Benchmark · Mega
The megakernel deck
The agent fuses an entire model block into one kernel — 02_kimi_linear_decode is a Kimi-Linear W4A16 hybrid decode — graded on decode speedup over an optimized-PyTorch baseline, not a roofline fraction. One unlimited session per cell on RTX PRO 6000, H100, and B200.
methodology + notes
The metric. 19.35x means 19x faster than the reference baseline; tok/s is decode tokens per second. Higher is better for both, reported per GPU. The transcript is the headline artifact: the model's full optimization journey from baseline to final megakernel.
Ordering. Each run is a single autonomous session with unlimited wall-clock; models self-terminate well under three hours. Models group by valid passes, then order by best decode speedup. The flagged badge counts audited sessions that failed the reward-hack or megakernel-authenticity review; it never changes the order.
Artifacts. The problem's reference and baseline are self-hosted. Browse the run index for transcripts and solutions, or the mega benchmark source.
