slurm-sdk 0.4.5

Pythonic SDK for Slurm.

Python Slurm SDK

A developer-friendly SDK to define, submit, and manage Slurm jobs in Python with native array job support, fluent dependency APIs, and container packaging.

Quick Start

Simple Task with Dependencies

from slurm import Cluster, task

@task(time="00:01:00", mem="1G")
def preprocess(dataset: str) -> str:
    return f"processed_{dataset}"

@task(time="00:05:00", mem="4G", cpus_per_task=4)
def train_model(data: str, lr: float) -> dict:
    return {"model": f"trained_on_{data}", "lr": lr, "accuracy": 0.95}

@task(time="00:01:00", mem="1G")
def evaluate(model: dict) -> str:
    return f"Model accuracy: {model['accuracy']}"

if __name__ == "__main__":
    # Local backend for offline development
    with Cluster(backend_type="local") as cluster:
        # Fluent dependency API with .after()
        prep_job = preprocess("dataset.csv")
        train_job = train_model.after(prep_job)(data=prep_job, lr=0.001)
        eval_job = evaluate.after(train_job)(model=train_job)

        eval_job.wait()
        print(eval_job.get_result())

Workflow Orchestration

Workflows run on the cluster and orchestrate multiple tasks, enabling complex pipelines that survive preemption and scheduling delays:

from slurm import Cluster, task, workflow
from slurm.workflow import WorkflowContext

@task(time="00:05:00", mem="2G")
def train_epoch(epoch: int, data_path: str) -> dict:
    return {"epoch": epoch, "loss": 1.0 / (epoch + 1), "checkpoint": f"ckpt_{epoch}.pt"}

@task(time="00:02:00", mem="1G")
def evaluate(checkpoint: str) -> float:
    return 0.85 + (int(checkpoint.split("_")[1].split(".")[0]) * 0.02)

@workflow(time="01:00:00", mem="512M")
def training_pipeline(epochs: int, data_path: str, ctx: WorkflowContext) -> dict:
    """Multi-epoch training with evaluation after each epoch."""
    results = []

    for epoch in range(epochs):
        # Submit training job
        train_job = train_epoch(epoch=epoch, data_path=data_path)
        train_result = train_job.get_result()

        # Submit eval after training completes
        eval_job = evaluate.after(train_job)(checkpoint=train_result["checkpoint"])
        accuracy = eval_job.get_result()

        results.append({"epoch": epoch, "loss": train_result["loss"], "accuracy": accuracy})

    return {"results": results, "best_accuracy": max(r["accuracy"] for r in results)}

if __name__ == "__main__":
    cluster = Cluster(
        backend_type="ssh",
        hostname="login.hpc.example.com",
        username="myuser",
    )

    with cluster:
        # The workflow itself runs as a job on the cluster
        job = training_pipeline(epochs=3, data_path="/data/train")
        job.wait()
        print(job.get_result())

Array Jobs for Parallel Processing

from slurm import Cluster, task

@task(time="00:02:00", mem="2G", cpus_per_task=2)
def process_chunk(chunk_id: int, start: int, end: int) -> dict:
    return {"chunk": chunk_id, "sum": sum(range(start, end))}

@task(time="00:01:00", mem="1G")
def aggregate(results: list) -> int:
    return sum(r["sum"] for r in results)

if __name__ == "__main__":
    with Cluster(backend_type="local") as cluster:
        # Map task over items to create native SLURM array job
        chunks = [
            {"chunk_id": i, "start": i * 1000, "end": (i + 1) * 1000}
            for i in range(10)
        ]
        array_job = process_chunk.map(chunks)

        # Aggregate results from array job
        final = aggregate.after(array_job)(results=array_job.get_results())
        final.wait()
        print(f"Total: {final.get_result()}")

SSH Backend with Container Packaging

from slurm import Cluster, task

@task(
    time="00:10:00",
    mem="8G",
    packaging="container",
    packaging_platform="linux/amd64",
    packaging_push=True,
    packaging_registry="myregistry.io/myproject/",
)
def compute_intensive_task(n: int) -> float:
    import numpy as np
    return np.mean(np.random.random(n))

if __name__ == "__main__":
    cluster = Cluster(
        backend_type="ssh",
        hostname="login.hpc.example.com",
        username="myuser",
    )

    with cluster:
        job = compute_intensive_task(1_000_000)
        job.wait()
        print(job.get_result())

Documentation

• Contributing Guide - Development setup, testing, and contribution guidelines
• AGENTS.md - Guide for AI agents working with this codebase
• Changelog - Release history and changes
• examples/ - Complete working examples including parallelization patterns

Running on an ARM Mac (Apple Silicon)

You can build and publish x86_64 (``linux/amd64```) images from an ARM Mac by enabling QEMU emulation inside your container runtime. The example below shows the required one-time setup for Podman; Docker Desktop users can follow a similar flow using docker buildx (no additional configuration usually needed).

1. Install QEMU locally. This provides the binary emulators that Podman will load into its virtual machine:

   brew install qemu
   

2. Make sure the Podman machine is created and running. If you have not initialised it yet:

   podman machine init --now
   

3. Register the QEMU static binaries inside the Podman machine so it can run amd64 containers. This step reboots the VM and only needs to be performed once (repeat if you recreate the machine):

   podman machine ssh sudo rpm-ostree install qemu-user-static
   podman machine ssh sudo systemctl reboot
   

4. After the machine restarts, confirm cross-building works:

   podman build --platform linux/amd64 -t test-amd64 .
   podman run --rm --platform linux/amd64 test-amd64 uname -m
   

   The final command should print x86_64 even though you are on an ARM host.

With QEMU registered you can now build the example container in this project using the provided Slurmfile (platform = "linux/amd64") and push it to your registry before submitting jobs to an x86 cluster.