Just a small remark (not the solution to your problem, as I maintain the solution I mentioned previously, but rather a simple remark): if you want to simulate the NVT ensemble as you said in the initial message, it’s better to use fix nvt instead of fix nve + fix langevin. The way the Langevin thermostat affects the equations of motion is apparently not such that allows you to truly sample the NVT ensemble (see Difference between nve+langevin and nvt dynamics! - #2 by akohlmey).