Skip to content

Use narrow over indexing in hadacore_transform to prep for ABI stable #28756

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
merged 1 commit into from
Nov 15, 2025
Merged

Use narrow over indexing in hadacore_transform to prep for ABI stable #28756

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
2 changes: 1 addition & 1 deletion csrc/quantization/hadamard/hadacore/hadamard_transform_cuda.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -802,7 +802,7 @@ torch::Tensor hadacore_transform(torch::Tensor& x, bool inplace) {
});

if (numel % 256 != 0) {
out = out.index({torch::indexing::Slice(0, numel / had_size)});
out = out.narrow(0, 0, numel / had_size);
Copy link
Contributor

@gemini-code-assist gemini-code-assist bot Nov 14, 2025

Choose a reason for hiding this comment

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

critical

While this change to use narrow is correct for ABI stability, it operates on the out tensor which contains uninitialized data when inplace=false. This is because the run_fht kernel at line 801 is called to perform an in-place operation on x, but when inplace=false, out is a separate tensor that is never populated with the result. This causes the function to return garbage data.

To fix this critical bug, the kernel should be called to write to out:

// at line 801
hadacore::run_fht<SCALAR_TYPE>(x.data_ptr(), out.data_ptr(), x.numel(), had_size, stream);

Additionally, the existing tests only seem to cover the inplace=true path. Please add a test case with inplace=false to verify the fix and prevent future regressions.

Copy link
Collaborator

@zou3519 zou3519 Nov 14, 2025

Choose a reason for hiding this comment

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

Um, I think gemini is trolling. If there is a problem, then it was pre-existing

}

if (inplace && out.data_ptr() != x.data_ptr()) {
Expand Down