Buckets:

rtrm's picture
download
raw
26.1 kB
<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;큰 모델 인스턴스화&quot;,&quot;local&quot;:&quot;instantiating-a-big-model&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;샤딩된 체크포인트&quot;,&quot;local&quot;:&quot;sharded-checkpoints&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;저(低)메모리 로딩&quot;,&quot;local&quot;:&quot;low-memory-loading&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2}],&quot;depth&quot;:1}">
<link href="/docs/transformers/main/ko/_app/immutable/assets/0.e3b0c442.css" rel="modulepreload">
<link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/entry/start.9aa88961.js">
<link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/chunks/scheduler.9bc65507.js">
<link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/chunks/singletons.9eec45c3.js">
<link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/chunks/index.3b203c72.js">
<link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/chunks/paths.566078f7.js">
<link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/entry/app.84fb67c3.js">
<link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/chunks/index.707bf1b6.js">
<link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/nodes/0.1c99376b.js">
<link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/chunks/each.e59479a4.js">
<link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/nodes/8.bd599368.js">
<link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/chunks/Tip.c2ecdbf4.js">
<link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/chunks/CodeBlock.54a9f38d.js">
<link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/chunks/EditOnGithub.922df6ba.js"><!-- HEAD_svelte-u9bgzb_START --><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;큰 모델 인스턴스화&quot;,&quot;local&quot;:&quot;instantiating-a-big-model&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;샤딩된 체크포인트&quot;,&quot;local&quot;:&quot;sharded-checkpoints&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;저(低)메모리 로딩&quot;,&quot;local&quot;:&quot;low-memory-loading&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2}],&quot;depth&quot;:1}"><!-- HEAD_svelte-u9bgzb_END --> <p></p> <h1 class="relative group"><a id="instantiating-a-big-model" 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="#instantiating-a-big-model"><span><svg class="" 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>큰 모델 인스턴스화</span></h1> <p data-svelte-h="svelte-bgpi4t">매우 큰 사전훈련된 모델을 사용하려면, RAM 사용을 최소화해야 하는 과제가 있습니다. 일반적인 PyTorch 워크플로우는 다음과 같습니다:</p> <ol data-svelte-h="svelte-od014n"><li>무작위 가중치로 모델을 생성합니다.</li> <li>사전훈련된 가중치를 불러옵니다.</li> <li>사전훈련된 가중치를 무작위 모델에 적용합니다.</li></ol> <p data-svelte-h="svelte-uxi99l">1단계와 2단계 모두 모델의 전체 버전을 메모리에 적재해야 하며, 대부분 문제가 없지만 모델이 기가바이트급의 용량을 차지하기 시작하면 복사본 2개가 RAM을 초과하여 메모리 부족 이슈를 야기할 수 있습니다. 더 심각한 문제는 분산 학습을 위해 <code>torch.distributed</code>를 사용하는 경우, 프로세스마다 사전훈련된 모델을 로드하고 복사본을 2개씩 RAM에 저장한다는 것입니다.</p> <div class="course-tip bg-gradient-to-br dark:bg-gradient-to-r before:border-green-500 dark:before:border-green-800 from-green-50 dark:from-gray-900 to-white dark:to-gray-950 border border-green-50 text-green-700 dark:text-gray-400"><p data-svelte-h="svelte-5v0cyc">무작위로 생성된 모델은 “비어 있는” (즉 그때 메모리에 있던 것으로 이뤄진) 텐서로 초기화되며 메모리 공간을 차지합니다. 초기화된 모델/파라미터의 종류에 적합한 분포(예: 정규 분포)에 따른 무작위 초기화는 가능한 한 빠르게 하기 위해 초기화되지 않은 가중치에 대해 3단계 이후에만 수행됩니다!</p></div> <p data-svelte-h="svelte-1xr3byf">이 안내서에서는 Transformers가 이 문제를 해결하기 위해 제공하는 솔루션을 살펴봅니다. 주의할 점은 아직 활발히 개발 중인 분야이므로 여기서 설명하는 API가 앞으로 약간 변경될 수 있다는 것입니다.</p> <h2 class="relative group"><a id="sharded-checkpoints" 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="#sharded-checkpoints"><span><svg class="" 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>샤딩된 체크포인트</span></h2> <p data-svelte-h="svelte-hfrx14">4.18.0 버전 이후, 10GB 이상의 공간을 차지하는 모델 체크포인트는 자동으로 작은 조각들로 샤딩됩니다. <code>model.save_pretrained(save_dir)</code>를 실행할 때 하나의 단일 체크포인트를 가지게 될 대신, 여러 부분 체크포인트(각각의 크기는 10GB 미만)와 매개변수 이름을 해당 파일에 매핑하는 인덱스가 생성됩니다.</p> <p data-svelte-h="svelte-tmoic1"><code>max_shard_size</code> 매개변수로 샤딩 전 최대 크기를 제어할 수 있으므로, 이 예제를 위해 샤드 크기가 작은 일반 크기의 모델을 사용하겠습니다: 전통적인 BERT 모델을 사용해 봅시다.</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 class="" 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> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> AutoModel
