Cudnn conv cache key patch (#81418) (#81418) (#81888)
Summary:
Fixes #81106
Patches on cudnn algo cache to consider the right memory_format used in descriptors, instead of blindly copy the memory_format on inputs.
Note that to be on the safe side, we could actually cache on all tensor strides instead. But given how we short-cut and align memory_format from pytorch tensor to cudnn descriptor, it suffice to have a single field in the cache.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/81418
Approved by: https://github.com/ngimel
Test Plan: contbuild & OSS CI, see https://hud.pytorch.org/commit/pytorch/pytorch/ce2ce3ae9674f1f385d7706fded8b8ab99a031d7
Reviewed By: DanilBaibak
Differential Revision: D37847747
Pulled By: DanilBaibak
fbshipit-source-id: 1e5583e29f911d0987b6ff959886697a4fc853c7
Co-authored-by: jjsjann123 <jiej@nvidia.com>