HUM AI AB Abhijeet Pendyala, Phd · TraceOpt Diagnosing a PyTorch DataLoader Bottleneck in a Real Training Run A ResNet-18 run wasted 43% of its GPU time on data loading. A three-line DataLoader change (num_workers, pin_memory, persistent_workers)…
HUM AI AB Abhijeet Pendyala, Phd · TraceOpt Two Clocks, One Training Step: How TraceML Measures PyTorch Performance A time.perf_counter() bracket and a CUDA event pair can disagree wildly about the same forward pass. Neither is wrong: they answer…
