List your company

Prefill Latency on Llama-3-8B 20 Mbps

1.549Prefill Latency (s)ASTRA
-7.0406850.93966108.92166.90034May 25, 2025

Evaluation results

MethodLinks
ASTRAG=12025.051.549
ASTRAG=162025.051.659
ASTRAG=322025.051.796
BPStrategy=Block Parallelism, Nb=42025.053.047
Llama-3-8B (Single-device)Device=Single-device2025.054.578
BPStrategy=Block Parallelism, Nb=82025.054.78
SPStrategy=Sequence Parallelism2025.0514.939
TPStrategy=Tensor Parallelism2025.05216.291