Can not work with train-val network in caffe??? cuda10.1

I build caffe-windows on win10, with cuda10.1 cudnn 7.5, RTX2080

I can use this build to train networks.

However, I can only train networks with the solver.prototxt only contain imformation for training, below is my solver.prototxt:
net : “newshuff-train.prototxt”

The base learning rate, momentum and the weight decay of the network.

#test_iter: 10
#test_interval: 1000
base_lr: 0.001
momentum: 0.9
weight_decay: 0.002

The learning rate policy

lr_policy: “step”
stepsize: 500
gamma: 0.8
display: 500
max_iter: 500000
snapshot: 1000
snapshot_prefix: “./models-eye/”
solver_mode: GPU
if i uncomm the test_iter and the test interval, i will got this error:
cudnn_conv_layer.cpp:53] Check failed: status == CUDNN_STATUS_SUCCESS (1 vs. 0) CUDNN_STATUS_NOT_INITIALIZED

how can i solve this problem? I just move to a new lab and the machine here use RTX2080, before in my previous lab they use 1080Ti and can train with both training and testing network.

can any body help, thanks!