Bit-TTT-Engine

Bit-TTT Engine: High-Performance Brain Core

1.58-bit Quantization + Test-Time Training (TTT) Implementation in Pure Rust.

日本語 / Japanese

---

✨ What is Bit-TTT?

Bit-TTT Engine combines two cutting-edge techniques:

• BitNet 1.58-bit Quantization: Extreme compression using ternary weights {-1, 0, +1}
• Test-Time Training (TTT): Adaptive attention replacement with online learning

The goal: Run 70B parameter models on 8-16GB VRAM with efficient inference.

📊 Current Status (2026-01)

Feature	Status	Description
Core Engine (cortex_rust)	✅ Complete	Candle-based neural network implementation
Training Pipeline	✅ Complete	End-to-end training in pure Rust
Streaming Inference	✅ Complete	~1100 tokens/sec on CPU
GUI Trainer	✅ Complete	Tauri-based visual training interface
Python Bindings (PyO3)	✅ Complete	Optional Python integration
Japanese Tokenizer	🚧 Planned	Phase 14
7B/70B Scaling	🚧 Planned	Phase 15
WASM/Browser Support	🚧 Planned	Phase 16

🏗️ Architecture

Bit-TTT Engine
├── crates/
│   ├── rust_engine/         # Core library (cortex_rust)
│   │   ├── layers/          # Neural network layers
│   │   │   ├── rms_norm.rs    # RMS Normalization
│   │   │   ├── bit_linear.rs  # 1.58-bit Linear Layer
│   │   │   ├── swiglu.rs      # SwiGLU MLP
│   │   │   └── ttt.rs         # TTT Layer
│   │   ├── model/           # Model architecture
│   │   │   ├── block.rs       # Transformer Block
│   │   │   ├── llama.rs       # BitLlama Model
│   │   │   └── config.rs      # Configuration
│   │   ├── python.rs        # PyO3 bindings
│   │   └── lib.rs           # Public API
│   │
│   └── bit_llama/           # CLI application
│       ├── train/           # Training module
│       │   ├── args.rs        # CLI arguments
│       │   ├── checkpoint.rs  # State management
│       │   └── training_loop.rs  # Main loop
│       ├── gui/             # Tauri GUI
│       └── inference.rs     # Inference engine
│
├── models/                  # Trained model checkpoints
├── data/                    # Training datasets
└── tools/                   # Utility scripts

🚀 Quick Start

Prerequisites

• Rust 1.70+
• (Optional) CUDA 11.8+ for GPU acceleration

1. Build

git clone https://github.com/imonoonoko/Bit-TTT-Engine.git
cd Bit-TTT-Engine
cargo build --release

2. Training

# Using launch script (Windows)
./launch_trainer.bat

# Manual training
cargo run --release --bin train_llama -- \
    --data data/TinyStories \
    --dim 256 \
    --layers 8 \
    --steps 10000 \
    --lr 3e-4

3. Inference

# Using launch script (Windows)
./launch_chat.bat

# Manual inference
cargo run --release --bin bit_llama -- \
    --model models/my_model \
    --prompt "Hello Bit-TTT!" \
    --max-tokens 100 \
    --temp 0.8

📖 Documentation

Document	Description
ARCHITECTURE.md	System design philosophy
ROADMAP.md	Future development plans
docs/SPECIFICATION.md	Technical specification
docs/CONTRIBUTING.md	Contribution guide

🛠️ Development Commands

# Run all tests
cargo test --workspace

# Check compilation
cargo check --workspace

# Format code
cargo fmt --all

# Run linter
cargo clippy --workspace

🐍 Python Integration (Optional)

cd crates/rust_engine
pip install maturin
maturin develop --release

import cortex_rust

config = cortex_rust.BitLlamaConfig(
    vocab_size=16384,
    hidden_dim=256,
    num_layers=8,
    inner_lr=0.1
)

model = cortex_rust.BitLlama(config, "model.safetensors", device="cuda")
logits = model.forward(token_id=42)

💖 Support

Solana Wallet: 13ui3nmE7smmK3Pk8wyKb7RE6wHyMJCcWgCeMRRdoory

---

Created by Project Bit-TTT • MIT License

This site is open source. Improve this page.