Buckets:

hf-doc-build
/

doc

Files
xet
hf-doc-build/doc / diffusers /main /tr /training /distributed_inference.html
mishig
about 15 hours ago
download
raw
110 kB
<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;Dağıtık çıkarım&quot;,&quot;local&quot;:&quot;distributed-inference&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Accelerate&quot;,&quot;local&quot;:&quot;accelerate&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;PyTorch Dağıtık&quot;,&quot;local&quot;:&quot;pytorch-distributed&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;device_map&quot;,&quot;local&quot;:&quot;devicemap&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Bağlam paralelliği&quot;,&quot;local&quot;:&quot;context-parallelism&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Halka Dikkat&quot;,&quot;local&quot;:&quot;ring-attention&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Ulysses Dikkat&quot;,&quot;local&quot;:&quot;ulysses-attention&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Birleşik Dikkat&quot;,&quot;local&quot;:&quot;unified-attention&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Ulysses Anything Dikkat&quot;,&quot;local&quot;:&quot;ulysses-anything-attention&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Ring Anything Attention&quot;,&quot;local&quot;:&quot;ring-anything-attention&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;parallel_config&quot;,&quot;local&quot;:&quot;parallelconfig&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3}],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Tensor paralelliği&quot;,&quot;local&quot;:&quot;tensor-parallelism&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Bir tensör paralelliği planı yazma&quot;,&quot;local&quot;:&quot;writing-a-tensor-parallelism-plan&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Colwise ve rowwise&quot;,&quot;local&quot;:&quot;colwise-and-rowwise&quot;,&quot;sections&quot;:[],&quot;depth&quot;:4},{&quot;title&quot;:&quot;Birleştirilmiş projeksiyonlar&quot;,&quot;local&quot;:&quot;fused-projections&quot;,&quot;sections&quot;:[],&quot;depth&quot;:4},{&quot;title&quot;:&quot;Dışarıda bırakılacaklar&quot;,&quot;local&quot;:&quot;what-to-leave-out&quot;,&quot;sections&quot;:[],&quot;depth&quot;:4},{&quot;title&quot;:&quot;Kısıtlamalar ve doğrulama&quot;,&quot;local&quot;:&quot;constraints-and-verification&quot;,&quot;sections&quot;:[],&quot;depth&quot;:4}],&quot;depth&quot;:3}],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Strateji seçimi&quot;,&quot;local&quot;:&quot;choosing-a-strategy&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2}],&quot;depth&quot;:1}"/>
		<link href="/docs/diffusers/main/en/_app/immutable/entry/start.DOtxR4Va.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/chunks/BPjc3X4I.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/chunks/BxrqFSCq.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/entry/app.CEihSODM.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/chunks/BxHPRFe5.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/chunks/f6ZfvXQ7.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/nodes/0.CM7zgfmq.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/chunks/CVE88MXi.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/nodes/5.D1LnFnkh.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/chunks/hylLzl4P.js" rel="modulepreload">
		<!--el5ux9--><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;Distributed inference&quot;,&quot;local&quot;:&quot;distributed-inference&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Accelerate&quot;,&quot;local&quot;:&quot;accelerate&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;PyTorch Distributed&quot;,&quot;local&quot;:&quot;pytorch-distributed&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;device_map&quot;,&quot;local&quot;:&quot;devicemap&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&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;:3},{&quot;title&quot;:&quot;Ulysses Attention&quot;,&quot;local&quot;:&quot;ulysses-attention&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Unified Attention&quot;,&quot;local&quot;:&quot;unified-attention&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Ulysses Anything Attention&quot;,&quot;local&quot;:&quot;ulysses-anything-attention&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Ring Anything Attention&quot;,&quot;local&quot;:&quot;ring-anything-attention&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;parallel_config&quot;,&quot;local&quot;:&quot;parallelconfig&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3}],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Tensor parallelism&quot;,&quot;local&quot;:&quot;tensor-parallelism&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Writing a tensor parallelism plan&quot;,&quot;local&quot;:&quot;writing-a-tensor-parallelism-plan&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Colwise and rowwise&quot;,&quot;local&quot;:&quot;colwise-and-rowwise&quot;,&quot;sections&quot;:[],&quot;depth&quot;:4},{&quot;title&quot;:&quot;Fused projections&quot;,&quot;local&quot;:&quot;fused-projections&quot;,&quot;sections&quot;:[],&quot;depth&quot;:4},{&quot;title&quot;:&quot;What to leave out&quot;,&quot;local&quot;:&quot;what-to-leave-out&quot;,&quot;sections&quot;:[],&quot;depth&quot;:4},{&quot;title&quot;:&quot;Constraints and verification&quot;,&quot;local&quot;:&quot;constraints-and-verification&quot;,&quot;sections&quot;:[],&quot;depth&quot;:4}],&quot;depth&quot;:3}],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Choosing a strategy&quot;,&quot;local&quot;:&quot;choosing-a-strategy&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="distributed-inference" 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="#distributed-inference"><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><!---->Dağıtık çıkarım<!----></span></h1><!--]--><!----> <p>Dağıtık çıkarım, iş yükünü birden fazla GPU arasında paylaştırır. Daha büyük modellerin belleğe sığdırılması için faydalı bir tekniktir ve daha yüksek bir aktarım hızı için birden çok istemciyi işleyebilir.</p> <p>Bu kılavuz, dağıtık çıkarım için <a href="https://huggingface.co/docs/accelerate/index" rel="nofollow">Accelerate</a> ve <a href="https://pytorch.org/tutorials/beginner/dist_overview.html" rel="nofollow">PyTorch Distributed</a> kullanmayı gösterecektir.</p> <!--[1--><h2 class="relative group"><a id="accelerate" 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="#accelerate"><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><!---->Accelerate<!----></span></h2><!--]--><!----> <p>Accelerate, kurulumu ele alarak kullanıcıların PyTorch kodlarına odaklanmasını sağlamak üzere, birden fazla hızlandırıcıda çıkarım ve eğitimi basitleştirmek için tasarlanmış bir kütüphanedir.</p> <p>Aşağıdaki komutla Accelerate'i yükleyin.</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> Kopyalandı</div><!----></button><!----></div> <pre class="language-bash "><!---->uv pip install accelerate<!----></pre></div><!----> <p>Dağıtık bir ortam oluşturmak için bir Python dosyasında <a href="https://huggingface.co/docs/accelerate/main/en/package_reference/state#accelerate.PartialState" rel="nofollow">accelerate.PartialState</a> sınıfını başlatın. <a href="https://huggingface.co/docs/accelerate/main/en/package_reference/state#accelerate.PartialState" rel="nofollow">accelerate.PartialState</a> sınıfı, süreç yönetimi, cihaz kontrolü ve dağıtımı ile süreç koordinasyonunu yönetir.</p> <p>Her bir işlem için bir GPU atamak üzere <a href="/docs/diffusers/main/tr/api/pipelines/overview#diffusers.DiffusionPipeline">DiffusionPipeline</a> nesnesini <code>accelerate.PartialState.device</code> konumuna taşıyın.</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> Kopyalandı</div><!----></button><!----></div> <pre class="language-py "><!----><span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> accelerate <span class="hljs-keyword">import</span> PartialState
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> DiffusionPipeline

