mugithi / google-terraform-pytorch-tpu Goto Github PK

During provisioning, the exists the need to clone code to the NFS share. This should be done from only one of the instances in the managed instance group

During training, the NFS share is used to store your code, accessible to all the instances. When you run the `_BUILD=destroy' command, then NFS share is destroy.

In order to protect the user from inadvertently deleting their work, there should be an automatic copy of all the content in the the NFS share to a persistant GCS bucket

Code sample

if [ -z $GCE_IMAGE_VERSION && $XLA_DOCKER_CONTAINER == 'true' ]
then 
    docker pull docker pull gcr.io/tpu-pytorch/xla:nightly_$GCE_IMAGE_VERSION
fi 

Complete example on how to kick off training on Roberta

The instance group not not getting created with _BUILD=create and _BUILD=destroy