Mixture of Block Attention (MoBA) Implementation Performance on New-feature tasks
38.7Session DurationPithTrain
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| PithTrainFramework=PithTrain2026.05 | 38.7 | 27.7 | 57 | 69,000 | 32,400 | |
| Megatron-LMFramework=Megatron-LM2026.05 | 61.6 | 49.5 | 134 | 120,900 | 53,800 | |
| TorchTitanFramework=TorchTitan2026.05 | 105.1 | 77.9 | 91 | 166,900 | 111,800 |