FractalWorkspace
A unified ecosystem and workspace for large-scale federated learning of Large Models & LLMS. Store compute as Wifi
What is FractalWorkspace?
A unified ecosystem and workspace for large-scale federated learning of Large Models & LLMS. Store compute as Wifi
Overview of FractalWorkspace
What it is
FractalWorkspace is a unified ecosystem and workspace for large-scale federated learning of Large Models and LLMs. It is a distributed compute orchestration and edge intelligence framework that bridges centralized model orchestration with decentralized edge intelligence. The framework treats edge devices as autonomous, resource-constrained nodes operating under human-centric constraints, aiming to harvest fragmented consumer compute across mobile edge devices while preserving user privacy and device health.
Key features
- 7.11x wire payload reduction (85.94% wire cut) for TinyLlama-1.1B, dropping parameter payload from 4,196 MB to 590 MB.
- Sub-60s transmission guarantee, slashing 10 Mbps upload duration from 321.2 seconds to 46.2 seconds, beating the 120s federated aggregation deadline.
- Zero accuracy collapse with 0.9878 mean cosine similarity preserved across all 22 layers, matching lossless 16-bit convergence.
- Zero mobile LMK crashes through native memory-mapped execution (mmap) and 85.6% reduction in radio transmission time.
- Two unified subsystems: FractalCore (Python control plane for multi-tenant isolation, dataset allocation, ATen graph compilation, memory-bounded slicing, federated weight aggregation, compute budgeting, and tokenized reward distribution) and FractalAndroid (native Kotlin execution engine for hardware-gated on-device training and memory-mapped inference).
Use cases
FractalWorkspace is designed for large-scale federated learning of Large Models and LLMs across consumer mobile edge devices. It addresses the uplink bottleneck, dynamic range collapse in low-bit quantization, and mobile memory wall (LMK) issues. Use cases include distributed training and inference of foundation models like TinyLlama, Llama-3, and Mistral on edge devices, with strict privacy and device health preservation.
Getting started
See the repository README for installation and quick-start instructions.
License and pricing
The license is not stated. The project is categorized as open source.
Based on the project's public README and repository metadata.