Skip to content

[BugFix][MoE] Make xe2 grouped-GEMM tile scheduler cudagraph-capture-safe #457

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

Open
AleksVidmantas wants to merge 2 commits into
base: main
Choose a base branch
from
+14 −13
Open

[BugFix][MoE] Make xe2 grouped-GEMM tile scheduler cudagraph-capture-safe #457

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
da7b095
[BugFix][MoE] Make xe2 grouped-GEMM tile scheduler cudagraph-capture-…
AleksVidmantas Jul 10, 2026
00a477e
Merge branch 'main' into n6a0/capture-safe
AleksVidmantas Jul 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
27 changes: 14 additions & 13 deletions csrc/xpu/grouped_gemm/xe_2/grouped_gemm_xe2.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -103,14 +103,13 @@ CUTE_DEVICE void MoEGEMM(
int group_range = item.get_group_range(1);
int local_id = item.get_local_linear_id();

if (group_id == 0 && local_id == 0) {
auto atm = sycl::atomic_ref<
int,
sycl::memory_order::relaxed,
sycl::memory_scope::device,
sycl::access::address_space::global_space>(atomic_buffer[0]);
atm.store(0);
}
// Grid-stride scheduling (below) replaces the persistent atomic work-steal,
// so there is no shared global tile counter to reset here. The previous
// in-kernel reset was un-barriered and ordered only by eager launch timing;
// under SYCL-graph capture/replay it raced the tile steals, producing
// out-of-bounds tile coordinates and a GPU fault at batch>1. Tile coverage is
// unchanged.
(void)atomic_buffer; // unused after grid-stride; retained in the signature

int pre_rows = 0;
int pre_tiles = 0;
Expand Down Expand Up @@ -205,11 +204,13 @@ CUTE_DEVICE void MoEGEMM(
mma);
}

if (local_id == 0) {
slm_mem[0] = cutlass::atomicAdd(atomic_buffer, 1);
}
item.barrier(sycl::access::fence_space::local_space);
group_id = group_range + slm_mem[0];
// Grid-stride tile assignment: each workgroup deterministically processes
// tiles g, g + group_range, g + 2*group_range, ... No global atomic and
// no cross-workgroup state, so it is safe under SYCL-graph capture/replay
// at any batch size; coverage is identical to the previous atomic
// work-steal (every tile computed exactly once).
(void)slm_mem; // unused after grid-stride; retained in the signature
group_id += group_range;
group_m_id = (group_id * wg_tile_n) / gemm_n_pad;
}
pre_rows = cumsum_rows_for_experts;
Expand Down