Transformer for Jet Tagging Implemented on Versal AI Engines.

Transformer-based models have demonstrated strong performance in jet tagging, a crucial task at the CERN Large Hadron Collider (LHC). However, deploying these complex models within the LHC's low-latency, resource-constrained trigger systems poses a significant engineering challenge. Researchers have tackled this by presenting an initial implementation of a quantized, integer-only transformer specifically designed for jet tagging. This model is mapped onto AMD Versal AI Engines (AIEs), with dense and multi-head attention layers configured to run efficiently on AIE tiles. A key contribution is a reusable software framework that simplifies this deployment. It represents transformer layers as composable AIE building blocks and automatically generates the necessary Vitis graph code from a high-level Python model description. This open-source framework provides a robust foundation for future research and development in deploying advanced AI models on reconfigurable hardware.