ashvardanian
Python test_intersect fails on an AWS c7g instance (Graviton3)
Looks like we have a fail on graviton 3. A t4g.small instance is fine, but with c7g the pytest fails on main-dev
================================================================ short test summary info ================================================================
SKIPPED [600] python/test.py:42: SciPy Mahalanobis distance returned NaN due to `sqrt` of a negative number
SKIPPED [300] python/test.py:388: Problems inferring the tolerance bounds for numerical errors
FAILED python/test.py::test_intersect[10-100-uint32-2-100] - AssertionError: Missing [ 6 26 40 41 51 54 73 76 98 119 153 167 189] from [ 3 6 7 3 7 19 21 23 25 26 25 30 31 34 35 36 ...
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! stopping after 1 failures !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
=============================================== 1 failed, 3918 passed, 900 skipped, 193 warnings in 4.84s ===============================================
The setup script:
sudo apt-get update && sudo apt-get install cmake build-essential libjemalloc-dev g++-13 gcc-13 -y
git clone https://github.com/ashvardanian/SimSIMD.git
cd SimSIMD/
git checkout main-dev
sudo apt install python3.12-venv python3-dev -y
source ~/env/bin/activate
pip install pytest pytest-repeat setuptools numpy scipy
python setup.py install --force
pytest python/test.py -s -x -Wd
You can skip the test and see the rest pass with pytest python/test.py -s -x -Wd -k "not test_intersect"
Yes, that's most likely an algorithmic issues. I am actively working on a much improved set intersection algorithm (https://github.com/ashvardanian/SimSIMD/commit/aba39868075e3dffca0daecd2ebb0f9528feea1a and in subsequent commits) and would postpone resolving this specific issue until we merge the new solution 🤗
For now we can just comment it as experimental and redirect from Arm to serial code every time. What do you think, @MarkReedZ?
Are we okay just leaving this as-is and using -k "not test_intersect" ? Assuming no one else will bump into this until the new code is in.
For now we can just comment it as experimental and redirect from Arm to serial code every time.
This way no -k is needed 🤗