This project is a fork of bbycroft/llm-viz by
Brendan Bycroft — the original interactive 3D LLM visualization. The WebGL renderer,
the block-layout engine, the camera/interaction system, and the overall design are his work.
At the time of forking, the upstream repository did not include an explicit open-source license. Under default copyright law that means the upstream code is all rights reserved by the original author. This fork is published for educational purposes as a companion to the AI Inference Engineer 2026 course, with prominent attribution.
This fork does not — and legally cannot — relicense the upstream code. If you wish to redistribute, host this publicly on an ongoing basis, fork it further, or use it commercially, please contact Brendan Bycroft for permission regarding the upstream portions first.
The following work was added on top of the upstream renderer and is original to this fork:
- Re-targeting the model from GPT-2 to the modern dense decoder-only architecture (GQA, RoPE, RMSNorm, SwiGLU, bias-free FFN) used by Qwen 2.5 and Llama 3.3.
- Three switchable model presets (Qwen 2.5 7B, Qwen 2.5 72B, Llama 3.3 70B).
- The left dense decoder-only forward-pass explainer, synced to the 3D model on hover.
- The roofline visualization (memory- vs compute-bound), calibrated to NVIDIA H200.
- The tensor-parallelism explorer (TP = 1/2/4/8): 3D weight sharding, per-GPU stats, all-reduce / communication modeling, and the disaggregated prefill/decode callout.
The fork authors offer these additions freely for educational and non-commercial use. This permission applies only to the newly-added code listed above, not to the upstream work.
Built as a companion to AI Inference Engineer 2026, part of the open AI Hardware Engineer Roadmap.