Buckets:
| <meta charset="utf-8" /><meta name="hf:doc:metadata" content="{"title":"모델 로드하기","local":"loading-models","sections":[{"title":"모델과 구성","local":"models-and-configurations","sections":[],"depth":2},{"title":"모델 클래스","local":"model-classes","sections":[],"depth":2},{"title":"대규모 모델","local":"large-models","sections":[{"title":"분할된 체크포인트","local":"sharded-checkpoints","sections":[],"depth":3},{"title":"대형 모델 추론","local":"big-model-inference","sections":[],"depth":3},{"title":"모델 데이터 타입","local":"model-data-type","sections":[],"depth":3}],"depth":2},{"title":"커스텀 모델","local":"custom-models","sections":[],"depth":2}],"depth":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.5e38b416.js"> | |
| <link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/chunks/scheduler.53228c21.js"> | |
| <link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/chunks/singletons.0f395d7b.js"> | |
| <link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/chunks/index.e93d0901.js"> | |
| <link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/chunks/paths.6e1a5eaf.js"> | |
| <link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/entry/app.240e8e3b.js"> | |
| <link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/chunks/preload-helper.cb103237.js"> | |
| <link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/chunks/index.3db2ce32.js"> | |
| <link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/nodes/0.8e23ad79.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/125.5c2b79e3.js"> | |
| <link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/chunks/CopyLLMTxtMenu.1327b590.js"> | |
| <link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/chunks/MermaidChart.svelte_svelte_type_style_lang.49b88d99.js"> | |
| <link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/chunks/Youtube.b27f3a4a.js"> | |
| <link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/chunks/CodeBlock.ada04ea6.js"> | |
| <link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/chunks/HfOption.be649c8b.js"><!-- HEAD_svelte-u9bgzb_START --><meta name="hf:doc:metadata" content="{"title":"모델 로드하기","local":"loading-models","sections":[{"title":"모델과 구성","local":"models-and-configurations","sections":[],"depth":2},{"title":"모델 클래스","local":"model-classes","sections":[],"depth":2},{"title":"대규모 모델","local":"large-models","sections":[{"title":"분할된 체크포인트","local":"sharded-checkpoints","sections":[],"depth":3},{"title":"대형 모델 추론","local":"big-model-inference","sections":[],"depth":3},{"title":"모델 데이터 타입","local":"model-data-type","sections":[],"depth":3}],"depth":2},{"title":"커스텀 모델","local":"custom-models","sections":[],"depth":2}],"depth":1}"><!-- HEAD_svelte-u9bgzb_END --> <p></p> <div class="items-center shrink-0 min-w-[100px] max-sm:min-w-[50px] justify-end ml-auto flex" style="float: right; margin-left: 10px; display: inline-flex; position: relative; z-index: 10;"><div 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>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> </div> <h1 class="relative group"><a id="loading-models" 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="#loading-models"><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>모델 로드하기</span></h1> <p data-svelte-h="svelte-9iqtfr">Transformers는 한 줄의 코드로 사용할 수 있는 많은 사전 훈련된 모델을 제공합니다. 모델 클래스와 <a href="/docs/transformers/main/ko/main_classes/model#transformers.PreTrainedModel.from_pretrained">from_pretrained()</a> 메소드가 필요합니다.