Flash-MSA: Accelerating Million-Token Training with Sparse Attention Kernels