Fractal-Compute-Orchestrations/FractalWorkspace

What it solves

Fractal addresses the computational and memory limits of running large machine learning models on a single device. It lets organizations use a fleet of Android phones as a distributed compute cluster, performing model training and inference without sending raw user data to a central server. It also protects devices by pausing work if battery or temperature thresholds are exceeded.

How it works

It has two main parts: a Python-based central server (FractalCore) and an Android app (FractalAndroid). For training, the server splits a dataset into segments and dispatches them to phones. Each phone trains on its local data using TFLite and sends back only the model weight updates. The server averages these updates (FedAvg) to improve the global model. For inference of large models (e.g., Llama 3 8B), the server slices the model into partitions no larger than 150MB. Each phone memory-maps its partition and passes intermediate activations to the next phone over a local Wi-Fi connection. Strict contracts enforce size limits and require a battery level above 50%, active charging, and temperature below 40°C before any work starts.

Who it's for

Developers and organizations that want to leverage underutilized mobile devices for machine learning tasks, especially when preserving user privacy and device health is critical. It is also for Android users who are willing to contribute compute in exchange for token rewards, as the system includes a token-based reward mechanism.

Highlights

• Federated learning with zero data ingress – raw datasets never leave the devices.
• Pipeline parallelism for large model inference across multiple phones.
• Hardware-aware operation: pauses work when thermals or battery are at risk.
• Strict contracts ensure partitions and network payloads stay within limits.
• Security features include stateless token authentication and replay protection.