Buckets:

hf-doc-build
/

doc

Files

xet

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

HuggingFaceDocBuilder

about 6 hours ago

download

raw

57.7 kB

	<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{"title":"디버깅","local":"debugging","sections":[{"title":"Multi-GPU 네트워크 문제 디버그","local":"multigpu-network-issues-debug","sections":[],"depth":2},{"title":"언더플로 및 오버플로 감지","local":"underflow-and-overflow-detection","sections":[{"title":"특정 배치의 절댓값 최소 및 최대 값 추적","local":"specific-batch-absolute-min-and-max-value-tracing","sections":[],"depth":3}],"depth":2}],"depth":1}"/>
	<link href="/docs/transformers/main/ko/_app/immutable/entry/start.nq1UffQ_.js" rel="modulepreload">
	<link href="/docs/transformers/main/ko/_app/immutable/chunks/e6UN-XtK.js" rel="modulepreload">
	<link href="/docs/transformers/main/ko/_app/immutable/chunks/BkyfusOu.js" rel="modulepreload">
	<link href="/docs/transformers/main/ko/_app/immutable/entry/app.vNitWgjs.js" rel="modulepreload">
	<link href="/docs/transformers/main/ko/_app/immutable/chunks/CDRgxHoV.js" rel="modulepreload">
	<link href="/docs/transformers/main/ko/_app/immutable/chunks/1EV5EJPW.js" rel="modulepreload">
	<link href="/docs/transformers/main/ko/_app/immutable/chunks/Cm2Dm0ZX.js" rel="modulepreload">
	<link href="/docs/transformers/main/ko/_app/immutable/chunks/DsnmJJEf.js" rel="modulepreload">
	<link href="/docs/transformers/main/ko/_app/immutable/nodes/0.CA_KJU7-.js" rel="modulepreload">
	<link href="/docs/transformers/main/ko/_app/immutable/chunks/BoVSaMb7.js" rel="modulepreload">
	<link href="/docs/transformers/main/ko/_app/immutable/nodes/13.DI15YLQD.js" rel="modulepreload">
	<link href="/docs/transformers/main/ko/_app/immutable/chunks/C529Avzz.js" rel="modulepreload">
	<link href="/docs/transformers/main/ko/_app/immutable/chunks/_7Q5EL3T.js" rel="modulepreload">
	<!--xyche3--><meta name="hf:doc:metadata" content="{"title":"디버깅","local":"debugging","sections":[{"title":"Multi-GPU 네트워크 문제 디버그","local":"multigpu-network-issues-debug","sections":[],"depth":2},{"title":"언더플로 및 오버플로 감지","local":"underflow-and-overflow-detection","sections":[{"title":"특정 배치의 절댓값 최소 및 최대 값 추적","local":"specific-batch-absolute-min-and-max-value-tracing","sections":[],"depth":3}],"depth":2}],"depth":1}"/><!---->
	<link href="/docs/transformers/main/ko/_app/immutable/assets/0.tn0RQdqM.css" rel="modulepreload"> <!--[--><!--[0--><!--[--><!--[0--><!--[--><p></p> <div class="items-center shrink-0 min-w-[100px] max-sm:min-w-[50px] justify-end ml-auto flex" style="float: 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> <!--[-1--><!--]--></div><!----> <!--[0--><h1 class="relative group"><a id="debugging" 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="#debugging"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span>디버깅</span></h1><!--]--><!----> <!--[1--><h2 class="relative group"><a id="multigpu-network-issues-debug" 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="#multigpu-network-issues-debug"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span>Multi-GPU 네트워크 문제 디버그</span></h2><!--]--><!----> <p><code>DistributedDataParallel</code> 및 다중 GPU를 사용하여 훈련하거나 추론할 때, 프로세스 및/또는 노드 간의 상호 통신 문제가 발생하는 경우, 다음 스크립트를 사용하여 네트워크 문제를 진단할 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 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-bash "><!---->wget https://raw.githubusercontent.com/huggingface/transformers/main/scripts/distributed/torch-distributed-gpu-test.py<!----></pre></div><!----> <p>예를 들어, 2개의 GPU가 상호 작용하는 방식을 테스트하려면 다음을 실행하세요:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Copied</div><!----></button><!----></div> <pre class="language-bash "><!---->python -m torch.distributed.run --nproc_per_node 2 --nnodes 1 torch-distributed-gpu-test.py<!----></pre></div><!----> <p>두 프로세스가 서로 통신하고 GPU 메모리를 할당하는 경우, 각각 “OK” 상태를 출력합니다.</p> <p>더 많은 GPU 또는 노드의 경우 스크립트의 인수를 조정하면 됩니다.</p> <p>진단 스크립트 내에서 더 많은 세부 정보와 SLURM 환경에서 실행하는 방법에 대한 레시피를 찾을 수 있습니다.</p> <p>추가적인 디버그 수준은 다음과 같이 <code>NCCL_DEBUG=INFO</code> 환경 변수를 추가하는 것입니다:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Copied</div><!----></button><!----></div> <pre class="language-bash "><!---->NCCL_DEBUG=INFO python -m torch.distributed.run --nproc_per_node 2 --nnodes 1 torch-distributed-gpu-test.py<!----></pre></div><!----> <p>이렇게 하면 NCCL 관련 디버그 정보가 많이 출력되며, 문제가 보고된 경우에는 인터넷에서 검색할 수 있습니다. 또는 출력을 해석하는 방법을 잘 모르는 경우 로그 파일을 이슈에 공유할 수 있습니다.</p> <!--[1--><h2 class="relative group"><a id="underflow-and-overflow-detection" 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="#underflow-and-overflow-detection"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span>언더플로 및 오버플로 감지</span></h2><!--]--><!----> <blockquote class="tip"><p>이 기능은 현재 PyTorch에서만 사용할 수 있습니다.</p><!----></blockquote><!----> <blockquote class="tip"><p>다중 GPU 훈련을 위해서는 DDP (<code>torch.distributed.launch</code>)가 필요합니다.</p><!----></blockquote><!----> <blockquote class="tip"><p>이 기능은 <code>nn.Module</code>을 기반으로 하는 모델과 함께 사용할 수 있습니다.</p><!----></blockquote><!----> <p><code>loss=NaN</code>이 나타나거나 모델이 <code>inf</code> 또는 <code>nan</code>으로 인해 다른 이상한 동작을 하는 경우, 언더플로 또는 오버플로의 첫 번째 발생 위치와 그 원인을 파악해야 합니다. 다행히도 이를 자동으로 감지하는 특수 모듈을 활성화하여 쉽게 알아낼 수 있습니다.</p> <p><a href="/docs/transformers/main/ko/main_classes/trainer#transformers.Trainer">Trainer</a>를 사용하는 경우, 다음을 기존의 명령줄 인수에 추가하면 됩니다.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Copied</div><!----></button><!----></div> <pre class="language-bash "><!---->--debug underflow_overflow<!----></pre></div><!----> <p>또는 <a href="/docs/transformers/main/ko/main_classes/trainer#transformers.TrainingArguments">TrainingArguments</a> 객체를 생성할 때 <code>debug="underflow_overflow"</code>를 전달합니다.</p> <p>자체 훈련 루프나 다른 Trainer를 사용하는 경우, 다음과 같이 수행할 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 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-python "><!----><span class="hljs-keyword">from</span> transformers.debug_utils <span class="hljs-keyword">import</span> DebugUnderflowOverflow

