[BUG] UMAP results broken compared to scanpy

[j-bac]

Describe the bug
Given same UMAP and neighbors graph search parameters, UMAP results look quite different between rapids_singlecell UMAP:

And scanpy.tl.umap:

Steps/Code to reproduce bug
In general UMAPs seem more unstable on GPU and the visualization is often affected by few distant and isolated points. I am not sure whether the problem comes from different kNN search results or different UMAP implementation

Expected behavior
Similar UMAPs to scanpy implementation.

[Intron7]

Thanks for raising this! This is related to #196, where we discussed differences between cuML’s UMAP and umap-learn. Neither rapids-singlecell nor Scanpy directly implement the UMAP algorithm; both rely on these underlying libraries. Some differences are expected due to variations in initialization, optimization strategies, and parallelism in cuML.

If you believe this is an issue with cuML’s UMAP itself rather than its usage in rapids-singlecell, I’d recommend raising it directly with the cuML team at rapidsai/cuml. If there’s a specific issue affecting integration, feel free to provide a minimal reproducible example.

Closing this for now, but happy to reopen if there's something actionable on our end.

[wphicks]

With rapidsai/raft#2568 merged, this may have been addressed in the cuML 25.04 nightly release. If not, please do let us know over at https://github.com/rapidsai/cuml/issues, and the RAPIDS team will look into it.