Skip to content

fix(sparsemax): preserve low-precision gradients by subtracting the mean in fp32 - #1609

Open
wqh-creator wants to merge 1 commit into
linkedin:mainfrom
wqh-creator:fix/sparsemax-backward-fp32-mean
Open

wqh-creator wants to merge 1 commit into
linkedin:mainfrom
wqh-creator:fix/sparsemax-backward-fp32-mean

Commits

  1. Commits on Nov 9, 2026