Tag
Introduces Flash-MSA, the first performant open-source training kernels for MiniMax Sparse Attention on Hopper and Blackwell GPUs, enabling efficient million-token training.