Skip to content

[HIP] [Bugfix] topk_gating kernels: trap when vec_size does not divide num_experts #6323

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
truong-v wants to merge 3 commits into
base: main
Choose a base branch
from
Open

[HIP] [Bugfix] topk_gating kernels: trap when vec_size does not divide num_experts #6323

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
8a194b4
[HIP] topk_gating kernels: trap when vec_size does not divide num_exp…
truong-v Oct 9, 2026
d54a07b
[HIP] topk_gating: shorten the vec_size check comment
truong-v Oct 13, 2026
56a7882
Merge branch 'main' into fix/hip-topk-gating-vec-tail
yzhou103 Oct 23, 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
6 changes: 6 additions & 0 deletions csrc/include/topk_gating_kernels.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -960,6 +960,10 @@ __global__ void topk_gating_kernel(
static constexpr int vec_size = opus::vector_traits<f32vec>::size();
using vec_i = opus::vector_t<cktype_i, vec_size>;
const int num_experts_vec = num_experts / vec_size;
// Top-k scans whole vectors only, so a partial last vector would be skipped.
// topk_gating_launch picks f32vec so that vec_size divides num_experts.
if(num_experts % vec_size != 0)
__builtin_trap();

// Step 1: load + score function
// For softmax, bias is NOT added here -- it's added AFTER normalization
Expand Down Expand Up @@ -1143,6 +1147,8 @@ __global__ void topk_gating_kernel_smem_n(
static constexpr int vec_size = opus::vector_traits<f32vec>::size();
using vec_i = opus::vector_t<cktype_i, vec_size>;
const int num_experts_vec = num_experts / vec_size;
if(num_experts % vec_size != 0)
__builtin_trap();

// Guard: trailing tokens of the last block may be out of range.
const bool valid = token_idx < num_tokens;
Expand Down
Loading