If your model fails to converge during a large-scale training run at Zyphra, what is your systematic approach to debugging the issue?
