stilt.ExecutionConfig

class stilt.ExecutionConfig(*, backend: Literal['local', 'slurm']='local', n_workers: Annotated[int, Ge(ge=1)] = 1, cpus: Annotated[int, Ge(ge=1)] = 1, time: str | int | None = None, mem: str | None = None, partition: str | None = None, account: str | None = None, qos: str | None = None, timeout: Annotated[int | None, Ge(ge=1)] = None, keep_workdir: bool = False, array_parallelism: Annotated[int | None, Ge(ge=1)] = None, setup: list[str] = <factory>, slurm: dict[str, typing.Any]=<factory>) → None

Where a project’s receptors run, and with what resources.

With backend: local the receptors run on this machine as one task. With backend: slurm they are split among n_workers tasks of one Slurm job array. Either way a task runs cpus receptors at once. The Slurm settings are ignored by a local run, so a config can hold them and be run either way. timeout and keep_workdir apply to both. None of these settings change a result.

Parameters

Parameter	Description	Default
backend	Where to run: on this machine (local) or as a Slurm job array (slurm).	'local'
n_workers	Number of Slurm array tasks, which the receptors are split evenly among. A local run is one task.	1
cpus	Receptors each task runs at once: processes on this machine for a local run, CPUs per array task on Slurm.	1
time	Slurm time limit per task, as HH:MM:SS, D-HH:MM:SS, or minutes. Required with backend: slurm: without it a task gets the partition’s default, which can be days.	None
mem	Memory per Slurm task, such as 8G.	None
partition	Slurm partition.	None
account	Slurm account.	None
qos	Slurm quality of service.	None
timeout	Time limit for one transport model run, in seconds. A run that exceeds it is stopped and recorded as a failed simulation, and the worker moves on. Unset waits indefinitely, so a hung run can hold a worker until its job ends.	None
keep_workdir	Keep every run’s workdir, as the transport model left it, under scratch/ in the output directory. A failed run’s is always kept.	False
array_parallelism	Most Slurm tasks running at once. Unset leaves it to Slurm.	None
setup	Shell commands run in each Slurm task before the worker, such as module load lines. The worker itself runs with the Python that submitted it.	[]
slurm	Any other sbatch options by name, such as exclude: node1 or constraint: skl. true writes the bare flag.	{}