[BUG] Passing 'groups' parameter to rapids_singlecell.tl.rank_genes_groups gives all 0 scores/NA LFCs

[dguin]

Describe the bug

When calling rsc.tl.rank_genes_groups with the groups=[...] parameter and method='t-test' or method='wilcoxon', the resulting scores, p-values, and log fold changes are all zero or NaN (scores are okay for wilcoxon). Calling the same function on the same AnnData without the groups= parameter (computing all groups) returns correct, non-degenerate values for every group including the one that was previously broken. The method='wilcoxon' path is partially affected: scores compute correctly with groups=, but pvals, pvals_adj, and logfoldchanges are still all zero / NaN.

scanpy's sc.tl.rank_genes_groups on the identical AnnData with the same groupby and groups= arguments returns correct values, so the data itself is valid.

Steps/Code to reproduce bug

import cupy as cp
import cupyx.scipy.sparse as cusp
import anndata as ad
import numpy as np
import pandas as pd
import rapids_singlecell as rsc
import scanpy as sc

n_cells, n_genes = 5000, 1000
X = cusp.random(n_cells, n_genes, density=0.1, dtype=cp.float32, format='csr')
obs = pd.DataFrame({
    'group': pd.Categorical(np.random.choice(list('ABCDE'), n_cells))
})
adata = ad.AnnData(X=X, obs=obs)
rsc.pp.normalize_total(adata)
rsc.pp.log1p(adata)

# Works: scores, pvals, lfc all populated correctly for every group
rsc.tl.rank_genes_groups(adata, groupby='group', method='t-test', use_raw=False)
print("All groups (no `groups=`):")
print(pd.DataFrame(adata.uns['rank_genes_groups']['scores']).head())
print(pd.DataFrame(adata.uns['rank_genes_groups']['pvals']).head())
print(pd.DataFrame(adata.uns['rank_genes_groups']['logfoldchanges']).head())

# Broken: scores all 0, pvals all 0, logfoldchanges all NaN
rsc.tl.rank_genes_groups(adata, groupby='group', groups=['A'],
                         method='t-test', use_raw=False)
print("\nWith `groups=['A']`:")
print(pd.DataFrame(adata.uns['rank_genes_groups']['scores']).head())
print(pd.DataFrame(adata.uns['rank_genes_groups']['pvals']).head())
print(pd.DataFrame(adata.uns['rank_genes_groups']['logfoldchanges']).head())

# scanpy's equivalent on the same data works fine for comparison
adata_cpu = adata.copy()
rsc.get.anndata_to_CPU(adata_cpu)
sc.tl.rank_genes_groups(adata_cpu, groupby='group', groups=['A'],
                        method='t-test', use_raw=False)
print("\nscanpy with `groups=['A']`:")
print(pd.DataFrame(adata_cpu.uns['rank_genes_groups']['scores']).head())

Expected behavior

Calling rsc.tl.rank_genes_groups with groups=[X] should produce the same scores, p-values, adjusted p-values, and log fold changes for group X as calling it without the groups= argument and indexing into the result for group X. Currently the two code paths disagree:

• Without groups=: correct values for all groups.
• With groups=[X]: scores 0, pvals 0, logfoldchanges NaN (t-test); scores correct but pvals/lfc still 0/NaN (wilcoxon).

Environment details (please complete the following information):

• Environment location: Cloud (Lambda Labs / Lambda Cloud GPU VM)
• Linux Distro/Architecture: Ubuntu 22.04 amd64
• GPU Model/Driver: NVIDIA H100 PCIe/Driver Version: 580.105.08
• CUDA: cu13
• Method of Rapids install: pip
  • pip install 'rapids-singlecell[rapids-cu12]' --extra-index-url=https://pypi.nvidia.com
  • Also tried pip install 'rapids-singlecell[rapids-cu13]' --extra-index-url=https://pypi.nvidia.com which gives a similar output

[Intron7]

@dguin thank you for reporting this. it will will be fixed with #651