pipeline = DiffusionPipeline.from_pretrained(
    <span class="hljs-string">&quot;Qwen/Qwen-Image&quot;</span>, dtype=torch.float16
)
distributed_state = PartialState()
pipeline.to(distributed_state.device)<!----></pre></div><!----> <p>İstemcileri işlem sayısı arasında otomatik olarak dağıtmak için <a href="https://huggingface.co/docs/accelerate/main/en/package_reference/state#accelerate.PartialState.split_between_processes" rel="nofollow">split_between_processes</a> yardımcısını bir bağlam yöneticisi olarak kullanın.</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> Kopyalandı</div><!----></button><!----></div> <pre class="language-py "><!----><span class="hljs-keyword">with</span> distributed_state.split_between_processes([<span class="hljs-string">&quot;a dog&quot;</span>, <span class="hljs-string">&quot;a cat&quot;</span>]) <span class="hljs-keyword">as</span> prompt:
    result = pipeline(prompt).images[<span class="hljs-number">0</span>]
    result.save(<span class="hljs-string">f&quot;result_<span class="hljs-subst">{distributed_state.process_index}</span>.png&quot;</span>)<!----></pre></div><!----> <p>Betik çalıştırmak için <code>accelerate launch</code> komutunu çağırın ve kullanılacak GPU sayısını belirlemek için <code>--num_processes</code> argümanını kullanın.</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> Kopyalandı</div><!----></button><!----></div> <pre class="language-bash "><!---->accelerate launch run_distributed.py --num_processes=2<!----></pre></div><!----> <blockquote class="tip"><p>Birden fazla GPU üzerinde çıkarım (inference) çalıştırmak için bu minimal <a href="https://gist.github.com/sayakpaul/cfaebd221820d7b43fae638b4dfa01ba" rel="nofollow">script</a> örneğine bakın. Daha fazla bilgi edinmek için <a href="https://huggingface.co/docs/accelerate/en/usage_guides/distributed_inference#distributed-inference-with-accelerate" rel="nofollow">Distributed Inference with 🤗 Accelerate</a> kılavuzuna göz atın.</p></blockquote> <!--[1--><h2 class="relative group"><a id="pytorch-distributed" 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="#pytorch-distributed"><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><!---->PyTorch Dağıtık<!----></span></h2><!--]--><!----> <p>PyTorch <a href="https://pytorch.org/docs/stable/generated/torch.nn.parallel.DistributedDataParallel.html" rel="nofollow">DistributedDataParallel</a>, aynı modeli her cihazda çoğaltarak farklı veri topluluklarını paralel olarak işlemek için <a href="https://huggingface.co/spaces/nanotron/ultrascale-playbook?section=data_parallelism" rel="nofollow">data parallelism</a> (veri paralelliği) sağlar.</p> <p>Dağıtık işlem grubunu oluşturmak ve her GPU için çıkarım süreçlerini başlatmak üzere bir Python dosyasına <code>torch.distributed</code> ve <code>torch.multiprocessing</code> içe aktarın.</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> Kopyalandı</div><!----></button><!----></div> <pre class="language-py "><!----><span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">import</span> torch.distributed <span class="hljs-keyword">as</span> dist
<span class="hljs-keyword">import</span> torch.multiprocessing <span class="hljs-keyword">as</span> mp

<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> DiffusionPipeline

pipeline = DiffusionPipeline.from_pretrained(
    <span class="hljs-string">&quot;Qwen/Qwen-Image&quot;</span>, dtype=torch.float16,
)<!----></pre></div><!----> <p><a href="https://pytorch.org/docs/stable/distributed.html?highlight=init_process_group#torch.distributed.init_process_group" rel="nofollow">init_process_group</a> ile çıkarım için bir fonksiyon oluşturun. Bu yöntem, backend türü, mevcut sürecin <code>rank</code> değeri ve katılan süreç sayısı olan <code>world_size</code> (örneğin, 2 GPU için <code>world_size=2</code>) ile dağıtık bir ortam oluşturur.</p> <p>Hattı <code>rank</code>'a taşıyın ve her bir sürece bir GPU atamak için <code>get_rank</code> kullanın. Her süreç farklı bir istemciyi işler.</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> Kopyalandı</div><!----></button><!----></div> <pre class="language-py "><!----><span class="hljs-keyword">def</span> <span class="hljs-title function_">run_inference</span>(<span class="hljs-params">rank, world_size</span>):
    dist.init_process_group(<span class="hljs-string">&quot;nccl&quot;</span>, rank=rank, world_size=world_size)

    pipeline.to(rank)

    <span class="hljs-keyword">if</span> torch.distributed.get_rank() == <span class="hljs-number">0</span>:
        prompt = <span class="hljs-string">&quot;a dog&quot;</span>
    <span class="hljs-keyword">elif</span> torch.distributed.get_rank() == <span class="hljs-number">1</span>:
        prompt = <span class="hljs-string">&quot;a cat&quot;</span>

    image = sd(prompt).images[<span class="hljs-number">0</span>]
    image.save(<span class="hljs-string">f&quot;./<span class="hljs-subst">{<span class="hljs-string">&#x27;_&#x27;</span>.join(prompt)}</span>.png&quot;</span>)<!----></pre></div><!----> <p><code>world_size</code> içinde tanımlanan süreç sayısını oluşturmak için <a href="https://pytorch.org/docs/stable/multiprocessing.html#torch.multiprocessing.spawn" rel="nofollow">mp.spawn</a> kullanın.</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> Kopyalandı</div><!----></button><!----></div> <pre class="language-py "><!----><span class="hljs-keyword">def</span> <span class="hljs-title function_">main</span>():
    world_size = <span class="hljs-number">2</span>
    mp.spawn(run_inference, args=(world_size,), nprocs=world_size, join=<span class="hljs-literal">True</span>)


