List your company

Prefill Latency on Llama-3-8B 10 Mbps

1.563Prefill Latency (s)ASTRA
-15.61256100.32247216.2575332.19253May 25, 2025

Evaluation results

MethodLinks
ASTRAG=12025.051.563
ASTRAG=162025.051.661
ASTRAG=322025.051.94
Llama-3-8B (Single-device)Device=Single-device2025.054.578
BPStrategy=Block Parallelism, Nb=42025.054.642
BPStrategy=Block Parallelism, Nb=82025.058.011
SPStrategy=Sequence Parallelism2025.0528.256
TPStrategy=Tensor Parallelism2025.05430.952