List your company
ResearchFormerly WizWand

PrefixLM Attention Performance on Llama2-7B (8k Context, 32/32 Heads)

163.7TFLOPS (PrefixLM Attention)CuBridge
8.594448.862289.13129.3978May 6, 2026
Updated 6mo ago

Evaluation results

MethodLinks
CuBridgeImplementation=CuBridge2026.05163.7
FlexAttentionImplementation=PyTorch FlexAttention2026.05142.77
Qimeng AttnImplementation=Qimeng Attention2026.05122.44
TorchImplementation=Standard PyTorch2026.0514.56