Error all cuda capable devices are busy or unavailable on remote server

I need to use synchronized BatchNorm so I downloaded the project from and runned their to test if everything works fine. The error “all cuda_capable devices are busy or unavailable” appears as soon as the training procedure starts, i’m not sure what the error is depending on. I’m using a remote server and can not reset GPUs by myself for the moment, they are pretty empty. Here are some information about my environment:
Ubuntu: 16.04
cuda: 8.0
Pytorch: 0.4.1
gcc: 5.5
gpu compute mode: default