pytorch

c28bb464 - Fix test_mem_efficient_attention_vs_math_ref_grads tolerance from test_transformers.py (#108094)

Fix test_mem_efficient_attention_vs_math_ref_grads tolerance from test_transformers.py (#108094) Tolerance currently too low, triggering test failures via numerical mismatch in NVIDIA internal testing for certain H100, A16, A40 configs. cc: @ptrblck @eqy Pull Request resolved: https://github.com/pytorch/pytorch/pull/108094 Approved by: https://github.com/eqy, https://github.com/msaroufim