model           =gpt2
total batch size=40
train num epochs=10
fp16            =True
max seq length  =40
eval_acc = 0.6864357017028773
eval_loss = 2.5627794878823416