Buckets:

hf-doc-build
/

doc-dev

Files

xet

hf-doc-build/doc-dev / lerobot /pr_3545 /en /hardware_guide.html

HuggingFaceDocBuilder

2 months ago

download

raw

28.1 kB

	<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{"title":"Compute HW Guide for LeRobot Training","local":"compute-hw-guide-for-lerobot-training","sections":[{"title":"Memory by policy group","local":"memory-by-policy-group","sections":[],"depth":2},{"title":"Training time","local":"training-time","sections":[{"title":"Common scenarios","local":"common-scenarios","sections":[],"depth":3},{"title":"Multi-GPU matters a lot","local":"multi-gpu-matters-a-lot","sections":[],"depth":3},{"title":"Schedule and checkpoints","local":"schedule-and-checkpoints","sections":[],"depth":3}],"depth":2},{"title":"Where to run","local":"where-to-run","sections":[{"title":"Hugging Face Jobs","local":"hugging-face-jobs","sections":[],"depth":3}],"depth":2}],"depth":1}">
	<link href="/docs/lerobot/pr_3545/en/_app/immutable/assets/0.e3b0c442.css" rel="modulepreload">
	<link rel="modulepreload" href="/docs/lerobot/pr_3545/en/_app/immutable/entry/start.e180c007.js">
	<link rel="modulepreload" href="/docs/lerobot/pr_3545/en/_app/immutable/chunks/scheduler.eb244325.js">
	<link rel="modulepreload" href="/docs/lerobot/pr_3545/en/_app/immutable/chunks/singletons.def024d3.js">
	<link rel="modulepreload" href="/docs/lerobot/pr_3545/en/_app/immutable/chunks/index.3c23fb4b.js">
	<link rel="modulepreload" href="/docs/lerobot/pr_3545/en/_app/immutable/chunks/paths.ce94f7d3.js">
	<link rel="modulepreload" href="/docs/lerobot/pr_3545/en/_app/immutable/entry/app.d3999ecd.js">
	<link rel="modulepreload" href="/docs/lerobot/pr_3545/en/_app/immutable/chunks/preload-helper.d0246088.js">
	<link rel="modulepreload" href="/docs/lerobot/pr_3545/en/_app/immutable/chunks/index.3fe63ad3.js">
	<link rel="modulepreload" href="/docs/lerobot/pr_3545/en/_app/immutable/nodes/0.8137c5fa.js">
	<link rel="modulepreload" href="/docs/lerobot/pr_3545/en/_app/immutable/chunks/each.e59479a4.js">
	<link rel="modulepreload" href="/docs/lerobot/pr_3545/en/_app/immutable/nodes/22.f8d95854.js">
	<link rel="modulepreload" href="/docs/lerobot/pr_3545/en/_app/immutable/chunks/CopyLLMTxtMenu.520da618.js">
	<link rel="modulepreload" href="/docs/lerobot/pr_3545/en/_app/immutable/chunks/MermaidChart.svelte_svelte_type_style_lang.593b9ec0.js">
	<link rel="modulepreload" href="/docs/lerobot/pr_3545/en/_app/immutable/chunks/CodeBlock.4284ca46.js"><!-- HEAD_svelte-u9bgzb_START --><meta name="hf:doc:metadata" content="{"title":"Compute HW Guide for LeRobot Training","local":"compute-hw-guide-for-lerobot-training","sections":[{"title":"Memory by policy group","local":"memory-by-policy-group","sections":[],"depth":2},{"title":"Training time","local":"training-time","sections":[{"title":"Common scenarios","local":"common-scenarios","sections":[],"depth":3},{"title":"Multi-GPU matters a lot","local":"multi-gpu-matters-a-lot","sections":[],"depth":3},{"title":"Schedule and checkpoints","local":"schedule-and-checkpoints","sections":[],"depth":3}],"depth":2},{"title":"Where to run","local":"where-to-run","sections":[{"title":"Hugging Face Jobs","local":"hugging-face-jobs","sections":[],"depth":3}],"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 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> </div> <h1 class="relative group"><a id="compute-hw-guide-for-lerobot-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="#compute-hw-guide-for-lerobot-training"><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>Compute HW Guide for LeRobot Training</span></h1> <p data-svelte-h="svelte-n84mhz">Rough sizing for training a LeRobot policy: how much VRAM each policy needs, what training time looks like, and where to run when local hardware isn’t enough.</p> <p data-svelte-h="svelte-tsnza0">The numbers below are <strong>indicative</strong> — order-of-magnitude figures for picking hardware, not exact predictions. Throughput depends heavily on dataset I/O, image resolution, batch size, and number of GPUs.