Out of memory or OOM
A memory request cannot be satisfied. An OOM alone does not establish a leak; investigate the failed request and the workload state. See the OOM guide.
Allocation and tensor lifetime
An allocation reserves memory for use. A tensor’s lifetime describes how long it remains alive; surviving across steps can be intentional or a clue worth investigating. Memory timelines help inspect these patterns.
Allocated and reserved memory
Allocated tensor memory and allocator-reserved memory have different scopes. Reusable cached space can remain reserved. Read the measurement guide before comparing totals.
Peak memory
The highest observed usage during a measurement interval. A peak is meaningful only with its device, workload phase, and measurement method.
Call stack
The chain of function calls associated with an event. Allocation stacks provide code context; they do not by themselves establish why an object stayed alive.
Preflight check
An evaluation before committing to a longer run. It exercises selected configurations and inputs, so passing does not prove that every future run will succeed. Use the preflight checklist.
Checkpoint and restart
Saving training state and restoring it later to continue work. The state required depends on the training setup; a checkpoint does not by itself diagnose a failure. Compare diagnostic workflows.