	debug_overflow = DebugUnderflowOverflow(model)<!----></pre></div><!----> <p><a href="/docs/transformers/main/ko/internal/trainer_utils#transformers.debug_utils.DebugUnderflowOverflow">DebugUnderflowOverflow</a>는 모델에 후크를 삽입하여 각 forward 호출 직후에 입력 및 출력 변수 및 해당 모듈의 가중치를 테스트합니다. 활성화나 가중치의 최소한 하나의 요소에서 <code>inf</code> 또는 <code>nan</code>이 감지되면 프로그램이 어설트되고 다음과 같은 보고서가 출력됩니다. (이 예제는 fp16 혼합 정밀도에서 <code>google/mt5-small</code>에서 캡처된 것입니다):</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Copied</div><!----></button><!----></div> <pre class=" "><!----><span class="hljs-attribute">Detected</span> inf/nan during batch_number=<span class="hljs-number">0</span>
	<span class="hljs-attribute">Last</span> <span class="hljs-number">21</span> forward frames:
	<span class="hljs-attribute">abs</span> min abs max metadata
	<span class="hljs-attribute">encoder</span>.block.<span class="hljs-number">1</span>.layer.<span class="hljs-number">1</span>.DenseReluDense.dropout Dropout
	<span class="hljs-attribute">0</span>.<span class="hljs-number">00</span>e+<span class="hljs-number">00</span> <span class="hljs-number">2</span>.<span class="hljs-number">57</span>e+<span class="hljs-number">02</span> input[<span class="hljs-number">0</span>]
	<span class="hljs-attribute">0</span>.<span class="hljs-number">00</span>e+<span class="hljs-number">00</span> <span class="hljs-number">2</span>.<span class="hljs-number">85</span>e+<span class="hljs-number">02</span> output<span class="hljs-meta">
	[...]</span>
	<span class="hljs-attribute">encoder</span>.block.<span class="hljs-number">2</span>.layer.<span class="hljs-number">0</span> T5LayerSelfAttention
	<span class="hljs-attribute">6</span>.<span class="hljs-number">78</span>e-<span class="hljs-number">04</span> <span class="hljs-number">3</span>.<span class="hljs-number">15</span>e+<span class="hljs-number">03</span> input[<span class="hljs-number">0</span>]
	<span class="hljs-attribute">2</span>.<span class="hljs-number">65</span>e-<span class="hljs-number">04</span> <span class="hljs-number">3</span>.<span class="hljs-number">42</span>e+<span class="hljs-number">03</span> output[<span class="hljs-number">0</span>]
	<span class="hljs-attribute">None</span> output[<span class="hljs-number">1</span>]
	<span class="hljs-attribute">2</span>.<span class="hljs-number">25</span>e-<span class="hljs-number">01</span> <span class="hljs-number">1</span>.<span class="hljs-number">00</span>e+<span class="hljs-number">04</span> output[<span class="hljs-number">2</span>]
	<span class="hljs-attribute">encoder</span>.block.<span class="hljs-number">2</span>.layer.<span class="hljs-number">1</span>.layer_norm T5LayerNorm
	<span class="hljs-attribute">8</span>.<span class="hljs-number">69</span>e-<span class="hljs-number">02</span> <span class="hljs-number">4</span>.<span class="hljs-number">18</span>e-<span class="hljs-number">01</span> weight
	<span class="hljs-attribute">2</span>.<span class="hljs-number">65</span>e-<span class="hljs-number">04</span> <span class="hljs-number">3</span>.<span class="hljs-number">42</span>e+<span class="hljs-number">03</span> input[<span class="hljs-number">0</span>]
	<span class="hljs-attribute">1</span>.<span class="hljs-number">79</span>e-<span class="hljs-number">06</span> <span class="hljs-number">4</span>.<span class="hljs-number">65</span>e+<span class="hljs-number">00</span> output
	<span class="hljs-attribute">encoder</span>.block.<span class="hljs-number">2</span>.layer.<span class="hljs-number">1</span>.DenseReluDense.wi_0 Linear
	<span class="hljs-attribute">2</span>.<span class="hljs-number">17</span>e-<span class="hljs-number">07</span> <span class="hljs-number">4</span>.<span class="hljs-number">50</span>e+<span class="hljs-number">00</span> weight
	<span class="hljs-attribute">1</span>.<span class="hljs-number">79</span>e-<span class="hljs-number">06</span> <span class="hljs-number">4</span>.<span class="hljs-number">65</span>e+<span class="hljs-number">00</span> input[<span class="hljs-number">0</span>]