</p> <h2 class="relative group"><a id="memory-by-policy-group" 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="#memory-by-policy-group"><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>Memory by policy group</span></h2> <p data-svelte-h="svelte-1kf932r">Policies cluster by backbone size; the groupings below give a single VRAM envelope per group instead of repeating numbers per policy. Memory scales roughly linearly with batch size; AdamW (the LeRobot default) carries optimizer state that adds ~30–100% over a forward+backward pass alone.</p> <table data-svelte-h="svelte-1o9058m"><thead><tr><th>Group</th> <th>Policies</th> <th align="right">Peak VRAM (BS 8, AdamW)</th> <th>Suitable starter GPUs</th></tr></thead> <tbody><tr><td>Light BC</td> <td><code>act</code>, <code>vqbet</code>, <code>tdmpc</code></td> <td align="right">~2–6GB</td> <td>Laptop GPU (RTX 3060), L4, A10G</td></tr> <tr><td>Diffusion</td> <td><code>diffusion</code>, <code>multi_task_dit</code></td> <td align="right">~8–14GB</td> <td>RTX 4070+ / L4 / A10G</td></tr> <tr><td>Small VLA</td> <td><code>smolvla</code></td> <td align="right">~10–16GB</td> <td>RTX 4080+ / L4 / A10G</td></tr> <tr><td>Large VLA</td> <td><code>pi0</code>, <code>pi0_fast</code>, <code>pi05</code>, <code>xvla</code>, <code>wall_x</code></td> <td align="right">~24–40GB</td> <td>A100 40 GB+ (24 GB tight at BS 1)</td></tr> <tr><td>Multimodal</td> <td><code>groot</code>, <code>eo1</code></td> <td align="right">~24–40GB</td> <td>A100 40 GB+</td></tr> <tr><td>RL</td> <td><code>sac</code></td> <td align="right">config-dep.</td> <td>See <a href="./hilserl">HIL-SERL guide</a></td></tr></tbody></table> <p data-svelte-h="svelte-18wcriq">Memory-bound? Drop the batch size (~linear), use gradient accumulation to recover effective batch, or for SmolVLA leave <code>freeze_vision_encoder=True</code>.</p> <h2 class="relative group"><a id="training-time" 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="#training-time"><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>Training time</span></h2> <p data-svelte-h="svelte-yk9vuk">Robotics imitation learning typically converges in <strong>5–10 epochs over the dataset</strong>, not hundreds of thousands of raw steps. Once you know your epoch count, wall-clock is essentially:</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="language-text "><!-- HTML_TAG_START -->total_frames = sum of frames over all episodes # 50 ep × 30 fps × 30 s ≈ 45,000
	steps_per_epoch = ceil(total_frames / (num_gpus × batch_size))
	total_steps = epochs × steps_per_epoch
	wall_clock ≈ total_steps × per_step_time<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1if0129">Per-step time depends on the policy and the GPU. The numbers in the table below are anchors — pick the row closest to your setup and scale linearly with <code>total_steps</code> if you train longer or shorter.</p> <h3 class="relative group"><a id="common-scenarios" 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="#common-scenarios"><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>Common scenarios</span></h3> <p data-svelte-h="svelte-1hsppy8">Indicative wall-clock for <strong>5 epochs on a ~50-episode dataset (~45k frames at 30 fps × 30 s)</strong>, default optimizer (AdamW), 640×480 images:</p> <table data-svelte-h="svelte-13l1lpi"><thead><tr><th>Setup</th> <th>Policy</th> <th>Batch</th> <th align="right">Wall-clock</th></tr></thead> <tbody><tr><td>Single RTX 4090 / RTX 3090 (24 GB)</td> <td><code>act</code></td> <td>8</td> <td align="right">~30–60min</td></tr> <tr><td>Single RTX 4090 / RTX 3090 (24 GB)</td> <td><code>diffusion</code></td> <td>8</td> <td align="right">~2–4h</td></tr> <tr><td>Single L4 / A10G (24 GB)</td> <td><code>act</code></td> <td>8</td> <td align="right">~1–2h</td></tr> <tr><td>Single L4 / A10G (24 