Hugging Face's logo

Buckets:
hf-doc-build
/
doc-dev

Files
xet
hf-doc-build/doc-dev / transformers /main /es /debugging.html
rtrm's picture
rtrm
download
raw
55.2 kB
<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;Debugging&quot;,&quot;local&quot;:&quot;debugging&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Debug de problemas de Network multi-GPU&quot;,&quot;local&quot;:&quot;debug-de-problemas-de-network-multi-gpu&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Detección de Underflow y Overflow&quot;,&quot;local&quot;:&quot;detección-de-underflow-y-overflow&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Rastreo de valores mínimos y máximos absolutos de batches específicos&quot;,&quot;local&quot;:&quot;rastreo-de-valores-mínimos-y-máximos-absolutos-de-batches-específicos&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3}],&quot;depth&quot;:2}],&quot;depth&quot;:1}">
 <link href="/docs/transformers/main/es/_app/immutable/assets/0.e3b0c442.css" rel="modulepreload">
 <link rel="modulepreload" href="/docs/transformers/main/es/_app/immutable/entry/start.b4301699.js">
 <link rel="modulepreload" href="/docs/transformers/main/es/_app/immutable/chunks/scheduler.36a0863c.js">
 <link rel="modulepreload" href="/docs/transformers/main/es/_app/immutable/chunks/singletons.a6ab9d4a.js">
 <link rel="modulepreload" href="/docs/transformers/main/es/_app/immutable/chunks/index.733708bb.js">
 <link rel="modulepreload" href="/docs/transformers/main/es/_app/immutable/chunks/paths.815a7736.js">
 <link rel="modulepreload" href="/docs/transformers/main/es/_app/immutable/entry/app.eaf9d9af.js">
 <link rel="modulepreload" href="/docs/transformers/main/es/_app/immutable/chunks/index.f891bdb2.js">
 <link rel="modulepreload" href="/docs/transformers/main/es/_app/immutable/nodes/0.e9e5a018.js">
 <link rel="modulepreload" href="/docs/transformers/main/es/_app/immutable/chunks/each.e59479a4.js">
 <link rel="modulepreload" href="/docs/transformers/main/es/_app/immutable/nodes/12.46802909.js">
 <link rel="modulepreload" href="/docs/transformers/main/es/_app/immutable/chunks/Tip.a8272f7f.js">
 <link rel="modulepreload" href="/docs/transformers/main/es/_app/immutable/chunks/CodeBlock.3ec784ea.js">
 <link rel="modulepreload" href="/docs/transformers/main/es/_app/immutable/chunks/EditOnGithub.a58e27a9.js"><!-- HEAD_svelte-u9bgzb_START --><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;Debugging&quot;,&quot;local&quot;:&quot;debugging&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Debug de problemas de Network multi-GPU&quot;,&quot;local&quot;:&quot;debug-de-problemas-de-network-multi-gpu&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Detección de Underflow y Overflow&quot;,&quot;local&quot;:&quot;detección-de-underflow-y-overflow&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Rastreo de valores mínimos y máximos absolutos de batches específicos&quot;,&quot;local&quot;:&quot;rastreo-de-valores-mínimos-y-máximos-absolutos-de-batches-específicos&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3}],&quot;depth&quot;:2}],&quot;depth&quot;:1}"><!-- HEAD_svelte-u9bgzb_END --> <p></p> <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 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>Debugging</span></h1> <h2 class="relative group"><a id="debug-de-problemas-de-network-multi-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="#debug-de-problemas-de-network-multi-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>Debug de problemas de Network multi-GPU</span></h2> <p data-svelte-h="svelte-b2gmoe">Cuando entrenas o infieres con <code>DistributedDataParallel</code> y varias GPUs, si encuentras problemas de intercomunicación entre procesos y/o nodos, puedes usar el siguiente script para diagnosticar problemas de red.</p> <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->wget https://raw.githubusercontent.com/huggingface/transformers/main/scripts/distributed/torch-distributed-gpu-test.py<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1936y28">Por ejemplo, para probar cómo interactúan 2 GPUs, haz lo siguiente:</p> <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->python -m torch.distributed.run --nproc_per_node 2 --nnodes 1 torch-distributed-gpu-test.py<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1m3rn6q">Si ambos procesos pueden hablar entre sí y asignar la memoria de la GPU, cada uno imprimirá un status OK.</p> <p data-svelte-h="svelte-1qjrb86">Para más GPUs o nodos, ajusta los argumentos en el script.</p> <p data-svelte-h="svelte-1ku6vfi">Encontrarás muchos más detalles dentro del script de diagnóstico e incluso una receta de cómo ejecutarlo en un entorno SLURM.