Bit-TTT Engine: High-Performance Brain Core

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

This package provides Python bindings for the Bit-TTT Engine, allowing you to run ultra-light ternary LLMs with real-time adaptation.

✨ Features

Ultra-Light: Runs large LLMs on cheap hardware using 1.58-bit (ternary) weights.
Adaptive (TTT): Learns while inferring, adapting to context in real-time.
Pure Rust: High performance with minimal dependencies.

🚀 Installation

pip install bit-ttt-engine

💻 Usage

import cortex_rust
import json

# Initialize Configuration
config = cortex_rust.BitLlamaConfig(
    vocab_size=32000,
    hidden_dim=512,
    num_layers=12,
    inner_lr=0.001
)

# Initialize Model (Inference)
model = cortex_rust.BitLlama(
    config=config,
    checkpoint_path="path/to/model.safetensors",
    device="cpu", # or "cuda"
    tokenizer_path="path/to/tokenizer.json"
)

# Generate Text
output = model.generate(prompt="Hello, world!", max_tokens=50)
print(output)

🏗️ Training (TTT)

trainer = cortex_rust.PyTrainer(
    config=config,
    checkpoint_path="path/to/model.safetensors",
    device="cuda"
)

# Single training step
loss = trainer.train_step(input_ids=[...], targets=[...])
print(f"Loss: {loss}")

# Save checkpoint
trainer.save_checkpoint("model_updated.safetensors")

📖 Documentation

For more details, please visit the GitHub repository.

🙏 Acknowledgments

This project incorporates ideas and techniques inspired by the DroPE method published by Sakana AI.

💖 License

MIT License

cortex_rust 0.6.0