model = AutoModel.from_pretrained(<span class="hljs-string">&quot;google-bert/bert-base-cased&quot;</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1akqon3"><code>save_pretrained()</code>을 사용하여 모델을 저장하면, 모델의 구성과 가중치가 들어있는 두 개의 파일이 있는 새 폴더가 생성됩니다:</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 class="" 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> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">import</span> os
<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">import</span> tempfile
<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">with</span> tempfile.TemporaryDirectory() <span class="hljs-keyword">as</span> tmp_dir:
<span class="hljs-meta">... </span> model.save_pretrained(tmp_dir)
<span class="hljs-meta">... </span> <span class="hljs-built_in">print</span>(<span class="hljs-built_in">sorted</span>(os.listdir(tmp_dir)))
[<span class="hljs-string">&#x27;config.json&#x27;</span>, <span class="hljs-string">&#x27;pytorch_model.bin&#x27;</span>]<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-148f27j">이제 최대 샤드 크기를 200MB로 사용해 봅시다:</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 class="" 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> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">with</span> tempfile.TemporaryDirectory() <span class="hljs-keyword">as</span> tmp_dir:
<span class="hljs-meta">... </span> model.save_pretrained(tmp_dir, max_shard_size=<span class="hljs-string">&quot;200MB&quot;</span>)
<span class="hljs-meta">... </span> <span class="hljs-built_in">print</span>(<span class="hljs-built_in">sorted</span>(os.listdir(tmp_dir)))
[<span class="hljs-string">&#x27;config.json&#x27;</span>, <span class="hljs-string">&#x27;pytorch_model-00001-of-00003.bin&#x27;</span>, <span class="hljs-string">&#x27;pytorch_model-00002-of-00003.bin&#x27;</span>, <span class="hljs-string">&#x27;pytorch_model-00003-of-00003.bin&#x27;</span>, <span class="hljs-string">&#x27;pytorch_model.bin.index.json&#x27;</span>]<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1fn1hkf">모델의 구성에 더해, 세 개의 다른 가중치 파일과 파라미터 이름과 해당 파일의 매핑이 포함된 <code>index.json</code> 파일을 볼 수 있습니다. 이러한 체크포인트는 <code>from_pretrained()</code> 메서드를 사용하여 완전히 다시 로드할 수 있습니다:</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 class="" 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> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">with</span> tempfile.TemporaryDirectory() <span class="hljs-keyword">as</span> tmp_dir:
<span class="hljs-meta">... </span> model.save_pretrained(tmp_dir, max_shard_size=<span class="hljs-string">&quot;200MB&quot;</span>)
<span class="hljs-meta">... </span> new_model = AutoModel.from_pretrained(tmp_dir)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-dbx6i5">큰 모델의 경우 이러한 방식으로 처리하는 주된 장점은 위에서 보여준 흐름의 2단계에서, 각 샤드가 이전 샤드 다음에 로드되므로 메모리 사용량이 모델 크기와 가장 큰 샤드의 크기를 초과하지 않는다는 점입니다.</p> <p data-svelte-h="svelte-1s70s7e">이 인덱스 파일은 키가 체크포인트에 있는지, 그리고 해당 가중치가 어디에 저장되어 있는지를 결정하는 데 사용됩니다. 이 인덱스를 json과 같이 로드하고 딕셔너리를 얻을 수 있습니다:</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 class="" 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> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">import</span> json
<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">with</span> tempfile.TemporaryDirectory() <span class="hljs-keyword">as</span> tmp_dir:
<span class="hljs-meta">... </span> model.save_pretrained(tmp_dir, max_shard_size=<span class="hljs-string">&quot;200MB&quot;</span>)
<span class="hljs-meta">... </span> <span class="hljs-keyword">with</span> <span class="hljs-built_in">open</span>(os.path.join(tmp_dir, <span class="hljs-string">&quot;pytorch_model.bin.index.json&quot;</span>), <span class="hljs-string">&quot;r&quot;</span>) <span class="hljs-keyword">as</span> f:
