SOTA · attention-kernel TFLOPs · H100 SXM5, BF16, seq=16K
Kernels rerun on a sealed H100 node with pinned driver, pinned CUDA version, seeded RNG.
loading...
loading...
click to expand · verify re-checks sigs + inclusion
Tools · live
Kernels, tokenizers, loaders, quantizers, eval harnesses. All data from .
SOTA · attention-kernel TFLOPs · H100 SXM5, BF16, seq=16K
Kernels rerun on a sealed H100 node with pinned driver, pinned CUDA version, seeded RNG.