部署配方 · Hugging Face (LeRobot)

SmolVLA

与我们实测时执行的命令、版本号一字不差。把 <hardware> 换成显卡 id（rtx-4090 / rtx-5090 / a800-80g）。通用步骤（镜像、磁盘、nohup）见配方总览。

1 · 环境

transformers 随 lerobot · torch >= 2.7（独立 venv） · lerobot[smolvla]

python -m venv /root/envs/lerobot && /root/envs/lerobot/bin/pip install -U pip
/root/envs/lerobot/bin/pip install "lerobot[smolvla]"

2 · 权重

lerobot/smolvla_base · 1 GB
HuggingFaceTB/SmolVLM2-500M-Video-Instruct · 1 GB · VLM 底座，策略加载时会再拉一次

export HF_ENDPOINT=https://hf-mirror.com HF_HOME=/root/autodl-tmp/hf HF_HUB_DISABLE_XET=1
python -c "from huggingface_hub import snapshot_download as s; s('lerobot/smolvla_base')"
python -c "from huggingface_hub import snapshot_download as s; s('HuggingFaceTB/SmolVLM2-500M-Video-Instruct')"

{T.run}

{r.run}

{r['load_' + L] &&

{T.load} · {r['load_' + L]}

}

{T.pit}

{(r.pitfalls || []).map((p: any) => ())}

{T.sym}	{T.fix}
{p[L]}	{p['fix_' + L]}

{T.meas}

{T.src}

{rows.length ?

{rows.map((x) => ())}

{T.hwc}	{T.prec}	{T.p50}	{T.p95}	{T.vram}	{T.n}	{T.date}
{x.h.name}	{x.prec}	{fmtMs(x.p50)}	{fmtMs(x.p95)}	{x.vram != null ? x.vram + ' GB' : '—'}	{x.n}	{x.date}

{T.nomeas}

}