Language Model Training Efficiency on UltraChat
3,008.3ThroughputHSAP
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| HSAPContext Length=16K, Topology (intraSPxinterSP)=8x12026.06 | 3,008.3 | 1.85 | |
| HSAPContext Length=64K, Topology (intraSPxinterSP)=1x82026.06 | 2,994.03 | 0.52 | |
| BatchingContext Length=16K, Topology (intraSPxinterSP)=8x12026.06 | 2,541.83 | 1.56 | |
| PackingContext Length=16K, Topology (intraSPxinterSP)=8x12026.06 | 2,530.93 | 1.56 | |
| PackingContext Length=64K, Topology (intraSPxinterSP)=1x82026.06 | 1,251.3 | 0.22 | |
| BatchingContext Length=64K, Topology (intraSPxinterSP)=1x82026.06 | 593.16 | 0.22 |