Hugging Face's logo

Buckets:
hf-doc-build
/
doc-dev

Files
xet
hf-doc-build/doc-dev / lerobot /pr_4380 /en /multi_gpu_training.html
HuggingFaceDocBuilder's picture
HuggingFaceDocBuilder
download
raw
43.7 kB
<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;Multi-GPU Training&quot;,&quot;local&quot;:&quot;multi-gpu-training&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Installation&quot;,&quot;local&quot;:&quot;installation&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Launching&quot;,&quot;local&quot;:&quot;launching&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Batch semantics, learning rate, and steps&quot;,&quot;local&quot;:&quot;batch-semantics-learning-rate-and-steps&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Sharded training (FSDP)&quot;,&quot;local&quot;:&quot;sharded-training-fsdp&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Wrap units&quot;,&quot;local&quot;:&quot;wrap-units&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;HSDP&quot;,&quot;local&quot;:&quot;hsdp&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3}],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Checkpoints&quot;,&quot;local&quot;:&quot;checkpoints&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Converting DCP checkpoints&quot;,&quot;local&quot;:&quot;converting-dcp-checkpoints&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Resuming&quot;,&quot;local&quot;:&quot;resuming&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3}],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Notes&quot;,&quot;local&quot;:&quot;notes&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2}],&quot;depth&quot;:1}"/>
 <link href="/docs/lerobot/main/en/_app/immutable/entry/start.BigjVlyv.js" rel="modulepreload">
 <link href="/docs/lerobot/main/en/_app/immutable/chunks/xA_erHoZ.js" rel="modulepreload">
 <link href="/docs/lerobot/main/en/_app/immutable/chunks/BeOBUSJC.js" rel="modulepreload">
 <link href="/docs/lerobot/main/en/_app/immutable/entry/app.wSiz0M_s.js" rel="modulepreload">
 <link href="/docs/lerobot/main/en/_app/immutable/chunks/DqTJsk7W.js" rel="modulepreload">
 <link href="/docs/lerobot/main/en/_app/immutable/chunks/BwNHWMUY.js" rel="modulepreload">
 <link href="/docs/lerobot/main/en/_app/immutable/chunks/DsnmJJEf.js" rel="modulepreload">
 <link href="/docs/lerobot/main/en/_app/immutable/chunks/BboA8w9n.js" rel="modulepreload">
 <link href="/docs/lerobot/main/en/_app/immutable/nodes/0.DCYzYRwi.js" rel="modulepreload">
 <link href="/docs/lerobot/main/en/_app/immutable/chunks/BUzkCu0d.js" rel="modulepreload">
 <link href="/docs/lerobot/main/en/_app/immutable/nodes/47.2EdNaIij.js" rel="modulepreload">
 <link href="/docs/lerobot/main/en/_app/immutable/chunks/DBfi0TBS.js" rel="modulepreload">
 <!--110g5q8--><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;Multi-GPU Training&quot;,&quot;local&quot;:&quot;multi-gpu-training&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Installation&quot;,&quot;local&quot;:&quot;installation&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Launching&quot;,&quot;local&quot;:&quot;launching&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Batch semantics, learning rate, and steps&quot;,&quot;local&quot;:&quot;batch-semantics-learning-rate-and-steps&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Sharded training (FSDP)&quot;,&quot;local&quot;:&quot;sharded-training-fsdp&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Wrap units&quot;,&quot;local&quot;:&quot;wrap-units&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;HSDP&quot;,&quot;local&quot;:&quot;hsdp&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3}],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Checkpoints&quot;,&quot;local&quot;:&quot;checkpoints&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Converting DCP checkpoints&quot;,&quot;local&quot;:&quot;converting-dcp-checkpoints&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Resuming&quot;,&quot;local&quot;:&quot;resuming&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3}],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Notes&quot;,&quot;local&quot;:&quot;notes&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2}],&quot;depth&quot;:1}"/><!---->
 <link href="/docs/lerobot/main/en/_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="multi-gpu-training" 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="#multi-gpu-training"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span>Multi-GPU Training</span></h1><!--]--><!----> <p>LeRobot trains on multiple GPUs through <a href="https://huggingface.co/docs/accelerate" rel="nofollow">Hugging Face Accelerate</a>. Three data-parallel layouts are supported:</p> <table><thead><tr><th>Layout</th><th>What it does</th><th>Config</th></tr></thead><tbody><tr><td><strong>DDP</strong></td><td>Replicates the full model on every GPU</td><td>default on any multi-GPU launch</td></tr><tr><td><strong>FSDP</strong></td><td>Shards parameters, gradients, and optimizer state across GPUs</td><td><code>--parallelism.dp_shard=N</code></td></tr><tr><td><strong>HSDP</strong></td><td>Shards within groups of GPUs, replicates across groups</td><td><code>--parallelism.dp_replicate=R --parallelism.dp_shard=S</code></td></tr></tbody></table> <!