Skip to content

lote

One command plane for your machines.

A DGX over ssh, a slurm cluster, a pbs supercomputer, the spare box under the desk. Each runs jobs its own way, so the same experiment needs pueue here, sbatch there, and qsub somewhere else. lote is the plane that sits over all of them. You pick a machine, it rsyncs your repo there, runs your job under whatever scheduler that host has, and pulls the results back. One job script runs anywhere.

  • One submit

    lote submit <host> job.sh ships the repo, launches the job, and prints a handle. No matter the scheduler.

  • Any scheduler

    lote detects each host's queue and adapts. pueue on a plain box, sbatch on slurm, qsub on pbs.

  • State across hosts

    lote ps and lote reconcile track every run across every machine from your laptop.

  • Just ssh

    Targets are your ~/.ssh/config host aliases. No agent to install, no control daemon to babysit.

Installation

pip install lote            # the lote command, ready to use
chefe add lote --pypi       # or pull it into a chefe project

lote dispatches its on-host executor over ssh as chefe run lote exec ..., so each host needs chefe (the sibling env tool), which lote setup installs and which brings up pixi on first run.

lote.toml
[sync]
include = ["research", "toolbox", "pixi.toml"]   # what a compute node needs
exclude = ["**/.venv", "**/__pycache__", "**/*.ckpt"]   # heavy, regenerable

lote setup miyabi                 # onboard a host: probe, sync, build the env
lote submit miyabi train.sh       # ship the repo and launch, prints a handle
lote ps                           # recent runs across every machine
lote pull <handle>                # rsync the results back

lote is early (0.0.x)

The config format and commands may still change.

Next