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

41 points | by rawsh 2 days ago

5 comments