Is this a new feature, an improvement, or a change to existing functionality?
New Feature
How would you describe the priority of this feature request
Medium
Please provide a clear description of problem this feature solves
cugraph.louvain produces non-deterministic partitions across repeated runs on the same graph, same RAPIDS version, same GPU. cugraph.leiden already accepts a random_state argument, but cugraph.louvain exposes no equivalent, so there is no way to obtain reproducible Louvain results.
cugraph version: 26.04.00 (also seen on earlier releases)
Reproducibility evidence. Running Louvain 6× on a fixed ~2k-node weighted graph (community-detection step of a gene-program pipeline): - 5/6 runs produced an identical partition - 1/6 collapsed ~15 communities into their neighbors (different community count)
The same workload under cugraph.leiden(random_state=42) was bit-identical across all 6 runs. The downstream effect in our case was a ~20% swing in the number of detected communities from one Louvain run to the next.
Describe your ideal solution
A way to make cugraph.louvain reproducible, ideally:
random_state parameter on cugraph.louvain (and pylibcugraph.louvain), matching the existing cugraph.leiden signature, to seed any RNG-driven tie-breaking; and/or- A documented deterministic mode — since Louvain's variation also stems from parallel vertex-move ordering and non-associative floating-point reductions of modularity gains, a seed alone may not be sufficient. An optional flag that fixes traversal order / uses deterministic reductions (accepting a performance cost) would let users trade speed for reproducibility when needed.
API parity with cugraph.leiden is the minimum bar; even just documenting which sources of non-determinism a seed would and would not address would help.
Describe any alternatives you have considered
No response
Additional context
Use case: community detection on weighted similarity graphs in a single-cell genomics pipeline, where reproducible runs are required for published results. Happy to share a minimal reproducer graph if useful.
Code of Conduct
Is this a new feature, an improvement, or a change to existing functionality?
New Feature
How would you describe the priority of this feature request
Medium
Please provide a clear description of problem this feature solves
cugraph.louvainproduces non-deterministic partitions across repeated runs on the same graph, same RAPIDS version, same GPU.cugraph.leidenalready accepts arandom_stateargument, butcugraph.louvainexposes no equivalent, so there is no way to obtain reproducible Louvain results.cugraph version: 26.04.00 (also seen on earlier releases)
Reproducibility evidence. Running Louvain 6× on a fixed ~2k-node weighted graph (community-detection step of a gene-program pipeline): - 5/6 runs produced an identical partition - 1/6 collapsed ~15 communities into their neighbors (different community count)
The same workload under
cugraph.leiden(random_state=42)was bit-identical across all 6 runs. The downstream effect in our case was a ~20% swing in the number of detected communities from one Louvain run to the next.Describe your ideal solution
A way to make
cugraph.louvainreproducible, ideally:random_stateparameter oncugraph.louvain(andpylibcugraph.louvain), matching the existingcugraph.leidensignature, to seed any RNG-driven tie-breaking; and/orAPI parity with
cugraph.leidenis the minimum bar; even just documenting which sources of non-determinism a seed would and would not address would help.Describe any alternatives you have considered
No response
Additional context
Use case: community detection on weighted similarity graphs in a single-cell genomics pipeline, where reproducible runs are required for published results. Happy to share a minimal reproducer graph if useful.
Code of Conduct