contentis commented on issue #49437:
URL: https://github.com/apache/arrow/issues/49437#issuecomment-5347007240
I started to look into this and from what I can tell most of the reported
issues arrise from "legacy" handling of CUDA context in numba-cuda<0.28 where
`.value` was being used vs. the new (arguably better way) of using the
CUContext objects from the cuda.bindings.- Writing a compatibility wrapper for
both representations should resolve this. However, my honest preference would
be to require numba-cuda >=0.28 and rely entirely on native CUcontext to keep
the code lean and clean.
Supporting both would require some handler like
```
def unwrap_numba_handle(handle):
if hasattr(handle, "value"): # legacy
return handle.value
return get_cuda_native_handle(handle) # new >= 0.28
def wrap_numba_handle(address):
....
```
Would this be acceptable, or are there strong reasons to retain
compatibility with numba-cuda<0.28?
Similarly, I think it may be worth dropping CUDA 11.x coverage and focusing
on CUDA 12.x and 13.x instead, especially since current numba-cuda no longer
provides a CUDA 11 extra.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
