https://blog.wongcw.com/2019/03/06/%e8%b0%b7%e6%ad%8c%e9%96%8b%e6%ba%90gpipe-%e8%a8%93%e7%b7%b4%e6%9b%b4%e5%a4%a7%e6%a8%a1%e5%9e%8b%e3%80%81%e2%80%8b%e2%80%8b%e2%80%8b%e4%b8%8d%e8%aa%bf%e6%95%b4%e8%b6%85%e5%8f%83%e6%93%b4%e5%b1%95/
谷歌開源GPipe 訓練更大模型、​​​不調整超參擴展性能