srunx 2.0.0

Slurm job workflow management

srunx

A unified CLI, web dashboard, and Python API for SLURM job management.

Stop juggling sbatch scripts, squeue loops, and SSH sessions.

• Submit & manage SLURM jobs from CLI, browser, or Python
• Orchestrate multi-step workflows with YAML and dependency graphs
• Monitor GPU availability and job states with Slack notifications
• SSH remote — submit jobs, sync files, and browse remote clusters from your laptop
• Container-native — Pyxis, Apptainer, and Singularity support built in

Installation

Requires Python 3.12+ and access to a SLURM cluster (local or via SSH).

uv add srunx             # with uv (recommended)
pip install srunx        # or with pip

The web dashboard and Slack notifications are included in the base install — no extras required.

For AI agent integration (MCP server), add the mcp extra:

uv add "srunx[mcp]"

Quick Start

Submit a job, wait for it, and view the logs — end to end:

# 1. Submit (use -- to separate srunx flags from the command)
$ srunx sbatch --name training --gpus-per-node 2 --conda ml_env -- python train.py
✅ Submitted job training (id=847291)

# 2. Follow until completion
$ srunx watch jobs 847291
⠋ 847291 training  PENDING  →  RUNNING  →  COMPLETED (4m 12s)

# 3. Inspect output
$ srunx tail 847291 -n 20

Or describe the whole pipeline once and let srunx drive it:

srunx flow run workflow.yaml

Why srunx?

Instead of stitching together sbatch, squeue, SSH, and a pipeline runner, srunx offers one coherent surface that covers the day-to-day SLURM loop.

Capability	srunx	submitit	simple-slurm	Snakemake
CLI for submit / status / cancel	✅	❌	❌	⚠️ partial
Python API	✅	✅	✅	✅
Web dashboard	✅	❌	❌	❌
Workflow DAG with dependencies	✅	❌	❌	✅
Inter-job value passing (load-time)	✅	❌	❌	⚠️ via files
Matrix parameter sweeps	✅	⚠️ manual	❌	⚠️ via wildcards
GPU availability monitoring	✅	❌	❌	❌
SSH remote submit + file sync	✅	❌	❌	❌
Container support (Pyxis / Apptainer / Singularity)	✅	⚠️ limited	❌	⚠️ via rules
Slack notifications	✅	❌	❌	⚠️ plugin

If you need full-featured scientific workflow tooling, Snakemake / Nextflow are still the right call. srunx targets the sweet spot of "SLURM + a few dependencies + a nice UI" without Airflow-scale infrastructure.

CLI

Command	Description
srunx sbatch <script> / srunx sbatch --wrap "<cmd>"	Submit a SLURM job
srunx squeue	List jobs in queue (use -j <id> for a single job's state)
srunx scancel <id>	Cancel a job
srunx tail <id>	View / stream job logs
srunx sinfo	Display GPU availability
srunx watch jobs|resources|cluster	Watch for state changes / resource availability
srunx flow	Run / validate YAML workflows
srunx flow run --arg KEY=VALUE	Override workflow args from the CLI
srunx flow run --sweep KEY=V1,V2 --max-parallel N	Ad-hoc matrix parameter sweep
srunx ssh	Remote SLURM operations over SSH
srunx sacct	Show job execution history
srunx sreport	Generate job execution report
srunx config	Manage configuration
srunx template	Manage job templates
srunx ui	Launch the web dashboard

Web Dashboard

A dashboard for visual cluster management. Connect to your SLURM cluster over SSH and manage jobs, workflows, and resources from a browser.

srunx ui                # -> http://127.0.0.1:8000
srunx ui --port 3000    # custom port

Jobs

Browse, search, filter, and cancel jobs.

Workflow DAG

Visualize job dependencies. Run workflows directly from the UI.

Resources

GPU and node availability per partition.

Explorer

Browse remote files via SSH mounts. Shell scripts can be submitted as sbatch jobs directly from the file tree.

Workflow Orchestration

Define pipelines in YAML. Jobs run as soon as their dependencies complete — independent branches execute in parallel automatically.

name: experiment
args:
  model: "bert-base-uncased"
  output_dir: "/outputs/{{ model }}"

jobs:
  - name: preprocess
    command: ["python", "preprocess.py", "--out", "{{ output_dir }}/data"]
    exports:
      DATA_PATH: "{{ output_dir }}/data/processed.parquet"

  - name: train
    command: ["python", "train.py", "--model", "{{ model }}", "--data", "{{ deps.preprocess.DATA_PATH }}"]
    depends_on: [preprocess]
    gpus_per_node: 2
    environment:
      container:
        image: nvcr.io/nvidia/pytorch:24.01-py3
        mounts:
          - /data:/data
    exports:
      MODEL_PATH: "{{ output_dir }}/models/best.pt"

  - name: evaluate
    command: ["python", "eval.py", "--model", "{{ deps.train.MODEL_PATH }}"]
    depends_on: [train]

What this shows off:

• args with Jinja2 — reusable, parameterized pipelines ({{ model }}, {{ output_dir }})
• Inter-job exports — parents declare exports:; children read them via {{ deps.<parent>.<key> }}, fully resolved at workflow load time (no runtime env files)
• Containers per job — Pyxis / Apptainer / Singularity are first-class (environment.container)
• Dependency-driven scheduling — evaluate blocks on train; parallel branches run automatically

Run it:

srunx flow run workflow.yaml              # execute
srunx flow run workflow.yaml --dry-run    # show plan only
srunx flow run workflow.yaml --from train # resume / partial execution

Retry with retry: N and retry_delay: <seconds> per job.

