LLM Training Performance on Qwen models on A6000 (48GB VRAM) 2.5 (train)
56.82TFLOPSMegaTrain
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| MegaTrainModel=Qwen2.5-14B, Max BS=92026.04 | 56.82 | 44.64 | 104.1 | 641 | |
| MegaTrainModel=Qwen2.5-7B, Max BS=122026.04 | 55.73 | 44.74 | 56.7 | 1,219 | |
| MegaTrainModel=Qwen2.5-3B, Max BS=152026.04 | 49.7 | 46.74 | 38.3 | 2,153 | |
| ZeRO-3 OffloadModel=Qwen2.5-7B, Max BS=12026.04 | 27.55 | 20.83 | — | — | |
| ZeRO-3 OffloadModel=Qwen2.5-3B, Max BS=12026.04 | 23.89 | 20.33 | — | — |