HPC - Leonardo

Sample configuration: CINECA Leonardo

.env.leonardo
# SSH / cluster login node access
SLURM_EDGE_NODE_HOST=login01-ext.leonardo.cineca.it  # valid login nodes are login01-ext, login02-ext, login05-ext, login07-ext
SLURM_EDGE_NODE_PORT=22
SLURM_EDGE_NODE_USER=your_leonardo_username
# SLURM_EDGE_NODE_PASSWORD=password            # specify either password or ssh key path
SLURM_EDGE_NODE_KEY_PATH=/Users/you/.ssh/id_ed25519_vsc5

# Deployment settings
SLURM_DEPLOYMENT_BASE_PATH=/leonardo/home/userexternal/your_leonardo_username/dagster-slurm
SLURM_PARTITION=boost_usr_prod
# SLURM_QOS=boost_qos_bprod                   # Queue for large jobs (65 to 256 nodes). Leave out for normal queue (for up to 64 nodes).
# SLURM_QOS=boost_qos_dbg                     # Development queue with higher priority (max. 2 nodes, max. 30 minutes)
# SLURM_RESERVATION=reservation_name          # optional reservation (if active)
SLURM_SUPERCOMPUTER_SITE=leonardo

# Dagster deployment selector
DAGSTER_DEPLOYMENT=production_supercomputer

Leonardo requires you to have an active project allocation; the permitted partitions depend on your account type. Verify your entitlements on the cluster:

sacctmgr show assoc where user=$USER format=Cluster,Account%20,Partition,QOS%60
sinfo -s

If the output does not list batch, pick one of the partitions above that appears in both commands. GPU queues also need the matching QoS (e.g. dcgpuqos). Training accounts typically use /leonardo/home/usertrain/<user>/… for storage—adjust SLURM_DEPLOYMENT_BASE_PATH accordingly. If your site enforces OTP/Kerberos, configure a local ~/.ssh/config entry or a bastion jump host—dagster-slurm will reuse that setup automatically.

With the variables defined, restart your Dagster code location. To dry-run against the real scheduler while still allowing on-the-fly environment packaging, point DAGSTER_DEPLOYMENT=staging_supercomputer and run:

pixi run start-staging-supercomputer

For production workloads you should publish the environment bundle ahead of time (e.g. via CI using python scripts/deploy_environment.py). Once you export the uploaded path as CI_DEPLOYED_ENVIRONMENT_PATH, switch to DAGSTER_DEPLOYMENT=production_supercomputer and start Dagster:

pixi run start-production-supercomputer

The production preset refuses to start if CI_DEPLOYED_ENVIRONMENT_PATH is missing, ensuring clusters never build environments during business-critical runs.

For additional information about Leonardo, see the Leonardo documentation.

Sample configuration: CINECA Leonardo​

Sample configuration: CINECA Leonardo