Train More with Less: Fewer GPUs, Less Data Movement, Better Cluster UtilizationHyungyo KimApoorve Mohanet al.2026PyTorch Conference 2026Poster