Buckets:

hf-doc-build
/

doc

Files
xet
hf-doc-build/doc / diffusers /main /ko /optimization /tpu.html
mishig
about 12 hours ago
download
raw
25.8 kB
<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;Google TPU&quot;,&quot;local&quot;:&quot;google-tpu&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Eager 모드&quot;,&quot;local&quot;:&quot;eager-mode&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;컴파일 모드&quot;,&quot;local&quot;:&quot;compiled-mode&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Tensor 병렬성&quot;,&quot;local&quot;:&quot;tensor-parallelism&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2}],&quot;depth&quot;:1}"/>
		<link href="/docs/diffusers/main/en/_app/immutable/entry/start.vZc-dOUR.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/chunks/Bafuik5c.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/chunks/Bupkjr9f.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/entry/app.B6c3Ml05.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/chunks/CyvZ3FS5.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/chunks/l0QnnOED.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/chunks/DnVIJfC5.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/chunks/DsnmJJEf.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/nodes/0.BQlx092V.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/chunks/CoSrdQax.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/nodes/340.B664tf49.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/chunks/B4AMoTAu.js" rel="modulepreload">
		<!--vjbfci--><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;Google TPU&quot;,&quot;local&quot;:&quot;google-tpu&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Eager mode&quot;,&quot;local&quot;:&quot;eager-mode&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Compiled mode&quot;,&quot;local&quot;:&quot;compiled-mode&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Tensor parallelism&quot;,&quot;local&quot;:&quot;tensor-parallelism&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="google-tpu" 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="#google-tpu"><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><!---->Google TPU<!----></span></h1><!--]--><!----> <p>Tensor Processing Units(TPU)는 Google의 머신러닝 가속기로, Cloud TPU(v6e, v7x)로 제공됩니다.
Diffusers는 TPU용 PyTorch 백엔드인 <a href="https://github.com/google-pytorch/torch_tpu/" rel="nofollow">TorchTPU</a>를 통해 TPU에서 실행되며,
최소한의 코드 변경으로 가능합니다.</p> <p>세 가지 실행 모드가 사용할 수 있습니다:</p> <table><thead><tr><th>모드</th><th>상수</th><th>활성화 방법</th><th>The input is a single word "Notes".

참고</th></tr></thead><tbody><tr><td>엄격한 eager (기본값)</td><td><code>EagerMode.DEFER_NEVER</code></td><td><code>pipe.to("tpu")</code></td><td>Operations dispatched one at a time, asynchronous</td></tr><tr><td>최적화된 eager</td><td><code>EagerMode.DEFER_AND_FUSE</code></td><td><code>TORCH_TPU_DEFER_AND_FUSE=1</code></td><td>작업이 지연되고 더 큰 그래프로 융합됩니다. 엄격한 eager보다 빠르지만 컴파일이 더 오래 걸립니다</td></tr><tr><td>컴파일</td><td>—</td><td><code>torch.compile(module, backend="tpu")</code></td><td><code>TpuBackend</code>를 사용한 AOT 컴파일</td></tr></tbody></table> <p><a href="https://github.com/google-pytorch/torch_tpu/" rel="nofollow">TorchTPU 설치 가이드</a>를 따르세요. 설치 후 <code>import torch</code>를 실행하면 자동으로 로드되고 <code>"tpu"</code> 디바이스가 등록되므로, <code>pipe.to("tpu")</code>만 변경하면 됩니다. <code>TORCH_DEVICE_BACKEND_AUTOLOAD=0</code>로 백엔드 자동 로딩을 비활성화한 경우에만 <code>import torch_tpu</code>를 추가하세요.</p> <!--[1--><h2 class="relative group"><a id="eager-mode" 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="#eager-mode"><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><!---->Eager 모드<!----></span></h2><!--]--><!----> <p>FLUX.1-schnell은 단일 v6e 칩에 한 번에 모두 들어가지 않으므로, <a href="/docs/diffusers/main/ko/api/pipelines/overview#diffusers.DiffusionPipeline.enable_model_cpu_offload">enable_model_cpu_offload()</a>를 사용하여 각 모델을 실행되는 동안에만 TPU로 이동하세요. <code>"tpu"</code> 장치를 자동으로 감지합니다.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5   text-gray-600  " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> 복사됨</div><!----></button><!----></div> <pre class="language-python "><!----><span class="hljs-keyword">import</span> torch

