Open-source tool for visualizing transformer attention weights
September 8, 2026
This tool visualizes LLM attention by calculating attention weights scaled by value vector magnitude across all heads and layers. Users can hover over generated tokens to see which previous tokens influenced the output via opacity mapping.
HOW THIS AFFECTS YOU
●
builderYou can use this to debug how specific tokens influence your model's generation patterns.
●
researcherThis provides a simplified heuristic for inspecting cross-layer attention aggregation.