrun sets LOCAL_RANK to this process's GPU index on its node</span> | |
| local_rank = <span class="hljs-built_in">int</span>(os.environ[<span class="hljs-string">"LOCAL_RANK"</span>]) | |
| device = torch.device(<span class="hljs-string">f"cuda:<span class="hljs-subst">{local_rank}</span>"</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">"black-forest-labs/FLUX.1-dev"</span>, dtype=torch.bfloat16 | |
| ).to(device) | |
| pipeline.transformer.set_attention_backend(<span class="hljs-string">"_native_cudnn"</span>) | |
| cp_config = ContextParallelConfig(ring_degree=world_size) | |
| pipeline.transformer.enable_parallelism(config=cp_config) | |
| prompt = <span class="hljs-string">""" | |
| 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 | |
| """</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">"output.png"</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">"__main__"</span>: | |
| main()<!----></pre></div><!----> <p>Сохраните скрипт как <code>ring_attention.py</code> и запустите его с помощью <a href="https://docs.pytorch.org/docs/stable/elastic/run.html" rel="nofollow">torchrun</a> или другого совместимого с PyTorch средства запуска. Установите значение параметра <code>--nproc-per-node</code> равным количеству GPU.</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> Скопировано</div><!----></button><!----></div> <pre class="language-bash "><!---->torchrun --nproc-per-node 2 ring_attention.py<!----></pre></div><!----> <p>Чтобы попробовать другие варианты, замените вызов <code>enable_parallelism</code> в <code>ring_attention.py</code> на тот, что приведён в их разделе.</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>Установите <code>ulysses_degree</code> равным количеству GPU и передайте конфигурацию в <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> Скопировано</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><!---->Единое внимание<!----></span></h2><!--]--><!----> <p><a href="https://huggingface.co/papers/2405.07719" rel="nofollow">Unified Sequence Parallelism</a> объединяет Ring Attention и Ulysses Attention. Сначала выполняется all-to-all для перераспределения heads и tokens, затем на полученном результате запускается Ring Attention, после чего all-to-all выполняется в обратном направлении для восстановления исходной раскладки. Ulysses параллелизует по attention heads, а Ring — по последовательности, поэтому вместе они позволяют использовать больше GPU, чем Ulysses в одиночку.</p> <p>Используйте Unified Attention, если у вас есть как минимум 4 GPU. Установите значения <code>ulysses_degree</code> и <code>ring_degree</code> больше 1. GPU образуют двумерную сетку из <code>ulysses_degree * ring_degree</code> устройств, при этом каждый GPU входит в одну группу Ulysses и одну группу Ring.</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> Скопировано</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>Бенчмарк ниже сравнивает Ulysses, Ring и Unified Attention на 4 GPU H100 с помощью <a href="https://github.com/huggingface/diffusers/pull/12693#issuecomment-3694727532" rel="nofollow">этого скрипта</a>. Для каждого изображения используется 50 шагов инференса.</p> <table><thead><tr><th>CP Backend</th><th>Время / Изображение (мс)</th><th>Шагов / сек</th><th>Пиковая память (ГБ)</th></tr></thead><tbody><tr><td>ulysses</td><td>6670.789</td><td>7.50</td><td>33.85</td></tr><tr><td>кольцо</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 обеспечивает максимальную пропускную способность, но он не может использовать больше GPU, чем количество головок внимания в модели. Unified Attention снимает это ограничение.</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 Attention требует, чтобы длина последовательности и количество головок внимания делились нацело на <code>ulysses_degree</code>. <a href="https://github.com/huggingface/diffusers/pull/12996" rel="nofollow">Ulysses Anything Attention</a> снимает оба этих требования. Его нельзя использовать вместе с Unified Attention, поэтому <code>ring_degree</code> должен быть равен 1.</p> <p>Установите <code>ulysses_degree</code> в значение больше 1 и <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> Скопировано</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>Добавьте бэкенд <code>gloo</code> для тензоров CPU в <a href="https://docs.pytorch.org/docs/stable/distributed.html#torch.distributed.init_process_group" rel="nofollow">init_process_group</a>. Это позволяет избежать принудительной синхронизации CUDA, вызываемой передачей данных с хоста на устройство и с устройства на хост.