Buckets:

hf-doc-build
/

doc

Files

xet

hf-doc-build/doc / transformers /main /ko /perf_infer_gpu_one.html

HuggingFaceDocBuilder

about 1 hour ago

download

raw

41 kB

	<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{"title":"단일 GPU에서 효율적인 추론","local":"efficient-inference-on-a-single-gpu","sections":[{"title":"FP4 혼합 정밀도 추론을 위한 bitsandbytes 통합","local":"bitsandbytes-integration-for-fp4-mixedprecision-inference","sections":[{"title":"요구 사항","local":"requirements-for-fp4-mixedprecision-inference","sections":[],"depth":3},{"title":"FP4 모델 실행 - 단일 GPU 설정 - 빠른 시작","local":"running-fp4-models-single-gpu-setup-quickstart","sections":[],"depth":3},{"title":"FP4 모델 실행 - 다중 GPU 설정","local":"running-fp4-models-multi-gpu-setup","sections":[],"depth":3},{"title":"고급 사용법","local":"advanced-usage","sections":[],"depth":3}],"depth":2},{"title":"Int8 혼합 정밀도 행렬 분해를 위한 bitsandbytes 통합","local":"bitsandbytes-integration-for-int8-mixedprecision-matrix-decomposition","sections":[{"title":"요구 사항","local":"requirements-for-int8-mixedprecision-matrix-decomposition","sections":[],"depth":3},{"title":"혼합 Int8 모델 실행 - 단일 GPU 설정","local":"running-mixedint8-models-single-gpu-setup","sections":[],"depth":3},{"title":"혼합 Int8 모델 실행 - 다중 GPU 설정","local":"running-mixedint8-models-multi-gpu-setup","sections":[],"depth":3},{"title":"Colab 데모","local":"colab-demos","sections":[],"depth":3}],"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.acfca1c5.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.0cc5f498.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.dab99b64.js">
	<link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/entry/app.f0c7a87a.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/0.3e3ee3d5.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/133.648aaab2.js">
	<link rel="modulepreload" href="/docs/transformers/main/ko/_app/immutable/chunks/Tip.d62b0755.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/CodeBlock.ada04ea6.js"><!-- HEAD_svelte-u9bgzb_START --><meta name="hf:doc:metadata" content="{"title":"단일 GPU에서 효율적인 추론","local":"efficient-inference-on-a-single-gpu","sections":[{"title":"FP4 혼합 정밀도 추론을 위한 bitsandbytes 통합","local":"bitsandbytes-integration-for-fp4-mixedprecision-inference","sections":[{"title":"요구 사항","local":"requirements-for-fp4-mixedprecision-inference","sections":[],"depth":3},{"title":"FP4 모델 실행 - 단일 GPU 설정 - 빠른 시작","local":"running-fp4-models-single-gpu-setup-quickstart","sections":[],"depth":3},{"title":"FP4 모델 실행 - 다중 GPU 설정","local":"running-fp4-models-multi-gpu-setup","sections":[],"depth":3},{"title":"고급 사용법","local":"advanced-usage","sections":[],"depth":3}],"depth":2},{"title":"Int8 혼합 정밀도 행렬 분해를 위한 bitsandbytes 통합","local":"bitsandbytes-integration-for-int8-mixedprecision-matrix-decomposition","sections":[{"title":"요구 사항","local":"requirements-for-int8-mixedprecision-matrix-decomposition","sections":[],"depth":3},{"title":"혼합 Int8 모델 실행 - 단일 GPU 설정","local":"running-mixedint8-models-single-gpu-setup","sections":[],"depth":3},{"title":"혼합 Int8 모델 실행 - 다중 GPU 설정","local":"running-mixedint8-models-multi-gpu-setup","sections":[],"depth":3},{"title":"Colab 데모","local":"colab-demos","sections":[],"depth":3}],"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="efficient-inference-on-a-single-gpu" 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="#efficient-inference-on-a-single-gpu"><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>단일 GPU에서 효율적인 추론</span></h1> <p data-svelte-h="svelte-1m7ldpj">이 가이드 외에도, <a href="perf_train_gpu_one">단일 GPU에서의 훈련 가이드</a>와 <a href="perf_infer_cpu">CPU에서의 추론 가이드</a>에서도 관련 정보를 찾을 수 있습니다.</p> <h2 class="relative group"><a id="bitsandbytes-integration-for-fp4-mixedprecision-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="#bitsandbytes-integration-for-fp4-mixedprecision-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>FP4 혼합 정밀도 추론을 위한 bitsandbytes 통합</span></h2> <p data-svelte-h="svelte-sxo64f"><code>bitsandbytes</code>를 설치하면 GPU에서 손쉽게 모델을 압축할 수 있습니다. FP4 양자화를 사용하면 원래의 전체 정밀도 버전과 비교하여 모델 크기를 최대 8배 줄일 수 있습니다. 아래에서 시작하는 방법을 확인하세요.</p> <blockquote class="tip"><p data-svelte-h="svelte-1wakscf">이 기능은 다중 GPU 설정에서도 사용할 수 있습니다.</p></blockquote> <h3 class="relative group"><a id="requirements-for-fp4-mixedprecision-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="#requirements-for-fp4-mixedprecision-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>요구 사항</span></h3> <ul data-svelte-h="svelte-dwg1v5"><li><p>최신 <code>bitsandbytes</code> 라이브러리
	<code>pip install bitsandbytes>=0.39.0</code></p></li> <li><p>최신 <code>accelerate</code>를 소스에서 설치
	<code>pip install git+https://github.com/huggingface/accelerate.git</code></p></li> <li><p>최신 <code>transformers</code>를 소스에서 설치
	<code>pip install git+https://github.com/huggingface/transformers.git</code></p></li></ul> <h3 class="relative group"><a id="running-fp4-models-single-gpu-setup-quickstart" 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="#running-fp4-models-single-gpu-setup-quickstart"><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>FP4 모델 실행 - 단일 GPU 설정 - 빠른 시작</span></h3> <p data-svelte-h="svelte-60dj5">다음 코드를 실행하여 단일 GPU에서 빠르게 FP4 모델을 실행할 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, BitsAndBytesConfig

