benchmark

8a4b1540 - Apply same 'pick_grad' on generating fp64 reference outputs (#111593)

Apply same 'pick_grad' on generating fp64 reference outputs (#111593) Summary: To lower memory consumption for inference mode. X-link: https://github.com/pytorch/pytorch/pull/111593 Approved by: https://github.com/msaroufim, https://github.com/thiagocrepaldi ghstack dependencies: #111867 Reviewed By: izaitsevfb Differential Revision: D50626494 fbshipit-source-id: c36dc14493c1013ef87c12001c23bb454396eaff