Apple Silicon: Accelerate or openBLAS for Numpy

[mxp0wr]

What kind of feature would you like to request?

Additional function parameters / changed functionality / changed defaults?

Please describe your wishes

After naively updated my apple OS to tahoe, my UMAPs and leiden clustering result in slightly different outputs. I learned from chatGPT that this is probably due to an update of the accelerate BLAS that my Numpy accesses. I could use OpenBLAS which would be independent from Apple updates, but performs even more distinct.

What is the general recommendation for scanpy usage?

[flying-sheep]

That’s a good question, and I’d like to figure this out to give guidance, but I don’t understand what you mean by “learned from ChatGPT” – LLMs are bullshit, so unless you verified that, it’s a bit more likely to be correct than not, but effectively a guess.

numpy 2.x does access Apple’s Accelerate BLAS/LAPACK starting from macOS 14 (Sonoma), but I didn’t find anything about it changing behavior between macOS versions, do you have a link that confirms that?

Did you change your environment in any other way that could cause that? like switching from numpy 1 to numpy 2 or so?

[mxp0wr]

Thanks for the reply!
I agree that this of course is only a guess. But I did not change the environment. The output of the scRNAseq analysis was just altered by the update. ChatGPT helped me elucidate that numpy in the used env accessed the user/local/lib OpenBLAS ILP64. It suggested me to link numpy against conda-forge LP64 OpenBLAS ( via "libblas=*=*openblas"), integrating the BLAS into my environment.
It was NOT linked to accelerate before (got this wrong, haven't tried that at all).
Unfortunately, I don't have a backup of before the Tahoe update.

Bottom line: The Tahoe update somehow changed the UMAP distribution of cells slightly and Leiden clustering (left vs middle UMAP). As the BLAS linked against UMAP was in system library, it could have been changed by the update. To ensure reproducibility in future, I linked it against conda-forge BLAS in the env, which alter the clustering even more, but makes it hopefully stable in future.

Presentation1.pdf