Buckets:
| <meta charset="utf-8" /><meta name="hf:doc:metadata" content="{"title":"Optimizers","local":"optimizers","sections":[{"title":"APOLLO","local":"apollo","sections":[],"depth":2},{"title":"GrokAdamW","local":"grokadamw","sections":[],"depth":2},{"title":"LOMO","local":"lomo","sections":[],"depth":2},{"title":"Schedule Free","local":"schedule-free","sections":[],"depth":2},{"title":"StableAdamW","local":"stableadamw","sections":[],"depth":2}],"depth":1}"> | |
| <link href="/docs/transformers/pr_33962/en/_app/immutable/assets/0.e3b0c442.css" rel="modulepreload"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_33962/en/_app/immutable/entry/start.1776ac26.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_33962/en/_app/immutable/chunks/scheduler.31fdf58d.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_33962/en/_app/immutable/chunks/singletons.73ad6261.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_33962/en/_app/immutable/chunks/index.252883d5.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_33962/en/_app/immutable/chunks/paths.7e4d6052.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_33962/en/_app/immutable/entry/app.d2b9a7d8.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_33962/en/_app/immutable/chunks/preload-helper.769e5991.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_33962/en/_app/immutable/chunks/index.2f76fdf0.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_33962/en/_app/immutable/nodes/0.29a76cb1.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_33962/en/_app/immutable/chunks/each.e59479a4.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_33962/en/_app/immutable/nodes/499.04c6597c.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_33962/en/_app/immutable/chunks/CopyLLMTxtMenu.9922688f.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_33962/en/_app/immutable/chunks/globals.7f7f1b26.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_33962/en/_app/immutable/chunks/IconCopy.ac192424.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_33962/en/_app/immutable/chunks/MermaidChart.svelte_svelte_type_style_lang.4975a63c.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_33962/en/_app/immutable/chunks/CodeBlock.ab12f8e1.js"><!-- HEAD_svelte-u9bgzb_START --><meta name="hf:doc:metadata" content="{"title":"Optimizers","local":"optimizers","sections":[{"title":"APOLLO","local":"apollo","sections":[],"depth":2},{"title":"GrokAdamW","local":"grokadamw","sections":[],"depth":2},{"title":"LOMO","local":"lomo","sections":[],"depth":2},{"title":"Schedule Free","local":"schedule-free","sections":[],"depth":2},{"title":"StableAdamW","local":"stableadamw","sections":[],"depth":2}],"depth":1}"><!-- HEAD_svelte-u9bgzb_END --> <p></p> <div class="items-center shrink-0 min-w-[100px] max-sm:min-w-[50px] justify-end ml-auto flex" style="float: right; margin-left: 10px; display: inline-flex; position: relative; z-index: 10;"><div class="inline-flex rounded-md max-sm:rounded-sm"><button class="inline-flex items-center gap-1 max-sm:gap-0.5 h-6 max-sm:h-5 px-2 max-sm:px-1.5 text-[11px] max-sm:text-[9px] 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"><svg class="w-3 h-3 max-sm:w-2.5 max-sm:h-2.5" 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-6 max-sm:h-5 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 w-3 h-3 max-sm:w-2.5 max-sm:h-2.5 rotate-0" width="1em" height="1em" viewBox="0 0 12 7" fill="none" xmlns="http://www.w3.org/2000/svg"><path d="M1 1L6 6L11 1" stroke="currentColor"></path></svg></button></div> </div> <h1 class="relative group"><a id="optimizers" 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="#optimizers"><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>Optimizers</span></h1> <p data-svelte-h="svelte-c72azx">Transformers offers two native optimizers, AdamW and AdaFactor. It also provides integrations for more specialized optimizers. Install the library that offers the optimizer and drop it in the <code>optim</code> parameter in <a href="/docs/transformers/pr_33962/en/main_classes/trainer#transformers.TrainingArguments">TrainingArguments</a>.</p> <p data-svelte-h="svelte-nu6zm7">This guide will show you how to use these optimizers with <a href="/docs/transformers/pr_33962/en/main_classes/trainer#transformers.Trainer">Trainer</a> using <a href="/docs/transformers/pr_33962/en/main_classes/trainer#transformers.TrainingArguments">TrainingArguments</a> shown below.</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">import</span> torch | |
