slurmomatic 0.1.0

custom decorator to turn any function slurm-deployable. Can submit multiple jobs and multiple job_arrays

Slurmomatic

A lightweight Python decorator to conditionally submit functions as SLURM jobs (or job arrays), falling back to local execution when SLURM is not available.

🚀 Key Features

• 📦 Drop-in simple: Decorate any function with @slurmify(...).
• 🔍 Auto-detects SLURM: Will submit jobs via SLURM if available, otherwise runs locally.
• ⚙️ Unified interface: Same code works on your laptop or cluster — no changes needed.
• 🧠 Smart job control: Supports both individual job submission and SLURM job arrays.

🔧 Requirements

• Python 3.10+
• submitit

---

🧠 Usage

Step 1: Import

from slurmomatic import slurmify, batch

Step 2: Decorate your function

Each decorated function must accept a use_slurm: bool argument.

---

✅ Example 1: Submitting a SLURM Job Array

from slurmomatic import slurmify

@slurmify(slurm_array_parallelism=True, timeout_min=20)
def train(a: int, b: int, use_slurm: bool = False):
    print(f"Training with a={a}, b={b}")

# Run job array of 5 parallel job_arrays
train([1, 2, 3, 4, 5], [10, 20, 30, 40, 50], use_slurm=True)

---

✅ Example 2: Submitting Multiple Individual Jobs

from slurmomatic import slurmify

@slurmify(timeout_min=10)
def run_experiment(seed: int, use_slurm: bool = False):
    print(f"Running experiment with seed={seed}")

for seed in range(5):
    run_experiment(seed, use_slurm=True)

Each call submits its own SLURM job (or runs locally).

---

✅ Example 3: Submitting Multiple Batches with Job Arrays

from slurmomatic import slurmify, batch

@slurmify(slurm_array_parallelism=10, timeout_min=30)
def evaluate(x: int, y: int, use_slurm: bool = False):
    print(f"Evaluating with x={x}, y={y}")
    # Prepare large input lists

xs = list(range(1000))
ys = [1] * 1000

# Submit in batches of 200 using job arrays
for x_batch, y_batch in batch(200, xs, ys):
    evaluate(x_batch, y_batch, use_slurm=True)

This submits 5 SLURM job arrays, each with 200 jobs.

---

📦 @slurmify(...) Parameters

You can pass any SLURM submitit parameters directly to the decorator:

@slurmify(timeout_min=30, cpus_per_task=4, gpus_per_node=1, partition="gpu")

Special key:

slurm_array_parallelism=10 → Triggers job array mode.

---

🧰 batch(batch_size: int, *args)

Utility to chunk long input lists into mini-batches.

from slurmomatic import batch

for a_batch, b_batch in batch(100, list_a, list_b):
    train(a_batch, b_batch, use_slurm=True)

---

🛡️ Notes

✅ If SLURM is not available (sinfo not found or no job ID in environment), the jobs run locally using submitit.LocalExecutor.

📜 License

MIT