trunk/c262fc64bd3a40834366246d22e9e7575f219a16: [CUDA] Capture Python launch stacks for CUDA graphs (#198419)
- 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
Add record_py_stacks to CUDA graph annotation_config and expose live stack lookup and gzip JSON export. CUPTI node-creation callbacks record bounded user launch stacks, sharing strings within a capture. The graph-owned registry from the preceding commit handles rekeying and cleanup without scanning other live graphs; only source-less aliases need additional stack handling here. By default, omit framework and generat…
Summary assembled by rule from the sources below