Skip to content

[PyTorch] Fix cross_entropy_forward stride guard for non-contiguous input #2746

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
yaox12 merged 3 commits into from
Mar 10, 2026
+22 −1
Merged

[PyTorch] Fix cross_entropy_forward stride guard for non-contiguous input #2746

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
cde3c91
Fix cross_entropy_forward stride guard for non-contiguous input
Bias92 Mar 9, 2026
bbe845b
Add regression test for non-contiguous transposed input
Bias92 Mar 9, 2026
c03ceb0
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot] Mar 9, 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
21 changes: 21 additions & 0 deletions tests/pytorch/test_parallel_cross_entropy.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -167,3 +167,24 @@ def test_ignore_idx_reduced_loss(self):
reduce_loss=True,
ignore_idx=True,
)


def test_non_contiguous_transposed_input():
"""Regression test: stride(-2) != shape[-1] should not produce wrong results."""
s, b, v = 4, 2, 8
torch.manual_seed(42)
logits = torch.randn(s, b, v, device="cuda")
target = torch.randint(0, v, (b, s), device="cuda")

logits_transposed = logits.transpose(0, 1) # stride(-2) != shape[-1]
logits_contiguous = logits_transposed.contiguous()

assert logits_transposed.stride(-1) == 1
assert logits_transposed.stride(-2) != logits_transposed.shape[-1]

loss_t = parallel_cross_entropy(logits_transposed, target, 0.0, False, None)
loss_c = parallel_cross_entropy(logits_contiguous, target, 0.0, False, None)

assert torch.allclose(
loss_t, loss_c
), f"Non-contiguous transposed input gave wrong results: {loss_t} vs {loss_c}"
2 changes: 1 addition & 1 deletion transformer_engine/pytorch/triton/cross_entropy.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -49,7 +49,7 @@ def cross_entropy_forward(
n_non_ignore = torch.zeros(1, dtype=torch.int64, device=_input.device)

# ensure _input and target are contiguous in the last dimension
if _input.stride(-1) != 1:
if _input.stride(-1) != 1 or _input.stride(-2) != _input.shape[-1]:
Copy link
Contributor

@greptile-apps greptile-apps bot Mar 9, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

No regression test for the fixed scenario

The fix correctly catches tensors where stride(-1) == 1 but stride(-2) != shape[-1] (e.g., a tensor of shape (SQ, B, V) transposed to (B, SQ, V) via .transpose(0, 1)). However, no test is added to cover this specific case.

The existing test_swapped_input in tests/pytorch/test_parallel_cross_entropy.py creates a naturally-shaped (SQ, batch, vocab) tensor (contiguous, stride(-2) == V), so it does not exercise the bug path described in the PR.

A minimal regression test would look like:

def test_non_contiguous_transposed_input(self):
    """Regression test for stride(-2) != shape[-1] on non-contiguous input."""
    self.generate_iters(3)
    self.generate_infra(True, 0)
    for _ in range(self.iters):
        batch, SQ, vocab = 2, 64, 1024
        # shape (SQ, batch, vocab) transposed → (batch, SQ, vocab)
        # stride(-1)==1 but stride(-2)==batch*vocab != vocab  ← the old guard missed this
        x = torch.rand((SQ, batch, vocab), dtype=torch.float32, device="cuda").transpose(0, 1)
        assert x.stride(-1) == 1 and x.stride(-2) != x.shape[-1]  # confirm the guard is needed
        self.input_test = x
        ...  # drive through one_iteration_test

Without a test, this class of regression can silently reappear.

_input = _input.contiguous()
if target.stride(-1) != 1:
target = target.contiguous()
Expand Down
Loading