Fix 2.8 issue per sample grad #4871

Job	Run time
pytorch_tutorial_build_worker (3, 15, linux.g5.4xlarge.nvidia.gpu)	23m 24s
pytorch_tutorial_build_worker (2, 15, linux.g5.4xlarge.nvidia.gpu)	37m 49s
pytorch_tutorial_build_worker (6, 15, linux.g5.4xlarge.nvidia.gpu)	41m 18s
pytorch_tutorial_build_worker (7, 15, linux.g5.4xlarge.nvidia.gpu)	23m 28s
pytorch_tutorial_build_worker (10, 15, linux.g5.4xlarge.nvidia.gpu)	15m 43s
pytorch_tutorial_build_worker (15, 15, linux.g5.4xlarge.nvidia.gpu)	21m 8s
pytorch_tutorial_build_worker (8, 15, linux.g5.4xlarge.nvidia.gpu)	26m 34s
pytorch_tutorial_build_worker (1, 15, linux.g5.12xlarge.nvidia.gpu)	16m 8s
pytorch_tutorial_build_worker (13, 15, linux.g5.4xlarge.nvidia.gpu)	17m 18s
pytorch_tutorial_build_worker (9, 15, linux.g5.4xlarge.nvidia.gpu)	20m 59s
pytorch_tutorial_build_worker (11, 15, linux.g5.4xlarge.nvidia.gpu)	15m 58s
pytorch_tutorial_build_worker (5, 15, linux.g5.4xlarge.nvidia.gpu)	24m 30s
pytorch_tutorial_build_worker (12, 15, linux.g5.4xlarge.nvidia.gpu)	17m 33s
pytorch_tutorial_build_worker (14, 15, linux.g5.4xlarge.nvidia.gpu)	29m 29s
pytorch_tutorial_build_worker (4, 15, linux.g5.4xlarge.nvidia.gpu)	35m 18s
pytorch_tutorial_build_manager	-1s
	6h 6m 36s

Provide feedback