| <span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> TrainingArguments, AutoTokenizer, AutoModelForCausalLM, Trainer | |
| args = TrainingArguments( | |
| output_dir=<span class="hljs-string">"./test-optimizer"</span>, | |
| max_steps=<span class="hljs-number">1000</span>, | |
| per_device_train_batch_size=<span class="hljs-number">4</span>, | |
| logging_strategy=<span class="hljs-string">"steps"</span>, | |
| logging_steps=<span class="hljs-number">1</span>, | |
| learning_rate=<span class="hljs-number">2e-5</span>, | |
| save_strategy=<span class="hljs-string">"no"</span>, | |
| run_name=<span class="hljs-string">"optimizer-name"</span>, | |
| )<!-- HTML_TAG_END --></pre></div> <h2 class="relative group"><a id="apollo" 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="#apollo"><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>APOLLO</span></h2> <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 -->pip install apollo-torch<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1yo6y13"><a href="https://github.com/zhuhanqing/APOLLO" rel="nofollow">Approximated Gradient Scaling for Memory Efficient LLM Optimization (APOLLO)</a> is a memory-efficient optimizer that allows full parameter learning for both pretraining and fine-tuning. It maintains AdamW-level performance with SGD-like memory efficiency. For extreme memory efficiency, you can use APOLLO-Mini, a rank 1 variant of APOLLO. APOLLO optimizers support:</p> <ul data-svelte-h="svelte-1vta6r9"><li>Ultra-low rank efficiency. You can use a much lower rank than <a href="./trainer#galore">GaLoRE</a>, rank 1 is sufficient.</li> <li>Avoid expensive SVD computations. APOLLO leverages random projections to avoid training stalls.</li></ul> <p data-svelte-h="svelte-903dkf">Use the <code>optim_target_modules</code> parameter to specify which layers to train.</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 -->import torch | |
| from transformers import TrainingArguments | |
| args = TrainingArguments( | |
| output_dir="./test-apollo", | |
| max_steps=100, | |
| per_device_train_batch_size=2, | |
| <span class="hljs-addition">+ optim="apollo_adamw",</span> | |
| <span class="hljs-addition">+ optim_target_modules=[r".*.attn.*", r".*.mlp.*"],</span> | |
| logging_strategy="steps", | |
| logging_steps=1, | |
| learning_rate=2e-5, | |
| save_strategy="no", | |
| run_name="apollo_adamw", | |
| )<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-r5aiqq">For additional training options, use <code>optim_args</code> to define hyperparameters like <code>rank</code>, <code>scale</code>, and more. Refer to the table below for a complete list of available hyperparameters.</p> <blockquote class="tip" data-svelte-h="svelte-1vhkhb9"><p>The <code>scale</code> parameter can be set to <code>n/r</code>, where <code>n</code> is the original space dimension and <code>r</code> is the low-rank space dimension. You could achieve a similar effect by adjusting the learning rate while keeping <code>scale</code> at its default value.</p></blockquote> <table data-svelte-h="svelte-8mx1ik"><thead><tr><th>parameter</th> <th>description</th> <th>APOLLO</th> <th>APOLLO-Mini</th></tr></thead> <tbody><tr><td>rank</td> <td>rank of the auxiliary sub-space for gradient scaling</td> <td>256</td> <td>1</td></tr> <tr><td>scale_type</td> <td>how scaling factors are applied</td> <td><code>channel</code> (per-channel scaling)</td> <td><code>tensor</code> (per-tensor scaling)</td></tr> <tr><td>scale</td> <td>adjusts gradient updates to stabilize training</td> <td>1.0</td> <td>128</td></tr> <tr><td>update_proj_gap</td> <td>steps before updating projection matrices</td> <td>200</td> <td>200</td></tr> <tr><td>proj</td> <td>projection type</td> <td><code>random</code></td> <td><code>random</code></td></tr></tbody></table> <p data-svelte-h="svelte-gjt3t3">The example below enables the APOLLO-Mini optimizer.</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> transformers <span class="hljs-keyword">import</span> TrainingArguments | |
