LLM Visualization
A 3D interactive visualization of a GPT-style large language model network running inference, showing a small trained model sorting letters and rendering larger networks such as GPT-2.
LLM Visualization displays a working 3D model of a GPT-style network topology similar to GPT-2 and GPT-3. The first network shown with working weights is a tiny model that sorts a short list of the letters A, B and C, the demo example model from Andrej Karpathy's minGPT implementation.
The renderer also supports visualizing arbitrarily sized networks and works with the smaller GPT-2 size, though its weights are not downloaded because of their size, hundreds of megabytes. The same repository hosts the author's personal homepage and other projects, including a work-in-progress 2D digital schematic editor for simulating a RISC-V CPU that is not yet exposed on the site.
It is written in TypeScript, runs locally with yarn then yarn dev, and is MIT licensed except for the personal homepage and third-party assets.