Skip to content

[Bug] Fix w4afp8 moe kernel #9392

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
zhyncs merged 2 commits into from
Aug 21, 2025
Merged

[Bug] Fix w4afp8 moe kernel #9392

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
ae9aeef
Fix a w4afp8 moe kernel bug.
yuhyao Aug 20, 2025
6334884
Merge branch 'main' into fix/w4a8-moe-kernel
AniZpZ Aug 20, 2025
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
4 changes: 4 additions & 0 deletions ...ss_extensions/gemm/collective/sm90_mma_array_tma_gmma_rs_warpspecialized_mixed_input_.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1488,6 +1488,10 @@ struct CollectiveMmaArrayMixedInput<
template <class... TMs>
CUTLASS_DEVICE void
tensormaps_cp_fence_release(TensorMapStorage& shared_tensormaps, cute::tuple<TMs...> const& input_tensormaps) {
if (cute::elect_one_sync()) {
cute::tma_desc_commit_group();
cute::tma_desc_wait_group();
}
Comment on lines +1491 to +1494
Copy link
Copy Markdown
Contributor

@gemini-code-assist gemini-code-assist bot Aug 20, 2025

Choose a reason for hiding this comment

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

critical

The cute::tma_desc_wait_group() function is a collective operation that all threads in the warp must execute to synchronize. Placing it inside the if (cute::elect_one_sync()) block means only the elected thread will perform the wait, while other threads in the warp will proceed without synchronization. This can lead to a race condition where other threads use stale TMA descriptors, or a deadlock. To fix this, the tma_desc_wait_group() call should be moved outside of the conditional block.

    if (cute::elect_one_sync()) {
      cute::tma_desc_commit_group();
    }
    cute::tma_desc_wait_group();

// Entire warp must do this (i.e. it's aligned)
tma_descriptor_cp_fence_release(get<0>(input_tensormaps), shared_tensormaps.smem_tensormap_A);
tma_descriptor_cp_fence_release(get<1>(input_tensormaps), shared_tensormaps.smem_tensormap_B);
Expand Down
Loading