To deploy an a3-megagpu-8g cluster running Slurm on Google Cloud, please follow these instructions.
Selective deployment and teardown for this blueprint are documented centrally. See examples/machine-learning/README.md for full details.
Example (deploy only the primary group for this blueprint):
./gcluster deploy -d a3mega-slurm-deployment.yaml a3mega-slurm-blueprint.yaml --only primarya3mega-slurm-gcsfuse-lssd-blueprint.yaml reflects best practices for using GCSFuse for ML workloads. It is configured to mount GCS buckets on two mountpoints on a3-mega nodes. Use the gcs_bucket variable to specify a GCS bucket to mount, or leave the variable empty to mount all available buckets dynamically.
The /gcs mountpoint enables parallel downloads, intended for reading/writing checkpoints, logs, application outputs, model serving, or loading large files (e.g. squashfs files). The read-only /gcs-ro mountpoint disables parallel downloads and enables the list cache, intended for reading training data. Parallel downloads are not recommended for training workloads; see GCSFuse documentation for details.