</p> <p data-svelte-h="svelte-deyufk">Un nivel adicional de debug es agregar la variable de entorno <code>NCCL_DEBUG=INFO</code> de la siguiente manera:</p> <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->NCCL_DEBUG=INFO python -m torch.distributed.run --nproc_per_node 2 --nnodes 1 torch-distributed-gpu-test.py<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-114ng1i">Esto mostrará mucha información de debug relacionada con NCCL, que luego puedes buscar online si encuentras que reporta algún problema. O si no estás seguro de cómo interpretar el output, puedes compartir el archivo de log en un Issue.</p> <h2 class="relative group"><a id="detección-de-underflow-y-overflow" 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="#detección-de-underflow-y-overflow"><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>Detección de Underflow y Overflow</span></h2> <div class="course-tip bg-gradient-to-br dark:bg-gradient-to-r before:border-green-500 dark:before:border-green-800 from-green-50 dark:from-gray-900 to-white dark:to-gray-950 border border-green-50 text-green-700 dark:text-gray-400"><p data-svelte-h="svelte-ymsays">Esta función está disponible actualmente sólo para PyTorch.</p></div> <div class="course-tip bg-gradient-to-br dark:bg-gradient-to-r before:border-green-500 dark:before:border-green-800 from-green-50 dark:from-gray-900 to-white dark:to-gray-950 border border-green-50 text-green-700 dark:text-gray-400"><p data-svelte-h="svelte-1svbjk8">Para el entrenamiento multi-GPU, requiere DDP (<code>torch.distributed.launch</code>).</p></div> <div class="course-tip bg-gradient-to-br dark:bg-gradient-to-r before:border-green-500 dark:before:border-green-800 from-green-50 dark:from-gray-900 to-white dark:to-gray-950 border border-green-50 text-green-700 dark:text-gray-400"><p data-svelte-h="svelte-16216u7">Esta función puede utilizarse con cualquier modelo basado en <code>nn.Module</code>.</p></div> <p data-svelte-h="svelte-7sit9h">Si empiezas a obtener <code>loss=NaN</code> o el modelo muestra algún otro comportamiento anormal debido a <code>inf</code> o <code>nan</code> en
activations o weights hay que descubrir dónde se produce el primer underflow o overflow y qué lo ha provocado. Por suerte
puedes lograrlo fácilmente activando un módulo especial que hará la detección automáticamente.</p> <p data-svelte-h="svelte-j3cpcs">Si estás usando <code>Trainer</code>, solo necesitas añadir:</p> <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->--debug underflow_overflow<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-15ev0q2">a los argumentos normales de la línea de comandos, o pasar <code>debug=&quot;underflow_overflow&quot;</code> al crear el objeto <code>TrainingArguments</code>.</p> <p data-svelte-h="svelte-wqtp9c">Si estás usando tu propio bucle de entrenamiento u otro Trainer puedes lograr lo mismo con:</p> <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">from</span> .debug_utils <span class="hljs-keyword">import</span> DebugUnderflowOverflow
debug_overflow = DebugUnderflowOverflow(model)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1jo5kre"><code>DebugUnderflowOverflow</code> inserta hooks en el modelo que inmediatamente después de cada forward
testeará las variables de input y output y también los weights del módulo correspondiente. Tan pronto como se detecte <code>inf</code> o
<code>nan</code> se detecta en al menos un elemento de las activations o weights, el programa afirmará e imprimirá un informe
como este (esto fue capturado con <code>google/mt5-small</code> bajo fp16 mixed precision):</p> <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-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<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-x04aaj">El output del ejemplo se ha recortado en el centro por razones de brevedad.</p> <p data-svelte-h="svelte-4r89ln">La segunda columna muestra el valor del elemento más grande en términos absolutos, por lo que si observas con detenimiento los últimos fotogramas,
los inputs y outputs estaban en el rango de <code>1e4</code>. Así que cuando este entrenamiento se hizo con fp16 mixed precision,
el último paso sufrió overflow (ya que bajo <code>fp16</code> el mayor número antes de <code>inf</code> es <code>64e3</code>). Para evitar overflows en
<code>fp16</code> las activations deben permanecer muy por debajo de <code>1e4</code>, porque <code>1e4 * 1e4 = 1e8</code> por lo que cualquier matrix multiplication con
grandes activations va a llevar a una condición de overflow numérico.</p> <p data-svelte-h="svelte-6zf2z9">Al principio del output puedes descubrir en qué número de batch se produjo el problema (aquí <code>Detected inf/nan during batch_number=0</code> significa que el problema se produjo en el primer batch).</p> <p data-svelte-h="svelte-1opunmv">Cada frame del informe comienza declarando la entrada completamente calificada para el módulo correspondiente que este frame está reportando.
