单机多卡训练报错Tensor holds no memory. operator < lookup_table_v2 > error
Created by: xiaom5248
版本信息
Paddle version: 1.7.2
Paddle With CUDA: True
OS: CentOS 6.3
Python version: 2.7.14
CUDA version: 10.0.130
cuDNN version: None.None.None
Nvidia driver version: 418.39
单机单卡训练正常,单机多卡训练报错,报错信息如下:
Python错误栈信息:
网络结构:
训练代码:
参数PARALLEL为False时可正常训练,为True时报上述错误