</p> <p data-svelte-h="svelte-1wcp6eg"><a href="/docs/transformers/main/ko/main_classes/model#transformers.PreTrainedModel.from_pretrained">from_pretrained()</a>를 호출하여 Hugging Face <a href="https://hf.co/models" rel="nofollow">Hub</a>에 저장된 모델의 가중치와 구성을 다운로드하고 로드하세요.</p> <blockquote data-svelte-h="svelte-1pa0gpc"><p>[!TIP][from_pretrained()](/docs/transformers/main/ko/main_classes/model#transformers.PreTrainedModel.from_pretrained) 메소드는 <a href="https://hf.co/docs/safetensors/index" rel="nofollow">safetensors</a> 파일 형식으로 저장된 가중치가 있으면 이를 로드합니다. 전통적으로 PyTorch 모델 가중치는 보안에 취약한 것으로 알려진 <a href="https://docs.python.org/3/library/pickle.html" rel="nofollow">pickle</a> 유틸리티로 직렬화됩니다. Safetensor 파일은 더 안전하고 로드 속도가 빠릅니다.</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 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="language-py "><!-- HTML_TAG_START --><span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> AutoModelForCausalLM | |
| model = AutoModelForCausalLM.from_pretrained(<span class="hljs-string">"meta-llama/Llama-2-7b-hf"</span>, torch_dtype=<span class="hljs-string">"auto"</span>, device_map=<span class="hljs-string">"auto"</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-jyvk8d">이 가이드는 모델을 불러오는 방법, 다양한 로딩 방식, 매우 큰 모델에서 발생할 수 있는 메모리 문제를 해결하는 방법, 그리고 사용자 정의 모델을 불러오는 방법을 설명합니다.</p> <h2 class="relative group"><a id="models-and-configurations" 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="#models-and-configurations"><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>모델과 구성</span></h2> <p data-svelte-h="svelte-1a8qjay">모든 모델에는 은닉 레이어 수, 어휘 사전 크기, 활성화 함수 등과 같은 특정 속성이 포함된 <code>configuration.py</code> 파일이 있습니다. 또한 각 레이어의 정의와 각각의 레이어 안에서 일어나는 수학적 연산을 정의하는 <code>modeling.py</code> 파일도 있습니다. <code>modeling.py</code> 파일은 <code>configuration.py</code>에 정의된 모델 속성을 바탕으로 모델을 구축합니다. 이 단계에서는 아직 학습되지 않은 무작위 가중치를 가진 상태이기 때문에, 의미 있는 출력을 얻기 위해서는 학습이 필요합니다.</p> <blockquote class="tip" data-svelte-h="svelte-skuufl"><p><em>아키텍처(Architecture)</em>는 모델의 골격을 의미하고 <em>체크포인트(checkpoint)</em>는 주어진 아키텍처에 대한 모델의 가중치를 의미합니다. 예를 들어, <a href="./model_doc/bert">BERT</a>는 아키텍처이고 <a href="https://huggingface.co/google-bert/bert-base-uncased" rel="nofollow">google-bert/bert-base-uncased</a>는 해당 아키텍처의 체크포인트(checkpoint)입니다. <em>모델</em>이라는 용어는 아키텍처 및 체크포인트(checkpoint)와 혼용하여 사용되는 것을 볼 수 있습니다.</p></blockquote> <p data-svelte-h="svelte-nk8v77">로드할 수 있는 모델은 일반적으로 두 가지 타입이 있습니다.</p> <ol data-svelte-h="svelte-1f7201i"><li>은닉 상태를 출력하는 <a href="/docs/transformers/main/ko/model_doc/auto#transformers.AutoModel">AutoModel</a> 또는 <a href="/docs/transformers/main/ko/model_doc/llama2#transformers.LlamaModel">LlamaModel</a>과 같은 기본 모델입니다.</li> <li>특정 작업을 수행하기 위해 특정 <em>헤드</em>가 붙은 <a href="/docs/transformers/main/ko/model_doc/auto#transformers.AutoModelForCausalLM">AutoModelForCausalLM</a> 또는 <a href="/docs/transformers/main/ko/model_doc/llama2#transformers.LlamaForCausalLM">LlamaForCausalLM</a>과 같은 모델입니다.</li></ol> <p data-svelte-h="svelte-1fj0089">각 모델 타입마다, 각각의 기계학습 프레임워크(PyTorch, TensorFlow, Flax)를 위한 별도의 클래스가 있습니다. 사용 중인 프레임워크에 해당하는 접두어(prefix)를 선택하세요.</p> <div class="flex space-x-2 items-center my-1.5 mr-8 h-7 !pl-0 -mx-3 md:mx-0"><div class="flex items-center border rounded-lg px-1.5 py-1 leading-none select-none text-smd border-gray-800 bg-black dark:bg-gray-700 text-white">PyTorch </div><div class="flex items-center border rounded-lg px-1.5 py-1 leading-none select-none text-smd text-gray-500 cursor-pointer opacity-90 hover:text-gray-700 dark:hover:text-gray-200 hover:shadow-sm">TensorFlow </div><div class="flex items-center border rounded-lg px-1.5 py-1 leading-none select-none text-smd text-gray-500 cursor-pointer opacity-90 hover:text-gray-700 dark:hover:text-gray-200 hover:shadow-sm">Flax </div></div> <div class="language-select"><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="language-py "><!-- HTML_TAG_START --><span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> AutoModelForCausalLM, MistralForCausalLM | |