	<span class="hljs-attribute">2</span>.<span class="hljs-number">68</span>e-<span class="hljs-number">06</span> <span class="hljs-number">3</span>.<span class="hljs-number">70</span>e+<span class="hljs-number">01</span> output
	<span class="hljs-attribute">encoder</span>.block.<span class="hljs-number">2</span>.layer.<span class="hljs-number">1</span>.DenseReluDense.wi_1 Linear
	<span class="hljs-attribute">8</span>.<span class="hljs-number">08</span>e-<span class="hljs-number">07</span> <span class="hljs-number">2</span>.<span class="hljs-number">66</span>e+<span class="hljs-number">01</span> weight
	<span class="hljs-attribute">1</span>.<span class="hljs-number">79</span>e-<span class="hljs-number">06</span> <span class="hljs-number">4</span>.<span class="hljs-number">65</span>e+<span class="hljs-number">00</span> input[<span class="hljs-number">0</span>]
	<span class="hljs-attribute">1</span>.<span class="hljs-number">27</span>e-<span class="hljs-number">04</span> <span class="hljs-number">2</span>.<span class="hljs-number">37</span>e+<span class="hljs-number">02</span> output
	<span class="hljs-attribute">encoder</span>.block.<span class="hljs-number">2</span>.layer.<span class="hljs-number">1</span>.DenseReluDense.dropout Dropout
	<span class="hljs-attribute">0</span>.<span class="hljs-number">00</span>e+<span class="hljs-number">00</span> <span class="hljs-number">8</span>.<span class="hljs-number">76</span>e+<span class="hljs-number">03</span> input[<span class="hljs-number">0</span>]
	<span class="hljs-attribute">0</span>.<span class="hljs-number">00</span>e+<span class="hljs-number">00</span> <span class="hljs-number">9</span>.<span class="hljs-number">74</span>e+<span class="hljs-number">03</span> output
	<span class="hljs-attribute">encoder</span>.block.<span class="hljs-number">2</span>.layer.<span class="hljs-number">1</span>.DenseReluDense.wo Linear
	<span class="hljs-attribute">1</span>.<span class="hljs-number">01</span>e-<span class="hljs-number">06</span> <span class="hljs-number">6</span>.<span class="hljs-number">44</span>e+<span class="hljs-number">00</span> weight
	<span class="hljs-attribute">0</span>.<span class="hljs-number">00</span>e+<span class="hljs-number">00</span> <span class="hljs-number">9</span>.<span class="hljs-number">74</span>e+<span class="hljs-number">03</span> input[<span class="hljs-number">0</span>]
	<span class="hljs-attribute">3</span>.<span class="hljs-number">18</span>e-<span class="hljs-number">04</span> <span class="hljs-number">6</span>.<span class="hljs-number">27</span>e+<span class="hljs-number">04</span> output
	<span class="hljs-attribute">encoder</span>.block.<span class="hljs-number">2</span>.layer.<span class="hljs-number">1</span>.DenseReluDense T5DenseGatedGeluDense
	<span class="hljs-attribute">1</span>.<span class="hljs-number">79</span>e-<span class="hljs-number">06</span> <span class="hljs-number">4</span>.<span class="hljs-number">65</span>e+<span class="hljs-number">00</span> input[<span class="hljs-number">0</span>]
	<span class="hljs-attribute">3</span>.<span class="hljs-number">18</span>e-<span class="hljs-number">04</span> <span class="hljs-number">6</span>.<span class="hljs-number">27</span>e+<span class="hljs-number">04</span> output
	<span class="hljs-attribute">encoder</span>.block.<span class="hljs-number">2</span>.layer.<span class="hljs-number">1</span>.dropout Dropout
	<span class="hljs-attribute">3</span>.<span class="hljs-number">18</span>e-<span class="hljs-number">04</span> <span class="hljs-number">6</span>.<span class="hljs-number">27</span>e+<span class="hljs-number">04</span> input[<span class="hljs-number">0</span>]
