Skip to content

[Fix] gemm: Keep Kimi-K3's tiny GEMM shapes on SIMT at every batch size #48732

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
Fridge003 merged 2 commits into from
Nov 6, 2026
+6 −1
Merged

[Fix] gemm: Keep Kimi-K3's tiny GEMM shapes on SIMT at every batch size #48732

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
121544f
[Fix] gemm: Keep Kimi-K3's tiny GEMM shapes on SIMT at every batch size
Fridge003 Nov 5, 2026
0f51ae7
Merge remote-tracking branch 'origin/main' into fix/kimi-k3-tiny-gemm…
Fridge003 Nov 5, 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
7 changes: 6 additions & 1 deletion python/sglang/kernels/ops/gemm/tiny_gemm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -75,8 +75,13 @@ class MMASchedule(NamedTuple):
# _simt_schedule.
# K <= 512 leaves MMA little K to split, outside the cost model.
_TUNED_MMA_MIN_M = {
(1536, 128): 8,
(128, 512): _NEVER,
# Kimi-K3's KDA gate and router projections (kimi_k3_tiny_gemm) stay on SIMT at
# every m: a decode row and a speculative-verify row of the same token must round
# alike, or the drafts are verified against different gates.
(144, 7168): _NEVER,
(896, 7168): _NEVER,
(1536, 128): _NEVER,
}
# 128x512 never takes MMA, and its 128 KB weight is cheap to copy: 4 rows of x per CTA
# beat the default 16 at every m (up to 2x at m = 16).
Expand Down
Loading