For further information on deploying an A3 Ultra cluster with Slurm, please see:
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 a3ultra-slurm-deployment.yaml a3ultra-slurm-blueprint.yaml --only primaryThis blueprint includes four Cloud Storage FUSE mounts to provide a simple and scalable way to manage data.
/gcsis a general purpose mount that can be used for shared tools and data./gcs-checkpointsis an optimized mount for writing and reading checkpoints. It uses the local SSD for caching and enables parallel downloads to improve performance./gcs-training-datais an optimized mount for reading training data. It can be further tuned if the training data fits fully within the local ssd storage./gcs-model-servingis an optimized mount for serving models, which downloads model weights in parallel to local ssd.
To test workloads directly on A3 Ultra VMs, you can deploy the a3ultra-vm.yaml:
- A configurable number of A3 Ultra VMs (default N=2)
- RDMA networking and GPU drivers pre-configured on our Ubuntu 22.04 Accelerator Image
- Additional software environment customization can be achieved by adding to the example startup-script
The VMs can be consumed from a reservation by modifying the reservation_name parameter in the a3ultra-vms module.
Cluster toolkit also supports DWS Flex-Start, Spot VMs, as well as reservations as ways to provision instances.
For more information on DWS Flex-Start in Slurm For more information on Spot VMs
We provide ways to enable the alternative provisioning models in the
a3ultra-slurm-deployment.yaml file.
To make use of these other models, replace a3u_reservation_name in the
deployment file with the variable of choice below.
a3u_enable_spot_vm: true for spot or a3u_dws_flex_enabled: true for DWS Flex-Start.