[NNC] Fix lowering of aten::pow #47795
Conversation
facebook-github-bot
added
cla signed
oncall: jit
💊 CI failures summary and remediationsAs of commit 21c6a78 (more details on the Dr. CI page):
🚧 2 fixed upstream failures:These were probably caused by upstream breakages that were already fixed.
Please rebase on the
|
|
|
||
| @torch.jit.script | ||
| def do_exp(x, y, z): | ||
| return ((x * y) * 2) * torch.pow(z, 2) |
There was a problem hiding this comment.
Nit: can the expression be simplified? I guess what really matters here is just torch.pow(z, 2)?
There was a problem hiding this comment.
It could be, but this is equivalent to the repro in the bug report.
There was a problem hiding this comment.
@nickgg has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.
Summary: NNC lowering of aten::pow assumes that the types of the exponent is either float or int cast to to float, which doesn't work great with double (or half for that matter). Fixes pytorch#47304 Pull Request resolved: pytorch#47795 Reviewed By: ZolotukhinM Differential Revision: D24904201 Pulled By: nickgg fbshipit-source-id: 43c3ea704399ebb36c33cd222db16c60e5b7ada5
NNC lowering of aten::pow assumes that the types of the exponent is either float or int cast to to float, which doesn't work great with double (or half for that matter).
Fixes #47304