nemo_gym.orchestration.executors.slurm_status
nemo_gym.orchestration.executors.slurm_status
How the Slurm executor looks at the jobs it queued.
One script per host asks Slurm about every job directory of every record at once: a single squeue for
the whole batch, a single sacct over every job id any of them has had, and the state files each job
directory holds. The cost grows with the number of hosts, not with the number of submissions. Jobs are
found by where they run, not by the recorded id, so a job started by hand, a resumed one, or one
resubmitted from another machine is still seen.
Module Contents
Functions
Data
API
The exit code of sacct’s code:signal pair.
The contents render_read_files_script printed, None for a file it did not find.
What render_status_script reported, as one ObservedStatus per job in refs.
The commands that print each readable file of paths, encoded so a file cannot break the output.
The commands that report the state of every job in refs, on one connection.
Slurm’s own state name as a JobState; a name this does not know is UNKNOWN.