Some errors of tensorflow calling multiple GPUs

Two servers GPU, run tensorflow edition bert I want two pieces GPU They all use . After reading other tutorials, I said there are two methods , stay windows server 2016 Tested on , With GPU-Z Tool execution GPU Observe , because windows server The task manager of is a castrated version, can't see GPU performance .

Method 1 ：

import os

os.environ["CUDA_VISIBLE_DEVICES"] = "1"# Specify the graphics card 1 function

### If you want to use multiple graphics cards, you can do it as follows , But not tested , My code is in estimator Written , It should be late GPU Deal with it by yourself GPU Resources can be multiple GPU parallel , I will run the computer with the first graphics card you set up in the following way ###

os.environ['CUDA_VISIBLE_DEVICES'] = '1,0',

 

Method 2 ：

config = tf.ConfigProto(device_count={'GPU': 0})

###GPU:0 Said the use of CPU Don't use GPU,GPU:1 Said the use of GPU###