[BUG] Fix device access check in build_sorted_mst

[tarang-jain]

X may be in managed memory.
Fixes rapidsai/cuml#6948

[jinsolp]

Did you run into an issue with the current check? Because we previously discussed this issue when we were adding a hotfix for UMAP for the 25.06 release. (link).

[tarang-jain]

@jinsolp I just saw the hotfix. Yes I did run into an issue. In the cuml accel tests for hdbscan, the host side connect_knn_graph will be called, but the Mutual Reachabilty distance across components is actually embedded inside the reduction op, which is not handled by the host side code path at the moment (seems to be hard coded to pairwise).

Secondly, there is an edge case arising because of this code path, which is not handled by raft's mst solver. That edge case is that the mst solver is called on a graph that has only a single edge. I am opening a PR in raft to handle this.

The second one is causing the cuML optional check (rapidsai/cuml#6948) to fail.

[tarang-jain]

How about a boolean to force one code path over the other (when the X matrix is device accessible)?

[jinsolp]

How about a boolean to force one code path over the other (when the X matrix is device accessible)?

I see, and I agree. I guess this is what this PR is doing?

[tarang-jain]

@jinsolp I'll update this PR. At this time, it is forcing the device route if X is device accessible, whereas for UMAP you prefer to use the host path despite it being device accessible. A boolean override variable is the only way to satisfy both these scenarios.

[jinsolp]

LGTM!

[tarang-jain]

/merge