cloud_build.md

Using Cloud Build to Deploy GPT-2 from Google Compute Engine

If you are using Google Compute Engine to train a GPT-2 model, it's more economical to build it in the cloud instead of downloading it, building it, then reuploading it.

This workflow moves the file from Google Compute Engine to Google Cloud Storage, then uses Cloud Builder to build the container and upload it to Cloud Registry.

From Google Compute Engine to Google Cloud Storage

First, create a bucket in Google Cloud Storage to save your model. Then give full scope permissions to your Google Compute Engine VM (you'll need to Stop and Edit it if it isn't already).

In the GCE VM, if you had to change the scope, you'll need to remove the cached gsutil:

rm -rf ~/.gsutil

Then you can copy the checkpoint folder to a GCS bucket of your choice.

gsutil -m cp -r checkpoint gs://<BUCKET>

Upload your app.py and Dockerfile to the same GCS bucket.

From Google Cloud Storage to Cloud Builder

The cloudbuild.yaml file will use Google Cloud Builder to build the container by copying the files from the GCS bucket, building them, then pushing to the Container Registry. On your local computer, replace the _BUCKET with your GCS bucket and _IMAGE with the destination name, then run:

gcloud builds submit --no-source --config=cloudbuild.yaml

The container should then appear in the Container Registry under the _IMAGE name!