internal: https://fb.workplace.com/groups/1075192433118967/permalink/1599802033991335/ Assuming flags don't change during compilation, we shouldn't allow incompatible configs to be set at torch.compile wrap time. Not in this PR: For flags that need to change during compilation, we'd have to be strict about where they can be used in the compile lifecycle Pull Request resolved: https://github.com/pytorch/pytorch/pull/146513 Approved by: https://github.com/williamwen42 Co-authored-by: Gabriel Ferns <gabeferns@meta.com>