| <span class="hljs-comment"># AutoClass 또는 모델별 클래스(model-specific class) 를 이용해 로드</span> | |
| model = AutoModelForCausalLM.from_pretrained(<span class="hljs-string">"mistralai/Mistral-7B-v0.1"</span>, torch_dtype=<span class="hljs-string">"auto"</span>, device_map=<span class="hljs-string">"auto"</span>) | |
| model = MistralForCausalLM.from_pretrained(<span class="hljs-string">"mistralai/Mistral-7B-v0.1"</span>, torch_dtype=<span class="hljs-string">"auto"</span>, device_map=<span class="hljs-string">"auto"</span>)<!-- HTML_TAG_END --></pre></div> </div> <h2 class="relative group"><a id="model-classes" 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="#model-classes"><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>모델 클래스</span></h2> <p data-svelte-h="svelte-5njp0p">사전 훈련된 모델을 가져오려면 모델에 가중치를 로드해야 합니다. 이는 Hugging Face Hub 또는 로컬 디렉터리에서 가중치를 받아들이는 <a href="/docs/transformers/main/ko/main_classes/model#transformers.PreTrainedModel.from_pretrained">from_pretrained()</a>를 호출하여 수행됩니다.</p> <p data-svelte-h="svelte-wu56qr">두 가지 모델 클래스로 <a href="./model_doc/auto">AutoModel</a> 클래스와 모델별 클래스가 있습니다.</p> <div class="flex space-x-2 items-center my-1.5 mr-8 h-7 !pl-0 -mx-3 md:mx-0"><div class="flex items-center border rounded-lg px-1.5 py-1 leading-none select-none text-smd border-gray-800 bg-black dark:bg-gray-700 text-white">AutoModel </div><div class="flex items-center border rounded-lg px-1.5 py-1 leading-none select-none text-smd text-gray-500 cursor-pointer opacity-90 hover:text-gray-700 dark:hover:text-gray-200 hover:shadow-sm">model-specific class </div></div> <div class="language-select"><iframe class="w-full xl:w-4/6 h-80" src="https://www.youtube-nocookie.com/embed/AhChOFRegn4" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture" allowfullscreen></iframe> <p data-svelte-h="svelte-tjbbe1"><a href="./model_doc/auto">AutoModel</a> 클래스는 정확한 모델 클래스 이름을 몰라도 아키텍처를 불러올 수 있는 편리한 방법입니다. 많은 모델이 제공되기 때문에, 이 클래스는 구성 파일을 기반으로 올바른 모델 클래스를 자동으로 선택해 줍니다. 원하는 작업과 사용하려는 체크포인트만 알고 있으면 됩니다.</p> <p data-svelte-h="svelte-13mx0dp">주어진 작업을 아키텍처가 지원하는 한, 모델이나 작업을 쉽게 전환할 수 있습니다.</p> <p data-svelte-h="svelte-1v4a24t">예를 들어, 동일한 모델을 서로 다른 작업에 사용할 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="language-py "><!-- HTML_TAG_START --><span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> AutoModelForCausalLM, AutoModelForSequenceClassification, AutoModelForQuestionAnswering | |
| <span class="hljs-comment"># 동일한 API를 3가지 다른 작업에 사용</span> | |
| model = AutoModelForCausalLM.from_pretrained(<span class="hljs-string">"meta-llama/Llama-2-7b-hf"</span>) | |
| model = AutoModelForSequenceClassification.from_pretrained(<span class="hljs-string">"meta-llama/Llama-2-7b-hf"</span>) | |
| model = AutoModelForQuestionAnswering.from_pretrained(<span class="hljs-string">"meta-llama/Llama-2-7b-hf"</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1333zve">다른 경우에는, 하나의 작업에 대해 여러 가지 모델을 빠르게 시험해보고 싶을 수도 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="language-py "><!-- HTML_TAG_START --><span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> AutoModelForCausalLM | |
| <span class="hljs-comment"># 동일한 API를 사용하여 3가지 다른 모델 로드</span> | |
