List your company

LLM Training Performance on Qwen models on RTX 3090 (24GB VRAM) 2.5 (train)

35.09TFLOPSMegaTrain
23.462826.481429.532.5186Apr 6, 2026

Evaluation results

MethodLinks
MegaTrainModel=Qwen2.5-7B, Max BS=52026.0435.0922.6356.7768
MegaTrainModel=Qwen2.5-3B, Max BS=72026.0433.1822.83251,792
MegaTrainModel=Qwen2.5-14B, Max BS=32026.0430.1921.1103.7341
ZeRO-3 OffloadModel=Qwen2.5-7B, Max BS=12026.0427.4920.83——
ZeRO-3 OffloadModel=Qwen2.5-3B, Max BS=12026.0423.9120.32——