	<span class="hljs-attribute">0</span>.<span class="hljs-number">00</span>e+<span class="hljs-number">00</span> inf output<!----></pre></div><!----> <p>예제 출력은 간략성을 위해 중간 부분이 잘려 있습니다.</p> <p>두 번째 열은 절대적으로 가장 큰 요소의 값이며, 따라서 마지막 몇 개의 프레임을 자세히 살펴보면 입력과 출력이 <code>1e4</code> 범위에 있음을 알 수 있습니다. 따라서 이 훈련은 <code>fp16</code> 혼합 정밀도로 수행될 때 가장 마지막 단계에서 오버플로우가 발생했습니다 (<code>fp16</code>에서 <code>inf</code> 이전의 가장 큰 숫자는 <code>64e3</code>입니다). <code>fp16</code> 아래에서 오버플로우를 피하기 위해서는 활성화는 <code>1e4</code>보다 훨씬 작아야 합니다. 왜냐하면 <code>1e4 * 1e4 = 1e8</code>이기 때문에 큰 활성화와의 행렬 곱은 수치적인 오버플로우 조건으로 이어질 것입니다.</p> <p>추적의 맨 처음에서 어느 배치 번호에서 문제가 발생했는지 알 수 있습니다 (여기서 <code>Detected inf/nan during batch_number=0</code>은 문제가 첫 번째 배치에서 발생했음을 의미합니다).</p> <p>각 보고된 프레임은 해당 프레임이 보고하는 해당 모듈에 대한 완전한 항목을 선언하며, 이 프레임만 살펴보면 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 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=" "><!----> <span class="hljs-attribute">encoder</span>.block.<span class="hljs-number">2</span>.layer.<span class="hljs-number">1</span>.layer_norm T5LayerNorm
	<span class="hljs-attribute">8</span>.<span class="hljs-number">69</span>e-<span class="hljs-number">02</span> <span class="hljs-number">4</span>.<span class="hljs-number">18</span>e-<span class="hljs-number">01</span> weight
	<span class="hljs-attribute">2</span>.<span class="hljs-number">65</span>e-<span class="hljs-number">04</span> <span class="hljs-number">3</span>.<span class="hljs-number">42</span>e+<span class="hljs-number">03</span> input[<span class="hljs-number">0</span>]
	<span class="hljs-attribute">1</span>.<span class="hljs-number">79</span>e-<span class="hljs-number">06</span> <span class="hljs-number">4</span>.<span class="hljs-number">65</span>e+<span class="hljs-number">00</span> output<!----></pre></div><!----> <p>여기서 <code>encoder.block.2.layer.1.layer_norm</code>은 인코더의 두 번째 블록의 첫 번째 레이어에 대한 레이어 정규화를 의미하며, <code>forward</code>의 특정 호출은 <code>T5LayerNorm</code>입니다.</p> <p>이 보고서의 마지막 몇 개 프레임을 살펴보겠습니다:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Copied</div><!----></button><!----></div> <pre class=" "><!----><span class="hljs-attribute">Detected</span> inf/nan during batch_number=<span class="hljs-number">0</span>
	<span class="hljs-attribute">Last</span> <span class="hljs-number">21</span> forward frames:
	<span class="hljs-attribute">abs</span> min abs max metadata<span class="hljs-meta">
	[...]</span>
	<span class="hljs-attribute">encoder</span>.block.<span class="hljs-number">2</span>.layer.<span class="hljs-number">1</span>.DenseReluDense.wi_0 Linear
	<span class="hljs-attribute">2</span>.<span class="hljs-number">17</span>e-<span class="hljs-number">07</span> <span class="hljs-number">4</span>.<span class="hljs-number">50</span>e+<span class="hljs-number">00</span> weight
	<span class="hljs-attribute">1</span>.<span class="hljs-number">79</span>e-<span class="hljs-number">06</span> <span class="hljs-number">4</span>.<span class="hljs-number">65</span>e+<span class="hljs-number">00</span> input[<span class="hljs-number">0</span>]
	<span class="hljs-attribute">2</span>.<span class="hljs-number">68</span>e-<span class="hljs-number">06</span> <span class="hljs-number">3</span>.<span class="hljs-number">70</span>e+<span class="hljs-number">01</span> output
	<span class="hljs-attribute">encoder</span>.block.<span class="hljs-number">2</span>.layer.<span class="hljs-number">1</span>.DenseReluDense.wi_1 Linear
	<span class="hljs-attribute">8</span>.<span class="hljs-number">08</span>e-<span class="hljs-number">07</span> <span class="hljs-number">2</span>.<span class="hljs-number">66</span>e+<span class="hljs-number">01</span> weight
	<span class="hljs-attribute">1</span>.<span class="hljs-number">79</span>e-<span class="hljs-number">06</span> <span class="hljs-number">4</span>.<span class="hljs-number">65</span>e+<span class="hljs-number">00</span> input[<span class="hljs-number">0</span>]
	<span class="hljs-attribute">1</span>.<span class="hljs-number">27</span>e-<span class="hljs-number">04</span> <span class="hljs-number">2</span>.<span class="hljs-number">37</span>e+<span class="hljs-number">02</span> output
