← 返回看板

🔄 Z-MAX 4090 环境恢复指南

2026-07-12 · 4090即将回收 · 下次快速启动

一、代码仓库

仓库分支地址
lerobot-smolvla-lewmain + webgithub.com/MikeBMW/lerobot-smolvla-lew
zmax-websitemaingithub.com/MikeBMW/zmax-website
Isaac-GR00Tmain (local)NVIDIA/Isaac-GR00T (补丁在 patches/)

二、环境搭建

2.1 Python 3.12

# 安装uv
curl -LsSf https://astral.sh/uv/install.sh | sh
# Python 3.12 路径
~/.local/share/uv/python/cpython-3.12-linux-x86_64-gnu/bin/python3.12
# 装包必须加
--break-system-packages

2.2 HuggingFace 镜像(必须!)

export HF_ENDPOINT=https://hf-mirror.com
export HF_HUB_DISABLE_XET=1  # 禁用Xet协议,走镜像
# 登录(需要token)
huggingface-cli login

2.3 模型权重下载

# SmolVLA base (873MB)
huggingface-cli download --local-dir /root/models/smolvla_base lerobot/smolvla_base
# SmolVLM2-500M (7.4GB)  
huggingface-cli download --local-dir /root/models/smolvlm2-500m HuggingFaceTB/SmolVLM2-500M-Video-Instruct
# GR00T N1.7-3B (6.9GB)
huggingface-cli download --local-dir /root/models/gr00t-n1.7-3b nvidia/GR00T-N1.7-3B
# Cosmos-Reason2-2B (4.6GB, 需HF授权)
huggingface-cli download --local-dir /root/models/cosmos-reason2-2b nvidia/Cosmos-Reason2-2B

2.4 Isaac-GR00T 环境

cd Isaac-GR00T
# 1. 打补丁
git apply zmax-website/patches/isaac-gr00t-4090-fixes.patch
# 2. uv sync
UV_HTTP_TIMEOUT=180 uv sync
# 3. 激活
source .venv/bin/activate

三、Sys2 推理服务

# 启动服务(Python 3.12)
cd /root/lerobot-smolvla-lew
export HF_ENDPOINT=https://hf-mirror.com
export HF_HUB_DISABLE_XET=1
~/.local/share/uv/python/cpython-3.12-linux-x86_64-gnu/bin/python3.12 sys2_prod_server.py
# 验证
curl localhost:50052/health
curl -X POST localhost:50052/predict -H 'Content-Type: application/json' -d '{"engine":"vtla","observation":{}}'

四、GR00T 推理(Mock HF API)

import huggingface_hub
huggingface_hub.model_info = lambda *a,**kw: type('o',(),{
    'sha':'9ce19a1','siblings':[],'id':'Cosmos',
    'tags':['qwen3'],'library_name':'transformers'})()

from gr00t.policy.gr00t_policy import Gr00tPolicy
policy = Gr00tPolicy(
    embodiment_tag="OXE_DROID_RELATIVE_EEF_RELATIVE_JOINT",
    model_path="/root/models/gr00t-n1.7-3b",
    device="cuda:0", strict=False)
# 观察格式: video{exterior_image_1_left,wrist_image_left}+state{eef_9d,joint_position,gripper_position}+language{annotation.language.language_instruction}

五、VLA-T 训练

cd /root/lerobot-smolvla-lew
export HF_ENDPOINT=https://hf-mirror.com
export HF_HUB_DISABLE_XET=1
~/.local/share/uv/python/cpython-3.12-linux-x86_64-gnu/bin/python3.12 zmax_full_pipeline.py

六、ECS 运维

七、补丁清单

文件内容
patches/0001-feat-web-Sys2-*.patchSys2 VLA-T+GR00T引擎+gRPC服务(初始)
patches/0002-Sys2-VLA-T-make_att_2d_masks-*.patch训练mask修复+生产服务+GUI嵌入
patches/isaac-gr00t-4090-fixes.patchdataclass kw_only+pyproject环境限制

八、网页清单

页面URL
主页/index.html
任务看板/kanban.html (cron自动刷新)
算力指标/performance.html
工厂调研/survey.html
项目复盘/retrospective.html
项目验收/acceptance.html
恢复指南/recovery.html (本页)