| model = AutoModelForCausalLM.from_pretrained(<span class="hljs-string">"meta-llama/Llama-2-7b-hf"</span>) | |
| model = AutoModelForCausalLM.from_pretrained(<span class="hljs-string">"mistralai/Mistral-7B-v0.1"</span>) | |
| model = AutoModelForCausalLM.from_pretrained(<span class="hljs-string">"google/gemma-7b"</span>)<!-- HTML_TAG_END --></pre></div> </div> <h2 class="relative group"><a id="large-models" 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="#large-models"><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>대규모 모델</span></h2> <p data-svelte-h="svelte-1fkzkq3">대규모 사전 훈련된 모델은 로드하는 데 많은 메모리가 필요합니다. 로드 과정은 다음을 포함합니다:</p> <ol data-svelte-h="svelte-1axci69"><li>무작위 가중치로 모델 생성</li> <li>사전 훈련된 가중치 로드</li> <li>사전 훈련된 가중치를 모델에 적용</li></ol> <p data-svelte-h="svelte-jr7u67">모델 가중치의 복사본 두 가지(무작위 가중치와 사전 훈련된 가중치)를 보관할 수 있는 충분한 메모리가 필요하며, 이는 보유한 하드웨어에 따라 불가능할 수 있습니다. 분산 학습 환경에서는 각 프로세스가 사전 훈련된 모델을 로드하기 때문에 이는 더욱 어려운 과제입니다.</p> <p data-svelte-h="svelte-1gpid5w">transformers는 빠른 초기화, 분할된 체크포인트, Accelerate의 <a href="https://hf.co/docs/accelerate/usage_guides/big_modeling" rel="nofollow">Big Model Inference</a> 기능, 그리고 더 낮은 비트 데이터 타입 지원을 통해 이러한 메모리 관련 문제들을 일부 줄여줍니다.</p> <h3 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>분할된 체크포인트</span></h3> <p data-svelte-h="svelte-1vk48rz"><a href="/docs/transformers/main/ko/main_classes/model#transformers.PreTrainedModel.save_pretrained">save_pretrained()</a> 메소드는 10GB보다 큰 체크포인트를 자동으로 샤드합니다.</p> <p data-svelte-h="svelte-bvo5rb">각 샤드(shard)는 이전 샤드가 로드된 후 순차적으로 로드되어, 메모리 사용량을 모델 크기와 가장 큰 샤드 크기로만 제한합니다.</p> <p data-svelte-h="svelte-et2duh"><code>max_shard_size</code> 매개변수는 각 샤드에 대해 기본적으로 5GB로 설정되어 있는데, 이는 메모리 부족 없이 무료 등급 GPU 인스턴스에서 더 쉽게 실행할 수 있기 때문입니다.</p> <p data-svelte-h="svelte-a65bc7">예를 들어, <a href="/docs/transformers/main/ko/main_classes/model#transformers.PreTrainedModel.save_pretrained">save_pretrained()</a>에서 <a href="https://hf.co/BioMistral/BioMistral-7B" rel="nofollow">BioMistral/BioMistral-7B</a>에 대한 분할된 체크포인트를 생성해보겠습니다.</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="language-py "><!-- HTML_TAG_START --><span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> AutoModel | |
| <span class="hljs-keyword">import</span> tempfile | |
| <span class="hljs-keyword">import</span> os | |
| model = AutoModel.from_pretrained(<span class="hljs-string">"biomistral/biomistral-7b"</span>) | |
| <span class="hljs-keyword">with</span> tempfile.TemporaryDirectory() <span class="hljs-keyword">as</span> tmp_dir: | |
| model.save_pretrained(tmp_dir, max_shard_size=<span class="hljs-string">"5GB"</span>) | |
| <span class="hljs-built_in">print</span>(<span class="hljs-built_in">sorted</span>(os.listdir(tmp_dir)))<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-wxpkyq"><a href="/docs/transformers/main/ko/main_classes/model#transformers.PreTrainedModel.from_pretrained">from_pretrained()</a>로 분할된 체크포인트를 다시 로드합니다.</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="language-py "><!-- HTML_TAG_START --><span class="hljs-keyword">with</span> tempfile.TemporaryDirectory() <span class="hljs-keyword">as</span> tmp_dir: | |
| model.save_pretrained(tmp_dir) | |
| new_model = AutoModel.from_pretrained(tmp_dir)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1w6yaz6">분할된 체크포인트는 <code>load_sharded_checkpoint()</code>로도 직접 불러올 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="language-py "><!-- HTML_TAG_START --><span class="hljs-keyword">from</span> transformers.trainer_utils <span class="hljs-keyword">import</span> load_sharded_checkpoint | |
