Recently, I’ve been learning the implementation of QAT with PyTorch 2.x. Up until now, it appears that the graph-based pt2e is becoming the predominant solution for QAT in the times ahead. Drawing upon my previous experiences, I have a crucial question: Unlike PTQ, users need to do a lot of debugging when using QAT, since QAT is essentially model training. However, it is widely acknowledged that debugging in graph-based models can be very challenging. So I want to ask whether PyTorch has any solutions or plans for this problem.