Memory Leak Hunt
Memory that only ever climbs is a reference you forgot to let go of.
Key Insight
A memory leak in PyTorch usually means you kept a reference to the loss or another piece of the computation graph across iterations, so the memory it holds can never be freed. Comparing memory snapshots taken over many steps exposes the slow, steady climb.
Why This Matters
A leak that adds a few megabytes per step runs fine for an hour and then crashes with an out-of-memory error. Snapshots turn that mysterious late crash into an obvious upward line you can trace back to its source.