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

pipe = FluxPipeline.from_pretrained(<span class="hljs-string">&quot;black-forest-labs/FLUX.1-schnell&quot;</span>, dtype=torch.bfloat16)
pipe.enable_model_cpu_offload()

image = pipe(
    prompt=<span class="hljs-string">&quot;a golden retriever surfing a wave, photorealistic&quot;</span>,
    height=<span class="hljs-number">1024</span>,
    width=<span class="hljs-number">1024</span>,
    num_inference_steps=<span class="hljs-number">4</span>,
    guidance_scale=<span class="hljs-number">0.0</span>,
).images[<span class="hljs-number">0</span>]

image.save(<span class="hljs-string">&quot;output.png&quot;</span>)<!----></pre></div><!----> <p>모델이 단일 칩에 너무 크거나, 여러 개의 칩을 보유하고 있어 낮은 지연 시간을 원한다면, 대신 모델을 칩에 걸쳐 샤딩하세요. <a href="#tensor-parallelism">Tensor parallelism</a> 섹션을 참조하세요.</p> <!--[1--><h2 class="relative group"><a id="compiled-mode" 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="#compiled-mode"><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><!---->컴파일 모드<!----></span></h2><!--]--><!----> <p>TorchTPU는 <code>"tpu"</code>를 <code>torch.compile</code> 백엔드 이름으로 등록합니다(내부적으로는 <code>TpuBackend</code>를 사용하므로), 따라서 각 구성 요소는 다른 <code>torch.compile</code> 대상과 동일한 방식으로 컴파일됩니다. 첫 번째 호출(워밍업)은 컴파일을 수행하기 때문에 느리지만, 동일한 shape을 가진 후속 호출은 컴파일된 그래프를 재사용합니다.</p> <blockquote class="important"><p>TorchTPU는 <strong>정적 shape</strong>를 요구하므로 <code>dynamic=False</code>를 전달해야 합니다. 새로운 <code>height</code> 또는 <code>width</code>가 해당 shape에 대해 한 번만 다시 컴파일되며, 이미 확인된 shape는 재사용됩니다. <code>num_inference_steps</code>를 변경해도 다시 컴파일되지 않습니다.</p></blockquote> <p>파이프라인 전체가 하나의 칩에 들어갈 경우, 이를 TPU로 이동하고 전체 transformer를 컴파일합니다. Stable Diffusion 3.5 Medium(~15GB in bf16)은 단일 v6e 칩에 들어갑니다.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5   text-gray-600  " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> 복사됨</div><!----></button><!----></div> <pre class="language-python "><!----><span class="hljs-keyword">import</span> torch

<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> StableDiffusion3Pipeline

pipe = StableDiffusion3Pipeline.from_pretrained(<span class="hljs-string">&quot;stabilityai/stable-diffusion-3.5-medium&quot;</span>, dtype=torch.bfloat16)
pipe.to(<span class="hljs-string">&quot;tpu&quot;</span>)
pipe.transformer.<span class="hljs-built_in">compile</span>(backend=<span class="hljs-string">&quot;tpu&quot;</span>, fullgraph=<span class="hljs-literal">True</span>, dynamic=<span class="hljs-literal">False</span>)

<span class="hljs-comment"># Warmup — triggers static graph compilation.</span>
pipe(prompt=<span class="hljs-string">&quot;warmup&quot;</span>, height=<span class="hljs-number">1024</span>, width=<span class="hljs-number">1024</span>, num_inference_steps=<span class="hljs-number">40</span>, guidance_scale=<span class="hljs-number">4.5</span>)

<span class="hljs-comment"># Later calls with the same shapes reuse the compiled graph.</span>
image = pipe(
    prompt=<span class="hljs-string">&quot;a golden retriever surfing a wave, photorealistic&quot;</span>,
    height=<span class="hljs-number">1024</span>,
    width=<span class="hljs-number">1024</span>,
    num_inference_steps=<span class="hljs-number">40</span>,
    guidance_scale=<span class="hljs-number">4.5</span>,
).images[<span class="hljs-number">0</span>]

image.save(<span class="hljs-string">&quot;output.png&quot;</span>)<!----></pre></div><!----> <p>파이프라인이 하나의 칩에 맞지 않는 경우:</p> <ul><li>여러 개의 칩을 사용하는 경우, 대신 <a href="#tensor-parallelism">tensor parallelism</a>으로 샤딩하세요. 모든 것이 TPU에 유지되며, <code>pipe.transformer.compile(...)</code>은 샤딩된 transformer에서도 동일한 방식으로 작동합니다.</li> <li><a href="/docs/diffusers/main/ko/api/pipelines/overview#diffusers.DiffusionPipeline.enable_model_cpu_offload">enable_model_cpu_offload()</a>를 사용하면 오프로드 훅을 <code>torch.compile</code>로 추적할 수 없습니다. 대신 <code>pipe.transformer.compile_repeated_blocks(backend="tpu", fullgraph=True, dynamic=False)</code>를 사용하여 transformer의 반복 블록만 컴파일하세요.</li></ul> <!--[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 병렬성<!----></span></h2><!--]--><!----> <p>하나의 칩에 너무 큰 모델을 여러 칩에 걸쳐 샤딩합니다. FLUX.2-dev의 텍스트 인코더(~48GB)와 트랜스포머(~64GB)는 각각 단일 칩을 초과하므로, 아래 예시에서는 둘 다 샤딩합니다:</p> <ul><li><code>parallel_config</code> 인수로 전달되는 <a href="/docs/diffusers/main/ko/api/models/overview#diffusers.ModelMixin.from_pretrained">from_pretrained()</a>에 <a href="/docs/diffusers/main/ko/api/parallel#diffusers.TensorParallelConfig">TensorParallelConfig</a>를 갖춘 transformer입니다. 각 rank는 샤딩된 가중치의 자기 자신의 슬라이스만 읽으므로 전체 모델은 절대 메모리에 로드되지 않습니다. 일반적인 TP 세부 사항(<code>_tp_plan</code>, colwise/rowwise)은 <a href="../training/distributed_inference#tensor-parallelism">Tensor parallelism</a> 가이드를 참조하세요.</li> <li>Transformers 자체의 <a href="https://huggingface.co/docs/transformers/perf_infer_gpu_multi" rel="nofollow">tensor parallelism</a>를 사용하여 텍스트 인코더를 구성하고, <a href="https://huggingface.co/docs/transformers/main/en/expert_parallelism#transformers.DistributedConfig" rel="nofollow">DistributedConfig</a>와 동일한 mesh를 전달합니다.</li></ul> <p>TPU에서는 <code>backend="tpu_dist"</code>로 프로세스 그룹을 초기화하고 <code>DeviceMesh("tpu", ...)</code>로 메시를 구성합니다.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5   text-gray-600  " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> 복사됨</div><!----></button><!----></div> <pre class="language-python "><!----><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">from</span> torch.distributed.device_mesh <span class="hljs-keyword">import</span> DeviceMesh
<span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> DistributedConfig, Mistral3ForConditionalGeneration