Si nos fijamos sólo en este frame:</p> <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --> <span class="hljs-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<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-dwkjw4">Aquí, <code>encoder.block.2.layer.1.layer_norm</code> indica que era una layer norm para la primera capa, del segundo
block del encoder. Y la call específica del <code>forward</code> es <code>T5LayerNorm</code>.</p> <p data-svelte-h="svelte-1l806z7">Veamos los últimos frames de ese informe:</p> <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-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<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-15sbs2p">El último frame informa para la función <code>Dropout.forward</code> con la primera entrada para el único input y la segunda para el
único output. Puedes ver que fue llamada desde un atributo <code>dropout</code> dentro de la clase <code>DenseReluDense</code>. Podemos ver
que ocurrió durante la primera capa, del segundo block, durante el primer batch. Por último, el mayor absoluto
elementos de input fue <code>6.27e+04</code> y el mismo para el output fue <code>inf</code>.</p> <p data-svelte-h="svelte-1h0cfpa">Puedes ver aquí, que <code>T5DenseGatedGeluDense.forward</code> resultó en output activations, cuyo valor máximo absoluto fue
alrededor de 62.7K, que está muy cerca del límite máximo de fp16 de 64K. En el siguiente frame tenemos <code>Dropout</code>, el cual renormaliza
los weights, después de poner a cero algunos de los elementos, lo que empuja el valor máximo absoluto a más de 64K, y obtenemos un
overflow (<code>inf</code>).</p> <p data-svelte-h="svelte-6rkr7t">Como puedes ver son los frames anteriores los que tenemos que mirar cuando los números empiezan a ser muy grandes para números fp16.</p> <p data-svelte-h="svelte-1odtud2">Combinemos el informe con el código de <code>models/t5/modeling_t5.py</code>:</p> <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">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__()
 self.wi_0 = nn.Linear(config.d_model, config.d_ff, bias=<span class="hljs-literal">False</span>)
 self.wi_1 = nn.Linear(config.d_model, config.d_ff, bias=<span class="hljs-literal">False</span>)
 self.wo = nn.Linear(config.d_ff, config.d_model, bias=<span class="hljs-literal">False</span>)
 self.dropout = nn.Dropout(config.dropout_rate)
 self.gelu_act = ACT2FN[<span class="hljs-string">&quot;gelu_new&quot;</span>]
<span class="hljs-keyword">def</span> <span class="hljs-title function_">forward</span>(<span class="hljs-params">self, hidden_states</span>):
 hidden_gelu = self.gelu_act(self.wi_0(hidden_states))
 hidden_linear = self.wi_1(hidden_states)
 hidden_states = hidden_gelu * hidden_linear
 hidden_states = self.dropout(hidden_states)
 hidden_states = self.wo(hidden_states)
 <span class="hljs-keyword">return</span> hidden_states<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-pypfv9">Ahora es fácil ver la call <code>dropout</code>, y también todas las calls anteriores.</p> <p data-svelte-h="svelte-14ofdnz">Dado que la detección se produce en un forward hook, estos informes se imprimen inmediatamente después de que cada <code>forward</code>
responda.</p> <p data-svelte-h="svelte-p0ujj7">Volviendo al informe completo, para actuar sobre él y arreglar el problema, tenemos que subir unos cuantos frames donde los números
empezaron a subir y probablemente cambiar al modo <code>fp32</code> aquí, para que los números no sufran overflow cuando se multipliquen
o al sumarlos. Por supuesto, puede haber otras soluciones. Por ejemplo, podríamos desactivar <code>amp</code> temporalmente si está
activado, después de mover el original <code>forward</code> dentro de un helper wrapper, así:</p> <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">def</span> <span class="hljs-title function_">_forward</span>(<span class="hljs-params">self, hidden_states</span>):
 hidden_gelu = self.gelu_act(self.wi_0(hidden_states))
 hidden_linear = self.wi_1(hidden_states)
 hidden_states = hidden_gelu * hidden_linear
 hidden_states = self.dropout(hidden_states)
 hidden_states = self.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>):
 <span class="hljs-keyword">if</span> torch.is_autocast_enabled():
 <span class="hljs-keyword">with</span> torch.cuda.amp.autocast(enabled=<span class="hljs-literal">False</span>):
 <span class="hljs-keyword">return</span> self._forward(hidden_states)
 <span class="hljs-keyword">else</span>:
 <span class="hljs-keyword">return</span> self._forward(hidden_states)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1nnn3eq">Como el detector automático sólo informa de los inputs y outputs de los frames completos, una vez que sepas dónde buscar, puedes
analizar también las etapas intermedias de una función específica de <code>forward</code>. En este caso, puede utilizar la función
