Skip to content

[MoE] Fix warp-shfl UB in grouped_topk renormalize #23774

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

Closed
Kangyan-Zhou wants to merge 1 commit into from
+7 −17
Closed

[MoE] Fix warp-shfl UB in grouped_topk renormalize #23774

Show file tree
Hide file tree
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
18 changes: 6 additions & 12 deletions python/sglang/jit_kernel/csrc/moe/grouped_topk.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -157,20 +157,14 @@ __global__ void grouped_topk_single_group_kernel(
__syncwarp();
}

// Phase 3: renormalize and write output
// Phase 3: renormalize and write output. `__shfl_xor_sync` requires every
// lane named in the mask (0xffffffff) to execute it, so we pad inactive
// lanes with 0 and run warp_sum_f32 on all 32 lanes uniformly.
float weight = (lane_id < topk) ? selected_weights[lane_id] : 0.0f;
float divisor = renormalize ? warp_sum_f32(weight) + 1e-20f : 1.0f;
if (lane_id < topk) {
float weight = selected_weights[lane_id];
float final_weight = weight * scaling_factor;

if (renormalize) {
// Warp-level sum of selected weights (only lanes < topk contribute)
float partial = (lane_id < topk) ? weight : 0.0f;
float total = warp_sum_f32(partial);
final_weight = weight * scaling_factor / (total + 1e-20f);
}

out_ids[lane_id] = selected_ids[lane_id];
out_vals[lane_id] = final_weight;
out_vals[lane_id] = weight * scaling_factor / divisor;
}
}

Expand Down
6 changes: 1 addition & 5 deletions test/registered/models/test_nvidia_nemotron_3_nano.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4,11 +4,7 @@
from sglang.test.kits.lm_eval_kit import LMEvalMixin
from sglang.test.server_fixtures.default_fixture import DefaultServerBase

register_cuda_ci(
est_time=564,
suite="stage-b-test-2-gpu-large",
disabled="Temporarily disabled; failing on main.",
)
register_cuda_ci(est_time=564, suite="stage-b-test-2-gpu-large")

NEMOTRON_3_NANO_THINKING_ARGS = [
"--trust-remote-code",
Expand Down
Loading