	model_name = <span class="hljs-string">"bigscience/bloom-2b5"</span>
	model_4bit = AutoModelForCausalLM.from_pretrained(model_name, device_map=<span class="hljs-string">"auto"</span>, quantization_config=BitsAndBytesConfig(load_in_4bit=<span class="hljs-literal">True</span>))<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1xe6z7w"><code>device_map</code>은 선택 사항입니다. 그러나 <code>device_map = 'auto'</code>로 설정하는 것이 사용 가능한 리소스를 효율적으로 디스패치하기 때문에 추론에 있어 권장됩니다.</p> <h3 class="relative group"><a id="running-fp4-models-multi-gpu-setup" 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="#running-fp4-models-multi-gpu-setup"><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>FP4 모델 실행 - 다중 GPU 설정</span></h3> <p data-svelte-h="svelte-11nsoog">다중 GPU에서 혼합 4비트 모델을 가져오는 방법은 단일 GPU 설정과 동일합니다(동일한 명령어 사용):</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 -->model_name = <span class="hljs-string">"bigscience/bloom-2b5"</span>
	model_4bit = AutoModelForCausalLM.from_pretrained(model_name, device_map=<span class="hljs-string">"auto"</span>, quantization_config=BitsAndBytesConfig(load_in_4bit=<span class="hljs-literal">True</span>))<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-7qj9fn">하지만 <code>accelerate</code>를 사용하여 각 GPU에 할당할 GPU RAM을 제어할 수 있습니다. 다음과 같이 <code>max_memory</code> 인수를 사용하세요:</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 -->max_memory_mapping = {<span class="hljs-number">0</span>: <span class="hljs-string">"600MB"</span>, <span class="hljs-number">1</span>: <span class="hljs-string">"1GB"</span>}
	model_name = <span class="hljs-string">"bigscience/bloom-3b"</span>
	model_4bit = AutoModelForCausalLM.from_pretrained(
	model_name, device_map=<span class="hljs-string">"auto"</span>, quantization_config=BitsAndBytesConfig(load_in_4bit=<span class="hljs-literal">True</span>), max_memory=max_memory_mapping
	)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1wq62m5">이 예에서는 첫 번째 GPU가 600MB의 메모리를 사용하고 두 번째 GPU가 1GB를 사용합니다.</p> <h3 class="relative group"><a id="advanced-usage" 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="#advanced-usage"><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>고급 사용법</span></h3> <p data-svelte-h="svelte-19t36l7">이 방법의 더 고급 사용법에 대해서는 <a href="main_classes/quantization">양자화</a> 문서 페이지를 참조하세요.</p> <h2 class="relative group"><a id="bitsandbytes-integration-for-int8-mixedprecision-matrix-decomposition" 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="#bitsandbytes-integration-for-int8-mixedprecision-matrix-decomposition"><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>Int8 혼합 정밀도 행렬 분해를 위한 bitsandbytes 통합</span></h2> <blockquote class="tip"><p data-svelte-h="svelte-1wakscf">이 기능은 다중 GPU 설정에서도 사용할 수 있습니다.</p></blockquote> <p data-svelte-h="svelte-1ivwma"><a href="https://huggingface.co/papers/2208.07339" rel="nofollow"><code>LLM.int8() : 8-bit Matrix Multiplication for Transformers at Scale</code></a> 논문에서 우리는 몇 줄의 코드로 Hub의 모든 모델에 대한 Hugging Face 통합을 지원합니다.
