On my M4 Max Studio, I encountered similar problems, but not on all kinds of models. But even a model convergence properly, it usually has worse R2 and RMSE than training the same model on CPU. I really want to fully utilize my M4 Max's GPU!!!