Exact full-vocabulary logit gradient for coppelius. The preceding exact
forward pass supplies loss[r] = logsumexp(logits[r]) - correct[r], so this
kernel reconstructs logsumexp as loss + correct and writes
(exp(logit - logsumexp) - one_hot(target)) / rows. Grid Y selects one of 48
vocabulary tiles, avoiding a source-level 48-way instruction expansion.
24constructor ExactCrossEntropyBackwardIdentityLengthInvalidThe compiler supplied declaration spans and resolved links from this source snapshot. This page does not assert that this file belongs to a checked closure.