Mike Ruberry

I added some inline comments for your review, @nkaretnikov. `addcdiv` is an interesting operation, and we may want to update its ATen behavior

> @mruberry do i need to report the [xla promotion failure](https://github.com/pytorch/pytorch/runs/7323967829?check_suite_focus=true) to the xla repo and skip this test on xla? or is `promote_integer_inputs_to_float`, which i'm enabling here for `addcdiv`,...

A difference between 1.9 and 1.9.1 seems very odd to me as only a few critical fixes where cherrypicked for 1.9.1. Or is the comparison between 1.8 and 1.9.1?

Thanks for debugging @IvanYashchuk, would you file a more detailed issue about what's causing the problem in the PyTorch GitHub so we can review it there?

We'll have an update on the TF32->FP32 default soon, too.

XLA and nvFuser failures look real and see suggested doc change inline. I suggest xfailing those tests for now and citing two issues.

Overall this looks good, but some of the test failures seem relevant and we should let @zou3519 have a look, too

Build errors are real on this pr, too. Should we delay removing torch.lstsq and update our current deprecation guidance if torch.lstsq is deterministic and torch.linalg.lstsq is not? If that's the...

> > Then `var` can and should remain prim. > > @ngimel, I understand that it _can_ remain prim, but I don't see why it _should_ remain to be prim....

Since the nan* reductions, like [nansum](https://pytorch.org/docs/master/generated/torch.nansum.html?highlight=nan#torch.nansum), are existing masked reductions we should be sure the semantics are equivalent. This proposal just allows the mask to be specified directly rather than...