BitNet-Agent: 1.58-bit C++ Neural Swarm Simulation

BitNet-Agent is a C++ implementation of a 1.58-bit (ternary) neural simulation utilizing a distributed agent mesh. The engine is optimized for AVX2 SIMD hardware and focuses on memory-efficient weight paging and parallel head processing.

Technical Specifications

• Bit-Packing Protocol: Weights and states are stored using 2-bit packing (4 values per uint8_t), achieving a 75% reduction in disk and memory footprint compared to 8-bit storage.
• Just-in-Time weight Paging: Individual agent weights are loaded from disk only during active inference and purged immediately after, capping global RAM usage at the size of a single agent's parameters plus metadata.
• Dynamic Node Mitosis: The simulation spawns or merges logical agents based on the bit-entropy of the input data stream.
• Parallel Head Processing: Multi-head attention operations are distributed across parallel CPU threads using std::async.
• Ternary Quantization: On-disk persistence utilizes a {-1, 0, 1} mapping for agent state serialization.

Implementation Workflow

1. Tokenization: Input stream is processed into localized character-integer mappings.
2. Mesh Expansion: Nodes are allocated dynamically based on complexity thresholds.
3. Council Inference: Sub-agents execute parallel weight-sum operations using AVX2 intrinsics.
4. Resonance Handoff: Inter-agent data transfer is gated by attention scores (scalar dot-product).
5. Serialization: Agent states are bit-packed and written to disk for persistence.

Performance Profile

Parameter	8-bit Baseline	1.58-bit (Packed)	Efficiency
Storage per 512 Weights	512 Bytes	128 Bytes	75% Reduction
RAM Scaling	O(N)	O(1) + Meta	Constant Floor
Execution Latency	~2.5 ms	0.6 - 0.9 ms	AVX2 Optimized

Usage

Build Requirements

• Windows OS
• MSVC Compiler (cl.exe)
• Python 3.x

1. Build Engine

Execute build.bat to compile with /arch:AVX2.

build.bat

2. Weight Conversion

Use the included bridge to prepare binary payloads from external float matrices.

python scripts/bridge.py --mock --output real_weights.bin

Development Protocol

Refer to docs/RULES.md for the technical constitution and phase-by-phase verification logs.

License

MIT License. See LICENSE for details.