[ANN] Numba 0.56.0 RC1

Dear all,

on behalf of the Numba crew I am happy to announce the availability of Numba 0.56.0 RC1 (Release Candidate 1) with a significant number of improvements!!

As usual, you can install this release candiate with conda from the numba channel on anaconda.org:

conda install -c numba numba=0.56.0rc1

Or alternatively install the wheels form PyPi with pip (note the need to install llvmlite individually):

# install llvmlite from anaconda.org first, no wheels on PyPi for 0.39.0rc1
pip install -i https://pypi.anaconda.org/numba/label/wheels/simple llvmlite==0.39.0rc1
# then install Numba
pip install numba==0.56.0rc1

You can find the release notes at: Release Notes — Numba 0.56.0rc1+0.gea8407264.dirty-py3.7-linux-x86_64.egg documentation

Please do test this release candidate with your libraries and applications and report back any issues you may encounter here, thank you!


cc @RC_Testers the RC1 is out now, please do test this and provide feedback about it working for you or not (mention any issues you encounter). Thank you!

Just tested on RBC and it’s working as expected.

1 Like

Tested the Bodo test suite with Python 3.8, 3.9, and 3.10. Everything is working as expected.

Dear all,

I am pleased to report that one of the test suites in my [ht](https://ht.readthedocs.io/) library for heat transfer (solving conduction and convection equations and so on) is again passing with this release candidate.

A single function with 17 arguments caused the error a Python release or so ago. That many arguments is a code smell to some people, but there are real engineering problems where that many arguments is a sensible choice. In this case, the multitude of arguments are turned into what are called “dimensionless numbers”, variable without units that can give greater confidence in predicting phenomenon outside measured experiments. In that correlation, the function takes both liquid and gas inputs for several properties, which double the number of inputs.

Thank you to everyone working on Numba for being so capable, and prioritizing fixing regressions so much! It makes Numba a very reliable tool.


Tested with latest poliastro, after adding NPY_DISABLE_CPU_FEATURES="AVX512_SKX" (because of Numerical differences in latest numba with NumPy 1.21 vs 1.22 · Issue #8141 · numba/numba · GitHub) it all works :rocket:

I tested set_parallel_chunksize which is beneficial, especially for computing the Mandelbrot set. Moreover, I tested both omp and tbb threading layers. Unfortunately, there’s no effect for set_parallel_chunksize using the omp layer.

I’m seeing a deprecation notice using the omp threading layer. There is an existing issue and commented there.