<span class="hljs-keyword">if</span> __name__ == <span class="hljs-string">&quot;__main__&quot;</span>:
    main()<!----></pre></div><!----> <p>Çıkarım betiğini çalıştırmak için <code>torchrun</code> çağırın ve kullanılacak GPU sayısını ayarlamak için <code>--nproc_per_node</code> argümanını kullanın.</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> Kopyalandı</div><!----></button><!----></div> <pre class="language-bash "><!---->torchrun --nproc_per_node=2 run_distributed.py<!----></pre></div><!----> <!--[1--><h2 class="relative group"><a id="devicemap" 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="#devicemap"><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><!---->device_map<!----></span></h2><!--]--><!----> <p><code>device_map</code> argümanı, model bileşenlerini ayrı GPU'lara otomatik olarak yerleştirerek dağıtık çıkarımı (inference) etkinleştirir. Bu, bir modelin tek bir GPU'ya sığmadığı durumlarda özellikle faydalıdır. Aşağıdaki örnekte gösterildiği gibi (iki GPU'nun mevcut olduğu varsayılır), belirli bir aşamada gerekli model bileşenlerini seçici olarak yüklemek ve boşaltmak için <code>device_map</code> kullanabilirsiniz.</p> <p><code>device_map="balanced"</code> değerini ayarlayarak metin kodlayıcılarını (text encoders) tüm kullanılabilir GPU'lara eşit şekilde dağıtabilirsiniz. Her bir metin kodlayıcı için tahsis edilecek maksimum bellek miktarını belirtmek için <code>max_memory</code> argümanını kullanabilirsiniz. Bellek kullanımını önlemek için başka hiçbir pipeline bileşeni yüklemeyin.</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> Kopyalandı</div><!----></button><!----></div> <pre class="language-py "><!----><span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> FluxPipeline
<span class="hljs-keyword">import</span> torch

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>

pipeline = FluxPipeline.from_pretrained(
    <span class="hljs-string">&quot;black-forest-labs/FLUX.1-dev&quot;</span>,
    transformer=<span class="hljs-literal">None</span>,
    vae=<span class="hljs-literal">None</span>,
    device_map=<span class="hljs-string">&quot;balanced&quot;</span>,
    max_memory={<span class="hljs-number">0</span>: <span class="hljs-string">&quot;16GB&quot;</span>, <span class="hljs-number">1</span>: <span class="hljs-string">&quot;16GB&quot;</span>},
    dtype=torch.bfloat16
)
<span class="hljs-keyword">with</span> torch.no_grad():
    <span class="hljs-built_in">print</span>(<span class="hljs-string">&quot;Encoding prompts.&quot;</span>)
    prompt_embeds, pooled_prompt_embeds, text_ids = pipeline.encode_prompt(
        prompt=prompt, prompt_2=<span class="hljs-literal">None</span>, max_sequence_length=<span class="hljs-number">512</span>
    )<!----></pre></div><!----> <p>Metin gömlemleri hesaplandıktan sonra, difüzyon dönüştürücüsü için yer açmak üzere GPU'dan kaldırın.</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> Kopyalandı</div><!----></button><!----></div> <pre class="language-py "><!----><span class="hljs-keyword">import</span> gc 

<span class="hljs-keyword">def</span> <span class="hljs-title function_">flush</span>():
    gc.collect()
    torch.cuda.empty_cache()
    torch.cuda.reset_max_memory_allocated()
    torch.cuda.reset_peak_memory_stats()

<span class="hljs-keyword">del</span> pipeline.text_encoder
<span class="hljs-keyword">del</span> pipeline.text_encoder_2
<span class="hljs-keyword">del</span> pipeline.tokenizer
<span class="hljs-keyword">del</span> pipeline.tokenizer_2
<span class="hljs-keyword">del</span> pipeline

flush()<!----></pre></div><!----> <p><code>device_map="auto"</code> değerini ayarlayarak modelin iki GPU üzerinde otomatik olarak dağıtılmasını sağlayın. Bu strateji, modeli önce en hızlı cihaza yerleştirir; gerekirse modeli CPU veya sabit disk gibi daha yavaş bir cihaza yerleştirir. Model parametrelerini daha yavaş cihazlarda saklamanın dezavantajı, daha yavaş çıkarım gecikmesidir.</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> Kopyalandı</div><!----></button><!----></div> <pre class="language-py "><!----><span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> AutoModel
<span class="hljs-keyword">import</span> torch 

transformer = AutoModel.from_pretrained(
    <span class="hljs-string">&quot;black-forest-labs/FLUX.1-dev&quot;</span>, 
    subfolder=<span class="hljs-string">&quot;transformer&quot;</span>,
    device_map=<span class="hljs-string">&quot;auto&quot;</span>,
    dtype=torch.bfloat16
)<!----></pre></div><!----> <blockquote class="tip"><p>Farklı modellerin cihazlar arasında nasıl dağıtıldığını görmek için <code>pipeline.hf_device_map</code> çalıştırın. Bu, model cihaz yerleşimini izlemek için kullanışlıdır. Ayrıca, dağılımını görmek için dönüştürücü model üzerinde <code>hf_device_map</code> çağrısını da yapabilirsiniz.</p></blockquote> <p>Dönüştürücü modelini hattın (pipeline) içine ekleyin ve latentleri üretmek için <code>output_type="latent"</code> değerini ayarlayın.</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> Kopyalandı</div><!----></button><!----></div> <pre class="language-py "><!---->pipeline = FluxPipeline.from_pretrained(
    <span class="hljs-string">&quot;black-forest-labs/FLUX.1-dev&quot;</span>,
    text_encoder=<span class="hljs-literal">None</span>,
    text_encoder_2=<span class="hljs-literal">None</span>,
    tokenizer=<span class="hljs-literal">None</span>,
    tokenizer_2=<span class="hljs-literal">None</span>,
    vae=<span class="hljs-literal">None</span>,
    transformer=transformer,
    dtype=torch.bfloat16
)

<span class="hljs-built_in">print</span>(<span class="hljs-string">&quot;Running denoising.&quot;</span>)
height, width = <span class="hljs-number">768</span>, <span class="hljs-number">1360</span>
latents = pipeline(
    prompt_embeds=prompt_embeds,
    pooled_prompt_embeds=pooled_prompt_embeds,
    num_inference_steps=<span class="hljs-number">50</span>,
    guidance_scale=<span class="hljs-number">3.5</span>,
    height=height,
    width=width,
    output_type=<span class="hljs-string">&quot;latent&quot;</span>,
).images<!----></pre></div><!----> <p>Hafızadan pipeline ve transformer'ı kaldırın ve latents'leri çözmek için bir VAE yükleyin. VAE genellikle tek bir cihaza yüklenebilecek kadar küçüktür.</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> Kopyalandı</div><!----></button><!----></div> <pre class="language-py "><!----><span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> AutoencoderKL
<span class="hljs-keyword">from</span> diffusers.image_processor <span class="hljs-keyword">import</span> VaeImageProcessor

vae = AutoencoderKL.from_pretrained(ckpt_id, subfolder=<span class="hljs-string">&quot;vae&quot;</span>, dtype=torch.bfloat16).to(<span class="hljs-string">&quot;cuda&quot;</span>)  <span class="hljs-comment"># or &quot;mps&quot;, &quot;xpu&quot;, &quot;cpu&quot;</span>
vae_scale_factor = <span class="hljs-number">2</span> ** (<span class="hljs-built_in">len</span>(vae.config.block_out_channels) - <span class="hljs-number">1</span>)
image_processor = VaeImageProcessor(vae_scale_factor=vae_scale_factor)

