Kernel Judge

qsa_moe_layer

One QSA sparse attention layer with its MoE block

Best score
no score yet
Score ceiling
6.62x
Distance to ceiling
not available
Baseline source
vLLM 0.33.0
Runs per step
12
Challenge commit
892390e7460a

All submissions

No submission has finished yet.

Ranking

No submission has passed yet.

Shapes

The best time of each shape comes from whichever ranked submission was fastest on it. Until a submission is ranked, baselines and the score ceiling come from the challenge's table, measured on GB200 GPUs.

ShapeBaselineBest timeFloorDistance to floor
batch 1, ctx 8192 111.8 us no time yet 11.60 us not available
batch 8, ctx 8192 178.3 us no time yet 32.70 us not available
batch 32, ctx 8192 283.8 us no time yet 66.10 us not available
batch 64, ctx 8192 326.8 us no time yet 87.30 us not available
batch 1, ctx 65536 118.0 us no time yet 11.80 us not available
batch 8, ctx 65536 185.5 us no time yet 34.60 us not available
batch 1, ctx 262144 122.0 us no time yet 12.60 us not available
batch 1, ctx 999000 130.2 us no time yet 15.50 us not available