Training job memory leak

It's the night before a quarterly model-refresh deadline. The distributed training job for the next ranking model has crashed three times in the last hour: each run gets ~60% through an epoch, then a worker hits CUDA OOM and the whole job dies, losing progress. Dashboards: per-GPU memory climbs across the run rather than staying flat, the dataset grew ~30% this quarter, and a recent commit added two new high-cardinality embedding features. There's no checkpoint newer than the start of the failing epoch. How do you get a good model trained before the deadline, and what do you fix?