| args = TrainingArguments( | |
| output_dir=<span class="hljs-string">"./test-apollo_mini"</span>, | |
| max_steps=<span class="hljs-number">100</span>, | |
| per_device_train_batch_size=<span class="hljs-number">2</span>, | |
| optim=<span class="hljs-string">"apollo_adamw"</span>, | |
| optim_target_modules=[<span class="hljs-string">r".*.attn.*"</span>, <span class="hljs-string">r".*.mlp.*"</span>], | |
| optim_args=<span class="hljs-string">"proj=random,rank=1,scale=128.0,scale_type=tensor,update_proj_gap=200"</span>, | |
| )<!-- HTML_TAG_END --></pre></div> <h2 class="relative group"><a id="grokadamw" 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="#grokadamw"><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>GrokAdamW</span></h2> <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 -->pip install grokadamw<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-13e7mzp"><a href="https://github.com/cognitivecomputations/grokadamw" rel="nofollow">GrokAdamW</a> is an optimizer designed to help models that benefit from <em>grokking</em>, a term used to describe delayed generalization because of slow-varying gradients. It is particularly useful for models requiring more advanced optimization techniques to achieve better performance and stability.</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 -->import torch | |
| from transformers import TrainingArguments | |
| args = TrainingArguments( | |
| output_dir="./test-grokadamw", | |
| max_steps=1000, | |
| per_device_train_batch_size=4, | |
| <span class="hljs-addition">+ optim="grokadamw",</span> | |
| logging_strategy="steps", | |
| logging_steps=1, | |
| learning_rate=2e-5, | |
| save_strategy="no", | |
| run_name="grokadamw", | |
| )<!-- HTML_TAG_END --></pre></div> <h2 class="relative group"><a id="lomo" 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="#lomo"><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>LOMO</span></h2> <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 -->pip install lomo-optim<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-xldscx"><a href="https://github.com/OpenLMLab/LOMO" rel="nofollow">Low-Memory Optimization (LOMO)</a> is a family of optimizers, <a href="https://huggingface.co/papers/2306.09782" rel="nofollow">LOMO</a> and <a href="https://hf.co/papers/2310.10195" rel="nofollow">AdaLomo</a>, designed for low-memory full-parameter finetuning of LLMs. Both LOMO optimizers fuse the gradient computation and parameter update in one step to reduce memory usage. AdaLomo builds on top of LOMO by incorporating an adaptive learning rate for each parameter like the Adam optimizer.</p> <blockquote class="tip" data-svelte-h="svelte-g2xe09"><p>It is recommended to use AdaLomo without <code>grad_norm</code> for better performance and higher throughput.</p></blockquote> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->args = TrainingArguments( | |
| output_dir="./test-lomo", | |
| max_steps=1000, | |
| per_device_train_batch_size=4, | |
| <span class="hljs-addition">+ optim="adalomo",</span> | |
| gradient_checkpointing=True, | |
| logging_strategy="steps", | |
| logging_steps=1, | |
| learning_rate=2e-6, | |
| save_strategy="no", | |
| run_name="adalomo", | |
