Prefill Latency on Llama-3-8B 500 Mbps
1.485Prefill Latency (s)BP
Evaluation Results
| Method | Links | |
|---|---|---|
| BPStrategy=Block Parallelism, Nb=42025.05 | 1.485 | |
| ASTRAG=12025.05 | 1.54 | |
| ASTRAG=162025.05 | 1.548 | |
| BPStrategy=Block Parallelism, Nb=82025.05 | 1.561 | |
| ASTRAG=322025.05 | 1.583 | |
| SPStrategy=Sequence Parallelism2025.05 | 2.052 | |
| Llama-3-8B (Single-device)Device=Single-device2025.05 | 4.578 | |
| TPStrategy=Tensor Parallelism2025.05 | 10.14 |