Training Throughput on DeepSeek-V2-Lite workload
114,600Training Throughput (tokens/s)PithTrain
Evaluation Results
| Method | Links | |
|---|---|---|
| PithTrainHardware=1×8-H100, Parallelism=PP2,DP1,CP1,EP4, SeqLen=2048, Precision=BF162026.05 | 114,600 | |
| Megatron-LMHardware=1×8-H100, Parallelism=PP2,DP1,CP1,EP4, SeqLen=2048, Precision=BF162026.05 | 107,300 | |
| TorchTitanHardware=1×8-H100, Parallelism=PP2,DP1,CP1,EP4, SeqLen=2048, Precision=BF162026.05 | 74,100 |