Muse Spark Open Kernel Challenge · decode layer

Layer Time Lapse

Every version of my decode layer, replayed as a CUDA graph on an H200 the way the judge times it, for all 12 scored shapes. Each load counts like a real decode step: 46 sliding-window layers plus 16 full-attention layers. Scrub through the versions to see where the time went and what each change moved.

Where the time goes, per load

Bar length is this version's time per layer, scaled so SGLang's baseline fills the track (the mark at the right end). The number is the load's speedup over SGLang.

Kernel timeline

One graph replay, the median of five. Time runs left to right on SGLang's scale for this shape; overlapping kernels (a side stream) sit on a second row.

Score by version

Speedup over SGLang, every load