Skip to content

adress flaky t-test #784

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
Intron7 merged 2 commits into from
Sep 2, 2026
Merged

adress flaky t-test #784

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
d0908e0
adress flaky t-test
Intron7 Sep 1, 2026
0d74b5e
Merge branch 'main' into address-flaky-t-test
Intron7 Sep 1, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
12 changes: 10 additions & 2 deletions tests/test_rank_genes_groups_ttest.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -68,8 +68,12 @@ def test_rank_genes_groups_ttest_matches_scanpy(reference, method, sparse):
for group in gpu_field.dtype.names:
gpu_values = np.asarray(gpu_field[group], dtype=float)
cpu_values = np.asarray(cpu_field[group], dtype=float)
# rtol=1e-12: group sums are accumulated with atomicAdd, whose
# nondeterministic ordering leaves ~1 ULP noise in mean/var that
# the t-CDF tail amplifies to ~1e-13 rel in pvals (measured);
# 1e-12 gives 10x margin over that noise floor.
np.testing.assert_allclose(
gpu_values, cpu_values, rtol=1e-13, atol=1e-15, equal_nan=True
gpu_values, cpu_values, rtol=1e-12, atol=1e-15, equal_nan=True
)

params = gpu_result["params"]
Expand Down Expand Up @@ -286,10 +290,14 @@ def test_rank_genes_groups_ttest_with_unsorted_groups(reference, method):
# Pick a group that's not the reference for comparison
test_group = "3" if reference != "3" else "0"
for field in ("scores", "logfoldchanges", "pvals", "pvals_adj"):
# rtol=1e-12: the two runs are independent GPU aggregations; atomicAdd
# ordering noise amplified through the t-CDF tail reaches ~1e-13 rel
# in pvals (measured), so the old 1e-13 sat exactly on the failure
# line and 1e-12 gives 10x margin.
np.testing.assert_allclose(
np.asarray(adata.uns["rank_genes_groups"][field][test_group], dtype=float),
np.asarray(bdata.uns["rank_genes_groups"][field][test_group], dtype=float),
rtol=1e-13,
rtol=1e-12,
atol=1e-15,
equal_nan=True,
)
Expand Down