SageAttention2++: A More Efficient Implementation of SageAttention2
This incremental improvement enhances efficiency for models in language, image, and video generation with minimal performance loss.
The paper tackles the quadratic time complexity of attention mechanisms by proposing SageAttention2++, which uses faster FP8 matrix multiplication instructions to accelerate SageAttention2, achieving a 3.9x speedup over FlashAttention while maintaining accuracy.
The efficiency of attention is critical because its time complexity grows quadratically with sequence length. SageAttention2 addresses this by utilizing quantization to accelerate matrix multiplications (Matmul) in attention. To further accelerate SageAttention2, we propose to utilize the faster instruction of FP8 Matmul accumulated in FP16. The instruction is 2x faster than the FP8 Matmul used in SageAttention2. Our experiments show that SageAttention2++ achieves a 3.9x speedup over FlashAttention while maintaining the same attention accuracy as SageAttention2. This means SageAttention2++ effectively accelerates various models, including those for language, image, and video generation, with negligible end-to-end metrics loss. The code will be available at https://github.com/thu-ml/SageAttention.