List your company

Paged Decode Attention Serving on Bucketed B1 trace

1.576GPU RatioPersistentKV
1.49721.53661.5761.6154Jun 25, 2026

Evaluation results

MethodLinks
PersistentKVSelected route=PersistentKV bucket, split 32, Page size=16, Hq=32, Hkv=8, d=128, Precision=FP162026.061.5761.3996.1044.398