warnings.warn is super spammy under Dynamo about pytorch HOT 5 OPEN

I debugged this some more:

• Python's warnings.warn has a cache (globals()['__warningregistry__']). It will only warn once as long as the warning was in the cache.
• The cache can be invalidated (you can see this in print(globals()['__warningregistry__']): the cache has a version number). It can be invalidated if any new filters are added, or if someone calls warnings.catch_warnings()
• Dynamo adds a new filter here and FakeTensor calls catch_warnings() here. These are for "hiding warnings"

Taking a step back: what should the behavior of the above script actually be? Should we expect torch.compile'd code to reproduce warnings faithfully, or is it fine if there are a lot more warnings than expected?

IMO we should reproduce the warnings faithfully (warning spam is bad UX), but this means that we need to "hide warnings" in a different way.

from pytorch.

This one is kind of funny -- if you use Dynamo, then we will never show you distributed warnings again, even in eager-mode :P

from pytorch.

I think we should reproduce the warnings faithfully. Pages and pages of spam could lead to bad UX.

from pytorch.

Haha ... I might have added that long time ago when I had lil understanding of what's really going on. We can remove these, some of the dynamo wrapped tests might fail.

from pytorch.

I'm attracted to beefed up warning handler context manager that restores the cache when done

from pytorch.