pytorch

3a5f8546 - [pytorch] fewer cuda sync in unique by using cub instead of thrust (#57323)

[pytorch] fewer cuda sync in unique by using cub instead of thrust (#57323) Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/57323 Use cub library instead of thrust to reduce # of cuda stream synchronize. Reviewed By: ngimel Differential Revision: D28088029 fbshipit-source-id: b616294cd776aa5643c153e172258a0153a42b6a