List your company

Prefill Latency on Llama-3-8B 50 Mbps

1.547Prefill Latency (s)ASTRA
-1.8890821.3044644.49867.69154May 25, 2025

Evaluation results

MethodLinks
ASTRAG=12025.051.547
ASTRAG=162025.051.595
ASTRAG=322025.051.661
BPStrategy=Block Parallelism, Nb=42025.052.085
BPStrategy=Block Parallelism, Nb=82025.052.773
Llama-3-8B (Single-device)Device=Single-device2025.054.578
SPStrategy=Sequence Parallelism2025.056.888
TPStrategy=Tensor Parallelism2025.0587.449