Skip to content

fix: matmul_all_reduce lock signal must use scope="sys" for xGMI visibility #477

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

Draft
Copilot wants to merge 2 commits into
base: main
Choose a base branch
from
Draft

fix: matmul_all_reduce lock signal must use scope="sys" for xGMI visibility #477

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
e336708
Initial plan
Copilot Mar 24, 2026
cc8cbcb
fix: use scope="sys" for lock signal atomic in matmul_all_reduce kernel
Copilot Mar 24, 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
2 changes: 1 addition & 1 deletion iris/ops/matmul_all_reduce.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -132,7 +132,7 @@ def _fused_matmul_all_reduce_kernel(
# Use atomic_xchg with release semantics to ensure memory ordering
tile_id = pid_m * num_tiles_n + pid_n
lock_ptr = locks + tile_id
tl.atomic_xchg(lock_ptr, 1, sem="release", scope="gpu") # Release ensures prior stores visible
tl.atomic_xchg(lock_ptr, 1, sem="release", scope="sys") # Release ensures prior stores visible to remote GPUs

# Create source view only when needed (aux_buffer is not None)
src_view = iris.x.make_tensor_view(aux_buffer, M, N, stride_cm, stride_cn)
Expand Down