<span class="hljs-keyword">with</span> torch.no_grad():
    <span class="hljs-built_in">print</span>(<span class="hljs-string">&quot;Running decoding.&quot;</span>)
    latents = FluxPipeline._unpack_latents(latents, height, width, vae_scale_factor)
    latents = (latents / vae.config.scaling_factor) + vae.config.shift_factor

    image = vae.decode(latents, return_dict=<span class="hljs-literal">False</span>)[<span class="hljs-number">0</span>]
    image = image_processor.postprocess(image, output_type=<span class="hljs-string">&quot;pil&quot;</span>)
    image[<span class="hljs-number">0</span>].save(<span class="hljs-string">&quot;split_transformer.png&quot;</span>)<!----></pre></div><!----> <p>Belirli bir aşamada ihtiyaç duyduğunuz modelleri seçici olarak yükleyip boşaltarak ve en büyük modelleri birden çok GPU arasında paylaştırarak, tüketici GPU'larında büyük modellerle çıkarım (inference) çalıştırılabilir.</p> <!--[1--><h2 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><!---->Bağlam paralelliği<!----></span></h2><!--]--><!----> <p><a href="https://huggingface.co/spaces/nanotron/ultrascale-playbook?section=context_parallelism" rel="nofollow">Bağlam paralelliği</a>, bellek kullanımını azaltmak için giriş dizilerini birden çok GPU arasında böler. Her GPU, dizinin kendi dilimini işler.</p> <p>Daha optimize bir dikkat (attention) arka ucuna geçmek için <a href="/docs/diffusers/main/tr/api/models/overview#diffusers.ModelMixin.set_attention_backend">set_attention_backend()</a> kullanın. Kullanılabilir arka uçların tam listesi için bu <a href="../optimization/attention_backends#available-backends">tabloya</a> bakın.</p> <p>Çoğu dikkat (attention) arka ucu, bağlam paralelliğiyle uyumludur. Bir arka uç uyumlu değilse bir <a href="https://github.com/huggingface/diffusers/issues/new" rel="nofollow">issue</a> açın.</p> <!--[2--><h3 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><!---->Halka Dikkat<!----></span></h3><!--]--><!----> <p>Key (K) ve value (V) temsilleri, cihazlar arasında <a href="https://huggingface.co/papers/2310.01889" rel="nofollow">Ring Attention</a> kullanarak iletişim kurar. Bu, her bölmenin diğer tüm token'ların K/V'lerini görmesini sağlar. Her GPU, yerel K/V'si için dikkat (attention) hesaplar ve bunu halkadaki bir sonraki GPU'ya iletir. Tek bir GPU tam diziyi barındırmadığından, iletişim gecikmesi azalır.</p> <p>Dönüştürücü modelinin <code>parallel_config</code> argümanına bir <a href="/docs/diffusers/main/tr/api/parallel#diffusers.ContextParallelConfig">ContextParallelConfig</a> geçirin. Yapılandırma, Ring Attention için kaç cihazın kullanılacağını belirleyen <code>ring_degree</code> argümanını destekler.</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> Kopyalandı</div><!----></button><!----></div> <pre class="language-py "><!----><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>)
    rank = dist.get_rank()
    device = torch.device(<span class="hljs-string">f&quot;cuda:<span class="hljs-subst">{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">f&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>Yukarıdaki betik, PyTorch ile uyumlu <a href="https://docs.pytorch.org/docs/stable/elastic/run.html" rel="nofollow">torchrun</a> gibi dağıtık bir başlatıcı ile çalıştırılmalıdır. <code>--nproc-per-node</code>, kullanılabilir GPU sayısına ayarlanır.</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> Kopyalandı</div><!----></button><!----></div> <pre class="language-shell "><!---->torchrun --nproc-per-node 2 above_script.py<!----></pre></div><!----> <!--[2--><h3 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 Dikkat<!----></span></h3><!--]--><!----> <p><a href="https://huggingface.co/papers/2309.14509" rel="nofollow">Ulysses Attention</a>, bir diziyi GPU'lar arasında böler ve <em>all-to-all</em> iletişimi gerçekleştirir (her cihaz, veri göndermek ve almak için her diğer cihazla etkileşime girer). Her GPU, yalnızca dikkat başlıklarının bir alt kümesi için tüm token'lara sahip olur. Her GPU, kendi başlığı için tüm token'lar üzerinde dikkati yerel olarak hesaplar, ardından bir sonraki katman için sonuçları token'lara göre yeniden gruplamak üzere başka bir all-to-all işlemi gerçekleştirir.</p> <p><a href="/docs/diffusers/main/tr/api/parallel#diffusers.ContextParallelConfig">ContextParallelConfig</a>, <code>ulysses_degree</code> argümanı aracılığıyla Ulysses Attention'ı destekler. Bu, Ulysses Attention için kaç cihazın kullanılacağını belirler.</p> <p><a href="/docs/diffusers/main/tr/api/parallel#diffusers.ContextParallelConfig">ContextParallelConfig</a> nesnesini <code>enable_parallelism()</code> fonksiyonuna iletin.</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> Kopyalandı</div><!----></button><!----></div> <pre class="language-py "><!----><span class="hljs-comment"># Depending on the number of GPUs available.</span>
pipeline.transformer.enable_parallelism(config=ContextParallelConfig(ulysses_degree=<span class="hljs-number">2</span>))<!----></pre></div><!----> <!--[2--><h3 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><!---->Birleşik Dikkat<!----></span></h3><!--]--><!----> <p><a href="https://huggingface.co/papers/2405.07719" rel="nofollow">Unified Sequence Parallelism</a>, uzun dizi işleme için verimli bir yaklaşım olarak Ring Attention ve Ulysses Attention'ı tek bir yöntemde birleştirir. Önce başlıkları ve dizi jetonlarını yeniden dağıtmak için Ulysses'in <em>all-to-all</em> iletişimini uygular, ardından yeniden dağıtılan verileri işlemek için Ring Attention kullanır ve son olarak orijinal düzeni geri getirmek için <em>all-to-all</em> işlemini tersine çevirir.</p> <p>Bu hibrit yaklaşım, her iki yöntemin güçlü yönlerinden faydalanır:</p> <ul><li><strong>Ulysses Attention</strong>, dikkat başlıkları arasında verimli bir şekilde paralelleştirme sağlar</li> <li><strong>Ring Attention</strong>, çok uzun dizileri minimal bellek maliyetiyle işler</li> <li>Birlikte, hem başlar hem de dizi boyutları boyunca 2B paralelleştirmeyi mümkün kılarlar</li></ul> <p><a href="/docs/diffusers/main/tr/api/parallel#diffusers.ContextParallelConfig">ContextParallelConfig</a>, hem <code>ulysses_degree</code> hem de <code>ring_degree</code> parametrelerini belirleyerek Unified Attention'ı destekler. Kullanılan toplam cihaz sayısı <code>ulysses_degree * ring_degree</code>'dir ve Ulysses ile Ring gruplarının dik (keseşmeyen) olduğu 2B bir ızgarada düzenlenir.