--[1--><h2 class="relative group"><a id="installation" 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="#installation"><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>Installation</span></h2><!--]--><!----> <p><code>accelerate</code> is included in the <code>training</code> extra:</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 "><!---->pip install <span class="hljs-string">&#x27;lerobot[training]&#x27;</span><!----></pre></div><!----> <!--[1--><h2 class="relative group"><a id="launching" 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="#launching"><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>Launching</span></h2><!--]--><!----> <p>Distributed training can be launched through both <code>torchrun</code> and <code>accelerate launch</code>. Accelerate is used as a plain launcher: it does not manage the training configuration, and every distributed training setting lives in LeRobot’s own config system.</p> <p>With <code>torchrun</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-bash "><!---->torchrun --nproc-per-node=2 $(<span class="hljs-built_in">which</span> lerobot-train) \
 --dataset.repo_id=<span class="hljs-variable">${HF_USER}</span>/my_dataset \
 --policy.type=act \
 --policy.repo_id=<span class="hljs-variable">${HF_USER}</span>/my_trained_policy \
 --output_dir=outputs/train/act_multi_gpu \
 --job_name=act_multi_gpu \
 --wandb.enable=<span class="hljs-literal">true</span><!----></pre></div><!----> <p>With <code>accelerate launch</code> (as a plain launcher):</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 "><!---->accelerate launch --num_processes=2 $(<span class="hljs-built_in">which</span> lerobot-train) \
 --dataset.repo_id=<span class="hljs-variable">${HF_USER}</span>/my_dataset \
 --policy.type=act \
 --policy.repo_id=<span class="hljs-variable">${HF_USER}</span>/my_trained_policy \
 --output_dir=outputs/train/act_multi_gpu \
 --job_name=act_multi_gpu \
 --wandb.enable=<span class="hljs-literal">true</span><!----></pre></div><!----> <p>With no <code>--parallelism.*</code> flags, a multi-process launch runs plain DDP. Multi-node runs use the standard <code>torchrun --nnodes/--node-rank/--rdzv-endpoint</code> flags (or <code>accelerate launch --num_machines/--machine_rank/--main_process_ip</code>).</p> <blockquote class="warning"><p>Accelerate’s YAML config files (<code>accelerate launch --config_file some.yaml</code>, <code>accelerate config</code>) are not supported. They configure the engine through environment variables, bypassing LeRobot’s configuration system, so <code>train_config.json</code> would no longer describe the settings a run actually used. <code>lerobot-train</code> therefore refuses to start when <a href="https://huggingface.co/docs/accelerate/usage_guides/fsdp" rel="nofollow">accelerate environment variables</a> are set. Put the settings in <code>--parallelism.*</code> / <code>--accelerator.*</code> flags instead, or set <code>LEROBOT_ALLOW_ACCELERATE_ENV=1</code> to acknowledge the override and proceed anyway.</p></blockquote> <!--[1--><h2 class="relative group"><a id="batch-semantics-learning-rate-and-steps" 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="#batch-semantics-learning-rate-and-steps"><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>Batch semantics, learning rate, and steps</span></h2><!--]--><!----> <p>Each of the <code>dp_replicate × dp_shard</code> data-parallel workers loads its own <code>--batch_size</code> micro-batch every step, so one training step consumes <code>batch_size × dp_world_size</code> samples, and <code>× gradient_accumulation_steps</code> of those go into each optimizer update:</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-attr">effective_batch_size</span> = batch_size × dp_world_size × gradient_accumulation_steps<!----></pre></div><!----> <p>The training banner prints this factorization at startup. <code>--steps</code> counts loop steps (micro-batches per worker), not optimizer updates.</p> <p>Gradient accumulation is a first-class flag:</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 "><!---->torchrun --nproc-per-node=2 $(<span class="hljs-built_in">which</span> lerobot-train) \
 --batch_size=8 --accelerator.gradient_accumulation.steps=4 ...<!----></pre></div><!----> <p><strong>LeRobot does not auto-scale the learning rate or the number of steps</strong> when the effective batch size grows. If you scale out and want equivalent training, please adjust manually, e.g. with 2 GPUs: double <code>--optimizer.lr</code> (linear scaling), or halve <code>--steps</code>.</p> <!--[1--><h2 class="relative group"><a id="sharded-training-fsdp" 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="#sharded-training-fsdp"><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>Sharded training (FSDP)</span></h2><!--]--><!----> <p>If a model is too large to train with DDP, shard it with FSDP2:</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 "><!---->torchrun --nproc-per-node=4 $(<span class="hljs-built_in">which</span> lerobot-train) \
