docs/architecture/ARCHITECTURE.md

LLARRI-O1 Architecture

Spanish version: ARCHITECTURE.es.md

Authorship

• Author: Lucas Ricardo Mella Chillemi (Segunda Cabeza)
• Coordinator: Alvaro (Segunda Cabeza)
• Date: 2026-01-07

Overview

LLARRI-O1 explores a parameter-sharing, fractal-like computation pattern:

• Inputs are projected into a hidden space.
• The hidden vector is split into 4 quadrants.
• Each quadrant is processed by a progressive fractal pipeline.
• Quadrants exchange information via cross-relations.
• Multiple “boxes” (processing stages) are connected via residual “keys”.

Current implementation focus (v4.0 HyperComprimido)

Key ideas:

• 6 boxes total: 3 primary processing boxes + 3 secondary processing boxes.
• 8 fractal levels: 2 → 4 → 8 → 16 → 32 → 64 → 128 → 256 (sequential).
• Binary cache at level 2 (CacheBinario) for fast lookup of basic operations.
• Secondary boxes include internal self-calculation between intermediate values.

Notes on scalability

• The maximum fractal level is constrained by quad_dim = hidden_dim // 4.
• Increasing hidden_dim increases quad_dim, which enables larger fractal levels (e.g. 256).
• GPU VRAM can still be a bottleneck during training due to optimizer states (AdamW) and activations.

Where to look

• Package: llarri_o1/
  • Config: llarri_o1/config.py
  • Model: llarri_o1/model.py
  • Modules: llarri_o1/modules/ (cache, niveles, relaciones, cajas, flujo)
  • Training: llarri_o1/training/trainer.py
  • Visualization: llarri_o1/visualization/diagrams.py
• Scripts: scripts/train.py
• Examples: examples/basic_usage.py
• v4 diagrams: diagrams/v4-current