Speaker
Description
AllShowers (arXiv:2601.11716) marks a significant step towards a universal model for calorimeter shower simulations in collider experiments. Unlike traditional surrogate models that train separate networks for each particle species, AllShowers unifies shower generation across multiple particle types in a single continuous normalizing flow with a Transformer architecture, producing realistic showers.
As we move towards the HL-LHC era, both dataset sizes and per-shower point-cloud lengths are expected to grow substantially, making training and inference scalability a central bottleneck for Transformer-based generative shower models, whose attention mechanism scales quadratically with point cloud size. In this work, we address this challenge by incorporating Block Sparse Attention, a library of sparse attention kernels supporting a range of sparse patterns - including token-granularity streaming attention, block-granularity streaming attention, and block-sparse attention - to substantially reduce the computational cost of the attention operation while preserving the inductive biases introduced by AllShowers' custom masking scheme. To accelerate sampling, we further integrate DPM-Solver++, a fast high-order ODE solver for diffusion/flow-based generative models, reducing the number of function evaluations needed at inference time.
Across our case studies, these optimizations yield up to a 5x reduction in training time and up to a 3x speedup in inference via DPM-Solver++ sampling, with negligible loss in shower fidelity. These results mark a significant step towards a scalable, universal model for calorimeter shower simulation capable of meeting the throughput demands of collider experiments in the HL-LHC era.