	<span class="hljs-attribute">encoder</span>.block.<span class="hljs-number">2</span>.layer.<span class="hljs-number">1</span>.DenseReluDense.wo Linear
	<span class="hljs-attribute">1</span>.<span class="hljs-number">01</span>e-<span class="hljs-number">06</span> <span class="hljs-number">6</span>.<span class="hljs-number">44</span>e+<span class="hljs-number">00</span> weight
	<span class="hljs-attribute">0</span>.<span class="hljs-number">00</span>e+<span class="hljs-number">00</span> <span class="hljs-number">9</span>.<span class="hljs-number">74</span>e+<span class="hljs-number">03</span> input[<span class="hljs-number">0</span>]
	<span class="hljs-attribute">3</span>.<span class="hljs-number">18</span>e-<span class="hljs-number">04</span> <span class="hljs-number">6</span>.<span class="hljs-number">27</span>e+<span class="hljs-number">04</span> output
	<span class="hljs-attribute">encoder</span>.block.<span class="hljs-number">2</span>.layer.<span class="hljs-number">1</span>.DenseReluDense T5DenseGatedGeluDense
	<span class="hljs-attribute">1</span>.<span class="hljs-number">79</span>e-<span class="hljs-number">06</span> <span class="hljs-number">4</span>.<span class="hljs-number">65</span>e+<span class="hljs-number">00</span> input[<span class="hljs-number">0</span>]
	<span class="hljs-attribute">3</span>.<span class="hljs-number">18</span>e-<span class="hljs-number">04</span> <span class="hljs-number">6</span>.<span class="hljs-number">27</span>e+<span class="hljs-number">04</span> output
	<span class="hljs-attribute">encoder</span>.block.<span class="hljs-number">2</span>.layer.<span class="hljs-number">1</span>.dropout Dropout
	<span class="hljs-attribute">3</span>.<span class="hljs-number">18</span>e-<span class="hljs-number">04</span> <span class="hljs-number">6</span>.<span class="hljs-number">27</span>e+<span class="hljs-number">04</span> input[<span class="hljs-number">0</span>]
	<span class="hljs-attribute">0</span>.<span class="hljs-number">00</span>e+<span class="hljs-number">00</span> inf output<!----></pre></div><!----> <p>마지막 프레임은 <code>Dropout.forward</code> 함수에 대한 보고입니다. 첫 번째 항목은 유일한 입력을 나타내고 두 번째 항목은 유일한 출력을 나타냅니다. 이 함수가 <code>DenseReluDense</code> 클래스 내부의 <code>dropout</code> 속성에서 호출된 것을 볼 수 있습니다. 이는 첫 번째 레이어의 두 번째 블록에서 첫 번째 배치 중에 발생했다는 것을 알 수 있습니다. 마지막으로, 절대적으로 가장 큰 입력 요소는 <code>6.27e+04</code>이고 출력도 마찬가지로 <code>inf</code>입니다.</p> <p>여기에서는 <code>T5DenseGatedGeluDense.forward</code>가 출력 활성화를 생성하는데, 절대적으로 가장 큰 값이 약 62.7K인 것을 볼 수 있습니다. 이 값은 fp16의 최대 제한인 64K에 매우 근접합니다. 다음 프레임에서는 일부 요소를 0으로 만든 후 가중치를 재정규화하는 <code>Dropout</code>이 있습니다. 이로 인해 절대 최대값이 64K를 초과하고 오버플로우(<code>inf</code>)가 발생합니다.</p> <p>보시다시피, fp16 숫자의 경우 숫자가 매우 커질 때 이전 프레임을 살펴보아야 합니다.</p> <p>보고서를 <code>models/t5/modeling_t5.py</code>의 코드와 일치시켜 보겠습니다.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Copied</div><!----></button><!----></div> <pre class="language-python "><!----><span class="hljs-keyword">class</span> <span class="hljs-title class_">T5DenseGatedGeluDense</span>(nn.Module):
	<span class="hljs-keyword">def</span> <span class="hljs-title function_">__init__</span>(<span class="hljs-params">self, config</span>):
	<span class="hljs-built_in">super</span>().__init__()
	<span class="hljs-variable language_">self</span>.wi_0 = nn.Linear(config.d_model, config.d_ff, bias=<span class="hljs-literal">False</span>)
	<span class="hljs-variable language_">self</span>.wi_1 = nn.Linear(config.d_model, config.d_ff, bias=<span class="hljs-literal">False</span>)
	<span class="hljs-variable language_">self</span>.wo = nn.Linear(config.d_ff, config.d_model, bias=<span class="hljs-literal">False</span>)
	<span class="hljs-variable language_">self</span>.dropout = nn.Dropout(config.dropout_rate)
	<span class="hljs-variable language_">self</span>.gelu_act = ACT2FN[<span class="hljs-string">"gelu_new"</span>]

