MiniMax open-sources MSA sparse attention kernel

MiniMax's MSA demonstrates that combining algorithmic sparsity with hardware-level co-design is key to deploying million-token context models without prohibitive hardware costs. Decoupled indexing uses a lightweight branch to run low-dimensional projection and Top-k selection, avoiding the quadratic bottleneck of dense attention. Hardware co-design featuring exp-free Top-k selection and KV-outer sparse attention ensures high Tensor Core utilization. Finally, integration into the MiniMax-M3 multimodal model shows the kernel is battle-tested and ready for production scaling.