labflow

Opinionated scientific workflow manager — reproducible experiments and transparent SLURM submission on heterogeneous HPC clusters. One experiment = one Python file + a dataclass config. Submitit + Jinja2 under the hood; Hydra config composition is an optional extra.

What it solves

1. NMSAT duplication pain. One experiment = one Python file + dataclass config. No more parameters/name.py + computations/name.py coupled by string matching.
2. Transparent HPC submission. labflow submit ou_stationary --system cluster-arm --nodes 4 renders the right Jinja2 template, respects your cluster's rules (GPU billing, parallel scratch, SSH keys, OMP_PROC_BIND=TRUE), and submits.
3. Env backend freedom. Conda, uv, poetry, venv, pixi, apptainer — each project can pick what fits. Backends implement a single resolver interface.
4. Provenance by default. Every run writes a manifest.json with git SHA, env hash, seed, hardware, wall time, container hash.

For whom

Computational-science labs running scientific Python on local workstations + HPC clusters, especially where experiments need sweeps, provenance, and portable submission.

Stack it's built for

• Simulation: NEST/NESTML, Arbor, Brian2, JAX + diffrax + optax, JAXley, NEURON, bmtk
• Inference/HPO: SBI + BayesFlow, Optuna, Ray Tune, Ax
• Analysis: numpy/scipy/pandas/sklearn, marimo notebooks
• HPC: SLURM via Submitit; Apptainer/Singularity containers; any SLURM cluster

Next steps

• Install →
• First experiment →