	이 방법은 <code>float16</code> 및 <code>bfloat16</code> 가중치에 대해 <code>nn.Linear</code> 크기를 2배로 줄이고, <code>float32</code> 가중치에 대해 4배로 줄입니다. 이는 절반 정밀도에서 이상치를 처리함으로써 품질에 거의 영향을 미치지 않습니다.</p> <p data-svelte-h="svelte-1tsrdqi"><img src="https://cdn-uploads.huggingface.co/production/uploads/1659861207959-62441d1d9fdefb55a0b7d12c.png" alt="HFxbitsandbytes.png"></p> <p data-svelte-h="svelte-1z8jha">Int8 혼합 정밀도 행렬 분해는 행렬 곱셈을 두 개의 스트림으로 분리합니다: (1) fp16로 곱해지는 체계적인 특이값 이상치 스트림 행렬(0.01%) 및 (2) int8 행렬 곱셈의 일반적인 스트림(99.9%). 이 방법을 사용하면 매우 큰 모델에 대해 예측 저하 없이 int8 추론이 가능합니다.
	이 방법에 대한 자세한 내용은 <a href="https://huggingface.co/papers/2208.07339" rel="nofollow">논문</a>이나 <a href="https://huggingface.co/blog/hf-bitsandbytes-integration" rel="nofollow">통합에 관한 블로그 글</a>에서 확인할 수 있습니다.</p> <p data-svelte-h="svelte-y2mgdg"><img src="https://cdn-uploads.huggingface.co/production/uploads/1660567469965-62441d1d9fdefb55a0b7d12c.gif" alt="MixedInt8.gif"></p> <p data-svelte-h="svelte-ewmhxc">커널은 GPU 전용으로 컴파일되어 있기 때문에 혼합 8비트 모델을 실행하려면 GPU가 필요합니다. 이 기능을 사용하기 전에 모델의 1/4(또는 모델 가중치가 절반 정밀도인 경우 절반)을 저장할 충분한 GPU 메모리가 있는지 확인하세요.
	이 모듈을 사용하는 데 도움이 되는 몇 가지 참고 사항이 아래에 나와 있습니다. 또는 <a href="#colab-demos">Google colab</a>에서 데모를 따라할 수도 있습니다.</p> <h3 class="relative group"><a id="requirements-for-int8-mixedprecision-matrix-decomposition" 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="#requirements-for-int8-mixedprecision-matrix-decomposition"><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>요구 사항</span></h3> <ul data-svelte-h="svelte-r9nrya"><li><code>bitsandbytes<0.37.0</code>을 사용하는 경우, 8비트 텐서 코어(Turing, Ampere 또는 이후 아키텍처 - 예: T4, RTX20s RTX30s, A40-A100)를 지원하는 NVIDIA GPU에서 실행하는지 확인하세요. <code>bitsandbytes>=0.37.0</code>을 사용하는 경우, 모든 GPU가 지원됩니다.</li> <li>올바른 버전의 <code>bitsandbytes</code>를 다음 명령으로 설치하세요:
	<code>pip install bitsandbytes>=0.31.5</code></li> <li><code>accelerate</code>를 설치하세요
	<code>pip install accelerate>=0.12.0</code></li></ul> <h3 class="relative group"><a id="running-mixedint8-models-single-gpu-setup" 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="#running-mixedint8-models-single-gpu-setup"><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>혼합 Int8 모델 실행 - 단일 GPU 설정</span></h3> <p data-svelte-h="svelte-1trjkap">필요한 라이브러리를 설치한 후 혼합 8비트 모델을 가져오는 방법은 다음과 같습니다:</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, BitsAndBytesConfig

