Mixed precision, gradient accumulation and checkpointing experiments

Check that accumulated gradients match a large batch, provoke and observe fp16 overflow with loss scaling, and measure checkpointing’s memory and time cost.

The exercise happens on Self-directed (PyTorch AMP). Canonical link: https://docs.pytorch.org/docs/stable/amp.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 with fp16 or bf16 support. Turns precision and memory knobs into observed behavior.