List your company

Attention Quantization Error Analysis on Flash Attention (seq=16384, head_dim=64)

0.49L2 Relative ErrorFlash+MSD
0.45320.70160.951.1984May 13, 2026

Evaluation results

MethodLinks
Flash+MSDSequence length (seq)=16384, Head dimension (head_dim)=64, KV Quantization type=INT8 KV with per-channel scale2026.050.4989.445.922.14.1
FlashSequence length (seq)=16384, Head dimension (head_dim)=64, KV Quantization type=INT8 KV with per-channel scale, mode=BF162026.051.3897.787.266.78.6
DequantSequence length (seq)=16384, Head dimension (head_dim)=64, KV Quantization type=INT8 KV with per-channel scale, mode=BF162026.051.4197.787.467.58.9