colonyml 0.1.1

Zero-config distributed ML training for CPU clusters

ColonyML

Zero-config distributed ML training for CPU clusters.

Pool any laptops into a training cluster — no IP addresses, no config files, no GPU required. Just:

pip install colonyml
colonyml join

---

The Problem

Training ML models on one laptop is slow. You might have multiple machines sitting idle. Existing distributed training tools require manual IP configuration, Linux, or expensive GPUs.

ColonyML fixes this. Any laptop. Any OS. Zero setup.

---

How It Works

Laptop 1 → colonyml join   (announces itself on network)
Laptop 2 → colonyml join   (finds Laptop 1 automatically)

Both laptops train together:
- Auto-discovery via mDNS (like AirDrop for ML)
- Ring AllReduce for gradient synchronization
- Adaptive batch sizing based on CPU power
- Gradient compression over WiFi (up to 50x)

---

Features

• Zero config — no IP addresses, no config files
• Auto-discovery — machines find each other via mDNS
• Ring AllReduce — efficient gradient averaging across nodes
• Adaptive scheduling — faster machines get bigger batches
• Gradient compression — up to 50x smaller over WiFi
• Any hardware — works on any CPU laptop, any OS

---

Installation

pip install colonyml

---

Status

Active Development — v0.1.0 is the initial release. Core features being built now. Star the repo to follow progress.

---

Roadmap

• v0.1.0 — PyPI release, project structure
• v0.2.0 — mDNS auto-discovery between machines
• v0.3.0 — gradient compression
• v0.4.0 — adaptive CPU-aware batch scheduler
• v0.5.0 — Ring AllReduce gradient synchronization
• v1.0.0 — stable release with full test coverage

---

Author

Jeevana Sai Gogineni MS Data Science · University of Maryland · GPA 4.0

GitHub: https://github.com/gjeevana27 LinkedIn: https://linkedin.com/in/jeevana-gogineni

---

License

MIT License — free to use, modify, and distribute.