We are inviting the boarder community and downstream projects to help us test Numba 0.52.0 and llvmlite 0.35.0 release candidates.
We would appreciate downstream projects to test our release candidates in their respective CI infrastructure and report regressions to our issue tracker.
We are also experimenting with the idea of getting sign-offs from the following opensource projects that have, in previous releases, reported regression issues:
Awkward
UMAP
librosa
RAPIDS
SDC
pydata sparse
clifford
Please comment in this thread if you can help with testing our packages and mention the project that you are associated with (if you can).
We are currently at RC2. Conda packages are in the numba channel. Wheels are available on PyPI.
Thanks in advance for everyone helping us test Numba.
I ran the Awkward Array integration tests, which includes all of our tests of Numba’s extension mechanism (extending it to recognize and correctly navigate Awkward Arrays). There were no errors.
Here’s the version information:
Python 3.8.6 | packaged by conda-forge | (default, Oct 7 2020, 19:08:05)
[GCC 7.5.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import numba
>>> numba.__version__
'0.52.0rc2'
I ran the poliastro test suite with numba 0.52.0rc2 and all tests passed without errors. I unfortunately don’t have the capacity to run proper time benchmarks today, but from a couple of quick runs with both versions I don’t see much difference in performance.
I’ve tested RBC with Numba 0.52 RC2 and there were no errors
~/git/rbc:rbc-38 (rbc) guilhermel $ python
Python 3.8.6 | packaged by conda-forge | (default, Oct 7 2020, 19:08:05)
[GCC 7.5.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import numba
nu>>> numba.__version__
'0.52.0rc2'
Please see the RC3 milestone for a list of changes: https://github.com/numba/numba/milestone/49?closed=1. The changes in RC3 are minor or opaque to user-facing API—adjustment of optimization passes, changes of internal C code.
I’ve tested locally with the RAPIDS libraries branch-0.17 (cuDF, cuSpatial, cuML, cuGraph) with single- and multi-GPU with 0.52.0RC3 and not encountered any issues.