A row uses one 256-thread CTA. Thread t visits vocabulary positions
t + 256*k; the first maximum tile is already read by the prefix. Derive
both tile traversals from the shape so a larger vocabulary cannot silently
reuse the legacy 48-tile image.
546def exactCrossEntropyRowsTileStrideBytes : Nat =
547 (naturalMultiply reductionSM86WarpSum256RequiredThreads 4)The compiler supplied declaration spans and resolved links from this source snapshot. This page does not assert that this file belongs to a checked closure.