Update GNN sampling post processing functions to optionally consider seed vertices in renumbering

[seunghwak]

Upon cugraph-pyg requests, this PR updates GNM sampling post processing functions to optionally consider seed vertices in renumbering.

Previously, seed vertices with 0 neighbors (to sample from) are ignored in renumbering. Once this PR is merged and if seed vertices is provided, seed vertices (regardless of whether they appear in the sampled edge list or not) will have a smaller vertex ID than vertices that first appear as seed vertices' neighbors or later hops.

Marked as non-breaking assuming that this functions are only internally used.

[alexbarghi-nv]

I'm tentatively approving this, will give final approval once I test

[ChuckHastings]

Should be able to merge once @alexbarghi-nv approves. Let us know how testing goes.

[alexbarghi-nv]

I'm seeing a bug where memory addresses are getting corrupted. Not sure if it's on my end or if it's something this PR introduced. Still troubleshooting. Will post a MRE if I am able to isolate it to something in this PR.

[seunghwak]

I'm seeing a bug where memory addresses are getting corrupted. Not sure if it's on my end or if it's something this PR introduced. Still troubleshooting. Will post a MRE if I am able to isolate it to something in this PR.

Have you set do_expensive_check to true?

[alexbarghi-nv]

I'm seeing a bug where memory addresses are getting corrupted. Not sure if it's on my end or if it's something this PR introduced. Still troubleshooting. Will post a MRE if I am able to isolate it to something in this PR.

Have you set do_expensive_check to true?

I'll try that

[alexbarghi-nv]

The error always occurs after the first call. The first call runs without any issues and produces the correct result. All subsequent calls either produce incorrect/garbage results, or fail with a memory access or frontier out of range error. When do_expensive_check is enabled, I get this error when trying to renumber the start vertices: Invalid input arguments: vertices have elements that are missing in (aggregate) renumber_map_labels.

[alexbarghi-nv]

It looks like it may be something I introduced in my PR

[alexbarghi-nv]

I couldn't replicate this with only the code in this PR

[alexbarghi-nv]

I'm seeing a bug where memory addresses are getting corrupted. Not sure if it's on my end or if it's something this PR introduced. Still troubleshooting. Will post a MRE if I am able to isolate it to something in this PR.

I have now isolated and fixed the bug. It was an issue with data being improperly copied in my PR.

[alexbarghi-nv]

Small change request. Otherwise good.

[alexbarghi-nv]

Approved, but we can't integrate this with dask+Python until that code gets refactored (#4358)

This is ready to go for the new distributed sampling framework, though, which just uses pylibcugraph and NCCL.

👍

[alexbarghi-nv]

/merge