SageAttention provides optimized attention kernels for Ampere, Ada, and Hopper GPUs that leverage quantization to speed up inference across language, vision, and video models. It includes INT8 quantization for QK^top with smoothing, FP8 quantization for PV computation, and a two-level accumulation strategy to maintain accuracy at FP8 precision. The library supports plug-and-play deployment across diverse transformer architectures.