cue-ml 1.0.0

A streamlined workload manager designed for Deep Learning research, optimizing GPU usage for individuals and small teams.

cue

Lightweight GPU job scheduler for deep learning researchers.

Stop waiting for GPUs. Stop writing bash scripts. Define your experiments in YAML and let cue handle the rest.

Why cue?

You have 4 GPUs and 50 experiments to run. Cue automatically queues your jobs, assigns available GPUs, and logs everything—no daemons, no databases, just Python.

Install

pip install cue-ml

Requires Python 3.10+ and Linux with NVIDIA drivers.

Quick Start

1. Create experiments.yaml:

config:
  log_dir: "logs/my_project"
  python_cmd: "python3"

experiments:
  - name: "bert-finetune"
    script: "train.py"
    runs:
      - args: "--model bert-base --lr 1e-4"
      - args: "--model bert-large --lr 2e-5"
        gpus: 2  # Request 2 GPUs for this run

  - name: "vision-test"
    script: "evaluate.py"
    runs:
      - args: "--dataset cifar10"

2. Run:

cue -p experiments.yaml

Cue launches a terminal dashboard showing job status, GPU usage, and queue depth. Each run's output is automatically logged.

Features

• Auto-GPU assignment – Detects idle GPUs and schedules jobs
• Multi-GPU support – Request 1, 2, 4, or 8 GPUs per run
• Live dashboard – Monitor everything in your terminal
• Smart logging – Organized stdout/stderr for every experiment
• No infrastructure – Just Python, no complex setup
• Simulation mode – Test configs without GPUs

Configuration

Key	Type	Description
config.log_dir	string	Output directory for logs (default: ./logs)
config.python_cmd	string	Python interpreter (default: system Python)
experiments	list	Experiment definitions
experiments[].script	string	Path to Python script
experiments[].runs	list	Configurations to execute
experiments[].runs[].args	string	Command-line arguments
experiments[].runs[].gpus	int	GPUs required (default: 1)

CLI Options

cue -p experiments.yaml [options]

Option	Description
-p, --path	Path to config file (required)
--gpus N	Default GPUs per job if not specified
--fail-fast	Stop queue if any job fails
--dry-run	Validate config without running
--simulate N	Test with N fake GPUs
--debug	Verbose error output

Testing Without GPUs

cue -p experiments.yaml --simulate 8

Simulates an 8-GPU machine for testing your pipeline.

Contributing

Contributions welcome! Fork, branch, commit, push, and open a PR.

License

MIT License – see LICENSE for details.