ciflow/inductor/196856: Gate deterministic grid-sample backward on deterministic algorithms
- PyTorch: 959 events in the last 90 days
- PyTorch: 948th Release in the last 90 days
- Previous: earlier the same day · ciflow/inductor/196343: Update
What happened
Retain the original native atomic and cuDNN backward paths by default, and select fixed-order input-gradient accumulation only in deterministic mode. Share the native grid-gradient math through a grid-only specialization. Cover off, strict, and warn-only dispatch, document the opt-in behavior, and include the new test file in the CODEOWNERS taxonomy. Test Plan: spin lint -a CODEOWNERS aten/src/ATen/native/cuda/GridS…
Summary assembled by rule from the sources below