[torch] set workspace size for cublas lt interface 1M (#73439)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/73439
Per discussion in https://github.com/pytorch/pytorch/issues/73328#issuecomment-1050422698
Workspace size is needed to get good cublas lt performance
Test Plan:
In PyTorch benchmark
python run.py nvidia_deeprecommender -d cuda -t train
Reviewed By: xuzhao9
Differential Revision: D34480690
fbshipit-source-id: 7a5fbbcf9e3503b6b08086612ff07ea6e2d8f748
(cherry picked from commit 158c498e106e411f00aecef3290bca46f33bbac9)
