eqy

I agree on the warning part, additionally we could perhaps format the output as a log and add it to debug logging in apex.

> Also, can there be multiple kernels per fusion key or, since all the inputs are forced to contiguous, there's a single kernel only, and tuning is runtime at the...

> > My understanding is that there would only be one compilation per fusion key; > > I don't think that's true, at the very least there could be generated...

Some ballpark numbers on the basic compile-cache-reuse workflow on the small workload in the tests; first execution is around 0.5-0.9s on V100, with about 30-60us of that being the actual...

Thanks for bringing this up; would it be possible to provide a repro snippet for this issue? In the meantime, does `pip uninstall apex` work for you?

Timely. I think this might also fix an issue I was seeing when testing with `CudaMallocAsync`. Is the intent to effectively leak the workspaces at shutdown to avoid crashes when...

I'm not familiar with `ghstack` and its usage with reverted PRs; do I still need to resubmit?

(referencing #85447 which includes the fixes)

Looking at an example mismatch, previously numpy returned `nan+infj` and now it returns `-infj+infj`

@Skylion007 do we anticipate #122527 landing in the near-future? caffe2 fixes might be involved and look something like #115719 which might be unnecessary if caffe2 is being removed soon