	<span class="hljs-keyword">def</span> <span class="hljs-title function_">forward</span>(<span class="hljs-params">self, hidden_states</span>):
	hidden_gelu = <span class="hljs-variable language_">self</span>.gelu_act(<span class="hljs-variable language_">self</span>.wi_0(hidden_states))
	hidden_linear = <span class="hljs-variable language_">self</span>.wi_1(hidden_states)
	hidden_states = hidden_gelu * hidden_linear
	hidden_states = <span class="hljs-variable language_">self</span>.dropout(hidden_states)
	hidden_states = <span class="hljs-variable language_">self</span>.wo(hidden_states)
	<span class="hljs-keyword">return</span> hidden_states<!----></pre></div><!----> <p>이제 <code>dropout</code> 호출과 이전의 모든 호출을 쉽게 확인할 수 있습니다.</p> <p>감지는 <code>forward</code> 후크에서 발생하므로, 이러한 보고서는 각 <code>forward</code>가 반환된 직후에 즉시 출력됩니다.</p> <p>전체 보고서로 돌아가서 문제에 대한 조치 및 수정을 하려면, 숫자가 증가하기 시작한 몇 개의 프레임 위로 이동해서 여기서 <code>fp32</code> 모드로 전환해야 합니다. 이렇게 해야 숫자가 곱해지거나 합쳐질 때 오버플로우되지 않을 가능성이 높습니다. 물론 다른 해결책도 있을 수 있습니다. 예를 들어, <code>amp</code>가 활성화된 경우 일시적으로 끄고 원래의 <code>forward</code>를 도우미 래퍼로 이동한 후 다음과 같이 할 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 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-python "><!----><span class="hljs-keyword">def</span> <span class="hljs-title function_">_forward</span>(<span class="hljs-params">self, hidden_states</span>):
	hidden_gelu = <span class="hljs-variable language_">self</span>.gelu_act(<span class="hljs-variable language_">self</span>.wi_0(hidden_states))
	hidden_linear = <span class="hljs-variable language_">self</span>.wi_1(hidden_states)
	hidden_states = hidden_gelu * hidden_linear
	hidden_states = <span class="hljs-variable language_">self</span>.dropout(hidden_states)
	hidden_states = <span class="hljs-variable language_">self</span>.wo(hidden_states)
	<span class="hljs-keyword">return</span> hidden_states