Parameter Sweeps

Run the same workflow across a matrix of hyperparameters without copying YAML. Each cell materializes into its own sbatch submission and is tracked independently.

name: train
args:
  lr: 0.01
  seed: 1

sweep:
  matrix:
    lr: [0.001, 0.01, 0.1]
    seed: [1, 2, 3]
  fail_fast: false
  max_parallel: 4

jobs:
  - name: train
    command: ["python", "train.py", "--lr", "{{ lr }}", "--seed", "{{ seed }}"]
    gpus_per_node: 1

Run it — or declare the axes ad-hoc on the command line:

srunx flow run train.yaml                                                # YAML-declared sweep
srunx flow run --sweep lr=0.001,0.01 --max-parallel 2 train.yaml          # ad-hoc
srunx flow run --sweep lr=0.001,0.01 --max-parallel 2 --dry-run train.yaml

Sweeps are a first-class concept across CLI, Web UI, and MCP. Web-triggered sweeps route cells through a bounded SlurmSSHExecutorPool against the configured SSH profile, while CLI and MCP runs use the local SLURM client by default. The Web UI surfaces per-cell progress with ETA, filter / sort, and per-cell cancellation.

Monitoring

# Monitor a job until completion
srunx watch jobs 12345

# Wait for GPUs, then submit
srunx watch resources --min-gpus 4
srunx sbatch --wrap "python train.py" --gpus-per-node 4

# Periodic cluster reports to Slack
srunx watch cluster --schedule 1h --notify $SLACK_WEBHOOK

Remote SSH

Keep your local editor workflow while running on the cluster:

# Submit to remote cluster (auto-syncs the script's mount first, then runs in-place)
srunx sbatch train.sh --profile dgx-server

# Manage connection profiles
srunx ssh profile add myserver --ssh-host dgx1

# Map local directories to remote and sync with rsync
srunx ssh profile mount add myserver workspace \
  --local ~/projects/ml-exp --remote /home/user/ml-exp
srunx ssh sync

• SSH config hosts, saved profiles, and proxy jump support
• Environment variable passthrough (--env KEY=VALUE, --env-local WANDB_API_KEY)
• File sync via rsync — auto-detects profile from current directory

Slack Notifications

Get notified when jobs finish — set SLACK_WEBHOOK_URL (or configure it in the web dashboard), then append --slack to any srunx flow run command. In Python, pass SlackCallback to the runner (see the Python API section below).

export SLACK_WEBHOOK_URL="https://hooks.slack.com/services/..."
srunx flow run workflow.yaml --slack

MCP Server

srunx ships an MCP server so Claude Code (and other MCP clients) can submit jobs, inspect the queue, and drive workflows over stdio. Install the extra and register the server with your client:

uv add "srunx[mcp]"
srunx-mcp                                                              # launch the stdio server directly

# Or register with Claude Code in one shot
claude mcp add --scope user srunx -- uvx --from 'srunx[mcp]' srunx-mcp

Once connected, the agent can call run_workflow with optional sweep and mount parameters:

run_workflow(
    yaml_path="train.yaml",
    sweep={"matrix": {"lr": [0.001, 0.01]}, "max_parallel": 2},
    mount="my-project",
)

Passing mount=<name> routes the run through the matching SSH profile mount, translating work_dir / log_dir into remote paths — so the agent can launch mount-aware submissions against a remote cluster without leaving the chat.

Python API

The full CLI surface is available as a Python library. Use it inside notebooks, existing Python pipelines, or custom tooling.

Submit and wait:

from srunx import Job, JobResource, JobEnvironment, Slurm

job = Job(
    name="training",
    command=["python", "train.py"],
    resources=JobResource(nodes=1, gpus_per_node=2, time_limit="4:00:00"),
    environment=JobEnvironment(conda="ml_env"),
)

client = Slurm()
completed = client.run(job)  # submit, poll, and return when terminal
print(completed.status, completed.job_id)

Fire-and-track:

submitted = client.submit(job)                 # returns Job with job_id populated
info = client.retrieve(submitted.job_id)       # poll status on demand
client.cancel(submitted.job_id)                # if you change your mind

Run a YAML workflow programmatically, with callbacks:

from srunx.callbacks import SlackCallback
from srunx.runner import WorkflowRunner

runner = WorkflowRunner.from_yaml(
    "workflow.yaml",
    callbacks=[SlackCallback(webhook_url="...")],
)
runner.run()                                    # blocks until the DAG finishes

Documentation

Full documentation at ksterx.github.io/srunx.

Development

git clone https://github.com/ksterx/srunx.git
cd srunx
uv sync --dev

# Full pre-commit quality gate
uv run pytest && uv run mypy . && uv run ruff check .

Contributions welcome — please open an issue or PR on GitHub.

License

Apache-2.0