[BUG] AttributeError: 'NoneType' object has no attribute 'implementation' when using dask_cugraph.louvain

[co-tran-sp]

Describe the bug

The error happened when I tried to perform dask cugraph louvain on a symmetric digraph. The reason I need to use dask cugraph louvain because my graph is very big to load into the graph class without dask and from_dask_cudf_edgelist only supports DiGraph.

AttributeError Traceback (most recent call last)
/tmp/ipykernel_7871/183532962.py in
11 G = cugraph.DiGraph()
12 G.from_dask_cudf_edgelist(df_0, source='source', destination='target',edge_attr = 'similarity', renumber=False)
---> 13 df, mod = dask_cugraph.louvain(G)
14
15 t2 = time.time() - t1

~/anaconda3/envs/rapids-21.10/lib/python3.8/site-packages/cugraph/dask/community/louvain.py in louvain(input_graph, max_iter, resolution)
88
89 ddf = input_graph.edgelist.edgelist_df
---> 90 vertex_partition_offsets = get_vertex_partition_offsets(input_graph)
91 num_verts = vertex_partition_offsets.iloc[-1]
92 num_edges = len(ddf)

~/anaconda3/envs/rapids-21.10/lib/python3.8/site-packages/cugraph/dask/common/input_utils.py in get_vertex_partition_offsets(input_graph)
227 def get_vertex_partition_offsets(input_graph):
228 import cudf
--> 229 renumber_vertex_count = input_graph.renumber_map.implementation.ddf.
230 map_partitions(len).compute()
231 renumber_vertex_cumsum = renumber_vertex_count.cumsum()

AttributeError: 'NoneType' object has no attribute 'implementation'

Steps/Code to reproduce bug
edges = 1000000
vertices = 100000
source = np.random.randint(vertices,size = edges)
target = np.random.randint(vertices,size = edges)
sim = np.random.uniform(0,1,edges)
df = cudf.DataFrame()
df['source'] = source
df['target'] = target
df['similarity'] = sim
df = dask_cudf.from_cudf(df,npartitions=2)
G = cugraph.DiGraph()
G.from_dask_cudf_edgelist(df, source='source', destination='target',edge_attr = 'similarity', renumber=False)
_, mod = dask_cugraph.louvain(G)

Expected behavior
Should be able to run louvain A
Environment overview (please complete the following information)

• Environment location: [aws ec2 g4dn.12xlarge]
• Method of cuGraph install: [conda]

Additional context
Add any other context about the problem here.

[BradReesWork]

@co-tran-sp thanks for submitting the issue. We will dive into solving it

[jnke2016]

Hi @co-tran-sp .
I was able to reproduce the error. I would like to confirm that the number of partitions is the same as the number workers your client hold. A quick fix to your problem would be to not set renumber to False as all our MNMG agos require renumbering (for performance optimization) which is by default set to True internally. A PR should be out soon to return a proper error message when renumber happens to be False.

[co-tran-sp]

Hi all, thanks for the replies. @jnke2016 I was able to go around the error. Will there be support for undirected graph class using from_dask_cudf_edgelist? The digraph louvain process returns me very low modularity comparing to undirected graph class.