transformer-atlas / visual map

From "Attention Is All You Need" to today's serving stack

Three ways to read the same 31 techniques:

Lineage

Architecture families as branches off one root, hybrids as merge points.

Pipeline

The Transformer branch, ordered by where each piece sits in a decoder-only forward pass — token input to production serving.

Timeline

Every technique, ordered by publication year.

Click any chip to open its folder in transformer-atlas; the small link opens the source paper.