<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> Flux2Pipeline, Flux2Transformer2DModel, TensorParallelConfig

dist.init_process_group(backend=<span class="hljs-string">&quot;tpu_dist&quot;</span>)
mesh = DeviceMesh(<span class="hljs-string">&quot;tpu&quot;</span>, <span class="hljs-built_in">list</span>(<span class="hljs-built_in">range</span>(dist.get_world_size())))

repo_id = <span class="hljs-string">&quot;black-forest-labs/FLUX.2-dev&quot;</span>
text_encoder = Mistral3ForConditionalGeneration.from_pretrained(
    repo_id,
    subfolder=<span class="hljs-string">&quot;text_encoder&quot;</span>,
    dtype=torch.bfloat16,
    distributed_config=DistributedConfig(tp_plan=<span class="hljs-string">&quot;auto&quot;</span>),
    device_mesh=mesh,
)
transformer = Flux2Transformer2DModel.from_pretrained(
    repo_id, subfolder=<span class="hljs-string">&quot;transformer&quot;</span>, dtype=torch.bfloat16, parallel_config=TensorParallelConfig(mesh=mesh)
)
pipe = Flux2Pipeline.from_pretrained(
    repo_id, text_encoder=text_encoder, transformer=transformer, dtype=torch.bfloat16
)
pipe.vae.to(<span class="hljs-string">&quot;tpu&quot;</span>)

image = pipe(
    prompt=<span class="hljs-string">&quot;a golden retriever surfing a wave, photorealistic&quot;</span>,
    num_inference_steps=<span class="hljs-number">28</span>,
    generator=torch.Generator(<span class="hljs-string">&quot;cpu&quot;</span>).manual_seed(<span class="hljs-number">0</span>),
).images[<span class="hljs-number">0</span>]
<span class="hljs-keyword">if</span> dist.get_rank() == <span class="hljs-number">0</span>:
    image.save(<span class="hljs-string">&quot;output.png&quot;</span>)<!----></pre></div><!----> <p>칩당 하나의 프로세스를 시작합니다. 호스트의 모든 TPU 칩 수를 사용하려면 <code>--nproc_per_node</code>를 설정합니다.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5   text-gray-600  " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> 복사됨</div><!----></button><!----></div> <pre class="language-bash "><!----><span class="hljs-built_in">eval</span> $(python -m torch_tpu._internal.distributed.launchers.singlehost_wrapper | sed <span class="hljs-string">&#x27;s/^/export /&#x27;</span>)
torchrun --nproc_per_node=8 flux2_tp.py<!----></pre></div><!----> <blockquote class="warning"><p>위 시작 명령은 TorchTPU 내부 모듈(<code>torch_tpu._internal</code>)을 사용하며, 이는 안정화된 API가 아니므로 향후 TorchTPU 릴리스에서 변경될 수 있습니다. 멀티 칩 작업을 시작하는 최신 방법은 <a href="https://github.com/google-pytorch/torch_tpu/" rel="nofollow">TorchTPU 저장소</a>에서 확인하세요.</p></blockquote> <a class="!text-gray-400 !no-underline text-sm flex items-center not-prose mt-4" href="https://github.com/huggingface/diffusers/blob/main/docs/source/en/optimization/tpu.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_1rjuniz = {
						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.vZc-dOUR.js"),
						import("/docs/diffusers/main/en/_app/immutable/entry/app.B6c3Ml05.js")
					]).then(([kit, app]) => {
						kit.start(app, element, {
							node_ids: [0, 340],
							data: [null,null],
							form: null,
							error: null
						});
					});
				}
			</script>
		
Xet Storage Details

Size:: 25.8 kB
Xet hash:: 4c3e4cd5eee8182d6c707971fd521d373c485f244e2e9962e6418ef4189496cd
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.