<code>ulysses_degree</code> ve <code>ring_degree</code> değerleri 1'den büyük olacak şekilde ayarlanmış <a href="/docs/diffusers/main/tr/api/parallel#diffusers.ContextParallelConfig">ContextParallelConfig</a> nesnesini <code>enable_parallelism()</code> fonksiyonuna iletin.</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> Kopyalandı</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><!----> <blockquote class="tip"><p>Unified Attention, 2D bir ızgara oluşturmak için yeterli sayıda cihaz olduğunda (en az 4 cihaz) kullanılmalıdır.</p></blockquote> <p>Ulysess, Ring ve Unified Attention için <a href="https://github.com/huggingface/diffusers/pull/12693#issuecomment-3694727532" rel="nofollow">bu betiği</a> kullanarak 4 H100 GPU'lu bir düğümde bir benchmark çalıştırdık. Sonuçlar aşağıda özetlenmiştir:</p> <table><thead><tr><th>CP Arka Ucu</th><th>Zaman / İterasyon (ms)</th><th>Adım / Saniye</th><th>Zirve Bellek (GB)</th></tr></thead><tbody><tr><td>ulysses</td><td>6670.789</td><td>7.50</td><td>33.85</td></tr><tr><td>halka</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>Yukarıdaki tablodan, Ulysses'in daha iyi bir veri akışı sağladığı açıkça görülüyor; ancak kullanabileceği cihaz sayısı, dikkat başlarının sayısıyla sınırlı kalmaya devam ediyor ve bu kısıtlama birleşik dikkat (unified attention) ile çözümleniyor.</p> <!--[2--><h3 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 Dikkat<!----></span></h3><!--]--><!----> <p>Varsayılan Ulysses Attention mekanizması, gizli durumların (hidden states) dizi uzunluğunun cihaz sayısına tam bölünmesini gerektirir. Bu durum, Ulysses Attention'ın pratik kullanımına önemli kısıtlamalar getirmektedir. <a href="https://github.com/huggingface/diffusers/pull/12996" rel="nofollow">Ulysses Anything Attention</a>, Ulysses Attention'ın bir varyantı olup keyfi dizi uzunluklarını ve keyfi dikkat başı (attention head) sayılarını destekleyerek, Ulysses Attention'ın pratik kullanımdaki çok yönlülüğünü artırır.</p> <p><a href="/docs/diffusers/main/tr/api/parallel#diffusers.ContextParallelConfig">ContextParallelConfig</a>, <code>ulysses_degree</code> ve <code>ulysses_anything</code> parametrelerini birlikte belirleyerek Ulysses Anything Attention desteği sunar. Ulysses Anything Attention'ın şu anda Unified Attention tarafından desteklenmediğini lütfen unutmayın. <code>ulysses_degree</code> değerini 1'den büyük olacak şekilde ve <code>ulysses_anything=True</code> olarak ayarlanmış <a href="/docs/diffusers/main/tr/api/parallel#diffusers.ContextParallelConfig">ContextParallelConfig</a> nesnesini <code>enable_parallelism()</code> fonksiyonuna iletin.</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> Kopyalandı</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>H2D ve D2H aktarımlarından kaynaklanan birden fazla zorunlu CUDA senkronizasyonunu önlemek için <code>init_process_group</code> içinde <strong>gloo</strong> arka ucunu ekleyin. Bu, iletişim gecikmesini önemli ölçüde azaltacaktır.</p></blockquote> <p>FLUX.1-dev için Ulysses, Ring, Unified Attention ve Ulysses Anything Attention ile <a href="https://github.com/huggingface/diffusers/pull/12996#issuecomment-3797695999" rel="nofollow">bu betiği</a> kullanarak 4 L20 GPU'lu bir düğümde bir benchmark çalıştırdık. Sonuçlar aşağıda özetlenmiştir:</p> <table><thead><tr><th>CP Arka Ucu</th><th>Zaman / İterasyon (ms)</th><th>Adım / Saniye</th><th>Zirve Bellek (GB)</th><th>Biçim (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>halka</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>başarısız</td><td>başarısız</td><td>başarısız</td><td>1008x1008</td></tr><tr><td>halka</td><td>başarısız</td><td>başarısız</td><td>başarısız</td><td>1008x1008</td></tr><tr><td>unified_balanced</td><td>başarısız</td><td>başarısız</td><td>başarısız</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>Yukarıdaki tablodan, Ulysses Anything Attention'ın standart Ulysses Attention ile aynı performansı korurken keyfi dizi uzunluklarıyla daha iyi uyumluluk sunduğu açıkça görülmektedir.</p> <!--[2--><h3 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></h3><!--]--><!----> <p>Varsayılan <a href="https://huggingface.co/papers/2310.01889" rel="nofollow">Ring Attention</a>, gizli durumların (hidden states) dizi uzunluğunun halka derecesine (ring degree) eşit olarak bölünebilmesini gerektirir. <a href="https://github.com/huggingface/diffusers/pull/13545#issuecomment-4302195582" rel="nofollow">Ring Anything Attention</a>, Ring Attention'ın keyfi (eşit olarak bölünemeyen) dizi uzunluklarını destekleyen bir varyantıdır. Her sıranın (rank) yerel KV'sini küresel maksimum dizi uzunluğuna kadar doldurur (padding), doldurulmuş KV tamponunu tüm toplar (all-gather) ve dikkat (attention) çalıştırılmadan önce her sıranın gerçek uzunluğuna geri keser.</p> <p>Ring Anything Attention, Unified Attention tarafından desteklenmez. Ring Anything Attention'i etkinleştirmek için <code>ring_degree > 1</code> ve <code>ring_anything=True</code> değerlerini ayarlayın.</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> Kopyalandı</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><!----> <blockquote class="tip"><p>Birden fazla zorunlu CUDA senkronizasyonunun H2D ve D2H aktarımlarından kaynaklanmasını önlemek için <a href="https://docs.pytorch.org/docs/stable/distributed.html#torch.distributed.init_process_group" rel="nofollow">init_process_group</a> fonksiyonuna <code>gloo</code> arka ucunu ekleyin.</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> Kopyalandı</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><!----> <blockquote class="note"><p>Ring Anything Attention şu anda yalnızca çıkarımı destekler ve <code>None</code> olmayan dikkat maskeleri desteklenmez. <code>attn_mask</code> <code>None</code> olmalıdır.</p></blockquote> <p>4 RTX 4090 (48GB) GPU'lu bir düğümde FLUX.1-dev referans değerleri aşağıda görülebilir.</p> <table><thead><tr><th>CP Arka Ucu</th><th>Zaman / İterasyon (ms)</th><th>Adım / Saniye</th><th>Zirve Bellek (GB)</th><th>Biçim (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>halka</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>başarısız</td><td>başarısız</td><td>başarısız</td><td>1008x1008</td></tr><tr><td>halka</td><td>başarısız</td><td>başarısız</td><td>başarısız</td><td>1008x1008</td></tr><tr><td>unified_balanced</td><td>başarısız</td><td>başarısız</td><td>başarısız</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>Yukarıdaki tablodan, Ring Anything Attention'ın standart Ring Attention ile karşılaştırılabilir performansını korurken keyfi dizi uzunluklarıyla uyumlu olduğunu görüyoruz.
Ring Anything Attention'ın motivasyonu ve ödünleşimleri hakkında daha fazla ayrıntı için <a href="https://github.com/huggingface/diffusers/pull/13545#issuecomment-4304104462" rel="nofollow">bu yorumu</a> inceleyin.</p> <!--[2--><h3 class="relative group"><a id="parallelconfig" 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="#parallelconfig"><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><!---->parallel_config<!----></span></h3><!--]--><!----> <p>Model başlatma sırasında bağlam paralelliğini etkinleştirmek için <code>parallel_config</code> parametresini geçirin.</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> Kopyalandı</div><!----></button><!----></div> <pre class="language-py "><!---->CKPT_ID = <span class="hljs-string">&quot;black-forest-labs/FLUX.1-dev&quot;</span>

