What Really Determines the Speed of Your PyTorch Code?

PyTorch GPU kernels launch asynchronously, so naïve Python timing measures CPU scheduling—not GPU work. This guide shows how to benchmark correctly using CUDA events, synchronization, warmups, and (optionally) L2 cache flushing, plus Triton’s do_bench and CUDA graphs to reduce CPU overhead. It also argues that realistic benchmarks must reflect production data patterns, illustrated with token routing imbalance in MoE grouped GEMM.