https://github.com/deskiziarecords/tungsten-alpha
Production JAX framework eliminating dynamic shape recompilation. 7-layer Adelic Riemannian stack: NRRB (threshold bridges), CSM (HLO hardening), ATC (adaptive JIT), ARSM (pure state), ASTC (telemetry), AFRH (FIM/nat-grad), AFRC (cluster transport). 100% JIT-stable, scales 10k+ Orbax→Triton ready. Tungsten Alpha OS
https://github.com/deskiziarecords/tungsten-alpha
custom-jax-primitives fisher-information-matrix hlo-fusion jax jit-compilation natural-gradient orbax riemannian-optimization sensor-swarms tpu triton-inference xla
Last synced: 2 months ago
JSON representation
Production JAX framework eliminating dynamic shape recompilation. 7-layer Adelic Riemannian stack: NRRB (threshold bridges), CSM (HLO hardening), ATC (adaptive JIT), ARSM (pure state), ASTC (telemetry), AFRH (FIM/nat-grad), AFRC (cluster transport). 100% JIT-stable, scales 10k+ Orbax→Triton ready. Tungsten Alpha OS
- Host: GitHub
- URL: https://github.com/deskiziarecords/tungsten-alpha
- Owner: deskiziarecords
- License: apache-2.0
- Created: 2026-03-11T16:18:38.000Z (4 months ago)
- Default Branch: main
- Last Pushed: 2026-03-11T17:52:10.000Z (4 months ago)
- Last Synced: 2026-03-11T21:52:17.316Z (4 months ago)
- Topics: custom-jax-primitives, fisher-information-matrix, hlo-fusion, jax, jit-compilation, natural-gradient, orbax, riemannian-optimization, sensor-swarms, tpu, triton-inference, xla
- Language: Python
- Homepage:
- Size: 63.5 KB
- Stars: 0
- Watchers: 0
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
- Zenodo: .zenodo.json
Awesome Lists containing this project
README
# TUNGSTEN ALPHA
## Adelic-Riemannian Operating System for Decentralized XLA
Tungsten Alpha is a high-performance compute engine and neural-symbolic operating system designed for JAX/XLA. It solves the "Discrete Singularity" problem—where gradients die at logical branches—by treating computation as a continuous, differentiable Riemannian manifold.
### 1. ARCHITECTURAL PILLARS
The system is built on a **7-Layer Adelic Stack**, ensuring that telemetry, logic, and memory are unified under a single geometric metric.
* **NRRB (Resurgence Bridge):** Employs Weierstrass Gaussian proxies to enable gradient flow through non-differentiable thresholds (if/else, bit-masks).
* **CSM (Crystalline Static Manifold):** Binds dynamic shapes into rigid, static HLO allocations for zero-jitter, 100% hardware utilization.
* **ATC (Trace Crystallizer):** Hardens adaptive logic into fixed XLA traces, eliminating recompilation latency.
* **ARSM (Recursive State):** Manages temporal continuity through recursive hidden state manifolds.
* **ASTC (Stochastic Ingest):** Maps non-stationary sensor streams into uniform 16D interaction lattices via SOS-DP.
* **AFRH (Fisher Regulator):** Computes the local Fisher Information Matrix (FIM) to stabilize learning via Natural Gradients.
* **AFRC (Riemann Connector):** Propagates the Fisher Metric across decentralized nodes using Levi-Civita Parallel Transport.
---
### 2. MATHEMATICAL CORE
**The Levi-Civita Connection ($\nabla$)**
To maintain coherence across a decentralized cluster, Tungsten Alpha implements Parallel Transport. The Fisher Metric $g_{ij}$ is moved across the manifold such that the information volume is conserved:
$$g' = \Omega g \Omega^T$$
Where $\Omega$ is an orthogonal connection matrix derived from the **Adelic-Fisher-Riemann-Connector (AFRC)**. This ensures that the "Curvature" of the intelligence remains invariant as it moves between sensor nodes.
---
### 3. PROJECT STRUCTURE
```text
Tungsten-Alpha/
├── README.md # Project Specification
├── main.py # Universal Orchestrator
├── flash_binary.py # Orbax Serialization Protocol
├── core/
│ ├── bridge.py # NRRB Implementation
│ ├── manifold.py # CSM Logic
│ ├── geometry.py # AFRH (Fisher Metric)
│ └── transport.py # AFRC (Parallel Transport)
└── visualization/
└── dashboard.py # Telemetry Dashboard
```
---
### 4. DEPLOYMENT AND VALIDATION
**Zero-Recompilation Hardening**
Tungsten Alpha is designed for production environments where latency is critical. By using `jax.jit` and `lax.fori_loop` within a static manifold, we achieve deterministic execution times.
**Verification:**
Run the internal validation suite to confirm HLO Fusion and Metric Stability:
```bash
python core/transport.py
```
Expected Output: `AFRC STATUS: PRODUCTION-READY ✓ HLO FUSION: TRUE ✓`
---
[dashboard](https://raw.githubusercontent.com/deskiziarecords/tungsten-alpha/refs/heads/main/tng-alpha.png)
Tungsten Alpha was built to move beyond the limitations of standard backpropagation. It understands the light by calculating its own curvature.
## LICENSE
Apache License 2.0. Developed for the future of decentralized AGI.
---
## Support My Work
If you enjoy my work and would like to support me, consider buying me a coffee!
[](https://www.buymeacoffee.com/hipotermiah)
### Author: J. ROBERTO JIMENEZ C. - tijuanapaint@gmail.com - @hipotermiah