cp_config = ContextParallelConfig(ring_degree=<span class="hljs-number">2</span>)
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><!----> <!--[1--><h2 class="relative group"><a id="tensor-parallelism" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#tensor-parallelism"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span><!---->Tensor paralelliği<!----></span></h2><!--]--><!----> <p><a href="https://huggingface.co/spaces/nanotron/ultrascale-playbook?section=tensor_parallelism" rel="nofollow">Tensor parallelism</a>, bir modelin ağırlık matrislerini cihazlar arasında parçalar. Her cihaz, her katman için sütun bazlı (<code>"colwise"</code>) veya satır bazlı (<code>"rowwise"</code>) bir dilim tutar, kısmi bir sonuç hesaplar ve katman sınırındaki <code>AllReduce</code>/<code>AllGather</code> işlemi tam çıktıyı yeniden oluşturur. Bağlam paralelliğinin aksine, cihaz başına <em>ağırlık</em> belleğini azaltır; bu da tek bir cihaza sığmayan modeller için faydalıdır.</p> <p>Modelin <a href="/docs/diffusers/main/tr/api/models/overview#diffusers.ModelMixin.from_pretrained">from_pretrained()</a> metodunun <code>parallel_config</code> argümanına bir <a href="/docs/diffusers/main/tr/api/parallel#diffusers.TensorParallelConfig">TensorParallelConfig</a> geçirin. <code>tp_degree</code>, parçalanacak cihaz sayısıdır ve modelin dikkat başı sayısına tam bölünmelidir. Model, bir <code>_tp_plan</code> (modül adı glob'larından <code>"colwise"</code>/<code>"rowwise"</code> stiline düz bir eşleme) tanımlamalıdır.</p> <p>Bu yöntemle yüklenirken kontrol noktası <em>okunurken</em> parçalanır: her rank, parçalanan her ağırlığın yalnızca kendi dilimini okur ve onu doğrudan kendi cihazına yerleştirir. Tam boyutlu hiçbir şey asla maddi olarak oluşturulmadığından, rank başına bellek <code>tp_degree</code> arttıkça düşer.</p> <p><code>enable_parallelism()</code> çağrısı yapılarak tam modeli yüklemeye kıyasla, bu yöntem daha hızlı yüklenir ve rank başına daha az CPU belleği kullanır; <code>tp_degree</code> arttıkça bu fark büyür. Aşağıdaki sayılar, 4x A10G (23GB) üzerinde <a href="https://huggingface.co/black-forest-labs/FLUX.2-dev" rel="nofollow">black-forest-labs/FLUX.2-dev</a> (yalnızca transformer, 32B parametre, bf16) için verilmiştir.</p> <table><thead><tr><th>tp_degree</th><th>yöntem</th><th>yüklenme süresi</th><th>zirve CPU/rank</th></tr></thead><tbody><tr><td>4</td><td><code>from_pretrained(parallel_config=...)</code></td><td>12.5s</td><td>6,8 GB</td></tr><tr><td>4</td><td><code>from_pretrained</code> + <code>enable_parallelism</code></td><td>30.4s</td><td>64,1 GB</td></tr></tbody></table> <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> Kopyalandı</div><!----></button><!----></div> <pre class="language-py "><!----><span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> torch <span class="hljs-keyword">import</span> distributed <span class="hljs-keyword">as</span> dist
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> DiffusionPipeline, Flux2Transformer2DModel, TensorParallelConfig

<span class="hljs-keyword">def</span> <span class="hljs-title function_">main</span>():
    dist.init_process_group(backend=<span class="hljs-string">&quot;nccl&quot;</span>)
    rank, world_size = dist.get_rank(), dist.get_world_size()
    device = torch.device(<span class="hljs-string">f&quot;cuda:<span class="hljs-subst">{rank}</span>&quot;</span>)
    torch.cuda.set_device(device)

