Fork自 PaddlePaddle / Paddle
* remove tensor copy in the update_loss_scaling op * not use thrust. * fix some cuda memory access error.
拖放文件到此处或点击上传
