Fix 2.8 issue per sample grad #4921

Job	Run time
pytorch_tutorial_build_worker (4, 15, linux.g5.4xlarge.nvidia.gpu)	31m 22s
pytorch_tutorial_build_worker (7, 15, linux.g5.4xlarge.nvidia.gpu)	23m 47s
pytorch_tutorial_build_worker (5, 15, linux.g5.4xlarge.nvidia.gpu)	24m 51s
pytorch_tutorial_build_worker (1, 15, linux.g5.12xlarge.nvidia.gpu)	16m 17s
pytorch_tutorial_build_worker (3, 15, linux.g5.4xlarge.nvidia.gpu)	23m 29s
pytorch_tutorial_build_worker (6, 15, linux.g5.4xlarge.nvidia.gpu)	41m 18s
pytorch_tutorial_build_worker (13, 15, linux.g5.4xlarge.nvidia.gpu)	17m 32s
pytorch_tutorial_build_worker (15, 15, linux.g5.4xlarge.nvidia.gpu)	21m 23s
pytorch_tutorial_build_worker (9, 15, linux.g5.4xlarge.nvidia.gpu)	21m 23s
pytorch_tutorial_build_worker (12, 15, linux.g5.4xlarge.nvidia.gpu)	18m 39s
pytorch_tutorial_build_worker (2, 15, linux.g5.4xlarge.nvidia.gpu)	38m 34s
pytorch_tutorial_build_worker (10, 15, linux.g5.4xlarge.nvidia.gpu)	15m 49s
pytorch_tutorial_build_worker (14, 15, linux.g5.4xlarge.nvidia.gpu)	29m 26s
pytorch_tutorial_build_worker (11, 15, linux.g5.4xlarge.nvidia.gpu)	16m 13s
pytorch_tutorial_build_worker (8, 15, linux.g5.4xlarge.nvidia.gpu)	26m 53s
pytorch_tutorial_build_manager	0s
	6h 6m 56s

Provide feedback