Understand how large language models really work — from tokens to systems.
An interactive playground for exploring LLM systems, inference workflows, and AI infrastructure through visualization and hands-on interaction.
🌐 Live Demo: https://skyliulu.github.io/LLM-Infra-Explorer/
Modern LLM systems are complex, opaque, and hard to reason about:
- What actually happens during inference?
- How does KV cache evolve over time?
- What do DP / TP / PP really change in execution?
- Why are optimizations like Flash Attention so effective?
This project helps you see and interact with these processes, instead of just reading about them.
- 🖥️ LLM Inference Visualization — step-by-step prefill/decode animation, KV cache lifecycle, MoE vs Dense architecture toggle, and temperature-controlled sampling
- 🔀 6D Parallel Strategy Explorer — interactive DP / TP / PP / CP / EP / ETP topology with real-time tensor slicing and GPU resource mapping
- ⚡ Flash Attention Walkthrough — Standard vs Flash Attention comparison, SRAM/HBM IO traffic tracking, and block-by-block tiled computation with causal mask skipping
- 🚀 Flash Decode — KV Cache splitting across SM compute units, parallel two-step reduction, and memory-wall-breaking decode optimization
- 🧬 Engram (DeepSeek) — n-gram conditional memory retrieval augmenting Transformer layers, with tensor flow visualization and hardware-level async prefetch timeline
- 🌲 Radix Cache — SGLang's radix-tree-based global KV cache sharing, with lazy prefix splitting, LRU eviction, and physical memory block reuse visualization
- 🔁 DP Attention — hybrid DP/TP attention flow, KV cache row-sharding strategy, and cross-rank communication trade-off visualization
git clone https://github.com/skyliulu/LLM-Infra-Explorer.git
cd LLM-Infra-Explorer
npm install
npm run devThis project is licensed under the GNU Affero General Public License v3.0 (AGPL-3.0). Commercial use is subject to the terms of this license. Any network-deployed modifications must also be open-sourced.