PyTorch training optimizations: 5× faster with GPU profiling and memory analysis