| <span class="hljs-keyword">with</span> tempfile.TemporaryDirectory() <span class="hljs-keyword">as</span> tmp_dir: | |
| model.save_pretrained(tmp_dir, max_shard_size=<span class="hljs-string">"5GB"</span>) | |
| load_sharded_checkpoint(model, tmp_dir)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1kt66yn"><a href="/docs/transformers/main/ko/main_classes/model#transformers.PreTrainedModel.save_pretrained">save_pretrained()</a> 메소드는 매개변수 이름을 저장된 파일에 매핑하는 인덱스 파일을 생성합니다. 인덱스 파일에는 <code>metadata</code>와 <code>weight_map</code>이라는 두 개의 키가 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="language-py "><!-- HTML_TAG_START --><span class="hljs-keyword">import</span> json | |
| <span class="hljs-keyword">with</span> tempfile.TemporaryDirectory() <span class="hljs-keyword">as</span> tmp_dir: | |
| model.save_pretrained(tmp_dir, max_shard_size=<span class="hljs-string">"5GB"</span>) | |
| <span class="hljs-keyword">with</span> <span class="hljs-built_in">open</span>(os.path.join(tmp_dir, <span class="hljs-string">"model.safetensors.index.json"</span>), <span class="hljs-string">"r"</span>) <span class="hljs-keyword">as</span> f: | |
| index = json.load(f) | |
| <span class="hljs-built_in">print</span>(index.keys())<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1wch5uu"><code>metadata</code> 키는 전체 모델 크기를 제공합니다.</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="language-py "><!-- HTML_TAG_START -->index[<span class="hljs-string">"metadata"</span>] | |
| {<span class="hljs-string">'total_size'</span>: <span class="hljs-number">28966928384</span>}<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1dyy9nc"><code>weight_map</code> 키는 각 매개변수를 저장된 샤드에 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="language-py "><!-- HTML_TAG_START -->index[<span class="hljs-string">"weight_map"</span>] | |
| {<span class="hljs-string">'lm_head.weight'</span>: <span class="hljs-string">'model-00006-of-00006.safetensors'</span>, | |
| <span class="hljs-string">'model.embed_tokens.weight'</span>: <span class="hljs-string">'model-00001-of-00006.safetensors'</span>, | |
| <span class="hljs-string">'model.layers.0.input_layernorm.weight'</span>: <span class="hljs-string">'model-00001-of-00006.safetensors'</span>, | |
| <span class="hljs-string">'model.layers.0.mlp.down_proj.weight'</span>: <span class="hljs-string">'model-00001-of-00006.safetensors'</span>, | |
| ... | |
| }<!-- HTML_TAG_END --></pre></div> <h3 class="relative group"><a id="big-model-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="#big-model-inference"><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>대형 모델 추론</span></h3> <blockquote class="tip" data-svelte-h="svelte-18wm6c8"><p>이 기능을 사용하려면 Accelerate v0.9.0 및 PyTorch v1.9.0 이상이 설치되어 있는지 확인하세요!</p></blockquote> <iframe class="w-full xl:w-4/6 h-80" src="https://www.youtube-nocookie.com/embed/MWCSGj9jEAo" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture" allowfullscreen></iframe> <p data-svelte-h="svelte-1kzh605"><a href="/docs/transformers/main/ko/main_classes/model#transformers.PreTrainedModel.from_pretrained">from_pretrained()</a>는 Accelerate의 <a href="https://hf.co/docs/accelerate/usage_guides/big_modeling" rel="nofollow">대형 모델 추론</a> 기능으로 강화되었습니다.</p> <p data-svelte-h="svelte-24dsu5">대형 모델 추론은 PyTorch <a href="https://pytorch.org/docs/main/meta.html" rel="nofollow">meta</a> 장치에서 <em>모델 스켈레톤</em>을 생성합니다. meta 장치는 실제 데이터를 저장하지 않고 메타데이터만 저장합니다.</p> <p data-svelte-h="svelte-11cr7gh">무작위로 초기화된 가중치는 사전 훈련된 가중치가 로드될 때만 생성되어 메모리에 동시에 모델의 두 복사본을 유지하는 것을 방지합니다. 