Skip to content

Fix warp illegal instruction in kimi k2 thinking PCG #15306

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
ispobock merged 1 commit into from
Dec 18, 2025
Merged

Fix warp illegal instruction in kimi k2 thinking PCG #15306

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
16 changes: 12 additions & 4 deletions sgl-kernel/csrc/moe/kimi_k2_moe_fused_gate.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -126,6 +126,9 @@ __global__ void kimi_k2_moe_fused_gate_kernel_small_token(
if (expert_id >= 0 && expert_id < NUM_EXPERTS) {
output_ptr[row_idx * topk + k] = shared_original_scores[expert_id];
indices_ptr[row_idx * topk + k] = expert_id;
} else {
output_ptr[row_idx * topk + k] = 0.0f;
indices_ptr[row_idx * topk + k] = 0;
}
}

Expand Down Expand Up @@ -219,11 +222,16 @@ __global__ void kimi_k2_moe_fused_gate_kernel(
}
}

if (lane_id == 0 && max_expert != -1) {
if (lane_id == 0) {
int64_t output_idx = row_idx * topk + k;
output_ptr[output_idx] = warp_original_scores[max_expert];
indices_ptr[output_idx] = max_expert;
warp_scores[max_expert] = -FLT_MAX;
if (max_expert != -1) {
output_ptr[output_idx] = warp_original_scores[max_expert];
indices_ptr[output_idx] = max_expert;
warp_scores[max_expert] = -FLT_MAX;
} else {
output_ptr[output_idx] = 0.0f;
indices_ptr[output_idx] = 0;
}
}

__syncwarp();
Expand Down
Loading