byted-technologies/github_
byted-technologies/github_ has 772 GitHub stars and 0 forks. GitHot measured 0 stars in 24 hours and -3 in 7 days. Last checked 2026-11-21 18:21 UTC. Last push 2026-11-15.
Crecimiento de estrellas medido
- in 24h
- 0
- in 7 days
- -3
- in 15 days
- Recopilando historial
- in 1 month
- Recopilando historial
Cambios netos en períodos que se solapan, no totales diarios. Los datos no disponibles se dejan en blanco. Las estrellas indican interés, no calidad.
Acerca de
Official Fractal Compute Orchestrations ecosystem hub (v3.1 non-commercial source-available): FractalWorkspace, FractalCore (v1.0.0), and FractalAndroid.
Herramientas de desarrolloREADME.md
Original
F R A C T A L
C O M P U T E O R C H E S T R A T I O N S
Decentralized Edge Compute Harvesting, Sub-Byte INT4 Federated Learning, and Split LLM Inference.
FractalWorkspace | FractalCore (v1.0.0) | FractalAndroid | Colab Quickstart
Executive Overview
Fractal Compute Orchestrations engineers a distributed edge intelligence stack designed to train and serve modern neural architectures across fragmented consumer mobile hardware while protecting battery longevity, SoC thermals, and user data sovereignty.
Across its coordinated runtime modules, FRACTAL addresses the memory and bandwidth bottlenecks of on-device AI:
- Outlier-Resilient Micro-Block INT4 Transport (
fractal_quant/FRQ4&FRQD): Confines high-kurtosis attention and projection outliers to 64-element blocks with error-feedback residual tracking, cutting TinyLlama-1.1B wire payloads by 7.1x (4,196 MB -> 590 MB) while preserving 0.9878 mean layer cosine similarity (where naive per-tensor INT4 collapses to0.8296). - Hardened Federated Orchestration & Byzantine Aggregation (
fractal_server): Enforces multi-tenant path sandboxing, constant-time bearer/token auth, FedAvg (weights/.frqd), Coordinate Median, Krum Byzantine filtering, and FedProx proximal stabilization. - Thermal-Gated Android Edge Runtime (
FractalAndroid): Executes on-device TFLite training and wire-compatibleInt4BlockCodec(FRQ4/FRQD) egress strictly when SoC thermals (<= 40.0 Cand non-severe thermal status), battery state-of-charge, charging policy, and Wi-Fi constraints pass. - Memory-Bounded Pipeline Parallel Slicing (
inference-model-maker): Isolates decoder partitions under a150 MBedge RAM envelope and compresses inter-node activations into INT8 frames bounded to4,100bytes.
Core Ecosystem Repositories
| Repository | Role | Key Capabilities & Stack |
|---|---|---|
FractalWorkspace |
Umbrella Architecture & Benchmarks | Unified multi-repo submodule workspace, reproducible TinyLlama/MobileNetV3 Pareto frontier evaluation charts, and system contracts. |
FractalCore |
Python SDK, CLI & Coordinator (v1.0.0) |
Installable fractal_quant SDK + fractal doctor/bench CLI, fractal_server federated coordinator, and ExecuTorch model slicer (pytest + Ruff CI). |
FractalAndroid |
Android Edge Client Runtime | Kotlin MVVM client, Int4BlockCodec bit-identical to Python fractal_quant, OperationControl thermal gating, and FractalPaths sandboxed storage. |
Quickstart Verification (CPU)
git clone https://github.com/Fractal-Compute-Orchestrations/FractalCore.git
cd FractalCore
pip install -e ".[dev]"
fractal doctor
fractal bench --layers 4 --hidden-size 256
python scripts/quickstart_demo.py
pytest -v
Empirical System Contracts
| Contract | Subsystem | Quantitative Guarantee |
|---|---|---|
| Micro-Block INT4 Wire Reduction | fractal_quant (FRQ4 / FRQD) & Android Int4BlockCodec |
7.11x wire reduction, >= 0.9878 mean cosine fidelity across outliers |
| Activation Frame Ceiling | Inter-Node Activation Streaming | <= 4,100 bytes INT8 quantized activation frame |
| Edge Partition Envelope | inference-model-maker Slicer |
<= 150 MB partition size for zero LMK memory eviction |
| SoC & Battery Telemetry Gate | Android OperationControl |
Battery temp <= 40.0 C, non-severe thermal status, configurable SoC & Wi-Fi gates |
| Tenant Storage Sandboxing | fractal_server & FractalPaths |
Canonical path containment + X-Auth-Token / Authorization: Bearer enforcement |
Governance & Commercial Licensing
FRACTAL and all companion repositories are maintained and owned by Fractal Compute Orchestrations.
- License: Proprietary Source-Available under the FRACTAL Proprietary Source-Available & Non-Commercial Restrictive License v3.1.
- Commercial Restriction: Free for non-monetized academic research, personal study, and non-commercial evaluation. Commercial deployment, startup/company derivatives, hosted SaaS/PaaS APIs, and revenue generation are strictly prohibited without a separately executed written Commercial License from Fractal Compute Orchestrations.
- Citation & Notice: See CITATION.cff and NOTICE.
- Contact: Fractal Compute Orchestrations (
[email protected])
Explorar proyectos relacionados
Cargando…