	<span class="hljs-keyword">import</span> torch


	<span class="hljs-keyword">def</span> <span class="hljs-title function_">forward</span>(<span class="hljs-params">self, hidden_states</span>):
	device_type = hidden_states.device.<span class="hljs-built_in">type</span>
	<span class="hljs-keyword">if</span> torch.is_autocast_enabled(device_type):
	<span class="hljs-keyword">with</span> torch.amp.autocast(device_type, enabled=<span class="hljs-literal">False</span>):
	<span class="hljs-keyword">return</span> <span class="hljs-variable language_">self</span>._forward(hidden_states)
	<span class="hljs-keyword">else</span>:
	<span class="hljs-keyword">return</span> <span class="hljs-variable language_">self</span>._forward(hidden_states)<!----></pre></div><!----> <p>자동 감지기는 전체 프레임의 입력과 출력에 대해서만 보고하므로, 어디를 살펴봐야 하는지 알면 특정 <code>forward</code> 함수의 중간 단계도 분석할 수 있습니다. 이 경우에는 <code>detect_overflow</code> 도우미 함수를 사용하여 원하는 위치에 감지기를 삽입할 수 있습니다. 예를 들어:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Copied</div><!----></button><!----></div> <pre class="language-python "><!----><span class="hljs-keyword">from</span> debug_utils <span class="hljs-keyword">import</span> detect_overflow


	<span class="hljs-keyword">class</span> <span class="hljs-title class_">T5LayerFF</span>(nn.Module):
	[...]

	<span class="hljs-keyword">def</span> <span class="hljs-title function_">forward</span>(<span class="hljs-params">self, hidden_states</span>):
	forwarded_states = <span class="hljs-variable language_">self</span>.layer_norm(hidden_states)
	detect_overflow(forwarded_states, <span class="hljs-string">"after layer_norm"</span>)
	forwarded_states = <span class="hljs-variable language_">self</span>.DenseReluDense(forwarded_states)
	detect_overflow(forwarded_states, <span class="hljs-string">"after DenseReluDense"</span>)
	<span class="hljs-keyword">return</span> hidden_states + <span class="hljs-variable language_">self</span>.dropout(forwarded_states)<!----></pre></div><!----> <p>여기서는 이를 추가하여 2개의 것을 추적하고 이제 <code>forwarded_states</code>의 <code>inf</code> 또는 <code>nan</code>이 중간에 감지되었는지를 추적합니다.</p> <p>실제로 위의 예제에서 각 호출이 <code>nn.Module</code>이기 때문에 탐지기가 이미 이를 보고합니다. 로컬에서 직접 계산하는 경우 이렇게 수행한다고 가정해 봅시다.</p> <p>또한, 자체 코드에서 디버거를 인스턴스화하는 경우 기본값에서 출력되는 프레임 수를 조정할 수 있습니다. 예를 들어:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Copied</div><!----></button><!----></div> <pre class="language-python "><!----><span class="hljs-keyword">from</span> transformers.debug_utils <span class="hljs-keyword">import</span> DebugUnderflowOverflow