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

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

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

<span class="hljs-keyword">if</span> __name__ == <span class="hljs-string">&quot;__main__&quot;</span>:
    main()<!----></pre></div><!----> <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> Kopyalandı</div><!----></button><!----></div> <pre class="language-shell "><!---->torchrun --nproc-per-node 4 tensor_parallel_flux.py<!----></pre></div><!----> <p><code>tp_degree</code>, yukarıdaki <code>world_size</code> değerinden alınır; bu nedenle <code>--nproc-per-node 4</code> komutu, transformer'ı 4 cihaz arasında paylaştırır.</p> <blockquote class="caution"><p><code>parallel_config</code> ile tensor-parallel yükleme, henüz <code>device_map</code>, <code>quantization_config</code>, <code>low_cpu_mem_usage=False</code>, <code>use_flashpack=True</code> veya safetensors olmayan ağırlıklarla desteklenmiyor; her biri sessizce tam checkpoint'i yüklemeye geri düşmek yerine hata fırlatıyor.</p> <p>Gergilik paralelliğinin (tensor parallelism) kuantizasyon, offloading veya LoRA adaptörleriyle birleştirilmesi henüz desteklenmediğinden, model nasıl bölünürse (shard) olsun bu durumlar yine de … hatası verir.</p> <p>Tensor-parallel modelleri bu şekilde yüklemenizi öneririz. Aynı yapılandırmayı kullanarak <code>enable_parallelism()</code> çağrısı yaparak bellekte zaten bulunan bir modeli dağitalayabilirsiniz (shard), ancak bu yöntem her şeyi önce yükleyip yeniden dağıtır, bu da her rank üzerinde tam checkpoint belleği maliyeti oluşturur.</p></blockquote> <p>Tensor-parallel bir modelin kaydedilmesi henüz desteklenmiyor ve <a href="/docs/diffusers/main/tr/api/models/overview#diffusers.ModelMixin.save_pretrained">save_pretrained()</a> bu durumda bir hata fırlatır. Modeli parçalara ayırmadan önce kaydedin.</p> <!--[2--><h3 class="relative group"><a id="writing-a-tensor-parallelism-plan" 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="#writing-a-tensor-parallelism-plan"><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><!---->Bir tensör paralelliği planı yazma<!----></span></h3><!--]--><!----> <p>Tensor paralelliği, yalnızca modül ad kümeleme kalıplarını bir bölme stiline eşleyen düz bir sınıf özelliği olan <code>_tp_plan</code> tanımlayan modellerde çalışır. Bir tane yazmak, büyük ölçüde gizli boyutu <em>genişleten</em> her projeksiyonu, onu geri <em>daraltan</em> projeksiyonla eşleştirmekle ilgilidir.</p> <p>Her anahtar en fazla bir <strong><code>*</code></strong> içerebilir ve bu karakterden önceki önek, tek bir girdinin tüm blokları kapsaması için bir <a href="https://pytorch.org/docs/stable/generated/torch.nn.ModuleList.html" rel="nofollow"><code>nn.ModuleList</code></a> olarak çözümlenmelidir. <code>*</code> içermeyen bir anahtar, modelin kendisine uygulanır. Yollar modele göre göreli olarak tanımlanır.</p> <!--[3--><h4 class="relative group"><a id="colwise-and-rowwise" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#colwise-and-rowwise"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span><!---->Colwise ve rowwise<!----></span></h4><!--]--><!----> <table><thead><tr><th>Stil</th><th>Parçalar</th><th>Her sıra</th><th>Kullanım alanları</th></tr></thead><tbody><tr><td><code>"colwise"</code></td><td>çıkış özellikleri (<code>weight</code> boyutu 0)</td><td>çıkışın bir dilimini hesaplar</td><td><code>to_q</code>, <code>to_k</code>, <code>to_v</code>, FFN iç projeksiyonu</td></tr><tr><td><code>"rowwise"</code></td><td>girdi özellikleri (<code>weight</code> boyutu 1)</td><td>kısmi bir toplam hesaplar</td><td><code>to_out.0</code>, FFN çıkış projeksiyonu</td></tr><tr><td><code>"rowwise_split_input"</code></td><td>girdi özellikleri (<code>weight</code> boyutu 1)</td><td>girdinin ve ağırlıkların kendi dilimine kullanarak kısmi bir sonuç hesaplar; <code>AllReduce</code> bu sonuçları ekler ve böylece her sıra tüm çıktı özelliklerini alır</td><td>örneğin zaman adımı gömümünden bir AdaLN modülasyon projeksiyonu gibi, replike bir aktivasyonu okuyan bağımsız bir projeksiyon</td></tr></tbody></table> <p>Her zaman bu sırayla eşleştirin. <code>"colwise"</code> projeksiyonu çıktısını parçalı (sharded) bırakır, ardından gelen <code>"rowwise"</code> projeksiyonu bu parçayı doğrudan tüketir ve blok sınırındaki tek bir <code>AllReduce</code> işlemi sonucu yeniden oluşturur. Çifti başka bir şekilde parçalamak, ortada bir toplama (gather) işlemi zorunlu kılar ve çok daha fazla iletişim gerektirir.</p> <p>Dikkat mekanizması için bu, her sıranın (rank) başların (head) bir alt kümesine sahip olduğu anlamına gelir; bu nedenle <code>tp_degree</code>, baş sayısına tam bölünmelidir. Kodlayıcı akış kopyaları (<code>add_q_proj</code>, <code>to_add_out</code>, <code>ff_context</code>), görüntü akışı karşılıklarıyla aynı deseni izler.</p> <!--[3--><h4 class="relative group"><a id="fused-projections" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#fused-projections"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span><!---->Birleştirilmiş projeksiyonlar<!----></span></h4><!--]--><!----> <p>Tek bir <code>Linear</code>, parçalanan boyut boyunca birden fazla mantıksal tensörü paketlediğinde, düz <code>"colwise"</code>/<code>"rowwise"</code> dilimleri birleştirmenin üzerinden doğrudan geçer ve parçaları hizasız bırakır. Bunun yerine, her paketlenmiş bloğu bağımsız olarak parçalayan <code>PackedColwiseParallel</code>/<code>PackedRowwiseParallel</code> kullanın.</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> Kopyalandı</div><!----></button><!----></div> <pre class="language-py "><!----><span class="hljs-comment"># in src/diffusers/models/transformers/your_model.py</span>
<span class="hljs-keyword">from</span> ...hooks.tensor_parallel <span class="hljs-keyword">import</span> PackedColwiseParallel, PackedRowwiseParallel<!----></pre></div><!----> <p><code>blocks</code>, toplamı paketlenmiş boyutu bölen oransal tam sayılardan oluşan bir listedir — eşit yarımlara sahip bir SwiGLU gate+up projeksiyonu için <code>[1, 1]</code> veya <code>mlp_ratio=3</code> ile birleşik Q+K+V+gate+up projeksiyonu için <code>[1, 1, 1, 3, 3]</code>.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5   text-gray-600  " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Kopyalandı</div><!----></button><!----></div> <pre class="language-py "><!----><span class="hljs-string">&quot;transformer_blocks.*.ff.linear_in&quot;</span>: PackedColwiseParallel([<span class="hljs-number">1</span>, <span class="hljs-number">1</span>]),<!----></pre></div><!----> <p>Blok boyutları yalnızca yapılandırmadan biliniyorsa, argümanı atlayın ve bunun yerine mutlak boyutları <code>__init__</code> sırasında <code>Linear</code> üzerinde <code>_tp_packed_col_blocks</code> veya <code>_tp_packed_row_blocks</code> olarak saklayın.</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> Kopyalandı</div><!----></button><!----></div> <pre class="language-py "><!----><span class="hljs-comment"># in the attention module&#x27;s __init__</span>