최대 메모리 사용량은 모델 크기만큼입니다.</p> <blockquote class="tip" data-svelte-h="svelte-1hnnr38"><p>장치 할당에 대한 자세한 내용은 <a href="https://hf.co/docs/accelerate/v0.33.0/en/concept_guides/big_model_inference#designing-a-device-map" rel="nofollow">장치 맵 설계하기</a>를 참조하세요.</p></blockquote> <p data-svelte-h="svelte-1auhvfy">대형 모델 추론의 두 번째 기능은 불러온 가중치가 모델 스켈레톤에 할당되는 방식과 관련이 있습니다. 모델 가중치는 사용 가능한 모든 디바이스에 분산되며, 가장 빠른 디바이스(보통 GPU)부터 시작해 나머지 가중치는 느린 디바이스(CPU 및 하드 디스크)로 순차적으로 할당됩니다.</p> <p data-svelte-h="svelte-pdqep3">두 기능을 결합하면 대형 사전 훈련된 모델의 메모리 사용량과 로딩 시간이 줄어듭니다.</p> <p data-svelte-h="svelte-l1gjl9">대형 모델 추론을 활성화하려면 <a href="https://github.com/huggingface/transformers/blob/026a173a64372e9602a16523b8fae9de4b0ff428/src/transformers/modeling_utils.py#L3061" rel="nofollow">device_map</a>을 <code>"auto"</code>로 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 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="language-py "><!-- HTML_TAG_START --><span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> AutoModelForCausalLM | |
| model = AutoModelForCausalLM.from_pretrained(<span class="hljs-string">"google/gemma-7b"</span>, device_map=<span class="hljs-string">"auto"</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-fs23vt"><code>device_map</code>을 사용하여 레이어를 원하는 디바이스에 수동으로 할당할 수도 있습니다. 레이어 전체가 동일한 디바이스에 할당되어 있다면, 해당 레이어의 모든 서브모듈이 어디에 배치되는지 일일이 지정할 필요는 없습니다.</p> <p data-svelte-h="svelte-1nvlmqp">모델이 각 디바이스에 어떻게 분산되어 있는지 확인하려면 <code>hf_device_map</code> 속성을 조회합니다.</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="language-py "><!-- HTML_TAG_START -->device_map = {<span class="hljs-string">"model.layers.1"</span>: <span class="hljs-number">0</span>, <span class="hljs-string">"model.layers.14"</span>: <span class="hljs-number">1</span>, <span class="hljs-string">"model.layers.31"</span>: <span class="hljs-string">"cpu"</span>, <span class="hljs-string">"lm_head"</span>: <span class="hljs-string">"disk"</span>} | |
| model.hf_device_map<!-- HTML_TAG_END --></pre></div> <h3 class="relative group"><a id="model-data-type" 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="#model-data-type"><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>모델 데이터 타입</span></h3> <p data-svelte-h="svelte-ois8e2">PyTorch 모델 가중치는 기본적으로 <code>torch.float32</code>로 초기화됩니다. <code>torch.float16</code>과 같이 다른 데이터 타입으로 모델을 로드하면 모델이 원하는 데이터 타입으로 다시 로드되기 때문에 추가 메모리가 필요합니다.</p> <p data-svelte-h="svelte-ztdkxa"><a href="https://pytorch.org/docs/stable/tensor_attributes.html#torch.dtype" rel="nofollow">torch_dtype</a> 매개변수를 명시적으로 설정하여 가중치를 두 번 로드하는 대신(<code>torch.float32</code> 후 <code>torch.float16</code>) 원하는 데이터 타입으로 모델을 직접 초기화하세요. 또한 <code>torch_dtype="auto"</code>를 설정하여 가중치를 저장된 데이터 타입으로 자동으로 로드할 수도 있습니다.</p> <div class="flex space-x-2 items-center my-1.5 mr-8 h-7 !pl-0 -mx-3 md:mx-0"><div class="flex items-center border rounded-lg px-1.5 py-1 leading-none select-none text-smd border-gray-800 bg-black dark:bg-gray-700 text-white">specific dtype </div><div class="flex items-center border rounded-lg px-1.5 py-1 leading-none select-none text-smd text-gray-500 cursor-pointer opacity-90 hover:text-gray-700 dark:hover:text-gray-200 hover:shadow-sm">auto dtype </div></div> <div class="language-select"><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="language-py "><!-- HTML_TAG_START --><span class="hljs-keyword">import</span> torch | |
| <span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> AutoModelForCausalLM | |
