[BUG] Lanczos solver isn't reproducible even with seed about raft HOT 3 CLOSED

Right now I'm sticking with random init. But will look into other places as well.

from raft.

@trivialfis, yeah that should work for now. Unfortunately, since it's too late to scrape through this and figure out the cause in time for 0.19, we should try to figure out what's going on here at some point because the spectral init has been shown to preserve more global structure than random init.

I found this bug by placing debug outputs in the spectral init for umap in cuml and found that executing it twice with the same random seed yielded different results. It could be as simple as a seed being overlooked somewhere, though I believe this was working before CTK 11.2 which makes me think something (or an assumption) has since broken, perhaps from one of the dependent libraries.

from raft.

Strangely, while reviewing rapidsai/cuml#3848 I no longer seem to be able to reproduce this issue using the spectral initialization, even w/ cuml's branch-0.20. I'm going to close this for now and reopen if I see the behavior again.

from raft.