Version
26.2
Which installation method(s) does this occur on?
Pip
Describe the bug.
In nx-cugraph==25.2, calling nx.betweenness_centrality() on a large directed graph (~75k nodes) without specifying k worked correctly within 15GB GPU VRAM. After upgrading to nx-cugraph-cu12 26.2.0, the same call raises an out-of-memory error unless k is explicitly passed.
The API states k=None means use all nodes as sources, and it should not silently require users to pass k to avoid OOM.
Minimum reproducible example
Working example:
!pip install --force-reinstall nx-cugraph-cu12==25.2.0 --extra-index-url https://pypi.nvidia.com
%env NX_CUGRAPH_AUTOCONFIG=True
import networkx as nx
print(f"using networkx version {nx.__version__}")
nx.config.warnings_to_ignore.add("cache")
G = pd.read_pickle(path)
nxcg_bc_results = nx.betweenness_centrality(G)
Not Working example:
!pip install nx-cugraph-cu12 --extra-index-url=https://pypi.nvidia.comRelevant log output
RuntimeError: non-success value returned from cugraph_betweenness_centrality: CUGRAPH_UNKNOWN_ERROR std::bad_alloc: out_of_memory: CUDA error (failed to allocate 17450397660 bytes) at: /__w/rmm/rmm/cpp/include/rmm/mr/cuda_memory_resource.hpp:51: cudaErrorMemoryAllocation out of memory
Other/Misc.
A large network with 75k nodes and 266261 edges to reproduce the error.
Code of Conduct
Version
26.2
Which installation method(s) does this occur on?
Pip
Describe the bug.
In nx-cugraph==25.2, calling
nx.betweenness_centrality()on a large directed graph (~75k nodes) without specifying k worked correctly within 15GB GPU VRAM. After upgrading tonx-cugraph-cu12 26.2.0, the same call raises an out-of-memory error unless k is explicitly passed.The API states k=None means use all nodes as sources, and it should not silently require users to pass k to avoid OOM.
Minimum reproducible example
Relevant log output
Other/Misc.
A large network with 75k nodes and 266261 edges to reproduce the error.
Code of Conduct