<span class="hljs-variable language_">self</span>.to_out._tp_packed_row_blocks = [<span class="hljs-variable language_">self</span>.inner_dim, <span class="hljs-variable language_">self</span>.mlp_hidden_dim]<!----></pre></div><!----> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5   text-gray-600  " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Kopyalandı</div><!----></button><!----></div> <pre class="language-py "><!----><span class="hljs-comment"># in the model&#x27;s _tp_plan</span>
<span class="hljs-string">&quot;single_transformer_blocks.*.attn.to_out&quot;</span>: PackedRowwiseParallel(),<!----></pre></div><!----> <p>Her iki biçim de <a href="https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/transformers/transformer_flux2.py" rel="nofollow"><code>transformer_flux2.py</code></a> içindeki tek bir planda görünür: çift akışlı SwiGLU çifti <code>blocks</code> değerini olduğu gibi geçirirken, birleştirilmiş tek akışlı QKV+MLP projeksiyonu, boyutları <code>mlp_ratio</code> ve <code>mlp_mult_factor</code> değerlerine bağlı olduğu için bu boyutları <code>Flux2ParallelSelfAttention.__init__</code> içindeki <code>Linear</code> sınıfında saklar.</p> <!--[3--><h4 class="relative group"><a id="what-to-leave-out" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#what-to-leave-out"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span><!---->Dışarıda bırakılacaklar<!----></span></h4><!--]--><!----> <p>Planın dışında kalan her şey, her sırada kopyalanır; bu, normalizasyon katmanları, AdaLN modülasyonu (<code>img_mod</code>/<code>txt_mod</code>), yama ve metin gömme değerleri ile son <code>norm_out</code>/<code>proj_out</code> için doğru seçimdir. Bunlar küçük olduğundan, bunları bölümlere ayırmak az miktarda bellek tasarrufu sağlar ancak iletişim yükü ekler.</p> <!--[3--><h4 class="relative group"><a id="constraints-and-verification" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#constraints-and-verification"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span><!---->Kısıtlamalar ve doğrulama<!----></span></h4><!--]--><!----> <ul><li><code>tp_degree</code>, <code>config.num_attention_heads</code>'i tam olarak bölmelidir.</li> <li>Her paketlenmiş blok, yalnızca toplamları değil, <em>tek tek</em> <code>tp_degree</code> değerine bölünebilir olmalıdır.</li></ul> <p>Her ikisi de herhangi bir ağırlık yüklenmeden veya bölünmeden önce <a href="/docs/diffusers/main/tr/api/models/overview#diffusers.ModelMixin.from_pretrained">from_pretrained()</a> ve <code>enable_parallelism()</code> tarafından doğrulanır.</p> <p>Yeni bir planı gözle değil, sayısal olarak doğrulayın: sabit bir tohumla tek bir cihazda üretim yapın, ardından tensor paralelliği altında tekrar üretin ve çıktıları karşılaştırın. Yanlış yerleştirilmiş bir <code>"colwise"</code>/<code>"rowwise"</code> genellikle yine de çalışır ve makul görünen ancak yanlış bir görüntü üretir.</p> <blockquote class="tip"><p>Benzer bir mimariye sahip mevcut bir plandan başlayın. <a href="/docs/diffusers/main/tr/api/models/qwenimage_transformer2d#diffusers.QwenImageTransformer2DModel">QwenImageTransformer2DModel</a> tamamen füzyon yapılmamış olup her giriş düz <code>"colwise"</code>/<code>"rowwise"</code> şeklindedir, <a href="/docs/diffusers/main/tr/api/models/flux_transformer#diffusers.FluxTransformer2DModel">FluxTransformer2DModel</a> tek bir paketli satır içi projeksiyon ekler ve <a href="/docs/diffusers/main/tr/api/models/flux2_transformer#diffusers.Flux2Transformer2DModel">Flux2Transformer2DModel</a> her iki paketli stili de kapsar — sırasıyla <a href="https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/transformers/transformer_qwenimage.py" rel="nofollow"><code>transformer_qwenimage.py</code></a>, <a href="https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/transformers/transformer_flux.py" rel="nofollow"><code>transformer_flux.py</code></a> ve <a href="https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/transformers/transformer_flux2.py" rel="nofollow"><code>transformer_flux2.py</code></a> dosyalarına bakın.</p></blockquote> <!--[1--><h2 class="relative group"><a id="choosing-a-strategy" 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="#choosing-a-strategy"><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><!---->Strateji seçimi<!----></span></h2><!--]--><!----> <p>Yukarıdaki stratejiler farklı sorunları çözer ve faydalı soru soyut olarak hangisinin en hızlı olduğu değil, neyin tükendiğidir.</p> <table><thead><tr><th>Strateji</th><th>Bölümler</th><th>Azaltır</th><th>Tek bir istem için gecikme</th><th>En iyi olduğu durum</th></tr></thead><tbody><tr><td><a href="#accelerate">Accelerate</a> / <a href="#pytorch-distributed">DDP</a></td><td>kopyalar arasında istemler</td><td>hiçbiri — her cihaz tam bir kopya tutar</td><td>değişmedi</td><td>model zaten sığıyor ve birçok isteminiz var</td></tr><tr><td><a href="#devicemap"><code>device_map</code></a></td><td>cihazlar arası bileşenler</td><td>ağırlık belleği</td><td>biraz daha kötü</td><td>model sığmıyor ve bağlantı yavaş</td></tr><tr><td><a href="#context-parallelism">Bağlam paralelliği</a></td><td>giriş dizisi</td><td>aktivasyon belleği</td><td>düşük</td><td>diziler uzun — yüksek çözünürlüklü veya video</td></tr><tr><td><a href="#tensor-parallelism">Tensor paralelliği</a></td><td>ağırlık matrisleri</td><td>ağırlık belleği</td><td>düşük</td><td>bir bileşenin ağırlıkları sığmıyor ve bağlantı hızlı</td></tr></tbody></table> <p>Bazı pratik yönergeler:</p> <ul><li><strong>Çoklu isteklerde (prompts) yüksek verimlilik, model zaten sığabiliyor.</strong> Veri paralelliği kullanın. Modeli değiştirmeden verimliliği doğrusal olarak ölçeklendiren tek strateji budur ve tek istek gecikmesini (latency) etkilemez.</li> <li><strong>Bir bileşenin ağırlıkları sığmıyor.</strong> Hem belleği hem de gecikmeyi düşürdüğü için önce tensör paralelliğine yönelin. Her blok sınırında iletişim kurduğu için NVLink gibi hızlı bir bağlantı ister; PCIe üzerinde katman başına bu trafik, kazandırdığı hesaplama gücünü aşabilir ve bu durumda <code>device_map</code> daha iyi bir seçim olur. <code>device_map</code>, bileşenlerin tek başına uygun olmasına rağmen toplu olarak çok büyük olduğu durumu da ele alır.</li> <li><strong>Problem ağırlıklarda değil, aktivasyonlarda.</strong> Bu, uzun dizi rejimidir — büyük görüntüler, çok sayıda kare — ve bağlam paralelliği doğrudan cevaptır. Yukarıdaki <a href="#ulysses-attention">Ulysses/Ring kıyaslamalarından</a> bir arka uç seçin.</li> <li><strong>Hem ağırlıklar hem de dizi çok büyük.</strong> Tensor ve bağlam paralelliğini birleştirin. <a href="/docs/diffusers/main/tr/api/parallel#diffusers.TensorParallelConfig">TensorParallelConfig</a>, her ikisinin de tek bir cihaz ızgarasını paylaşabilmesi için <code>mesh</code> argümanını kabul eder. Bu kombinasyon henüz doğrulanmadığından deneysel olarak ele alınmalıdır.</li></ul> <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/training/distributed_inference.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_12hgqlw = {
						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.DOtxR4Va.js"),
						import("/docs/diffusers/main/en/_app/immutable/entry/app.CEihSODM.js")
					]).then(([kit, app]) => {
						kit.start(app, element, {
							node_ids: [0, 5],
							data: [null,null],
							form: null,
							error: null
						});
					});
				}
			</script>
		
Xet Storage Details

Size:: 110 kB
Xet hash:: 1f8e9bfd025b5fa12ce2db60a55c10c663913502fee7315db2c239a53802e36b
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.