Inference Optimization for MiniMax Sparse Attention(together.ai)

https://www.together.ai/blog/serving-minimax-m3-for-efficient-inference-unlocking-1m-token-context-and-multimodality-without-regrets