Researchers from China Introduce INT-FlashAttention: INT8 Quantization Architecture Compatible with FlashAttention Improving the Inference Speed of FlashAttention on Ampere GPUs