</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> Скопировано</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">"cpu:gloo,cuda:nccl"</span>)<!----></pre></div><!----> <p>Бенчмарк ниже запускает FLUX.1-dev на 4 GPU L20 с помощью <a href="https://github.com/huggingface/diffusers/pull/12996#issuecomment-3797695999" rel="nofollow">этого скрипта</a>.</p> <table><thead><tr><th>CP Backend</th><th>Время / Итерация (мс)</th><th>Шагов / сек</th><th>Пиковая память (ГБ)</th><th>Форма (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>кольцо</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>сбой</td><td>сбой</td><td>сбой</td><td>1008x1008</td></tr><tr><td>кольцо</td><td>сбой</td><td>сбой</td><td>сбой</td><td>1008x1008</td></tr><tr><td>unified_balanced</td><td>сбой</td><td>сбой</td><td>сбой</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>При размере 1008x1008, когда длина последовательности не делится нацело, выполняется только Ulysses Anything Attention. При размере 1024x1024 его производительность такая же, как у стандартного Ulysses Attention.</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 требует, чтобы длина последовательности делилась нацело на <code>ring_degree</code>. <a href="https://github.com/huggingface/diffusers/pull/13545#issuecomment-4302195582" rel="nofollow">Ring Anything Attention</a> поддерживает любую длину последовательности. Он дополняет локальные K/V на каждом GPU до самой длинной локальной длины, выполняет all-gather для дополненных K/V и вырезает каждый блок обратно до его истинной длины перед вычислением внимания.</p> <p>Его нельзя использовать вместе с Unified Attention, поэтому <code>ulysses_degree</code> должно быть равно 1. Установите <code>ring_degree</code> больше 1 и <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> Скопировано</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 также выигрывает от использования бэкенда <code>gloo</code>, показанного в <a href="#ulysses-anything-attention">Ulysses Anything Attention</a>.</p> <blockquote class="note"><p>Ring Anything Attention поддерживает только вывод, и <code>attn_mask</code> должен быть <code>None</code>.</p></blockquote> <p>Бенчмарк ниже запускает FLUX.1-dev на 4 GPU RTX 4090 (48 ГБ).</p> <table><thead><tr><th>CP Backend</th><th>Время / Итерация (мс)</th><th>Шагов / сек</th><th>Пиковая память (ГБ)</th><th>Форма (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>кольцо</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>сбой</td><td>сбой</td><td>сбой</td><td>1008x1008</td></tr><tr><td>кольцо</td><td>сбой</td><td>сбой</td><td>сбой</td><td>1008x1008</td></tr><tr><td>unified_balanced</td><td>сбой</td><td>сбой</td><td>сбой</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 работает с разрешением 1008x1008 и демонстрирует производительность, сопоставимую со стандартным Ring Attention. Подробнее о мотивации и компромиссах см. в <a href="https://github.com/huggingface/diffusers/pull/13545#issuecomment-4304104462" rel="nofollow">этом комментарии</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><!---->Включить параллелизм контекста при загрузке<!----></span></h2><!--]--><!----> <p>Вместо вызова <code>enable_parallelism</code> после загрузки передайте <a href="/docs/diffusers/v0.43.0/ru/api/parallel#diffusers.ContextParallelConfig">ContextParallelConfig</a> в аргумент <code>parallel_config</code> метода <a href="/docs/diffusers/v0.43.0/ru/api/models/overview#diffusers.ModelMixin.from_pretrained">from_pretrained()</a>. Сначала инициализируйте группу процессов, так как <code>from_pretrained</code> вызывает ошибку, если <code>torch.distributed</code> не инициализирован. Запустите скрипт с помощью <code>torchrun</code>, как в примере <a href="#ring-attention">Ring Attention</a>.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Скопировано</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">"black-forest-labs/FLUX.1-dev"</span> | |
| dist.init_process_group(backend=<span class="hljs-string">"nccl"</span>) | |
| local_rank = <span class="hljs-built_in">int</span>(os.environ[<span class="hljs-string">"LOCAL_RANK"</span>]) | |
| device = torch.device(<span class="hljs-string">f"cuda:<span class="hljs-subst">{local_rank}</span>"</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">"transformer"</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:
- 44.5 kB
- Xet hash:
- 2a6b31b6dbbc28c796731cde375a837128428138d723ab173f56c83221e219af
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.