<span class="hljs-meta">... </span> index = json.load(f)
<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-built_in">print</span>(index.keys())
dict_keys([<span class="hljs-string">&#x27;metadata&#x27;</span>, <span class="hljs-string">&#x27;weight_map&#x27;</span>])<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-3pru07">메타데이터는 현재 모델의 총 크기만 포함됩니다. 앞으로 다른 정보를 추가할 계획입니다:</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 class="" 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> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-meta">&gt;&gt;&gt; </span>index[<span class="hljs-string">&quot;metadata&quot;</span>]
{<span class="hljs-string">&#x27;total_size&#x27;</span>: <span class="hljs-number">433245184</span>}<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-fyu3ex">가중치 맵은 이 인덱스의 주요 부분으로, 각 매개변수 이름(PyTorch 모델 <code>state_dict</code>에서 보통 찾을 수 있는)을 해당 파일에 매핑합니다:</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 class="" 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> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-meta">&gt;&gt;&gt; </span>index[<span class="hljs-string">&quot;weight_map&quot;</span>]
{<span class="hljs-string">&#x27;embeddings.LayerNorm.bias&#x27;</span>: <span class="hljs-string">&#x27;pytorch_model-00001-of-00003.bin&#x27;</span>,
<span class="hljs-string">&#x27;embeddings.LayerNorm.weight&#x27;</span>: <span class="hljs-string">&#x27;pytorch_model-00001-of-00003.bin&#x27;</span>,
...<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1gt3bpq">만약 <code>from_pretrained()</code>를 사용하지 않고 모델 내에서 이러한 샤딩된 체크포인트를 직접 가져오려면 (전체 체크포인트를 위해 <code>model.load_state_dict()</code>를 수행하는 것처럼), <code>load_sharded_checkpoint()</code>를 사용해야 합니다.</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 class="" 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> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">from</span> transformers.modeling_utils <span class="hljs-keyword">import</span> load_sharded_checkpoint
<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">with</span> tempfile.TemporaryDirectory() <span class="hljs-keyword">as</span> tmp_dir:
<span class="hljs-meta">... </span> model.save_pretrained(tmp_dir, max_shard_size=<span class="hljs-string">&quot;200MB&quot;</span>)
<span class="hljs-meta">... </span> load_sharded_checkpoint(model, tmp_dir)<!-- HTML_TAG_END --></pre></div> <h2 class="relative group"><a id="low-memory-loading" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#low-memory-loading"><span><svg class="" 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>저(低)메모리 로딩</span></h2> <p data-svelte-h="svelte-1w5jvsg">샤딩된 체크포인트는 위에서 언급한 작업 흐름의 2단계에서 메모리 사용량을 줄이지만, 저(低)메모리 설정에서 모델을 사용하기 위해 우리의 Accelerate 라이브러리를 기반으로 한 도구를 활용하는 것이 좋습니다.</p> <p data-svelte-h="svelte-1u81xgf">자세한 사항은 다음 가이드를 참조해주세요: <a href="../en/main_classes/model#large-model-loading">Accelerate로 대규모 모델 가져오기 (영문)</a></p> <a class="!text-gray-400 !no-underline text-sm flex items-center not-prose mt-4" href="https://github.com/huggingface/transformers/blob/main/docs/source/ko/big_models.md" target="_blank"><span data-svelte-h="svelte-1kd6by1">&lt;</span> <span data-svelte-h="svelte-x0xyl0">&gt;</span> <span data-svelte-h="svelte-1dajgef"><span class="underline ml-1.5">Update</span> on GitHub</span></a> <p></p>
<script>
{
__sveltekit_1hrx8 = {
assets: "/docs/transformers/main/ko",
base: "/docs/transformers/main/ko",
env: {}
};
const element = document.currentScript.parentElement;
const data = [null,null];
Promise.all([
import("/docs/transformers/main/ko/_app/immutable/entry/start.9aa88961.js"),
import("/docs/transformers/main/ko/_app/immutable/entry/app.84fb67c3.js")
]).then(([kit, app]) => {
kit.start(app, element, {
node_ids: [0, 8],
data,
form: null,
error: null
});
});
}
</script>

Xet Storage Details

Size:
26.1 kB
·
Xet hash:
970441aa37cf767df57ba0efc605d9d642a92d2a618e504131008cae69708682

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