How is L1 penalty term coefficient implemented?

Hi. I'm trying to reproduce the m_IoU performance on ADE20k and notice that your results are divided to HRnet-A/B using different L1 penalty term (lambda). But I couldn't find the lambda adjustment protocol in the code. Could you please tell me how is this lambda adjustment implemented? Thank you!