trunk/fdbb719fcefb8b6f1490212896544b62f711eaad: [CUDA] Make graphs own annotation registry cleanup (#198418)
- PyTorch: 1453 events in the last 90 days
- PyTorch: 1434th Release in the last 90 days
- Previous: earlier the same day · ciflow/torchtitan/198536: [DTensor] Migrate unbind sharding propagation to single-dim strategy
What happened
Annotation cleanup depended on an active Cuspy observer and only tracked IDs at instantiation, leaking captures without a profiler or deferred graphs that were never instantiated. Track capture, body, and executable IDs on the graph and purge their entries before notifying profiler destroy hooks. Store nodes in per-graph buckets so rekeying moves one bucket instead of copying all live graphs, and destruction only vi…
Summary assembled by rule from the sources below