activation-fusion

This paper presents CUTLASS-based kernels that fuse SwiGLU activation with GeMM at the tile level, achieving up to 2.47× speedup on NVIDIA H100 for efficient LLM inference.