Profile a training step and attribute its time

Attribute step time to data loading, host-to-device copies, compute kernels, communication and idle gaps using a profiler timeline, then fix the largest item.

The exercise happens on Self-directed (PyTorch profiler). Canonical link: https://docs.pytorch.org/tutorials/recipes/recipes/profiler_recipe.html. Link check: link responded (5 Oct 2026). Opened from this app 0 times; opening is not counted as an attempt.

Supports: Needs a GPU; a multi-GPU machine adds the communication part. Builds the habit of measuring before optimizing.