Candle dev here, we also support training/backdrop! We certainly focus on optimizing inference performance but hopefully that should improve the training efficiency too.