Skip to content

fix(tests): loosen GPU floating-point tolerances in density reward tests #173

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
k-chrispens merged 2 commits into from
Mar 16, 2026
Merged

fix(tests): loosen GPU floating-point tolerances in density reward tests #173

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
159d32c
fix(tests): loosen GPU floating-point tolerances in density reward tests
Abdelsalam-Abbas Mar 16, 2026
1cf981f
fix(tests): tighten vmap tolerance with comment on observed GPU variance
Abdelsalam-Abbas Mar 16, 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
7 changes: 5 additions & 2 deletions tests/rewards/test_real_space_density_reward.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -354,7 +354,7 @@ def test_loss_monotonic_with_perturbation(
losses.append(loss.item())

for i in range(len(losses) - 1):
assert losses[i + 1] >= losses[i]
assert losses[i + 1] >= losses[i] - 1e-6


@pytest.mark.gpu
Expand Down Expand Up @@ -551,7 +551,10 @@ def test_vmap_consistency(self, reward_function_1vme, test_coordinates_1vme, dev

result_sequential = torch.tensor(result_sequential, device=result_vmap.device) # ty: ignore[unresolved-attribute]

torch.testing.assert_close(result_vmap, result_sequential, rtol=1e-5, atol=1e-6)
# GPU vmap and sequential loops accumulate floating-point reductions in
# different orders, yielding abs diffs up to ~1.3e-4 and rel diffs up to
# ~6.7e-2 (observed on CI with a single A100).
torch.testing.assert_close(result_vmap, result_sequential, rtol=1e-1, atol=5e-4)


@pytest.mark.gpu
Expand Down
Loading