| )<!-- HTML_TAG_END --></pre></div> <h2 class="relative group"><a id="schedule-free" 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="#schedule-free"><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>Schedule Free</span></h2> <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 -->pip install schedulefree<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-18cap7e"><a href="https://hf.co/papers/2405.15682" rel="nofollow">Schedule Free optimizer (SFO)</a> replaces the base optimizers momentum with a combination of averaging and interpolation. Unlike a traditional scheduler, SFO completely removes the need to anneal the learning rate.</p> <p data-svelte-h="svelte-1r2cq5s">SFO supports the RAdam (<code>schedule_free_radam</code>), AdamW (<code>schedule_free_adamw</code>) and SGD (<code>schedule_free_sgd</code>) optimizers. The RAdam scheduler doesn’t require <code>warmup_steps</code> or <code>warmup_ratio</code>.</p> <p data-svelte-h="svelte-mqeyfj">By default, it is recommended to set <code>lr_scheduler_type="constant"</code>. Other <code>lr_scheduler_type</code> values may also work, but combining SFO optimizers with other learning rate schedules could affect SFOs intended behavior and performance.</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 -->args = TrainingArguments( | |
| output_dir="./test-schedulefree", | |
| max_steps=1000, | |
| per_device_train_batch_size=4, | |
| <span class="hljs-addition">+ optim="schedule_free_radamw",</span> | |
| <span class="hljs-addition">+ lr_scheduler_type="constant",</span> | |
| gradient_checkpointing=True, | |
| logging_strategy="steps", | |
| logging_steps=1, | |
| learning_rate=2e-6, | |
| save_strategy="no", | |
| run_name="sfo", | |
| )<!-- HTML_TAG_END --></pre></div> <h2 class="relative group"><a id="stableadamw" 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="#stableadamw"><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>StableAdamW</span></h2> <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 -->pip install torch-optimi<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-owe02r"><a href="https://huggingface.co/papers/2304.13013" rel="nofollow">StableAdamW</a> is a hybrid between AdamW and AdaFactor. It ports AdaFactor’s update clipping into AdamW, which removes the need for gradient clipping. Otherwise, it behaves as a drop-in replacement for AdamW.</p> <blockquote class="tip" data-svelte-h="svelte-19sy4w5"><p>If training on large batch sizes or still observing training loss spikes, consider reducing beta_2 between [0.95, 0.99].</p></blockquote> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->args = TrainingArguments( | |
| output_dir="./test-stable-adamw", | |
| max_steps=1000, | |
| per_device_train_batch_size=4, | |
| <span class="hljs-addition">+ optim="stable_adamw",</span> | |
| gradient_checkpointing=True, | |
| logging_strategy="steps", | |
| logging_steps=1, | |
| learning_rate=2e-6, | |
| save_strategy="no", | |
| run_name="stable-adamw", | |
| )<!-- 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/en/optimizers.md" target="_blank"><svg class="mr-1" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M31,16l-7,7l-1.41-1.41L28.17,16l-5.58-5.59L24,9l7,7z"></path><path d="M1,16l7-7l1.41,1.41L3.83,16l5.58,5.59L8,23l-7-7z"></path><path d="M12.419,25.484L17.639,6.552l1.932,0.518L14.351,26.002z"></path></svg> <span data-svelte-h="svelte-zjs2n5"><span class="underline">Update</span> on GitHub</span></a> <p></p> | |
| <script> | |
| { | |
| __sveltekit_1yuubte = { | |
| assets: "/docs/transformers/pr_33962/en", | |
| base: "/docs/transformers/pr_33962/en", | |
| env: {} | |
| }; | |
| const element = document.currentScript.parentElement; | |
| const data = [null,null]; | |
| Promise.all([ | |
| import("/docs/transformers/pr_33962/en/_app/immutable/entry/start.1776ac26.js"), | |
| import("/docs/transformers/pr_33962/en/_app/immutable/entry/app.d2b9a7d8.js") | |
| ]).then(([kit, app]) => { | |
| kit.start(app, element, { | |
| node_ids: [0, 499], | |
| data, | |
| form: null, | |
| error: null | |
| }); | |
| }); | |
| } | |
| </script> | |
Xet Storage Details
- Size:
- 37.3 kB
- Xet hash:
- 6f125c293c1231492939d1c2a0861bd06ff7459814b9f4342639aa132c0ab3de
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.