graph is the final output exposed at the python level. It looks like umap-learn runs the fuzzy simplicial operation (link, corresponding to our FuzzySimplSetImpl::symmetrize, then eliminates zeros, which becomes the final output.

I think if we doperform_thresholding here, we end up storing the graph after thresholding as the final graph, which seems different from that umap-learn has?
I understand that we need to do the thresholding before the embedding init, but think by doing this we end up with a different output graph (which should correspond to umap-learn's self.graph_). Please correct me if I'm wrong!

My thinking was that once the graph reaches the Python layer, we can no longer make assumptions about how the user intends to use it—so we should avoid modifying or potentially corrupting it. Since our goal is to trim the graph, that would require creating a copy, which could significantly increase VRAM usage. I acknowledge this approach doesn't exactly align with how umap-learn handles it, but I assumed that most users would primarily be interested in a fuzzy simplicial set graph that retains only the most important connections. I had also even considered performing the trimming before the symmetrization step, since the thrust operations there appears to cause a memory spike—if I’ve understood correctly.

However, I just realized that because the trimming threshold depends on the number of training epochs, trimming before storing the graph might work well with the UMAP estimator, but not as effectively with fuzzy_simplicial_set, which lacks access to the number of epochs. In that case, creating a copy might indeed be the only viable option—unless the graph is stored as a SciPy array on the Cython side?

In that case, creating a copy might indeed be the only viable option—unless the graph is stored as a SciPy array on the Cython side?

When Corey and I discussed this a month or two ago, IIRC we came to agreement that the graph_ attribute could (and probably should) be returned to the user on host. It's mostly there for introspection or for umap-learn compatibility. We do want to make sure the graph is being treated roughly the same as umap-learn does so zero-code-change methods that reference it (e.g. merging of estimators) work roughly the same. Since we never reference it ourselves, having it on host would be fine and would decrease the device memory pressure at this point.