Graham Markall
Graham Markall
@detrout - many thanks - the root cause is known - there's nothing to stop LLVM's RuntimeDyld component allocating memory for different objects that reference each other more than 4GB...
Somehow I missed the bot marking this a couple of weeks ago.
> Thanks for reopening. I have no idea how to help with this problem. Thanks for letting us know - it is quite a complex and frustrating general issue with...
From the apparent PC value (`0x0028000852800028` - similar as seen in https://github.com/numba/numba/issues/8738#issuecomment-1460047157) I think you are hitting https://github.com/llvm/llvm-project/issues/61402. To fix it, this commit: https://github.com/llvm/llvm-project/commit/2e1b838a889f9793d4bcd5dbfe10db9796b77143 would need backporting into Debian's LLVM...
Noted also in the triage discussion - we will need to hold a reference to referenced globals and closure variables to ensure their lifetime exceeds that of the dispatcher.
Are you caching the kernel?
No problem. A final patch implementing this would raise an error when attempting to cache a kernel that makes references to globals and/or closure variables.
@yxqd This is in the 0.59 milestone, so it should be part of that release.
0.59 seems to be aimed for 2023-12: https://github.com/numba/numba/issues/8971
@pentschev I'm a bit confused as to how you got that error. If you do `conda install cudatoolkit` it seems that you get 11.8 (because that's the last version before...