 --dataset.repo_id=<span class="hljs-variable">${HF_USER}</span>/my_dataset \
 --policy.type=&lt;your_policy&gt; \
 --parallelism.dp_shard=4 \
 --accelerator.mixed_precision=bf16 \
 --output_dir=outputs/train/my_policy_fsdp<!----></pre></div><!----> <p><code>--parallelism.dp_shard=-1</code> shards over however many processes the launcher started.</p> <!--[2--><h3 class="relative group"><a id="wrap-units" 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="#wrap-units"><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>Wrap units</span></h3><!--]--><!----> <p>FSDP shards the model in units (typically the repeated transformer block) and gathers one unit at a time during forward/backward. Policies declare their wrap units via <code>_fsdp_wrap_modules</code> on the policy class. For example, ACT declares <code>["ACTEncoderLayer", "ACTDecoderLayer"]</code> and FastWAM declares <code>["MoTLayer"]</code>. For a policy without a <code>_fsdp_wrap_modules</code> declaration, pass one of the flags below. You can specify the module class name explicitly, or use a size-based policy instead:</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 "><!---->--accelerator.fsdp.wrap_modules=<span class="hljs-string">&#x27;[&quot;MyTransformerBlock&quot;]&#x27;</span> <span class="hljs-comment"># explicit class names</span>
--accelerator.fsdp.min_num_params=1000000 <span class="hljs-comment"># or: wrap every submodule above 1M params</span><!----></pre></div><!----> <p>If a policy doesn’t declare <code>_fsdp_wrap_modules</code> and no <code>--accelerator.fsdp.wrap_modules</code> or <code>--accelerator.fsdp.min_num_params</code> is passed, the run fails at startup rather than silently wrapping only the root module (which would forfeit all sharding memory savings).</p> <p>Other sharding settings:</p> <ul><li><code>--accelerator.fsdp.reshard_after_forward</code>: whether to keep each unit’s parameters resident after forward.</li> <li><code>--accelerator.fsdp.cpu_offload</code>: keeps parameters, gradients and optimizer states on CPU.</li> <li><code>--accelerator.fsdp.ignored_modules</code>: a regex of module paths to keep unsharded.</li></ul> <!--[2--><h3 class="relative group"><a id="hsdp" 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="#hsdp"><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>HSDP</span></h3><!--]--><!----> <p>Hybrid Sharded Data Parallel: parameters, gradients and optimizer states are sharded across <code>dp_shard</code> ranks, and that sharding is replicated <code>dp_replicate</code> times. Parameter all-gathers and gradient reduce-scatters stay inside a shard group; only the all-reduce that synchronizes the replicas crosses between groups. The two degrees must multiply to the world size:</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 "><!----><span class="hljs-comment"># 16 GPUs = 2 nodes × 8: shard within each node, replicate across nodes</span>
torchrun --nnodes=2 --nproc-per-node=8 ... $(<span class="hljs-built_in">which</span> lerobot-train) \
 --parallelism.dp_replicate=2 --parallelism.dp_shard=8 ...<!----></pre></div><!----> <!--[1--><h2 class="relative group"><a id="checkpoints" 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="#checkpoints"><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>Checkpoints</span></h2><!--]--><!----> <p>Every checkpoint contains a <code>pretrained_model/</code> directory and a <code>training_state/</code> directory:</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-text "><!---->005000/ # the training step at that checkpoint
├── pretrained_model/
│ ├── config.json # policy config
│ ├── train_config.json # the full training config
│ ├── model.safetensors # full weights (checkpoint_format ∈ {safetensors, safetensors_dcp}, or any non-sharded run)
│ ├── pytorch_model_fsdp_0/ # DCP weight shards (checkpoint_format ∈ {dcp, safetensors_dcp})
│ ├── policy_preprocessor.json # preprocessor config (when the run has a preprocessor)
│ ├── policy_preprocessor_step_*.safetensors # state of the stateful preprocessor steps
│ ├── policy_postprocessor.json # postprocessor config (when the run has a postprocessor)
│ └── policy_postprocessor_step_*.safetensors # state of the stateful postprocessor steps
└── training_state/
 ├── training_step.json # step counter, topology, and batch semantics
 ├── rng_state.safetensors # rng states
 ├── scheduler_state.json # scheduler state (when the run has a scheduler)
 ├── optimizer_state.safetensors # full optimizer state (non-sharded runs)
 ├── optimizer_param_groups.json # optimizer param groups (non-sharded runs)
