seperate atol and rtol in correct check and read tol from os.env (#1072)
Summary:
Some backends may not have enough precision for some tests, so can we add these environment variables("TORCHBENCH_ATOL" and "TORCHBENCH_RTOL") to dynamicly contol tolerance of correctness_check?
Pull Request resolved: https://github.com/pytorch/benchmark/pull/1072
Reviewed By: FindHao
Differential Revision: D38437225
Pulled By: xuzhao9
fbshipit-source-id: f0ee653347484ec8d96233c10364427f4d6f90eb