cuda

RTX PRO 6000 Blackwell (SM120 · GDDR7 · 1.8 TB/s)● CUDA-only language gate · Triton / DSL = fail

Isolated CUDA-writing sibling of hard. Four hard problems: GLM-5.2 fused MoE (256 routed + 1 shared, top-8), DeepSeek NSA sparse attention, MegaQwen decode (improve baseline; decode-only at 2k–128k), and grid+MinGRU SPS. Hard and Mega prompts stay frozen.

Methodology and the full problem deck live in the spec. Every published cell is contamination-checked and reward-hack audited; full agent traces are on HuggingFace. Browse the run index for transcripts, submitted solutions, checks, and timing.