Can LLMs predict the convergence of Stochastic Gradient Descent?