GB)</td> <td><code>smolvla</code></td> <td>4</td> <td align="right">~3–6h</td></tr> <tr><td>Single A100 40 GB</td> <td><code>smolvla</code></td> <td>16</td> <td align="right">~1–2h</td></tr> <tr><td>Single A100 40 GB</td> <td><code>pi0</code> / <code>pi05</code></td> <td>4</td> <td align="right">~4–8h</td></tr> <tr><td>4× H100 80 GB cluster (<code>accelerate</code>)</td> <td><code>diffusion</code></td> <td>32</td> <td align="right">~30–60min</td></tr> <tr><td>4× H100 80 GB cluster (<code>accelerate</code>)</td> <td><code>smolvla</code></td> <td>32</td> <td align="right">~1–2h</td></tr> <tr><td>Apple Silicon M1/M2/M3 Max (MPS)</td> <td><code>act</code></td> <td>4</td> <td align="right">~6–14h</td></tr></tbody></table> <p data-svelte-h="svelte-k2eisr">These are order-of-magnitude figures. Real runs deviate by ±50% depending on image resolution, dataset I/O, dataloader threading, and exact GPU SKU. They are useful as “is this run going to take an hour or a day?” intuition, not as SLAs.</p> <h3 class="relative group"><a id="multi-gpu-matters-a-lot" 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-matters-a-lot"><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>Multi-GPU matters a lot</span></h3> <p data-svelte-h="svelte-1jjto0k"><code>accelerate launch --num_processes=N</code> is the easiest way to cut training time. Each optimizer step processes <code>N × batch_size</code> samples in roughly the same wall-clock as a single-GPU step, so 4 GPUs ≈ 4× speedup for compute-bound runs. See the <a href="./multi_gpu_training">Multi GPU training</a> guide for the full setup.</p> <p data-svelte-h="svelte-1kb6e0q">Reference data points on a 4×H100 80 GB cluster (<code>accelerate launch --num_processes=4</code>), 5000 steps, batch 32, AdamW, dataset <a href="https://huggingface.co/datasets/imstevenpmwork/super_poulain_draft" rel="nofollow"><code>imstevenpmwork/super_poulain_draft</code></a> (~50 episodes, ~640×480 images):</p> <table data-svelte-h="svelte-1g5geu3"><thead><tr><th>Policy</th> <th>Wall-clock</th> <th align="right"><code>update_s</code></th> <th align="right"><code>dataloading_s</code></th> <th>GPU util</th> <th>Notable flags</th></tr></thead> <tbody><tr><td><code>diffusion</code></td> <td>16m 17s</td> <td align="right">0.167</td> <td align="right">0.015</td> <td>~90%</td> <td>defaults (training from scratch)</td></tr> <tr><td><code>smolvla</code></td> <td>27m 49s</td> <td align="right">0.312</td> <td align="right">0.011</td> <td>~80%</td> <td><code>--policy.path=lerobot/smolvla_base</code>, <code>freeze_vision_encoder=false</code>, <code>train_expert_only=false</code></td></tr> <tr><td><code>pi05</code></td> <td>3h 41m</td> <td align="right">2.548</td> <td align="right">0.014</td> <td>~95%</td> <td><code>--policy.pretrained_path=lerobot/pi05_base</code>, <code>gradient_checkpointing=true</code>, <code>dtype=bfloat16</code>, vision encoder + expert trained</td></tr></tbody></table> <p data-svelte-h="svelte-kkfxpc">The <code>dataloading_s</code> vs. <code>update_s</code> ratio is the diagnostic that matters: when <code>dataloading_s</code> approaches <code>update_s</code>, more GPUs stop helping — your dataloader is the bottleneck and you should look at <code>--num_workers</code>, image resolution, and disk speed before adding compute.</p> <h3 class="relative group"><a id="schedule-and-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="#schedule-and-checkpoints"><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 and checkpoints</span></h3> <p data-svelte-h="svelte-7vfyut">If you shorten training (e.g. 5k–10k steps on a small dataset), also shorten the LR schedule with <code>--policy.scheduler_decay_steps≈--steps</code>. Otherwise the LR stays near its peak and never decays. Same for <code>--save_freq</code>.