Welcome to Arton. This project is the ultimate bridge between ML engineering and complete beginners. We are attacking the AI black box from both sides: building it from absolute scratch to understand the math, and using the fastest enterprise tools on the market to understand production.
We are building a Generative Pre-trained Transformer (GPT) model from absolute scratch.
- Zero Deep Learning Dependencies. No PyTorch, no TensorFlow.
- Pure Mojo. We get Python's readability but C/Rust's speed.
- Understand Everything. We build the Autograd Engine, the Multi-Head Attention mechanism, and the Training Loop out in the open.
Once you understand how the atomic model works, you need to know how the enterprise world deploys state-of-the-art models like Gemma 4.
- Modular MAX Engine. We use the Modular Accelerated Xecution (MAX) engine—the fastest inference engine on the market.
- Hardware Harmonization. Learn how MAX compiles models to run flawlessly across NVIDIA GPUs, AMD GPUs, and Apple Silicon without rewriting code.
Test the workbench by running:
pixi run mojo hello_world.mojo
Computers don't read English; they read numbers. We use prepare_data.mojo to download our dataset (Tiny Shakespeare, 1MB). Then we build a Tokenizer (a secret decoder ring) to translate letters into numbers.
Get the book by running:
pixi run mojo atomic_gpt/prepare_data.mojoSee the Tokenizer:pixi run mojo atomic_gpt/tokenizer.mojo
While we build the atomic model, we will simultaneously look at how Modular's MAX engine handles a massive Google model like Gemma.
See the deployment script:
pixi run python production_max/run_gemma.py
(More phases coming soon...)