model           =gpt2
total batch size=40
train num epochs=3
fp16            =True
max seq length  =40
eval_acc = 0.7935483870967742
eval_loss = 0.5006362595865803