Thank @ChengjieLuo for pointing out that when the constraint has small coefficient, the multiplier for constraints does not work well. The reason is that when the constraint is scaled down by $K$, which theoretically has no influence on the model, the effect of multiplier is scaled down by $K^2$ in numerics. Such issue should be fixed by choosing an acceptance according to the coefficients of the constraints.
Thank @ChengjieLuo for pointing out that when the constraint has small coefficient, the multiplier for constraints does not work well. The reason is that when the constraint is scaled down by $K$, which theoretically has no influence on the model, the effect of multiplier is scaled down by $K^2$ in numerics. Such issue should be fixed by choosing an acceptance according to the coefficients of the constraints.