pre-compute approximate pairwise distances between vectors to use in cuML UMAP

[choucavalier]

hey there! i'm having perf issues because cuml.UMAP does not scale well with the size of my dataset.

probably because it performs a greed KNN algorithm, thereby computing billions of distances.

do you know if it's possible to use TorchPQ to pre-compute a sparse distance matrix i can pass to cuml.UMAP here?

[DeMoriarty]

Hi, this is currently not possible with TorchPQ.