Fork自 PaddlePaddle / Paddle
restrict block num of layer_norm_grad cuda kernel to 128, test=develop
拖放文件到此处或点击上传