Repository navigation
Complete: CPU + V100 verified reproduction
V100 leg landed. Headline: Sudoku hard FULL official 2000-board test on a shared V100-SXM2 = 86.30% solved vs paper 86.1% (exact-configuration reproduction). XSum 1000 docs: R-L 27.78 (paper 28.0 full set). Latency ladder: 2.34 s/doc (HX370 CPU) -> 0.113 s/doc (V100 bs=1), 10.4 docs/s bs=32 under co-tenant pressure (conservative floor). Includes cross-platform fetcher with all three HF sources revision-pinned.