tahamajs/RR / Docs /visualizer /docs /CONDUIT_SYSTEM.md
tahamajs's picture
|
download
raw
2.16 kB
# 3D Dataflow Conduits & Inter-Module Wiring Engine
## 1. Conduits Overview
Modern transformer architectures are not isolated blocks; they are deeply connected networks of attention operators, multi-query head splits, gating mechanisms, and residual highways.
The visualizer computes code-accurate 3D cubic Bezier curves (`src/builders/conduit_builder.js`) that physically route token signals between operations.
```
Layer Input Bus Port Ring (◎)
│
┌──────────┴──────────┐
▼ ▼
[Query Projection] [Key/Value Split]
│ │
└──────────┬──────────┘
▼
[Attention Core]
│
▼
[Output Dense]
│
▼
Residual Addition Node (+) ◀──── [Bypass Highway Conduit]
```
---
## 2. Color-Coded Signal Bus Standards
| Signal Type | Color | Hex | Visual Role |
|---|---|---|---|
| **Input Stream** | Emerald Green | `#10b981` | Sequence entry from embedding lookup |
| **Baseline Feed** | Sky Blue | `#38bdf8` | Dense baseline forward connections |
| **Compressed Feed** | Mint Cyan | `#2dd4bf` | Low-rank / compressed forward connections |
| **GQA KV Bundle** | Amber Orange | `#f59e0b` | Shared Key/Value projections (Qwen) |
| **SwiGLU Branch** | Violet Purple | `#a855f7` | Bifurcated gate/up projection cables |
| **Residual Attn** | Deep Azure | `#0284c7` | Attention identity skip highway ($+X_{\text{in}}$) |
| **Residual FFN** | Royal Indigo | `#7c3aed` | Feed-forward identity skip highway ($+X_{\text{mid}}$) |
| **Output Head** | Crimson Rose | `#f43f5e` | LM Unembedding Softmax distribution |
---
## 3. Port Rings
At every layer entrance and exit, the engine instantiates a high-tech circular ingress/egress port ring (`THREE.RingGeometry`) indicating active tensor flow.
When a user clicks on any sublayer in the Deep Inspector or Component Pipeline, the associated conduits illuminate in bright emissive white to trace the signal pathway.

Xet Storage Details

Size:
2.16 kB
·
Xet hash:
abdb7ce2b6fe38d506d31135cdc4d03d7b18cece9adcfa1ae7a7fd70a7d85ac8

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.