slurmster 0.2.26

Minimal Slurm experiment runner with persistent SSH, YAML configs, and log streaming

slurmster

A minimal Python tool to run parameter-grid experiments on a Slurm cluster with persistent SSH, log streaming, and simple YAML configs.

Install

pip install slurmster

Features

• CLI with subcommands: submit, monitor, status, fetch, cancel, gui
• YAML config (explicitly provided via --config)
• Persistent SSH connection for low latency
• Per-run working directories on the remote side
• Automatic log redirection to stdout.log inside each run directory
• Live log streaming (and re-attach later)
• Local workspace to track runs and "fetched" state
• Cancel jobs from local machine
• Web-based GUI for easy management

CLI Usage

All commands follow this pattern:

slurmster --config <config.yaml> --user <username> --host <hostname> [options] <command>

Basic Commands

Submit experiments:

slurmster --config config.yaml --user myuser --host myhost submit

Monitor logs:

# Monitor by job ID:
slurmster --config config.yaml --user myuser --host myhost monitor --job 1234567

Check status:

slurmster --config config.yaml --user myuser --host myhost status

Fetch completed runs:

# Fetch all completed runs:
slurmster --config config.yaml --user myuser --host myhost fetch
# Or fetch a specific job:
slurmster --config config.yaml --user myuser --host myhost fetch --job 1234567

Cancel jobs:

# Cancel specific job:
slurmster --config config.yaml --user myuser --host myhost cancel --job 1234567
# or cancel all:
slurmster --config config.yaml --user myuser --host myhost cancel --all

Additional Options

• --password-env ENV_VAR: Use password from environment variable
• --key /path/to/key: Use SSH key file instead of password
• --port 22: Specify SSH port (default: 22)

For submit:

• --no-monitor: Don't automatically start monitoring after submission

For monitor:

• --from-start: Stream from beginning instead of last 100 lines
• --lines N: Number of trailing lines when attaching (default: 100)

For status:

• --all: Show all runs (default: only non-fetched)

For fetch:

• --job <job_id>: Only fetch a specific job by ID

Configuration File

Create a YAML config file (see example/config.yaml):

remote:
  base_dir: ~/experiments            # remote working root

files:
  push:
    - example/train.py               # any code/data files you need on remote
  fetch:
    - "model.pth"                   # optional; if omitted we fetch the entire run dir
    - "log.txt"

slurm:
  directives: |                      # SBATCH lines; placeholders allowed
    #SBATCH --job-name={base_dir}
    #SBATCH --partition=gpu
    #SBATCH --time=00:10:00
    #SBATCH --cpus-per-gpu=40
    #SBATCH --nodes=1
    #SBATCH --gres=gpu:1
    #SBATCH --mem=32G

run:
  command: |                         # your run command; placeholders allowed
    source venv/bin/activate
    python example/train.py --lr {lr} --epochs {epochs} --save_model "{run_dir}/model.pth" --log_file "{run_dir}/log.txt"

  # ONE of the following:
  grid:
    lr: [0.1, 0.01, 0.001]
    epochs: [1, 2, 5, 10]
  # experiments:
  #   - { lr: 0.1, epochs: 1 }
  #   - { lr: 0.001, epochs: 10 }

Placeholders

• {base_dir}: resolved remote base directory (e.g. /home/you/experiments)
• Any run parameter placeholder, e.g. {lr}, {epochs}
• {remote_dir}: the configured remote.base_dir
• {run_dir}: the per-run directory (under remote.base_dir/runs/{exp_name})

Local workspace

Under the .slurmster directory next to your config.yaml (<config-dir>/.slurmster/<user>@<host>/<sanitized-remote-base>), we store:

• runs.json — run registry (job id, exp name, fetched flag, etc.)
• results/<exp_name>_<job_id>/... — fetched run directories

GUI Usage

For a more user-friendly experience, you can use the web-based GUI:

slurmster --config config.yaml --user myuser --host myhost gui

Additional GUI options:

• --gui-port 8000: Set the HTTP port (default: 8000)
• --gui-bind 0.0.0.0: Set the bind interface (default: 0.0.0.0)
• --no-browser: Don't automatically open browser

The GUI provides:

Configuration Management:

• View and edit your current configuration
• See resolved placeholders and SLURM directives
• Modify files to push/fetch and run commands

Job Submission:

• Submit single jobs with custom parameters
• Submit grid jobs with parameter combinations
• Real-time parameter validation

Job Monitoring:

• View all jobs with their current status
• Monitor and browse job outputs in real-time
• Access job logs directly in the browser

Bulk Operations:

• Fetch all completed jobs at once
• Cancel multiple jobs
• Track job progress and completion status

The GUI automatically opens in your browser at http://localhost:8000 (or your specified port) and provides an intuitive interface for all slurmster functionality.

License

MIT — see LICENSE.