Running branch-and-bound optimization on GPUs with batched processing

Finding the best sparse generalized linear models (picking the right variables while fitting) is computationally hard because it mixes continuous and discrete choices—and GPUs haven't helped much since the algorithm traditionally runs sequentially. Liu and Lodi batch multiple search nodes together on GPUs using padding and custom kernels to handle irregular data shapes. On hard instances, they achieve 10–100× speedups with zero optimality gap and can extract the full Rashomon set (all near-optimal models) for downstream analysis like variable importance.