新实例(第1天) → 跑实验 → 第13天触发脚本
  → 备份数据到文件存储
  → API 创建新实例（同配置）
  → 从文件存储恢复数据
  → API 释放旧实例
  → 新实例接替，倒计时重置为14天

# 存为环境变量
export AUTODL_TOKEN="你的token"

#!/usr/bin/env python3
"""AutoDL 14天自动续命 — 重建实例并恢复环境"""
import requests, os, time, json

TOKEN = os.environ["AUTODL_TOKEN"]
HOST = "https://api.autodl.com"
HEADERS = {"Authorization": TOKEN, "Content-Type": "application/json"}

# ======== 按你的实例改这里 ========
REGION = "beijingDC2"          # 地区
GPU_SPEC = "pro6000-p"         # GPU规格ID（见API文档附录）
IMAGE_ID = "image-xxxxxxxxx"   # 镜像UUID（私有镜像列表里查）
INSTANCE_NAME = "auto-renew"
DATA_CENTER = ["beijingDC2"]   # 可多选
CUDA_MIN = 113                 # CUDA >= 11.3
GPU_COUNT = 1
DISK_EXPAND = 0               # 系统盘扩容(GB)
START_CMD = "sleep 1"          # 开机后自动执行的命令
# =================================


def _post(path, data=None):
    r = requests.post(f"{HOST}{path}", headers=HEADERS, json=data or {})
    return r.json()


def get_old_instance():
    """查当前实例列表，取最新一个"""
    r = _post("/api/v1/dev/instance/pro/list")
    items = r.get("data", {}).get("list", [])
    if not items:
        raise Exception("没找到实例，先去控制台手动创建一个")
    return items[0]


def release_instance(uuid):
    """先关机，再释放"""
    _post("/api/v1/dev/instance/pro/power_off", {"instance_uuid": uuid})
    time.sleep(10)
    _post("/api/v1/dev/instance/pro/release", {"instance_uuid": uuid})
    print(f"  已释放 {uuid}")


def create_instance():
    """用预设配置创建新实例"""
    data = {
        "data_center_list": DATA_CENTER,
        "req_gpu_amount": GPU_COUNT,
        "expand_system_disk_by_gb": DISK_EXPAND,
        "gpu_spec_uuid": GPU_SPEC,
        "image_uuid": IMAGE_ID,
        "cuda_v_from": CUDA_MIN,
        "instance_name": INSTANCE_NAME,
        "start_command": START_CMD,
    }
    r = _post("/api/v1/dev/instance/pro/create", data)
    info = r.get("data", {})
    uuid = info.get("instance_uuid")
    print(f"  新实例 {uuid} 创建成功")
    return info


def wait_until_running(uuid, timeout=300):
    """等实例开机就绪"""
    for _ in range(timeout // 10):
        r = _post("/api/v1/dev/instance/pro/status", {"instance_uuid": uuid})
        status = r.get("data", {}).get("status")
        if status == "running":
            return True
        time.sleep(10)
    raise Exception("实例启动超时")


def main():
    print("=== AutoDL 14天续命 ===")

    print("[1/4] 查旧实例...")
    old = get_old_instance()
    old_uuid = old["instance_uuid"]
    print(f"  旧实例: {old_uuid}")

    print("[2/4] 创建新实例...")
    new_info = create_instance()
    new_uuid = new_info["instance_uuid"]

    print("[3/4] 等新实例就绪...")
    wait_until_running(new_uuid)
    # 新实例开机后可通过 START_CMD 自动挂载文件存储、拉代码
    # 例如: START_CMD = "bash /root/setup.sh"

    print("[4/4] 释放旧实例...")
    release_instance(old_uuid)

    print(f"=== 完成，新实例 {new_uuid}，倒计时重置 ===")


if __name__ == "__main__":
    main()

#!/bin/bash
# setup.sh — 新实例首次启动时执行

# 挂载文件存储（提前在控制台创建好）
# 挂载后数据目录直接可用

# 装依赖（镜像已有 conda/pytorch 就跳过）
pip install transformers datasets accelerate -q

# 拉最新代码
cd /root && git clone git@github.com:you/project.git

# AutoDL 文件存储已自动挂载到 /root/autodl-fs
# 把训练输出、checkpoint、数据集放这里
ln -s /root/autodl-fs/checkpoints /root/project/checkpoints
ln -s /root/autodl-fs/datasets /root/project/data

# 编辑 crontab
crontab -e

# 每月 1 号和 14 号凌晨 3 点执行
0 3 1,14 * * AUTODL_TOKEN=xxx python3 /home/you/autodl-renew.py >> /home/you/autodl-renew.log 2>&1

# .github/workflows/autodl-renew.yml
name: AutoDL 14天续命
on:
  schedule:
    - cron: "0 3 1,14 * *"
  workflow_dispatch:
jobs:
  renew:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - run: pip install requests && python3 autodl-renew.py
        env:
          AUTODL_TOKEN: ${{ secrets.AUTODL_TOKEN }}