Created by: QiJune

We create global step variable in CPU, and than add a cast operator.

def _decay_step_counter():
    # the first global step is zero in learning rate decay
    global_step = nn.autoincreased_step_counter(
        counter_name='@LR_DECAY_COUNTER@', begin=0, step=1)
    global_step = tensor.cast(global_step, 'float32')
    return global_step

If we run the program in GPU, cast operator will have a CUDADeviceContext, and set the output of cast operator in GPU. It will cause segment fault.