	debug_overflow = DebugUnderflowOverflow(model, max_frames_to_save=<span class="hljs-number">100</span>)<!----></pre></div><!----> <!--[2--><h3 class="relative group"><a id="specific-batch-absolute-min-and-max-value-tracing" 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="#specific-batch-absolute-min-and-max-value-tracing"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span>특정 배치의 절댓값 최소 및 최대 값 추적</span></h3><!--]--><!----> <p>동일한 디버깅 클래스는 언더플로우/오버플로우 감지 기능이 꺼진 상태에서 배치별 추적에도 사용할 수 있습니다.</p> <p>예를 들어, 특정 배치의 각 <code>forward</code> 호출의 모든 구성 성분에 대한 절대 최솟값과 최댓값을 확인하고, 이를 배치 1과 3에 대해서만 수행하려면 다음과 같이 이 클래스를 인스턴스화합니다:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Copied</div><!----></button><!----></div> <pre class="language-python "><!---->debug_overflow = DebugUnderflowOverflow(model, trace_batch_nums=[<span class="hljs-number">1</span>, <span class="hljs-number">3</span>])<!----></pre></div><!----> <p>그러면 이제 배치 1과 3 전체가 언더플로우/오버플로우 감지기와 동일한 형식으로 추적됩니다.</p> <p>배치는 0부터 시작합니다.</p> <p>이는 프로그램이 특정 배치 번호 이후에 오작동하기 시작하는 것을 알고 있는 경우에 유용합니다. 그렇기 때문에 해당 영역으로 바로 이동할 수 있습니다. 이런 구성에 대한 샘플 축소된 출력은 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 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=" "><!----> * Starting batch number=1 *
	abs min abs max metadata
	shared Embedding
	1.01e<span class="hljs-string">-06</span> 7.92e<span class="hljs-string">+02</span> weight
	0.00e<span class="hljs-string">+00</span> 2.47e<span class="hljs-string">+04</span> input[0]
	5.36e<span class="hljs-string">-05</span> 7.92e<span class="hljs-string">+02</span> output
	[...]
	decoder.dropout Dropout
	1.60e<span class="hljs-string">-07</span> 2.27e<span class="hljs-string">+01</span> input[0]
	0.00e<span class="hljs-string">+00</span> 2.52e<span class="hljs-string">+01</span> output
	decoder T5Stack
	not a tensor output
	lm_head Linear
	1.01e<span class="hljs-string">-06</span> 7.92e<span class="hljs-string">+02</span> weight
	0.00e<span class="hljs-string">+00</span> 1.11e<span class="hljs-string">+00</span> input[0]
	6.06e<span class="hljs-string">-02</span> 8.39e<span class="hljs-string">+01</span> output
	T5ForConditionalGeneration
	not a tensor output

	* Starting batch number=3 *
	abs min abs max metadata
	shared Embedding
	1.01e<span class="hljs-string">-06</span> 7.92e<span class="hljs-string">+02</span> weight
	0.00e<span class="hljs-string">+00</span> 2.78e<span class="hljs-string">+04</span> input[0]
	5.36e<span class="hljs-string">-05</span> 7.92e<span class="hljs-string">+02</span> output
	[...]<!----></pre></div><!----> <p>여기에서는 모델의 forward 호출 수와 동일한 수의 프레임이 덤프되므로 많은 수의 프레임이 생성됩니다. 따라서 원하는 것일 수도 있고 아닐 수도 있습니다. 그러나 때로는 일반 디버거보다 디버깅 목적으로 더 쉽게 사용할 수 있습니다. 예를 들어, 문제가 배치 번호 150에서 시작하는 경우 149와 150의 추적을 덤프하고 숫자가 어디서부터 다르게 되었는지 비교할 수 있습니다.</p> <p>또한, 훈련을 중지할 배치 번호를 지정할 수도 있습니다. 다음과 같이 지정할 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 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-python "><!---->debug_overflow = DebugUnderflowOverflow(model, trace_batch_nums=[<span class="hljs-number">1</span>, <span class="hljs-number">3</span>], abort_after_batch_num=<span class="hljs-number">3</span>)<!----></pre></div><!----> <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/debugging.md" target="_blank"><svg class="mr-1" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M31,16l-7,7l-1.41-1.41L28.17,16l-5.58-5.59L24,9l7,7z"></path><path d="M1,16l7-7l1.41,1.41L3.83,16l5.58,5.59L8,23l-7-7z"></path><path d="M12.419,25.484L17.639,6.552l1.932,0.518L14.351,26.002z"></path></svg><!----> <span><span class="underline">Update</span> on GitHub</span></a><!----> <p></p><!--]--><!----><!--]--><!--]--><!--]--> <!--[-1--><!--]--><!--]-->

	<script>
	{
	__sveltekit_1du874l = {
	base: "/docs/transformers/main/ko",
	assets: "/docs/transformers/main/ko"
	};

	const element = document.currentScript.parentElement;

	Promise.all([
	import("/docs/transformers/main/ko/_app/immutable/entry/start.nq1UffQ_.js"),
	import("/docs/transformers/main/ko/_app/immutable/entry/app.vNitWgjs.js")
	]).then(([kit, app]) => {
	kit.start(app, element, {
	node_ids: [0, 13],
	data: [null,null],
	form: null,
	error: null
	});
	});
	}
	</script>

Xet Storage Details

Size:: 57.7 kB
Xet hash:: e68ab713b02588b2034e6cdde0a16af9ddb46cf4183d45600cf992962357b200

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