Skip to content

fix: pad ids_dst allocation to prevent MMQ stream-k OOB reads #13

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
davide221 merged 1 commit into from
May 28, 2026
Merged

fix: pad ids_dst allocation to prevent MMQ stream-k OOB reads #13

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
6 changes: 5 additions & 1 deletion ggml/src/ggml-cuda/mmq.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -175,7 +175,11 @@ void ggml_cuda_mul_mat_q(
GGML_ASSERT(ne1 == n_expert_used);

ggml_cuda_pool_alloc<int32_t> ids_src1(ctx.pool(), ne_get_rows);
ggml_cuda_pool_alloc<int32_t> ids_dst(ctx.pool(), ne_get_rows);
// Pad ids_dst by mmq_x to prevent OOB reads in the stream-k kernel.
// The kernel loads a full mmq_x-wide tile from ids_dst (line 3709 in mmq.cuh)
// without bounds-checking the load, only the write-back is bounded.
const int64_t mmq_x_pad = (int64_t)get_mmq_x_max_host(cc);
ggml_cuda_pool_alloc<int32_t> ids_dst(ctx.pool(), ne_get_rows + mmq_x_pad);
ggml_cuda_pool_alloc<int32_t> expert_bounds(ctx.pool(), ne02 + 1);

{
Expand Down
Loading