make_heap code object inequality at import time

[JesseRMeyer]

In attempting to go through the introductory material for UMAP, I hit the following exception with import umap:

<code object <module> at ..., file "/path/to/keras/lib/python3.12/path/to/pynndescent/sparse_nndescent.py", line 1> != <code object make_heap at ..., file "/path/to/keras/lib/python3.12/path/to/pynndescent/utils.py", line 171>.

I'm at a loss on how to debug this. Note that the venv here is called "keras".

WSL2
Python 3.12
scikit-learn 1.6
umap-learn 0.5.7
scipy 1.14.1
pynndescent 0.5.13
numpy 2.0.2
numba 0.60.0
keras 3.6.0

[lmcinnes]

I have never seen this before. I'll have to see if I can reproduce anything like it locally.

[JesseRMeyer]

A member of our group reports that a very similar setup with Python 3.10 works as expected (both on Native Windows and through WSL).

[erathorn]

I have the same issue on my side with the following specifications:

• Linux
• Python 3.12
• scikit-learn=1.6.0
• scipy=1.15.0
• umap-learn=0.5.7
• pynndescent=0.5.13
• numpy=2.0.2
• numba=0.60.0

I installed everything via mamba install -c conda-forge umap-learn.

Maybe this helps in tracking down the issue.

[ratthapon]

I have this problem when I run a file using debugger (VS Code). Is there any recommended debugger?

[flying-sheep]

Yeah, without being able to use a debugger, I can’t maintain anything …

This seems to be the asssertion that gets triggered: https://github.com/microsoft/debugpy/blob/39879bdeeae5908a8f3527c77ba252e0930afabc/src/debugpy/_vendored/pydevd/_pydevd_sys_monitoring/_pydevd_sys_monitoring_cython.pyx#L562

Workaround: microsoft/debugpy#1719

Current workaround is to downgrade either numba to 0.59.1 or python debugger to v2024.10.0

(I think using Python 3.11 also works)

Upstream issue: numba/numba#9760