cuGraph Dask edge-weight behavior in multi-GPU algorithms: clarification requested

Hi NVIDIA / RAPIDS Developer Community,

I raised a RAPIDS cuGraph GitHub issue asking for clarification on edge-weight behavior in Dask / multi-GPU cuGraph:

The main question is whether weighted-edge behavior in Dask cuGraph is:

  1. fully supported at graph-construction level,
  2. algorithm-dependent at execution level, or
  3. unsupported for specific distributed algorithms.

From the documentation, Dask graph construction appears to support weighted edge lists through from_dask_cudf_edgelist() using edge_attr or weight.

However, algorithm-level behavior appears to vary. Some algorithms, such as weighted SSSP, appear to use edge weights, while other distributed algorithms may ignore weights by design.

This matters for graph-AI use cases such as:

  • fraud detection
  • AML
  • mule-account detection
  • payment-risk analytics
  • transaction-network analysis

In these use cases, edge weights often represent transaction amount, frequency, risk score, or relationship strength.

I would appreciate guidance from RAPIDS/cuGraph users or maintainers on:

  1. Whether edge weights are fully preserved in Dask cuGraph graph construction.
  2. Which Dask / multi-GPU cuGraph algorithms use edge weights.
  3. Which algorithms intentionally ignore edge weights.
  4. Whether there is an official weighted-edge support matrix for single-GPU vs Dask / multi-GPU cuGraph.
  5. Whether this would be useful as a small regression/documentation test case.

GitHub issue:

Thanks for any clarification or pointers.