llm-viz is a 3D visualization tool that renders the architecture and inference process of a GPT-style language model network.
The tool addresses the challenge of understanding how transformer-based language models work by making their internal structure visible and interactive. It displays a three-dimensional model of a working GPT network topology, showing how data flows through the layers during inference. The visualization uses a tiny demonstration model trained to sort lists of letters, derived from a reference implementation, allowing users to see actual weights and computations in action rather than abstract diagrams.
The project suits developers and researchers who want to understand transformer internals through direct visualization rather than reading papers or code. It works best as an educational tool for grasping how GPT-style networks process information. The renderer can theoretically handle larger models like GPT-2 scale networks, though downloading full weights for such models is impractical due to file size. This is a learning and exploration tool rather than a production inference system.
The project is maintained as part of a personal portfolio repository alongside other non-trivial projects, sharing utility code across multiple initiatives. Development appears focused on the core visualization working correctly with the demonstration model rather than on expanding to production-scale networks. A CPU simulation project is also present in the repository but remains in early stages and is not yet exposed for public use.