| gemma = AutoModelForCausalLM.from_pretrained(<span class="hljs-string">"google/gemma-7b"</span>, torch_dtype=torch.float16)<!-- HTML_TAG_END --></pre></div> </div> <p data-svelte-h="svelte-3cvyec">모델을 처음부터 인스턴스화하는 경우, <code>torch_dtype</code> 파라미터는 <a href="/docs/transformers/main/ko/model_doc/auto#transformers.AutoConfig">AutoConfig</a>에서 설정할 수도 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="language-py "><!-- HTML_TAG_START --><span class="hljs-keyword">import</span> torch | |
| <span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> AutoConfig, AutoModel | |
| my_config = AutoConfig.from_pretrained(<span class="hljs-string">"google/gemma-2b"</span>, torch_dtype=torch.float16) | |
| model = AutoModel.from_config(my_config)<!-- HTML_TAG_END --></pre></div> <h2 class="relative group"><a id="custom-models" 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="#custom-models"><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>커스텀 모델</span></h2> <p data-svelte-h="svelte-1rx2dj3">커스텀 모델은 트랜스포머의 구성 및 모델링 클래스를 기반으로 구축되며, <a href="#autoclass">AutoClass</a> API를 지원하고 <a href="/docs/transformers/main/ko/main_classes/model#transformers.PreTrainedModel.from_pretrained">from_pretrained()</a>로 로드됩니다. 차이점은 모델링 코드가 트랜스포머에서 제공되는 것이 <em>아니라는</em> 점입니다.</p> <p data-svelte-h="svelte-1aj1pl8">커스텀 모델을 로드할 때는 특별히 주의해야 합니다. Hub에는 모든 저장소에 대한 <a href="https://hf.co/docs/hub/security-malware#malware-scanning" rel="nofollow">악성코드 스캔</a>이 포함되어 있지만, 여전히 실수로 악성코드를 실행하지 않도록 주의해야 합니다.</p> <p data-svelte-h="svelte-10cu1gn">커스텀 모델을 로드하려면 <a href="/docs/transformers/main/ko/main_classes/model#transformers.PreTrainedModel.from_pretrained">from_pretrained()</a>에서 <code>trust_remote_code=True</code>를 설정하세요.</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="language-py "><!-- HTML_TAG_START --><span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> AutoModelForImageClassification | |
| model = AutoModelForImageClassification.from_pretrained(<span class="hljs-string">"sgugger/custom-resnet50d"</span>, trust_remote_code=<span class="hljs-literal">True</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1xlbls8">추가적인 보안 조치로, 변경되었을 수도 있는 모델 코드를 로드하는 것을 피하기 위해 특정 리비전에서 커스텀 모델을 로드합니다. 커밋 해시는 모델의 <a href="https://hf.co/sgugger/custom-resnet50d/commits/main" rel="nofollow">커밋 기록</a>에서 복사할 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="language-py "><!-- HTML_TAG_START -->commit_hash = <span class="hljs-string">"ed94a7c6247d8aedce4647f00f20de6875b5b292"</span> | |
| model = AutoModelForImageClassification.from_pretrained( | |
| <span class="hljs-string">"sgugger/custom-resnet50d"</span>, trust_remote_code=<span class="hljs-literal">True</span>, revision=commit_hash | |
| )<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-pvk0aq">자세한 내용은 <a href="./custom_models">사용자 정의 모델 공유하기</a> 가이드를 참조하세요.</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/models.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 data-svelte-h="svelte-zjs2n5"><span class="underline">Update</span> on GitHub</span></a> <p></p> | |
| <script> | |
| { | |
| __sveltekit_igaaa = { | |
| 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.5e38b416.js"), | |
| import("/docs/transformers/main/ko/_app/immutable/entry/app.240e8e3b.js") | |
| ]).then(([kit, app]) => { | |
| kit.start(app, element, { | |
| node_ids: [0, 125], | |
| data, | |
| form: null, | |
| error: null | |
| }); | |
| }); | |
| } | |
| </script> | |
Xet Storage Details
- Size:
- 59.1 kB
- Xet hash:
- 330ff58f896a5dd2d73f2d00c39d079706c08d350a8abc350433ef47f9150f06
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.