Training Throughput on Qwen3 32B (train)
545.29Training Throughput (128K Seq Len)Ulysses
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| UlyssesHardware=16xH100s2026.02 | 545.29 | 370.7 | 217.04 | 117.02 | 59.98 | — | — | — | |
| UPipeHardware=16xH100s2026.02 | 483.29 | 339.56 | 204.46 | 113.26 | 59.56 | 40.42 | 29.97 | — | |
| RingHardware=16xH100s2026.02 | 418.39 | 308.88 | 194.44 | 110.27 | 58.45 | — | — | — | |
| FPDTHardware=16xH100s2026.02 | 286.4 | 217.85 | 151.91 | 95.88 | 55.41 | 38.86 | 27.66 | — | |
| Native PyTorchHardware=16xH100s2026.02 | 127.03 | 112.2 | 91.39 | — | — | — | — | — |