Espresso: Cost-Efficient Large Model Training by Exploiting GPU Heterogeneity in the Cloud | IEEE Conference Publication | IEEE Xplore