A deliberately small physical realization of Bob's complete objective.
One 256-thread block assigns one row to each thread. A thread walks that
row's 256 logits serially, writes its exact full-vocabulary cross-entropy,
and writes all 256 gradient elements. The schedule favors transparent
qualification over throughput and introduces no CPU learning path.
18constructor BobBigramLossGradientSM86ManifestValueThe compiler supplied declaration spans and resolved links from this source snapshot. This page does not assert that this file belongs to a checked closure.