función de ayuda <code>detect_overflow</code> para inyectar el detector donde quieras, por ejemplo:</p> <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">from</span> 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 = self.layer_norm(hidden_states)
 detect_overflow(forwarded_states, <span class="hljs-string">&quot;after layer_norm&quot;</span>)
 forwarded_states = self.DenseReluDense(forwarded_states)
 detect_overflow(forwarded_states, <span class="hljs-string">&quot;after DenseReluDense&quot;</span>)
 <span class="hljs-keyword">return</span> hidden_states + self.dropout(forwarded_states)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-ewhn06">Puedes ver que hemos añadido 2 de estos y ahora se trackea si <code>inf</code> o <code>nan</code> para <code>forwarded_states</code> fue detectado
en algún punto intermedio.</p> <p data-svelte-h="svelte-1psj6in">De hecho, el detector ya informa de esto porque cada una de las llamadas en el ejemplo anterior es un <code>nn.Module</code>, pero
digamos que si tuvieras algunos cálculos directos locales, así es como lo harías.</p> <p data-svelte-h="svelte-vdijwi">Además, si estás instanciando el debugger en tu propio código, puedes ajustar el número de frames impresos de
su valor por defecto, por ejemplo:</p> <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">from</span> .debug_utils <span class="hljs-keyword">import</span> DebugUnderflowOverflow
debug_overflow = DebugUnderflowOverflow(model, max_frames_to_save=<span class="hljs-number">100</span>)<!-- HTML_TAG_END --></pre></div> <h3 class="relative group"><a id="rastreo-de-valores-mínimos-y-máximos-absolutos-de-batches-específicos" 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="#rastreo-de-valores-mínimos-y-máximos-absolutos-de-batches-específicos"><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>Rastreo de valores mínimos y máximos absolutos de batches específicos</span></h3> <p data-svelte-h="svelte-12vhr9v">La misma clase de debugging se puede utilizar para el rastreo por batches con la función de detección de underflow/overflow desactivada.</p> <p data-svelte-h="svelte-4ky2ig">Digamos que quieres ver los valores mínimos y máximos absolutos de todos los ingredientes de cada call <code>forward</code> de un determinado
batch, y sólo hacerlo para los batches 1 y 3. Entonces instancias esta clase como:</p> <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->debug_overflow = DebugUnderflowOverflow(model, trace_batch_nums=[<span class="hljs-number">1</span>, <span class="hljs-number">3</span>])<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1deawea">Y ahora los batches 1 y 3 completos serán rastreados usando el mismo formato que el detector de underflow/overflow.</p> <p data-svelte-h="svelte-1d0mz9t">Los batches son 0-index.</p> <p data-svelte-h="svelte-1gv06dh">Esto es muy útil si sabes que el programa empieza a comportarse mal después de un determinado número de batch, para que puedas avanzar rápidamente
hasta esa área. Aquí hay un ejemplo de output recortado para tal configuración:</p> <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --> *** 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
[...]<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-k61p1e">Aquí obtendrás un gran número de frames mostrados - tantos como forward calls haya en tu modelo, por lo que puede o no ser lo que quieras, pero a veces puede ser más fácil de usar para debug que un debugger normal.
Por ejemplo, si un problema comienza a ocurrir en el batch 150. Entonces puedes mostrar las trazas de los batches 149 y 150 y comparar dónde
los números empezaron a divergir.</p> <p data-svelte-h="svelte-1vqwv9x">También puedes especificar el número de batch después del cual se debe detener el entrenamiento, con:</p> <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->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>)<!-- HTML_TAG_END --></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/es/debugging.md" target="_blank"><span data-svelte-h="svelte-1kd6by1">&lt;</span> <span data-svelte-h="svelte-x0xyl0">&gt;</span> <span data-svelte-h="svelte-1dajgef"><span class="underline ml-1.5">Update</span> on GitHub</span></a> <p></p>
<script>
 {
 __sveltekit_8nzhgd = {
 assets: "/docs/transformers/main/es",
 base: "/docs/transformers/main/es",
 env: {}
 };
 const element = document.currentScript.parentElement;
 const data = [null,null];
 Promise.all([
 import("/docs/transformers/main/es/_app/immutable/entry/start.b4301699.js"),
 import("/docs/transformers/main/es/_app/immutable/entry/app.eaf9d9af.js")
 ]).then(([kit, app]) => {
 kit.start(app, element, {
 node_ids: [0, 12],
 data,
 form: null,
 error: null
 });
 });
 }
 </script>

Xet Storage Details

Size:
55.2 kB
·
Xet hash:
ba63fdcec7d800bc925bd3ad51e7a1f6a8885f6f5a5e85a896cb70552cf7a71c

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