viable/strict/1790274788: Validate packed RNN batch sizes before backend dispatch (#197560)
- PyTorch: 1453 events in the last 90 days
- PyTorch: 1434th Release in the last 90 days
- Previous: earlier the same day · viable/strict/1790256510: [Test] Make test_foreach device-agnostic for OOT backends (#191179)
What happened
Issue Fixes ROCm/TheRock#8337 . Summary Validate packed RNN batch_sizes at the shared ATen entry points before selecting cuDNN/MIOpen or the native implementation. This preserves the CPU-only contract when backend capability probing falls back to the native PackedLayer and prevents it from dereferencing device memory on the host. Extend the existing regression test to cover both the normal backend dispatch and the n…
Summary assembled by rule from the sources below