Buckets:
| <meta charset="utf-8" /><meta name="hf:doc:metadata" content="{"title":"Debugging","local":"debugging","sections":[{"title":"Debug de problemas de Network multi-GPU","local":"debug-de-problemas-de-network-multi-gpu","sections":[],"depth":2},{"title":"Detección de Underflow y Overflow","local":"detección-de-underflow-y-overflow","sections":[{"title":"Rastreo de valores mínimos y máximos absolutos de batches específicos","local":"rastreo-de-valores-mínimos-y-máximos-absolutos-de-batches-específicos","sections":[],"depth":3}],"depth":2}],"depth":1}"/> | |
| <link href="/docs/transformers/main/es/_app/immutable/entry/start.Cex4YXjh.js" rel="modulepreload"> | |
| <link href="/docs/transformers/main/es/_app/immutable/chunks/D_b9q9jZ.js" rel="modulepreload"> | |
| <link href="/docs/transformers/main/es/_app/immutable/chunks/Dl9aQtdI.js" rel="modulepreload"> | |
| <link href="/docs/transformers/main/es/_app/immutable/entry/app.DNxQXwTy.js" rel="modulepreload"> | |
| <link href="/docs/transformers/main/es/_app/immutable/chunks/ByWXHg74.js" rel="modulepreload"> | |
| <link href="/docs/transformers/main/es/_app/immutable/chunks/NvhOjA8g.js" rel="modulepreload"> | |
| <link href="/docs/transformers/main/es/_app/immutable/chunks/BWsCTqVr.js" rel="modulepreload"> | |
| <link href="/docs/transformers/main/es/_app/immutable/chunks/DsnmJJEf.js" rel="modulepreload"> | |
| <link href="/docs/transformers/main/es/_app/immutable/nodes/0.DUkXUlQz.js" rel="modulepreload"> | |
| <link href="/docs/transformers/main/es/_app/immutable/chunks/DkDiSK5v.js" rel="modulepreload"> | |
| <link href="/docs/transformers/main/es/_app/immutable/nodes/12.FPfsb5MU.js" rel="modulepreload"> | |
| <link href="/docs/transformers/main/es/_app/immutable/chunks/Cy51lpuW.js" rel="modulepreload"> | |
| <link href="/docs/transformers/main/es/_app/immutable/chunks/71x9jLwp.js" rel="modulepreload"> | |
| <!--xyche3--><meta name="hf:doc:metadata" content="{"title":"Debugging","local":"debugging","sections":[{"title":"Debug de problemas de Network multi-GPU","local":"debug-de-problemas-de-network-multi-gpu","sections":[],"depth":2},{"title":"Detección de Underflow y Overflow","local":"detección-de-underflow-y-overflow","sections":[{"title":"Rastreo de valores mínimos y máximos absolutos de batches específicos","local":"rastreo-de-valores-mínimos-y-máximos-absolutos-de-batches-específicos","sections":[],"depth":3}],"depth":2}],"depth":1}"/><!----> | |
| <link href="/docs/transformers/main/es/_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>Debugging</span></h1><!--]--><!----> <!--[1--><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 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>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 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>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 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>Si ambos procesos pueden hablar entre sí y asignar la memoria de la GPU, cada uno imprimirá un status OK.</p> <p>Para más GPUs o nodos, ajusta los argumentos en el script.</p> <p>Encontrarás muchos más detalles dentro del script de diagnóstico e incluso una receta de cómo ejecutarlo en un entorno SLURM.</p> <p>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 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>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> <!--[1--><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 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><!--]--><!----> <blockquote class="tip"><p>Esta función está disponible actualmente sólo para PyTorch.</p><!----></blockquote><!----> <blockquote class="tip"><p>Para el entrenamiento multi-GPU, requiere DDP (<code>torch.distributed.launch</code>).</p><!----></blockquote><!----> <blockquote class="tip"><p>Esta función puede utilizarse con cualquier modelo basado en <code>nn.Module</code>.</p><!----></blockquote><!----> <p>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>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 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>a los argumentos normales de la línea de comandos, o pasar <code>debug="underflow_overflow"</code> al crear el objeto <code>TrainingArguments</code>.</p> <p>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 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> DebugUnderflowOverflow | |
| debug_overflow = DebugUnderflowOverflow(model)<!----></pre></div><!----> <p><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 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>El output del ejemplo se ha recortado en el centro por razones de brevedad.</p> <p>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>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>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 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>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>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 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>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>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>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>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 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>Ahora es fácil ver la call <code>dropout</code>, y también todas las calls anteriores.</p> <p>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>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 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>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 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>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>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>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 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> DebugUnderflowOverflow | |
| debug_overflow = DebugUnderflowOverflow(model, max_frames_to_save=<span class="hljs-number">100</span>)<!----></pre></div><!----> <!--[2--><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 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>La misma clase de debugging se puede utilizar para el rastreo por batches con la función de detección de underflow/overflow desactivada.</p> <p>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 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>Y ahora los batches 1 y 3 completos serán rastreados usando el mismo formato que el detector de underflow/overflow.</p> <p>Los batches son 0-index.</p> <p>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 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>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>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 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/es/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_y4vo64 = { | |
| base: "/docs/transformers/main/es", | |
| assets: "/docs/transformers/main/es" | |
| }; | |
| const element = document.currentScript.parentElement; | |
| Promise.all([ | |
| import("/docs/transformers/main/es/_app/immutable/entry/start.Cex4YXjh.js"), | |
| import("/docs/transformers/main/es/_app/immutable/entry/app.DNxQXwTy.js") | |
| ]).then(([kit, app]) => { | |
| kit.start(app, element, { | |
| node_ids: [0, 12], | |
| data: [null,null], | |
| form: null, | |
| error: null | |
| }); | |
| }); | |
| } | |
| </script> | |
Xet Storage Details
- Size:
- 56.8 kB
- Xet hash:
- baff456dacedbb6c9feb931e9804434711887f0624db6c990f2bca611bb2c09e
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.