Fork自 PaddlePaddle / Paddle
体验新版 GitCode,发现更多精彩内容 >>
restrict block num of layer_norm_grad cuda kernel to 128, test=develop