</p> <h2 class="relative group"><a id="where-to-run" 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="#where-to-run"><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>Where to run</span></h2> <p data-svelte-h="svelte-61jttz">VRAM is the first filter. Within a tier, pick by budget and availability — the <code>$</code>–<code>$$$$</code> columns are relative; check current pricing on the provider you actually use.</p> <table data-svelte-h="svelte-2oothf"><thead><tr><th>Class</th> <th>VRAM</th> <th>Tier</th> <th>Comfortable for</th></tr></thead> <tbody><tr><td>RTX 3090 / 4090 (consumer)</td> <td>24 GB</td> <td><code>$</code></td> <td>Light BC, Diffusion, SmolVLA. Tight for VLAs at batch 1.</td></tr> <tr><td>L4 / A10G (cloud)</td> <td>24 GB</td> <td><code>$–$$</code></td> <td>Same envelope; common on Google Cloud, RunPod, AWS <code>g5/g6</code>.</td></tr> <tr><td>A100 40 GB</td> <td>40 GB</td> <td><code>$$$</code></td> <td>Any policy at reasonable batch sizes.</td></tr> <tr><td>A100 80 GB / H100 80 GB</td> <td>80 GB</td> <td><code>$$$$</code></td> <td>Multi-GPU clusters; large batches for VLAs.</td></tr> <tr><td><strong>CPU only</strong></td> <td>—</td> <td>—</td> <td>Don’t train. Use Colab or rent a GPU.</td></tr></tbody></table> <h3 class="relative group"><a id="hugging-face-jobs" 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="#hugging-face-jobs"><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>Hugging Face Jobs</span></h3> <p data-svelte-h="svelte-aorm1f"><a href="https://huggingface.co/docs/hub/jobs" rel="nofollow">Hugging Face Jobs</a> lets you run training on managed HF infrastructure, billed by the second. The repo publishes a ready-to-use image: <strong><code>huggingface/lerobot-gpu:latest</code></strong>, rebuilt <strong>every night at 02:00 UTC from <code>main</code></strong> (<a href="https://github.com/huggingface/lerobot/blob/main/.github/workflows/docker_publish.yml" rel="nofollow"><code>docker_publish.yml</code></a>) — so it tracks the current state of the repo, not a tagged release.</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="language-bash "><!-- HTML_TAG_START -->hf <span class="hljs-built_in">jobs</span> run --flavor a10g-large huggingface/lerobot-gpu:latest \
	bash -c <span class="hljs-string">"nvidia-smi && lerobot-train \
	--policy.type=act --dataset.repo_id=<USER>/<DATASET> \
	--policy.repo_id=<USER>/act_<task> --batch_size=8 --steps=50000"</span><!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1biq3pv">Notes:</p> <ul data-svelte-h="svelte-1xtsfkk"><li>The leading <code>nvidia-smi</code> is a quick sanity check that CUDA is visible inside the container — useful to fail fast if the flavor or driver mismatched.</li> <li>The default Job timeout is 30 minutes; pass <code>--timeout 4h</code> (or longer) for real training.</li> <li><code>--flavor</code> maps onto the table above: <code>t4-small</code>/<code>t4-medium</code> (T4, ACT only), <code>l4x1</code>/<code>l4x4</code> (L4 24 GB), <code>a10g-small/large/largex2/largex4</code> (A10G 24 GB scaled out), <code>a100-large</code> (A100). For the current full catalogue + pricing see <a href="https://huggingface.co/docs/hub/jobs" rel="nofollow">https://huggingface.co/docs/hub/jobs</a>.</li></ul> <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/hardware_guide.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 data-svelte-h="svelte-zjs2n5"><span class="underline">Update</span> on GitHub</span></a> <p></p>

	<script>
	{
	__sveltekit_1s8w7mm = {
	assets: "/docs/lerobot/pr_3545/en",
	base: "/docs/lerobot/pr_3545/en",
	env: {}
	};

	const element = document.currentScript.parentElement;

	const data = [null,null];

	Promise.all([
	import("/docs/lerobot/pr_3545/en/_app/immutable/entry/start.e180c007.js"),
	import("/docs/lerobot/pr_3545/en/_app/immutable/entry/app.d3999ecd.js")
	]).then(([kit, app]) => {
	kit.start(app, element, {
	node_ids: [0, 22],
	data,
	form: null,
	error: null
	});
	});
	}
	</script>

Xet Storage Details

Size:: 28.1 kB
Xet hash:: a50e37aa275481f7caaf41e79a65e54e69a0a69d1138f5b29a9c74bf407964bc

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