 └── optimizer_0/ # DCP optimizer shards (sharded runs)<!----></pre></div><!----> <p>During single-GPU or DDP training, the pipeline serializes each state dict into a single file: <code>model.safetensors</code> for the model and <code>optimizer_state.safetensors</code> for the optimizer.</p> <p>During sharded training, the optimizer state is saved as DCP shards under <code>training_state/optimizer_0/</code>, and the layout of the model under <code>pretrained_model/</code> can be configured through <code>--checkpoint_format</code>:</p> <table><thead><tr><th><code>--checkpoint_format</code></th><th>Weights artifact</th><th>Use when</th></tr></thead><tbody><tr><td><code>safetensors</code> <em>(default)</em></td><td>single <code>model.safetensors</code> only</td><td>you want every checkpoint immediately loadable with <code>from_pretrained</code></td></tr><tr><td><code>dcp</code></td><td><code>pytorch_model_fsdp_0/</code> shard directory only</td><td>gathering the full weights makes saves and resumes too slow</td></tr><tr><td><code>safetensors_dcp</code></td><td>both</td><td>you want fast resume <em>and</em> immediately loadable checkpoints</td></tr></tbody></table> <p>Two things to know about gathered (<code>safetensors</code>) checkpoints from sharded runs:</p> <ul><li><strong>They store fp32 weights.</strong> Under mixed precision training, FSDP keeps an fp32 master copy, and the checkpoint saves the master copy to make sure training resumes consistently.</li> <li>The gather is collective (all ranks participate) but only the main process writes.</li></ul> <!--[2--><h3 class="relative group"><a id="converting-dcp-checkpoints" 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="#converting-dcp-checkpoints"><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>Converting DCP checkpoints</span></h3><!--]--><!----> <p><code>lerobot-convert-dcp</code> merges a DCP shard directory into a regular <code>model.safetensors</code>, offline and without GPUs:</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 "><!---->lerobot-convert-dcp --checkpoint_dir=outputs/train/run/checkpoints/005000
lerobot-convert-dcp --checkpoint_dir=... --delete_dcp=<span class="hljs-literal">true</span> --push_to_hub=<span class="hljs-variable">${HF_USER}</span>/my_policy<!----></pre></div><!----> <p><code>--push_to_hub</code> publishes the converted directory as a model repo.</p> <!--[2--><h3 class="relative group"><a id="resuming" 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="#resuming"><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>Resuming</span></h3><!--]--><!----> <p>Resume with <code>--resume=true --config_path=.../checkpoints/last/pretrained_model/train_config.json</code>. Resuming from a DCP checkpoint supports resharding the model and optimizer state to the <em>current</em> topology, which means you can resume with a different <code>dp_replicate/dp_shard</code> split. The data sampler can always resume at the right epoch and offset, but is only <em>sample-exact</em> when the world size and batch size match the original run (a warning is logged otherwise).</p> <blockquote class="note"><p>FSDP checkpoints written by LeRobot 0.6.x and earlier used a different on-disk layout (a gathered full optimizer state) and <strong>cannot be resumed</strong>.</p></blockquote> <!--[1--><h2 class="relative group"><a id="notes" 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="#notes"><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>Notes</span></h2><!--]--><!----> <ul><li>Checkpoint saves and end-of-training publishes are collective (every rank enters them). Gathered weights, sidecar files and Hub uploads are written by the main process alone.</li> <li>Metrics are reduced across ranks before logging: losses are averaged, and <code>samples/s</code> reports cluster-wide throughput.</li> <li>Learning-rate scheduling is stepped once per training step regardless of the number of processes (<code>step_scheduler_with_optimizer=False</code> is baked in).</li></ul> <p>For background on the underlying machinery, see the <a href="https://huggingface.co/docs/accelerate/usage_guides/fsdp" rel="nofollow">Accelerate FSDP guide</a>. To go deeper on large-scale training, check out the <a href="https://huggingface.co/spaces/nanotron/ultrascale-playbook" rel="nofollow">Ultrascale Playbook</a>.</p> <a class="!text-gray-400 !no-underline text-sm flex items-center not-prose mt-4" href="https://github.com/huggingface/lerobot/blob/main/docs/source/multi_gpu_training.mdx" 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_1ndbhru = {
 base: "/docs/lerobot/main/en",
 assets: "/docs/lerobot/main/en"
 };
 const element = document.currentScript.parentElement;
 Promise.all([
 import("/docs/lerobot/main/en/_app/immutable/entry/start.BigjVlyv.js"),
 import("/docs/lerobot/main/en/_app/immutable/entry/app.wSiz0M_s.js")
 ]).then(([kit, app]) => {
 kit.start(app, element, {
 node_ids: [0, 47],
 data: [null,null],
 form: null,
 error: null
 });
 });
 }
 </script>

Xet Storage Details

Size:
43.7 kB
·
Xet hash:
5fa313fea5a0c8e25ca49ce1623d799c291d3c14a11bc4ebd02b25f4b749436d

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