	model_name = <span class="hljs-string">"bigscience/bloom-2b5"</span>
	model_8bit = AutoModelForCausalLM.from_pretrained(model_name, quantization_config=BitsAndBytesConfig(load_in_8bit=<span class="hljs-literal">True</span>))<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-bekc1k">텍스트 생성의 경우:</p> <ul data-svelte-h="svelte-daelpd"><li><code>pipeline()</code> 함수 대신 모델의 <code>generate()</code> 메소드를 사용하는 것을 권장합니다. <code>pipeline()</code> 함수로는 추론이 가능하지만, 혼합 8비트 모델에 최적화되지 않았기 때문에 <code>generate()</code> 메소드를 사용하는 것보다 느릴 수 있습니다. 또한, nucleus 샘플링과 같은 일부 샘플링 전략은 혼합 8비트 모델에 대해 <code>pipeline()</code> 함수에서 지원되지 않습니다.</li> <li>입력을 모델과 동일한 GPU에 배치하는 것이 좋습니다.</li></ul> <p data-svelte-h="svelte-1md3fzz">다음은 간단한 예입니다:</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, AutoTokenizer, BitsAndBytesConfig

	model_name = <span class="hljs-string">"bigscience/bloom-2b5"</span>
	tokenizer = AutoTokenizer.from_pretrained(model_name)
	model_8bit = AutoModelForCausalLM.from_pretrained(model_name, quantization_config=BitsAndBytesConfig(load_in_8bit=<span class="hljs-literal">True</span>))

	prompt = <span class="hljs-string">"Hello, my llama is cute"</span>
	inputs = tokenizer(prompt, return_tensors=<span class="hljs-string">"pt"</span>).to(<span class="hljs-string">"cuda"</span>)
	generated_ids = model.generate(**inputs)
	outputs = tokenizer.batch_decode(generated_ids, skip_special_tokens=<span class="hljs-literal">True</span>)<!-- HTML_TAG_END --></pre></div> <h3 class="relative group"><a id="running-mixedint8-models-multi-gpu-setup" 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="#running-mixedint8-models-multi-gpu-setup"><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>혼합 Int8 모델 실행 - 다중 GPU 설정</span></h3> <p data-svelte-h="svelte-bnkuc8">다중 GPU에서 혼합 8비트 모델을 로드하는 방법은 단일 GPU 설정과 동일합니다(동일한 명령어 사용):</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 -->model_name = <span class="hljs-string">"bigscience/bloom-2b5"</span>
	model_8bit = AutoModelForCausalLM.from_pretrained(model_name, quantization_config=BitsAndBytesConfig(load_in_8bit=<span class="hljs-literal">True</span>))<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-7qj9fn">하지만 <code>accelerate</code>를 사용하여 각 GPU에 할당할 GPU RAM을 제어할 수 있습니다. 다음과 같이 <code>max_memory</code> 인수를 사용하세요:</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 -->max_memory_mapping = {<span class="hljs-number">0</span>: <span class="hljs-string">"1GB"</span>, <span class="hljs-number">1</span>: <span class="hljs-string">"2GB"</span>}
	model_name = <span class="hljs-string">"bigscience/bloom-3b"</span>
	model_8bit = AutoModelForCausalLM.from_pretrained(
	model_name, device_map=<span class="hljs-string">"auto"</span>, quantization_config=BitsAndBytesConfig(load_in_4bit=<span class="hljs-literal">True</span>), max_memory=max_memory_mapping
	)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1bub4cv">이 예시에서는 첫 번째 GPU가 1GB의 메모리를 사용하고 두 번째 GPU가 2GB를 사용합니다.</p> <h3 class="relative group"><a id="colab-demos" 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="#colab-demos"><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>Colab 데모</span></h3> <p data-svelte-h="svelte-1at89fq">이 방법을 사용하면 이전에 Google Colab에서 추론할 수 없었던 모델에 대해 추론할 수 있습니다.
	Google Colab에서 8비트 양자화를 사용하여 T5-11b(42GB in fp32)를 실행하는 데모를 확인하세요:</p> <p data-svelte-h="svelte-1yb5ek4"><a href="https://colab.research.google.com/drive/1YORPWx4okIHXnjW7MSAidXN29mPVNT7F?usp=sharing" rel="nofollow"><img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab: T5-11b demo"></a></p> <p data-svelte-h="svelte-u2br1q">또는 BLOOM-3B에 대한 데모를 확인하세요:</p> <p data-svelte-h="svelte-6z7881"><a href="https://colab.research.google.com/drive/1qOjXfQIAULfKvZqwCen8-MoWKGdSatZ4?usp=sharing" rel="nofollow"><img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab: BLOOM-3b demo"></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/perf_infer_gpu_one.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_16kp15g = {
	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.acfca1c5.js"),
	import("/docs/transformers/main/ko/_app/immutable/entry/app.f0c7a87a.js")
	]).then(([kit, app]) => {
	kit.start(app, element, {
	node_ids: [0, 133],
	data,
	form: null,
	error: null
	});
	});
	}
	</script>

Xet Storage Details

Size:: 41 kB
Xet hash:: aa4997787454333fed9a6679b5339a26f5bd65cc19d3a3eac8502a237881c3b7

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