Estimate, then measure, the training memory of a small transformer

Write a per-component memory estimate before running, then reconcile it with peak allocated memory and a memory snapshot.

The exercise happens on Self-directed (PyTorch memory tools). Canonical link: https://docs.pytorch.org/docs/stable/torch_cuda_memory.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. Prepares the “will this fit, and what would you change?” question.