midas-parsl-configs

Bundled + user-extensible Parsl configs for MIDAS pipelines.

This package owns the cluster Parsl configs that used to live under FF_HEDM/workflows/*Config.py and NF_HEDM/workflows/*Config.py, plus a registry + a SLURM/PBS submit-script-to-config generator so users don't have to hand-edit Python to bring a new cluster online.

What it gives you

from midas_parsl_configs import load_config, available_configs

available_configs()
# {'local': 'builtin', 'umich': 'builtin', 'polaris': 'builtin',
#  'mycluster': 'user:/Users/me/.midas/parsl_configs/myclusterConfig.py', ...}

cfg = load_config("polaris", n_nodes=4, n_cpus=64,
                  script_dir="/path/to/run")
parsl.load(config=cfg)

load_config sets the env vars the bundled configs read at import time (MIDAS_SCRIPT_DIR, nNodes, PROJECT_NAME, QUEUE_NAME, CONDA_LOC) before importing the module so you get a runnable Config back.

Bundled configs

Bringing your own cluster online

midas-parsl-configs generate /path/to/your_submit.sh --name mycluster

This parses #SBATCH / #PBS directives + the body's module load / conda activate / source lines, then writes ~/.midas/parsl_configs/myclusterConfig.py. From then on midas-parsl-configs list and any pipeline driver that resolves through the registry (midas-ff-pipeline, in particular) can use --machine mycluster.

Override the user dir with MIDAS_PARSL_CONFIGS_DIR if you want the file written somewhere else.

User configs override bundled configs of the same name — useful when you're iterating on a cluster's setup without losing the upstream default.

CLI

midas-parsl-configs list                                    # show every visible config + source
midas-parsl-configs generate <submit.sh> --name <short>     # write user config from submit script
midas-parsl-configs generate <submit.sh> --name x --print   # dry run, prints to stdout
midas-parsl-configs show <name>                             # repr() of the resolved Config
midas-parsl-configs path                                    # path to the user config dir

What generate translates

SLURM (#SBATCH):

• --partition / -p → partition
• --account / -A → scheduler_options (#SBATCH --account=…)
• --time / -t → walltime
• --nodes / -N → nodes_per_block
• --ntasks-per-node → cores_per_worker
• --cpus-per-task / -c → cores_per_worker
• --gres → scheduler_options + GPU count detection
• --constraint / --qos / --mem → passed through as scheduler_options
• --job-name / -J → executor label

PBS (#PBS):

• -q → queue
• -A → account
• -l walltime=… → walltime
• -l select=N:ncpus=M:ngpus=K → nodes_per_block + cores_per_worker + GPU count
• -l filesystems=… → scheduler_options
• -N → executor label

Anything we can't translate becomes a passthrough scheduler_options line so you can refine the generated module by hand if needed.